Add files using upload-large-folder tool

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/benchmark/benchmark_tf.py

@@ -0,0 +1,298 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+    Benchmarking the library on inference and training in PyTorch.
+"""
+
+
+import random
+import timeit
+from functools import wraps
+from typing import Callable, Optional
+
+from ..configuration_utils import PretrainedConfig
+from ..models.auto.modeling_tf_auto import TF_MODEL_MAPPING, TF_MODEL_WITH_LM_HEAD_MAPPING
+from ..utils import is_py3nvml_available, is_tf_available, logging
+from .benchmark_utils import (
+    Benchmark,
+    Memory,
+    MemorySummary,
+    measure_peak_memory_cpu,
+    start_memory_tracing,
+    stop_memory_tracing,
+)
+
+
+if is_tf_available():
+    import tensorflow as tf
+    from tensorflow.python.framework.errors_impl import ResourceExhaustedError
+
+    from .benchmark_args_tf import TensorFlowBenchmarkArguments
+
+if is_py3nvml_available():
+    import py3nvml.py3nvml as nvml
+
+logger = logging.get_logger(__name__)
+
+
+def run_with_tf_optimizations(do_eager_mode: bool, use_xla: bool):
+    def run_func(func):
+        @wraps(func)
+        def run_in_eager_mode(*args, **kwargs):
+            return func(*args, **kwargs)
+
+        @wraps(func)
+        @tf.function(experimental_compile=use_xla)
+        def run_in_graph_mode(*args, **kwargs):
+            return func(*args, **kwargs)
+
+        if do_eager_mode is True:
+            assert (
+                use_xla is False
+            ), "Cannot run model in XLA, if `args.eager_mode` is set to `True`. Please set `args.eager_mode=False`."
+            return run_in_eager_mode
+        else:
+            return run_in_graph_mode
+
+    return run_func
+
+
+def random_input_ids(batch_size: int, sequence_length: int, vocab_size: int) -> ["tf.Tensor"]:
+    rng = random.Random()
+    values = [rng.randint(0, vocab_size - 1) for i in range(batch_size * sequence_length)]
+    return tf.constant(values, shape=(batch_size, sequence_length), dtype=tf.int32)
+
+
+class TensorFlowBenchmark(Benchmark):
+    args: TensorFlowBenchmarkArguments
+    configs: PretrainedConfig
+    framework: str = "TensorFlow"
+
+    @property
+    def framework_version(self):
+        return tf.__version__
+
+    def _inference_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+        # initialize GPU on separate process
+        strategy = self.args.strategy
+        assert strategy is not None, "A device strategy has to be initialized before using TensorFlow."
+        _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
+        return self._measure_speed(_inference)
+
+    def _train_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+        strategy = self.args.strategy
+        assert strategy is not None, "A device strategy has to be initialized before using TensorFlow."
+        _train = self._prepare_train_func(model_name, batch_size, sequence_length)
+        return self._measure_speed(_train)
+
+    def _inference_memory(
+        self, model_name: str, batch_size: int, sequence_length: int
+    ) -> [Memory, Optional[MemorySummary]]:
+        # initialize GPU on separate process
+        if self.args.is_gpu:
+            tf.config.experimental.set_memory_growth(self.args.gpu_list[self.args.device_idx], True)
+        strategy = self.args.strategy
+        assert strategy is not None, "A device strategy has to be initialized before using TensorFlow."
+        _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
+        return self._measure_memory(_inference)
+
+    def _train_memory(
+        self, model_name: str, batch_size: int, sequence_length: int
+    ) -> [Memory, Optional[MemorySummary]]:
+        if self.args.is_gpu:
+            tf.config.experimental.set_memory_growth(self.args.gpu_list[self.args.device_idx], True)
+        strategy = self.args.strategy
+        assert strategy is not None, "A device strategy has to be initialized before using TensorFlow."
+
+        _train = self._prepare_train_func(model_name, batch_size, sequence_length)
+        return self._measure_memory(_train)
+
+    def _prepare_inference_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
+        config = self.config_dict[model_name]
+
+        if self.args.fp16:
+            raise NotImplementedError("Mixed precision is currently not supported.")
+
+        has_model_class_in_config = (
+            hasattr(config, "architectures")
+            and isinstance(config.architectures, list)
+            and len(config.architectures) > 0
+        )
+        if not self.args.only_pretrain_model and has_model_class_in_config:
+            try:
+                model_class = "TF" + config.architectures[0]  # prepend 'TF' for tensorflow model
+                transformers_module = __import__("transformers", fromlist=[model_class])
+                model_cls = getattr(transformers_module, model_class)
+                model = model_cls(config)
+            except ImportError:
+                raise ImportError(
+                    f"{model_class} does not exist. If you just want to test the pretrained model, you might want to"
+                    " set `--only_pretrain_model` or `args.only_pretrain_model=True`."
+                )
+        else:
+            model = TF_MODEL_MAPPING[config.__class__](config)
+
+        # encoder-decoder has vocab size saved differently
+        vocab_size = config.vocab_size if hasattr(config, "vocab_size") else config.encoder.vocab_size
+        input_ids = random_input_ids(batch_size, sequence_length, vocab_size)
+
+        @run_with_tf_optimizations(self.args.eager_mode, self.args.use_xla)
+        def encoder_decoder_forward():
+            return model(input_ids, decoder_input_ids=input_ids, training=False)
+
+        @run_with_tf_optimizations(self.args.eager_mode, self.args.use_xla)
+        def encoder_forward():
+            return model(input_ids, training=False)
+
+        _inference = encoder_decoder_forward if config.is_encoder_decoder else encoder_forward
+
+        return _inference
+
+    def _prepare_train_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
+        config = self.config_dict[model_name]
+
+        assert (
+            self.args.eager_mode is False
+        ), "Training cannot be done in eager mode. Please make sure that `args.eager_mode = False`."
+
+        if self.args.fp16:
+            raise NotImplementedError("Mixed precision is currently not supported.")
+
+        has_model_class_in_config = (
+            hasattr(config, "architectures")
+            and isinstance(config.architectures, list)
+            and len(config.architectures) > 0
+        )
+        if not self.args.only_pretrain_model and has_model_class_in_config:
+            try:
+                model_class = "TF" + config.architectures[0]  # prepend 'TF' for tensorflow model
+                transformers_module = __import__("transformers", fromlist=[model_class])
+                model_cls = getattr(transformers_module, model_class)
+                model = model_cls(config)
+            except ImportError:
+                raise ImportError(
+                    f"{model_class} does not exist. If you just want to test the pretrained model, you might want to"
+                    " set `--only_pretrain_model` or `args.only_pretrain_model=True`."
+                )
+        else:
+            model = TF_MODEL_WITH_LM_HEAD_MAPPING[config.__class__](config)
+
+        # encoder-decoder has vocab size saved differently
+        vocab_size = config.vocab_size if hasattr(config, "vocab_size") else config.encoder.vocab_size
+        input_ids = random_input_ids(batch_size, sequence_length, vocab_size)
+
+        @run_with_tf_optimizations(self.args.eager_mode, self.args.use_xla)
+        def encoder_decoder_train():
+            loss = model(input_ids, decoder_input_ids=input_ids, labels=input_ids, training=True)[0]
+            gradients = tf.gradients(loss, model.trainable_variables)
+            return gradients
+
+        @run_with_tf_optimizations(self.args.eager_mode, self.args.use_xla)
+        def encoder_train():
+            loss = model(input_ids, labels=input_ids, training=True)[0]
+            gradients = tf.gradients(loss, model.trainable_variables)
+            return gradients
+
+        _train = encoder_decoder_train if config.is_encoder_decoder else encoder_train
+
+        return _train
+
+    def _measure_speed(self, func) -> float:
+        with self.args.strategy.scope():
+            try:
+                if self.args.is_tpu or self.args.use_xla:
+                    # run additional 10 times to stabilize compilation for tpu
+                    logger.info("Do inference on TPU. Running model 5 times to stabilize compilation")
+                    timeit.repeat(func, repeat=1, number=5)
+
+                # as written in https://docs.python.org/2/library/timeit.html#timeit.Timer.repeat, min should be taken rather than the average
+                runtimes = timeit.repeat(
+                    func,
+                    repeat=self.args.repeat,
+                    number=10,
+                )
+
+                return min(runtimes) / 10.0
+            except ResourceExhaustedError as e:
+                self.print_fn(f"Doesn't fit on GPU. {e}")
+
+    def _measure_memory(self, func: Callable[[], None]) -> [Memory, MemorySummary]:
+        logger.info(
+            "Note that TensorFlow allocates more memory than "
+            "it might need to speed up computation. "
+            "The memory reported here corresponds to the memory "
+            "reported by `nvidia-smi`, which can vary depending "
+            "on total available memory on the GPU that is used."
+        )
+        with self.args.strategy.scope():
+            try:
+                if self.args.trace_memory_line_by_line:
+                    assert self.args.eager_mode, (
+                        "`args.eager_mode` is set to `False`. Make sure to run model in eager mode to measure memory"
+                        " consumption line by line."
+                    )
+                    trace = start_memory_tracing("transformers")
+
+                if self.args.is_tpu:
+                    # tpu
+                    raise NotImplementedError(
+                        "Memory Benchmarking is currently not implemented for TPU. Please disable memory benchmarking"
+                        " with `args.memory=False`"
+                    )
+                elif self.args.is_gpu:
+                    # gpu
+                    if not is_py3nvml_available():
+                        logger.warning(
+                            "py3nvml not installed, we won't log GPU memory usage. "
+                            "Install py3nvml (pip install py3nvml) to log information about GPU."
+                        )
+                        memory = "N/A"
+                    else:
+                        logger.info(
+                            "Measuring total GPU usage on GPU device. Make sure to not have additional processes"
+                            " running on the same GPU."
+                        )
+                        # init nvml
+                        nvml.nvmlInit()
+                        func()
+                        handle = nvml.nvmlDeviceGetHandleByIndex(self.args.device_idx)
+                        meminfo = nvml.nvmlDeviceGetMemoryInfo(handle)
+                        max_bytes_in_use = meminfo.used
+                        memory = Memory(max_bytes_in_use)
+                        # shutdown nvml
+                        nvml.nvmlShutdown()
+                else:
+                    # cpu
+                    if self.args.trace_memory_line_by_line:
+                        logger.info(
+                            "When enabling line by line tracing, the max peak memory for CPU is inaccurate in"
+                            " TensorFlow."
+                        )
+                        memory = None
+                    else:
+                        memory_bytes = measure_peak_memory_cpu(func)
+                        memory = Memory(memory_bytes) if isinstance(memory_bytes, int) else memory_bytes
+                if self.args.trace_memory_line_by_line:
+                    summary = stop_memory_tracing(trace)
+                    if memory is None:
+                        memory = summary.total
+                else:
+                    summary = None
+
+                return memory, summary
+            except ResourceExhaustedError as e:
+                self.print_fn(f"Doesn't fit on GPU. {e}")
+                return "N/A", None

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/benchmark/benchmark_utils.py

@@ -0,0 +1,913 @@
+# This file is adapted from the AllenNLP library at https://github.com/allenai/allennlp
+
+# Copyright 2020 The HuggingFace Team and the AllenNLP authors. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Utilities for working with the local dataset cache.
+"""
+
+import copy
+import csv
+import linecache
+import os
+import platform
+import sys
+import warnings
+from abc import ABC, abstractmethod
+from collections import defaultdict, namedtuple
+from datetime import datetime
+from multiprocessing import Pipe, Process, Queue
+from multiprocessing.connection import Connection
+from typing import Callable, Iterable, List, NamedTuple, Optional, Union
+
+from .. import AutoConfig, PretrainedConfig
+from .. import __version__ as version
+from ..utils import is_psutil_available, is_py3nvml_available, is_tf_available, is_torch_available, logging
+from .benchmark_args_utils import BenchmarkArguments
+
+
+if is_torch_available():
+    from torch.cuda import empty_cache as torch_empty_cache
+
+if is_tf_available():
+    from tensorflow.python.eager import context as tf_context
+
+if is_psutil_available():
+    import psutil
+
+if is_py3nvml_available():
+    import py3nvml.py3nvml as nvml
+
+if platform.system() == "Windows":
+    from signal import CTRL_C_EVENT as SIGKILL
+else:
+    from signal import SIGKILL
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+_is_memory_tracing_enabled = False
+
+BenchmarkOutput = namedtuple(
+    "BenchmarkOutput",
+    [
+        "time_inference_result",
+        "memory_inference_result",
+        "time_train_result",
+        "memory_train_result",
+        "inference_summary",
+        "train_summary",
+    ],
+)
+
+
+def separate_process_wrapper_fn(func: Callable[[], None], do_multi_processing: bool) -> Callable[[], None]:
+    """
+    This function wraps another function into its own separated process. In order to ensure accurate memory
+    measurements it is important that the function is executed in a separate process
+
+    Args:
+        - `func`: (`callable`): function() -> ... generic function which will be executed in its own separate process
+        - `do_multi_processing`: (`bool`) Whether to run function on separate process or not
+    """
+
+    def multi_process_func(*args, **kwargs):
+        # run function in an individual
+        # process to get correct memory
+        def wrapper_func(queue: Queue, *args):
+            try:
+                result = func(*args)
+            except Exception as e:
+                logger.error(e)
+                print(e)
+                result = "N/A"
+            queue.put(result)
+
+        queue = Queue()
+        p = Process(target=wrapper_func, args=[queue] + list(args))
+        p.start()
+        result = queue.get()
+        p.join()
+        return result
+
+    if do_multi_processing:
+        logger.info(f"Function {func} is executed in its own process...")
+        return multi_process_func
+    else:
+        return func
+
+
+def is_memory_tracing_enabled():
+    global _is_memory_tracing_enabled
+    return _is_memory_tracing_enabled
+
+
+class Frame(NamedTuple):
+    """
+    `Frame` is a NamedTuple used to gather the current frame state. `Frame` has the following fields:
+
+        - 'filename' (string): Name of the file currently executed
+        - 'module' (string): Name of the module currently executed
+        - 'line_number' (int): Number of the line currently executed
+        - 'event' (string): Event that triggered the tracing (default will be "line")
+        - 'line_text' (string): Text of the line in the python script
+    """
+
+    filename: str
+    module: str
+    line_number: int
+    event: str
+    line_text: str
+
+
+class UsedMemoryState(NamedTuple):
+    """
+    `UsedMemoryState` are named tuples with the following fields:
+
+        - 'frame': a `Frame` namedtuple (see below) storing information on the current tracing frame (current file,
+          location in current file)
+        - 'cpu_memory': CPU RSS memory state *before* executing the line
+        - 'gpu_memory': GPU used memory *before* executing the line (sum for all GPUs or for only `gpus_to_trace` if
+          provided)
+    """
+
+    frame: Frame
+    cpu_memory: int
+    gpu_memory: int
+
+
+class Memory(NamedTuple):
+    """
+    `Memory` NamedTuple have a single field `bytes` and you can get a human readable str of the number of mega bytes by
+    calling `__repr__`
+
+        - `byte` (integer): number of bytes,
+    """
+
+    bytes: int
+
+    def __repr__(self) -> str:
+        return str(bytes_to_mega_bytes(self.bytes))
+
+
+class MemoryState(NamedTuple):
+    """
+    `MemoryState` are namedtuples listing frame + CPU/GPU memory with the following fields:
+
+        - `frame` (`Frame`): the current frame (see above)
+        - `cpu`: CPU memory consumed at during the current frame as a `Memory` named tuple
+        - `gpu`: GPU memory consumed at during the current frame as a `Memory` named tuple
+        - `cpu_gpu`: CPU + GPU memory consumed at during the current frame as a `Memory` named tuple
+    """
+
+    frame: Frame
+    cpu: Memory
+    gpu: Memory
+    cpu_gpu: Memory
+
+
+class MemorySummary(NamedTuple):
+    """
+    `MemorySummary` namedtuple otherwise with the fields:
+
+        - `sequential`: a list of `MemoryState` namedtuple (see below) computed from the provided `memory_trace` by
+          subtracting the memory after executing each line from the memory before executing said line.
+        - `cumulative`: a list of `MemoryState` namedtuple (see below) with cumulative increase in memory for each line
+          obtained by summing repeated memory increase for a line if it's executed several times. The list is sorted
+          from the frame with the largest memory consumption to the frame with the smallest (can be negative if memory
+          is released)
+        - `total`: total memory increase during the full tracing as a `Memory` named tuple (see below). Line with
+          memory release (negative consumption) are ignored if `ignore_released_memory` is `True` (default).
+    """
+
+    sequential: List[MemoryState]
+    cumulative: List[MemoryState]
+    current: List[MemoryState]
+    total: Memory
+
+
+MemoryTrace = List[UsedMemoryState]
+
+
+def measure_peak_memory_cpu(function: Callable[[], None], interval=0.5, device_idx=None) -> int:
+    """
+    measures peak cpu memory consumption of a given `function` running the function for at least interval seconds and
+    at most 20 * interval seconds. This function is heavily inspired by: `memory_usage` of the package
+    `memory_profiler`:
+    https://github.com/pythonprofilers/memory_profiler/blob/895c4ac7a08020d66ae001e24067da6dcea42451/memory_profiler.py#L239
+
+    Args:
+        - `function`: (`callable`): function() -> ... function without any arguments to measure for which to measure
+          the peak memory
+
+        - `interval`: (`float`, `optional`, defaults to `0.5`) interval in second for which to measure the memory usage
+
+        - `device_idx`: (`int`, `optional`, defaults to `None`) device id for which to measure gpu usage
+
+    Returns:
+
+        - `max_memory`: (`int`) consumed memory peak in Bytes
+    """
+
+    def get_cpu_memory(process_id: int) -> int:
+        """
+        measures current cpu memory usage of a given `process_id`
+
+        Args:
+            - `process_id`: (`int`) process_id for which to measure memory
+
+        Returns
+
+            - `memory`: (`int`) consumed memory in Bytes
+        """
+        process = psutil.Process(process_id)
+        try:
+            meminfo_attr = "memory_info" if hasattr(process, "memory_info") else "get_memory_info"
+            memory = getattr(process, meminfo_attr)()[0]
+        except psutil.AccessDenied:
+            raise ValueError("Error with Psutil.")
+        return memory
+
+    if not is_psutil_available():
+        logger.warning(
+            "Psutil not installed, we won't log CPU memory usage. "
+            "Install Psutil (pip install psutil) to use CPU memory tracing."
+        )
+        max_memory = "N/A"
+    else:
+
+        class MemoryMeasureProcess(Process):
+
+            """
+            `MemoryMeasureProcess` inherits from `Process` and overwrites its `run()` method. Used to measure the
+            memory usage of a process
+            """
+
+            def __init__(self, process_id: int, child_connection: Connection, interval: float):
+                super().__init__()
+                self.process_id = process_id
+                self.interval = interval
+                self.connection = child_connection
+                self.num_measurements = 1
+                self.mem_usage = get_cpu_memory(self.process_id)
+
+            def run(self):
+                self.connection.send(0)
+                stop = False
+                while True:
+                    self.mem_usage = max(self.mem_usage, get_cpu_memory(self.process_id))
+                    self.num_measurements += 1
+
+                    if stop:
+                        break
+
+                    stop = self.connection.poll(self.interval)
+
+                # send results to parent pipe
+                self.connection.send(self.mem_usage)
+                self.connection.send(self.num_measurements)
+
+        while True:
+            # create child, parent connection
+            child_connection, parent_connection = Pipe()
+
+            # instantiate process
+            mem_process = MemoryMeasureProcess(os.getpid(), child_connection, interval)
+            mem_process.start()
+
+            # wait until we get memory
+            parent_connection.recv()
+
+            try:
+                # execute function
+                function()
+
+                # start parent connection
+                parent_connection.send(0)
+
+                # receive memory and num measurements
+                max_memory = parent_connection.recv()
+                num_measurements = parent_connection.recv()
+            except Exception:
+                # kill process in a clean way
+                parent = psutil.Process(os.getpid())
+                for child in parent.children(recursive=True):
+                    os.kill(child.pid, SIGKILL)
+                mem_process.join(0)
+                raise RuntimeError("Process killed. Error in Process")
+
+            # run process at least 20 * interval or until it finishes
+            mem_process.join(20 * interval)
+
+            if (num_measurements > 4) or (interval < 1e-6):
+                break
+
+            # reduce interval
+            interval /= 10
+
+        return max_memory
+
+
+def start_memory_tracing(
+    modules_to_trace: Optional[Union[str, Iterable[str]]] = None,
+    modules_not_to_trace: Optional[Union[str, Iterable[str]]] = None,
+    events_to_trace: str = "line",
+    gpus_to_trace: Optional[List[int]] = None,
+) -> MemoryTrace:
+    """
+    Setup line-by-line tracing to record rss mem (RAM) at each line of a module or sub-module. See `./benchmark.py` for
+    usage examples. Current memory consumption is returned using psutil and in particular is the RSS memory "Resident
+    Set Size” (the non-swapped physical memory the process is using). See
+    https://psutil.readthedocs.io/en/latest/#psutil.Process.memory_info
+
+    Args:
+        - `modules_to_trace`: (None, string, list/tuple of string) if None, all events are recorded if string or list
+          of strings: only events from the listed module/sub-module will be recorded (e.g. 'fairseq' or
+          'transformers.models.gpt2.modeling_gpt2')
+        - `modules_not_to_trace`: (None, string, list/tuple of string) if None, no module is avoided if string or list
+          of strings: events from the listed module/sub-module will not be recorded (e.g. 'torch')
+        - `events_to_trace`: string or list of string of events to be recorded (see official python doc for
+          `sys.settrace` for the list of events) default to line
+        - `gpus_to_trace`: (optional list, default None) list of GPUs to trace. Default to tracing all GPUs
+
+    Return:
+
+        - `memory_trace` is a list of `UsedMemoryState` for each event (default each line of the traced script).
+
+            - `UsedMemoryState` are named tuples with the following fields:
+
+                - 'frame': a `Frame` namedtuple (see below) storing information on the current tracing frame (current
+                  file, location in current file)
+                - 'cpu_memory': CPU RSS memory state *before* executing the line
+                - 'gpu_memory': GPU used memory *before* executing the line (sum for all GPUs or for only
+                  `gpus_to_trace` if provided)
+
+    `Frame` is a namedtuple used by `UsedMemoryState` to list the current frame state. `Frame` has the following
+    fields: - 'filename' (string): Name of the file currently executed - 'module' (string): Name of the module
+    currently executed - 'line_number' (int): Number of the line currently executed - 'event' (string): Event that
+    triggered the tracing (default will be "line") - 'line_text' (string): Text of the line in the python script
+
+    """
+    if is_psutil_available():
+        process = psutil.Process(os.getpid())
+    else:
+        logger.warning(
+            "Psutil not installed, we won't log CPU memory usage. "
+            "Install psutil (pip install psutil) to use CPU memory tracing."
+        )
+        process = None
+
+    if is_py3nvml_available():
+        try:
+            nvml.nvmlInit()
+            devices = list(range(nvml.nvmlDeviceGetCount())) if gpus_to_trace is None else gpus_to_trace
+            nvml.nvmlShutdown()
+        except (OSError, nvml.NVMLError):
+            logger.warning("Error while initializing communication with GPU. We won't perform GPU memory tracing.")
+            log_gpu = False
+        else:
+            log_gpu = is_torch_available() or is_tf_available()
+    else:
+        logger.warning(
+            "py3nvml not installed, we won't log GPU memory usage. "
+            "Install py3nvml (pip install py3nvml) to use GPU memory tracing."
+        )
+        log_gpu = False
+
+    memory_trace = []
+
+    def traceit(frame, event, args):
+        """
+        Tracing method executed before running each line in a module or sub-module Record memory allocated in a list
+        with debugging information
+        """
+        global _is_memory_tracing_enabled
+
+        if not _is_memory_tracing_enabled:
+            return traceit
+
+        # Filter events
+        if events_to_trace is not None:
+            if isinstance(events_to_trace, str) and event != events_to_trace:
+                return traceit
+            elif isinstance(events_to_trace, (list, tuple)) and event not in events_to_trace:
+                return traceit
+
+        if "__name__" not in frame.f_globals:
+            return traceit
+
+        # Filter modules
+        name = frame.f_globals["__name__"]
+        if not isinstance(name, str):
+            return traceit
+        else:
+            # Filter whitelist of modules to trace
+            if modules_to_trace is not None:
+                if isinstance(modules_to_trace, str) and modules_to_trace not in name:
+                    return traceit
+                elif isinstance(modules_to_trace, (list, tuple)) and all(m not in name for m in modules_to_trace):
+                    return traceit
+
+            # Filter blacklist of modules not to trace
+            if modules_not_to_trace is not None:
+                if isinstance(modules_not_to_trace, str) and modules_not_to_trace in name:
+                    return traceit
+                elif isinstance(modules_not_to_trace, (list, tuple)) and any(m in name for m in modules_not_to_trace):
+                    return traceit
+
+        # Record current tracing state (file, location in file...)
+        lineno = frame.f_lineno
+        filename = frame.f_globals["__file__"]
+        if filename.endswith(".pyc") or filename.endswith(".pyo"):
+            filename = filename[:-1]
+        line = linecache.getline(filename, lineno).rstrip()
+        traced_state = Frame(filename, name, lineno, event, line)
+
+        # Record current memory state (rss memory) and compute difference with previous memory state
+        cpu_mem = 0
+        if process is not None:
+            mem = process.memory_info()
+            cpu_mem = mem.rss
+
+        gpu_mem = 0
+        if log_gpu:
+            # Clear GPU caches
+            if is_torch_available():
+                torch_empty_cache()
+            if is_tf_available():
+                tf_context.context()._clear_caches()  # See https://github.com/tensorflow/tensorflow/issues/20218#issuecomment-416771802
+
+            # Sum used memory for all GPUs
+            nvml.nvmlInit()
+
+            for i in devices:
+                handle = nvml.nvmlDeviceGetHandleByIndex(i)
+                meminfo = nvml.nvmlDeviceGetMemoryInfo(handle)
+                gpu_mem += meminfo.used
+
+            nvml.nvmlShutdown()
+
+        mem_state = UsedMemoryState(traced_state, cpu_mem, gpu_mem)
+        memory_trace.append(mem_state)
+
+        return traceit
+
+    sys.settrace(traceit)
+
+    global _is_memory_tracing_enabled
+    _is_memory_tracing_enabled = True
+
+    return memory_trace
+
+
+def stop_memory_tracing(
+    memory_trace: Optional[MemoryTrace] = None, ignore_released_memory: bool = True
+) -> Optional[MemorySummary]:
+    """
+    Stop memory tracing cleanly and return a summary of the memory trace if a trace is given.
+
+    Args:
+        `memory_trace` (optional output of start_memory_tracing, default: None):
+            memory trace to convert in summary
+        `ignore_released_memory` (boolean, default: None):
+            if True we only sum memory increase to compute total memory
+
+    Return:
+
+        - None if `memory_trace` is None
+        - `MemorySummary` namedtuple otherwise with the fields:
+
+            - `sequential`: a list of `MemoryState` namedtuple (see below) computed from the provided `memory_trace` by
+              subtracting the memory after executing each line from the memory before executing said line.
+            - `cumulative`: a list of `MemoryState` namedtuple (see below) with cumulative increase in memory for each
+              line obtained by summing repeated memory increase for a line if it's executed several times. The list is
+              sorted from the frame with the largest memory consumption to the frame with the smallest (can be negative
+              if memory is released)
+            - `total`: total memory increase during the full tracing as a `Memory` named tuple (see below). Line with
+              memory release (negative consumption) are ignored if `ignore_released_memory` is `True` (default).
+
+    `Memory` named tuple have fields
+
+        - `byte` (integer): number of bytes,
+        - `string` (string): same as human readable string (ex: "3.5MB")
+
+    `Frame` are namedtuple used to list the current frame state and have the following fields:
+
+        - 'filename' (string): Name of the file currently executed
+        - 'module' (string): Name of the module currently executed
+        - 'line_number' (int): Number of the line currently executed
+        - 'event' (string): Event that triggered the tracing (default will be "line")
+        - 'line_text' (string): Text of the line in the python script
+
+    `MemoryState` are namedtuples listing frame + CPU/GPU memory with the following fields:
+
+        - `frame` (`Frame`): the current frame (see above)
+        - `cpu`: CPU memory consumed at during the current frame as a `Memory` named tuple
+        - `gpu`: GPU memory consumed at during the current frame as a `Memory` named tuple
+        - `cpu_gpu`: CPU + GPU memory consumed at during the current frame as a `Memory` named tuple
+    """
+    global _is_memory_tracing_enabled
+    _is_memory_tracing_enabled = False
+
+    if memory_trace is not None and len(memory_trace) > 1:
+        memory_diff_trace = []
+        memory_curr_trace = []
+
+        cumulative_memory_dict = defaultdict(lambda: [0, 0, 0])
+
+        for (
+            (frame, cpu_mem, gpu_mem),
+            (next_frame, next_cpu_mem, next_gpu_mem),
+        ) in zip(memory_trace[:-1], memory_trace[1:]):
+            cpu_mem_inc = next_cpu_mem - cpu_mem
+            gpu_mem_inc = next_gpu_mem - gpu_mem
+            cpu_gpu_mem_inc = cpu_mem_inc + gpu_mem_inc
+            memory_diff_trace.append(
+                MemoryState(
+                    frame=frame,
+                    cpu=Memory(cpu_mem_inc),
+                    gpu=Memory(gpu_mem_inc),
+                    cpu_gpu=Memory(cpu_gpu_mem_inc),
+                )
+            )
+
+            memory_curr_trace.append(
+                MemoryState(
+                    frame=frame,
+                    cpu=Memory(next_cpu_mem),
+                    gpu=Memory(next_gpu_mem),
+                    cpu_gpu=Memory(next_gpu_mem + next_cpu_mem),
+                )
+            )
+
+            cumulative_memory_dict[frame][0] += cpu_mem_inc
+            cumulative_memory_dict[frame][1] += gpu_mem_inc
+            cumulative_memory_dict[frame][2] += cpu_gpu_mem_inc
+
+        cumulative_memory = sorted(
+            cumulative_memory_dict.items(), key=lambda x: x[1][2], reverse=True
+        )  # order by the total CPU + GPU memory increase
+        cumulative_memory = [
+            MemoryState(
+                frame=frame,
+                cpu=Memory(cpu_mem_inc),
+                gpu=Memory(gpu_mem_inc),
+                cpu_gpu=Memory(cpu_gpu_mem_inc),
+            )
+            for frame, (cpu_mem_inc, gpu_mem_inc, cpu_gpu_mem_inc) in cumulative_memory
+        ]
+
+        memory_curr_trace = sorted(memory_curr_trace, key=lambda x: x.cpu_gpu.bytes, reverse=True)
+
+        if ignore_released_memory:
+            total_memory = sum(max(0, step_trace.cpu_gpu.bytes) for step_trace in memory_diff_trace)
+        else:
+            total_memory = sum(step_trace.cpu_gpu.bytes for step_trace in memory_diff_trace)
+
+        total_memory = Memory(total_memory)
+
+        return MemorySummary(
+            sequential=memory_diff_trace,
+            cumulative=cumulative_memory,
+            current=memory_curr_trace,
+            total=total_memory,
+        )
+
+    return None
+
+
+def bytes_to_mega_bytes(memory_amount: int) -> int:
+    """Utility to convert a number of bytes (int) into a number of mega bytes (int)"""
+    return memory_amount >> 20
+
+
+class Benchmark(ABC):
+    """
+    Benchmarks is a simple but feature-complete benchmarking script to compare memory and time performance of models in
+    Transformers.
+    """
+
+    args: BenchmarkArguments
+    configs: PretrainedConfig
+    framework: str
+
+    def __init__(self, args: BenchmarkArguments = None, configs: PretrainedConfig = None):
+        self.args = args
+        if configs is None:
+            self.config_dict = {
+                model_name: AutoConfig.from_pretrained(model_name) for model_name in self.args.model_names
+            }
+        else:
+            self.config_dict = dict(zip(self.args.model_names, configs))
+
+        warnings.warn(
+            f"The class {self.__class__} is deprecated. Hugging Face Benchmarking utils"
+            " are deprecated in general and it is advised to use external Benchmarking libraries "
+            " to benchmark Transformer models.",
+            FutureWarning,
+        )
+
+        if self.args.memory and os.getenv("TRANSFORMERS_USE_MULTIPROCESSING") == 0:
+            logger.warning(
+                "Memory consumption will not be measured accurately if `args.multi_process` is set to `False.` The"
+                " flag 'TRANSFORMERS_USE_MULTIPROCESSING' should only be disabled for debugging / testing."
+            )
+
+        self._print_fn = None
+        self._framework_version = None
+        self._environment_info = None
+
+    @property
+    def print_fn(self):
+        if self._print_fn is None:
+            if self.args.log_print:
+
+                def print_and_log(*args):
+                    with open(self.args.log_filename, "a") as log_file:
+                        log_file.write("".join(args) + "\n")
+                    print(*args)
+
+                self._print_fn = print_and_log
+            else:
+                self._print_fn = print
+        return self._print_fn
+
+    @property
+    @abstractmethod
+    def framework_version(self):
+        pass
+
+    @abstractmethod
+    def _inference_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+        pass
+
+    @abstractmethod
+    def _train_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
+        pass
+
+    @abstractmethod
+    def _inference_memory(
+        self, model_name: str, batch_size: int, sequence_length: int
+    ) -> [Memory, Optional[MemorySummary]]:
+        pass
+
+    @abstractmethod
+    def _train_memory(
+        self, model_name: str, batch_size: int, sequence_length: int
+    ) -> [Memory, Optional[MemorySummary]]:
+        pass
+
+    def inference_speed(self, *args, **kwargs) -> float:
+        return separate_process_wrapper_fn(self._inference_speed, self.args.do_multi_processing)(*args, **kwargs)
+
+    def train_speed(self, *args, **kwargs) -> float:
+        return separate_process_wrapper_fn(self._train_speed, self.args.do_multi_processing)(*args, **kwargs)
+
+    def inference_memory(self, *args, **kwargs) -> [Memory, Optional[MemorySummary]]:
+        return separate_process_wrapper_fn(self._inference_memory, self.args.do_multi_processing)(*args, **kwargs)
+
+    def train_memory(self, *args, **kwargs) -> [Memory, Optional[MemorySummary]]:
+        return separate_process_wrapper_fn(self._train_memory, self.args.do_multi_processing)(*args, **kwargs)
+
+    def run(self):
+        result_dict = {model_name: {} for model_name in self.args.model_names}
+        inference_result_time = copy.deepcopy(result_dict)
+        inference_result_memory = copy.deepcopy(result_dict)
+        train_result_time = copy.deepcopy(result_dict)
+        train_result_memory = copy.deepcopy(result_dict)
+
+        for c, model_name in enumerate(self.args.model_names):
+            self.print_fn(f"{c + 1} / {len(self.args.model_names)}")
+
+            model_dict = {
+                "bs": self.args.batch_sizes,
+                "ss": self.args.sequence_lengths,
+                "result": {i: {} for i in self.args.batch_sizes},
+            }
+            inference_result_time[model_name] = copy.deepcopy(model_dict)
+            inference_result_memory[model_name] = copy.deepcopy(model_dict)
+            train_result_time[model_name] = copy.deepcopy(model_dict)
+            train_result_memory[model_name] = copy.deepcopy(model_dict)
+
+            inference_summary = train_summary = None
+
+            for batch_size in self.args.batch_sizes:
+                for sequence_length in self.args.sequence_lengths:
+                    if self.args.inference:
+                        if self.args.memory:
+                            memory, inference_summary = self.inference_memory(model_name, batch_size, sequence_length)
+                            inference_result_memory[model_name]["result"][batch_size][sequence_length] = memory
+                        if self.args.speed:
+                            time = self.inference_speed(model_name, batch_size, sequence_length)
+                            inference_result_time[model_name]["result"][batch_size][sequence_length] = time
+
+                    if self.args.training:
+                        if self.args.memory:
+                            memory, train_summary = self.train_memory(model_name, batch_size, sequence_length)
+                            train_result_memory[model_name]["result"][batch_size][sequence_length] = memory
+                        if self.args.speed:
+                            time = self.train_speed(model_name, batch_size, sequence_length)
+                            train_result_time[model_name]["result"][batch_size][sequence_length] = time
+
+        if self.args.inference:
+            if self.args.speed:
+                self.print_fn("\n" + 20 * "=" + ("INFERENCE - SPEED - RESULT").center(40) + 20 * "=")
+                self.print_results(inference_result_time, type_label="Time in s")
+                self.save_to_csv(inference_result_time, self.args.inference_time_csv_file)
+                if self.args.is_tpu:
+                    self.print_fn(
+                        "TPU was used for inference. Note that the time after compilation stabilized (after ~10"
+                        " inferences model.forward(..) calls) was measured."
+                    )
+
+            if self.args.memory:
+                self.print_fn("\n" + 20 * "=" + ("INFERENCE - MEMORY - RESULT").center(40) + 20 * "=")
+                self.print_results(inference_result_memory, type_label="Memory in MB")
+                self.save_to_csv(inference_result_memory, self.args.inference_memory_csv_file)
+
+            if self.args.trace_memory_line_by_line:
+                self.print_fn("\n" + 20 * "=" + ("INFERENCE - MEMOMRY - LINE BY LINE - SUMMARY").center(40) + 20 * "=")
+                self.print_memory_trace_statistics(inference_summary)
+
+        if self.args.training:
+            if self.args.speed:
+                self.print_fn("\n" + 20 * "=" + ("TRAIN - SPEED - RESULTS").center(40) + 20 * "=")
+                self.print_results(train_result_time, "Time in s")
+                self.save_to_csv(train_result_time, self.args.train_time_csv_file)
+                if self.args.is_tpu:
+                    self.print_fn(
+                        "TPU was used for training. Note that the time after compilation stabilized (after ~10 train"
+                        " loss=model.forward(...) + loss.backward() calls) was measured."
+                    )
+
+            if self.args.memory:
+                self.print_fn("\n" + 20 * "=" + ("TRAIN - MEMORY - RESULTS").center(40) + 20 * "=")
+                self.print_results(train_result_memory, type_label="Memory in MB")
+                self.save_to_csv(train_result_memory, self.args.train_memory_csv_file)
+
+            if self.args.trace_memory_line_by_line:
+                self.print_fn("\n" + 20 * "=" + ("TRAIN - MEMOMRY - LINE BY LINE - SUMMARY").center(40) + 20 * "=")
+                self.print_memory_trace_statistics(train_summary)
+
+        if self.args.env_print:
+            self.print_fn("\n" + 20 * "=" + ("ENVIRONMENT INFORMATION").center(40) + 20 * "=")
+            self.print_fn("\n".join([f"- {prop}: {val}" for prop, val in self.environment_info.items()]) + "\n")
+
+        if self.args.save_to_csv:
+            with open(self.args.env_info_csv_file, mode="w", newline="") as csv_file:
+                writer = csv.writer(csv_file)
+                for key, value in self.environment_info.items():
+                    writer.writerow([key, value])
+
+        return BenchmarkOutput(
+            inference_result_time,
+            inference_result_memory,
+            train_result_time,
+            train_result_memory,
+            inference_summary,
+            train_summary,
+        )
+
+    @property
+    def environment_info(self):
+        if self._environment_info is None:
+            info = {}
+            info["transformers_version"] = version
+            info["framework"] = self.framework
+            if self.framework == "PyTorch":
+                info["use_torchscript"] = self.args.torchscript
+            if self.framework == "TensorFlow":
+                info["eager_mode"] = self.args.eager_mode
+                info["use_xla"] = self.args.use_xla
+            info["framework_version"] = self.framework_version
+            info["python_version"] = platform.python_version()
+            info["system"] = platform.system()
+            info["cpu"] = platform.processor()
+            info["architecture"] = platform.architecture()[0]
+            info["date"] = datetime.date(datetime.now())
+            info["time"] = datetime.time(datetime.now())
+            info["fp16"] = self.args.fp16
+            info["use_multiprocessing"] = self.args.do_multi_processing
+            info["only_pretrain_model"] = self.args.only_pretrain_model
+
+            if is_psutil_available():
+                info["cpu_ram_mb"] = bytes_to_mega_bytes(psutil.virtual_memory().total)
+            else:
+                logger.warning(
+                    "Psutil not installed, we won't log available CPU memory. "
+                    "Install psutil (pip install psutil) to log available CPU memory."
+                )
+                info["cpu_ram_mb"] = "N/A"
+
+            info["use_gpu"] = self.args.is_gpu
+            if self.args.is_gpu:
+                info["num_gpus"] = 1  # TODO(PVP) Currently only single GPU is supported
+                if is_py3nvml_available():
+                    nvml.nvmlInit()
+                    handle = nvml.nvmlDeviceGetHandleByIndex(self.args.device_idx)
+                    info["gpu"] = nvml.nvmlDeviceGetName(handle)
+                    info["gpu_ram_mb"] = bytes_to_mega_bytes(nvml.nvmlDeviceGetMemoryInfo(handle).total)
+                    info["gpu_power_watts"] = nvml.nvmlDeviceGetPowerManagementLimit(handle) / 1000
+                    info["gpu_performance_state"] = nvml.nvmlDeviceGetPerformanceState(handle)
+                    nvml.nvmlShutdown()
+                else:
+                    logger.warning(
+                        "py3nvml not installed, we won't log GPU memory usage. "
+                        "Install py3nvml (pip install py3nvml) to log information about GPU."
+                    )
+                    info["gpu"] = "N/A"
+                    info["gpu_ram_mb"] = "N/A"
+                    info["gpu_power_watts"] = "N/A"
+                    info["gpu_performance_state"] = "N/A"
+
+            info["use_tpu"] = self.args.is_tpu
+            # TODO(PVP): See if we can add more information about TPU
+            # see: https://github.com/pytorch/xla/issues/2180
+
+            self._environment_info = info
+        return self._environment_info
+
+    def print_results(self, result_dict, type_label):
+        self.print_fn(80 * "-")
+        self.print_fn(
+            "Model Name".center(30) + "Batch Size".center(15) + "Seq Length".center(15) + type_label.center(15)
+        )
+        self.print_fn(80 * "-")
+        for model_name in self.args.model_names:
+            for batch_size in result_dict[model_name]["bs"]:
+                for sequence_length in result_dict[model_name]["ss"]:
+                    result = result_dict[model_name]["result"][batch_size][sequence_length]
+                    if isinstance(result, float):
+                        result = round(1000 * result) / 1000
+                        result = "< 0.001" if result == 0.0 else str(result)
+                    else:
+                        result = str(result)
+                    self.print_fn(
+                        model_name[:30].center(30) + str(batch_size).center(15),
+                        str(sequence_length).center(15),
+                        result.center(15),
+                    )
+        self.print_fn(80 * "-")
+
+    def print_memory_trace_statistics(self, summary: MemorySummary):
+        self.print_fn(
+            "\nLine by line memory consumption:\n"
+            + "\n".join(
+                f"{state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+                for state in summary.sequential
+            )
+        )
+        self.print_fn(
+            "\nLines with top memory consumption:\n"
+            + "\n".join(
+                f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+                for state in summary.cumulative[:6]
+            )
+        )
+        self.print_fn(
+            "\nLines with lowest memory consumption:\n"
+            + "\n".join(
+                f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+                for state in summary.cumulative[-6:]
+            )
+        )
+        self.print_fn(f"\nTotal memory increase: {summary.total}")
+
+    def save_to_csv(self, result_dict, filename):
+        if not self.args.save_to_csv:
+            return
+        self.print_fn("Saving results to csv.")
+        with open(filename, mode="w") as csv_file:
+            assert len(self.args.model_names) > 0, f"At least 1 model should be defined, but got {self.model_names}"
+
+            fieldnames = ["model", "batch_size", "sequence_length"]
+            writer = csv.DictWriter(csv_file, fieldnames=fieldnames + ["result"])
+            writer.writeheader()
+
+            for model_name in self.args.model_names:
+                result_dict_model = result_dict[model_name]["result"]
+                for bs in result_dict_model:
+                    for ss in result_dict_model[bs]:
+                        result_model = result_dict_model[bs][ss]
+                        writer.writerow(
+                            {
+                                "model": model_name,
+                                "batch_size": bs,
+                                "sequence_length": ss,
+                                "result": ("{}" if not isinstance(result_model, float) else "{:.4f}").format(
+                                    result_model
+                                ),
+                            }
+                        )

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/add_new_model.py

@@ -0,0 +1,259 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+import shutil
+import warnings
+from argparse import ArgumentParser, Namespace
+from pathlib import Path
+from typing import List
+
+from ..utils import logging
+from . import BaseTransformersCLICommand
+
+
+try:
+    from cookiecutter.main import cookiecutter
+
+    _has_cookiecutter = True
+except ImportError:
+    _has_cookiecutter = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def add_new_model_command_factory(args: Namespace):
+    return AddNewModelCommand(args.testing, args.testing_file, path=args.path)
+
+
+class AddNewModelCommand(BaseTransformersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        add_new_model_parser = parser.add_parser("add-new-model")
+        add_new_model_parser.add_argument("--testing", action="store_true", help="If in testing mode.")
+        add_new_model_parser.add_argument("--testing_file", type=str, help="Configuration file on which to run.")
+        add_new_model_parser.add_argument(
+            "--path", type=str, help="Path to cookiecutter. Should only be used for testing purposes."
+        )
+        add_new_model_parser.set_defaults(func=add_new_model_command_factory)
+
+    def __init__(self, testing: bool, testing_file: str, path=None, *args):
+        self._testing = testing
+        self._testing_file = testing_file
+        self._path = path
+
+    def run(self):
+        warnings.warn(
+            "The command `transformers-cli add-new-model` is deprecated and will be removed in v5 of Transformers. "
+            "It is not actively maintained anymore, so might give a result that won't pass all tests and quality "
+            "checks, you should use `transformers-cli add-new-model-like` instead."
+        )
+        if not _has_cookiecutter:
+            raise ImportError(
+                "Model creation dependencies are required to use the `add_new_model` command. Install them by running "
+                "the following at the root of your `transformers` clone:\n\n\t$ pip install -e .[modelcreation]\n"
+            )
+        # Ensure that there is no other `cookiecutter-template-xxx` directory in the current working directory
+        directories = [directory for directory in os.listdir() if "cookiecutter-template-" == directory[:22]]
+        if len(directories) > 0:
+            raise ValueError(
+                "Several directories starting with `cookiecutter-template-` in current working directory. "
+                "Please clean your directory by removing all folders starting with `cookiecutter-template-` or "
+                "change your working directory."
+            )
+
+        path_to_transformer_root = (
+            Path(__file__).parent.parent.parent.parent if self._path is None else Path(self._path).parent.parent
+        )
+        path_to_cookiecutter = path_to_transformer_root / "templates" / "adding_a_new_model"
+
+        # Execute cookiecutter
+        if not self._testing:
+            cookiecutter(str(path_to_cookiecutter))
+        else:
+            with open(self._testing_file, "r") as configuration_file:
+                testing_configuration = json.load(configuration_file)
+
+            cookiecutter(
+                str(path_to_cookiecutter if self._path is None else self._path),
+                no_input=True,
+                extra_context=testing_configuration,
+            )
+
+        directory = [directory for directory in os.listdir() if "cookiecutter-template-" in directory[:22]][0]
+
+        # Retrieve configuration
+        with open(directory + "/configuration.json", "r") as configuration_file:
+            configuration = json.load(configuration_file)
+
+        lowercase_model_name = configuration["lowercase_modelname"]
+        generate_tensorflow_pytorch_and_flax = configuration["generate_tensorflow_pytorch_and_flax"]
+        os.remove(f"{directory}/configuration.json")
+
+        output_pytorch = "PyTorch" in generate_tensorflow_pytorch_and_flax
+        output_tensorflow = "TensorFlow" in generate_tensorflow_pytorch_and_flax
+        output_flax = "Flax" in generate_tensorflow_pytorch_and_flax
+
+        model_dir = f"{path_to_transformer_root}/src/transformers/models/{lowercase_model_name}"
+        os.makedirs(model_dir, exist_ok=True)
+        os.makedirs(f"{path_to_transformer_root}/tests/models/{lowercase_model_name}", exist_ok=True)
+
+        # Tests require submodules as they have parent imports
+        with open(f"{path_to_transformer_root}/tests/models/{lowercase_model_name}/__init__.py", "w"):
+            pass
+
+        shutil.move(
+            f"{directory}/__init__.py",
+            f"{model_dir}/__init__.py",
+        )
+        shutil.move(
+            f"{directory}/configuration_{lowercase_model_name}.py",
+            f"{model_dir}/configuration_{lowercase_model_name}.py",
+        )
+
+        def remove_copy_lines(path):
+            with open(path, "r") as f:
+                lines = f.readlines()
+            with open(path, "w") as f:
+                for line in lines:
+                    if "# Copied from transformers." not in line:
+                        f.write(line)
+
+        if output_pytorch:
+            if not self._testing:
+                remove_copy_lines(f"{directory}/modeling_{lowercase_model_name}.py")
+
+            shutil.move(
+                f"{directory}/modeling_{lowercase_model_name}.py",
+                f"{model_dir}/modeling_{lowercase_model_name}.py",
+            )
+
+            shutil.move(
+                f"{directory}/test_modeling_{lowercase_model_name}.py",
+                f"{path_to_transformer_root}/tests/models/{lowercase_model_name}/test_modeling_{lowercase_model_name}.py",
+            )
+        else:
+            os.remove(f"{directory}/modeling_{lowercase_model_name}.py")
+            os.remove(f"{directory}/test_modeling_{lowercase_model_name}.py")
+
+        if output_tensorflow:
+            if not self._testing:
+                remove_copy_lines(f"{directory}/modeling_tf_{lowercase_model_name}.py")
+
+            shutil.move(
+                f"{directory}/modeling_tf_{lowercase_model_name}.py",
+                f"{model_dir}/modeling_tf_{lowercase_model_name}.py",
+            )
+
+            shutil.move(
+                f"{directory}/test_modeling_tf_{lowercase_model_name}.py",
+                f"{path_to_transformer_root}/tests/models/{lowercase_model_name}/test_modeling_tf_{lowercase_model_name}.py",
+            )
+        else:
+            os.remove(f"{directory}/modeling_tf_{lowercase_model_name}.py")
+            os.remove(f"{directory}/test_modeling_tf_{lowercase_model_name}.py")
+
+        if output_flax:
+            if not self._testing:
+                remove_copy_lines(f"{directory}/modeling_flax_{lowercase_model_name}.py")
+
+            shutil.move(
+                f"{directory}/modeling_flax_{lowercase_model_name}.py",
+                f"{model_dir}/modeling_flax_{lowercase_model_name}.py",
+            )
+
+            shutil.move(
+                f"{directory}/test_modeling_flax_{lowercase_model_name}.py",
+                f"{path_to_transformer_root}/tests/models/{lowercase_model_name}/test_modeling_flax_{lowercase_model_name}.py",
+            )
+        else:
+            os.remove(f"{directory}/modeling_flax_{lowercase_model_name}.py")
+            os.remove(f"{directory}/test_modeling_flax_{lowercase_model_name}.py")
+
+        shutil.move(
+            f"{directory}/{lowercase_model_name}.mdx",
+            f"{path_to_transformer_root}/docs/source/en/model_doc/{lowercase_model_name}.mdx",
+        )
+
+        shutil.move(
+            f"{directory}/tokenization_{lowercase_model_name}.py",
+            f"{model_dir}/tokenization_{lowercase_model_name}.py",
+        )
+
+        shutil.move(
+            f"{directory}/tokenization_fast_{lowercase_model_name}.py",
+            f"{model_dir}/tokenization_{lowercase_model_name}_fast.py",
+        )
+
+        from os import fdopen, remove
+        from shutil import copymode, move
+        from tempfile import mkstemp
+
+        def replace(original_file: str, line_to_copy_below: str, lines_to_copy: List[str]):
+            # Create temp file
+            fh, abs_path = mkstemp()
+            line_found = False
+            with fdopen(fh, "w") as new_file:
+                with open(original_file) as old_file:
+                    for line in old_file:
+                        new_file.write(line)
+                        if line_to_copy_below in line:
+                            line_found = True
+                            for line_to_copy in lines_to_copy:
+                                new_file.write(line_to_copy)
+
+            if not line_found:
+                raise ValueError(f"Line {line_to_copy_below} was not found in file.")
+
+            # Copy the file permissions from the old file to the new file
+            copymode(original_file, abs_path)
+            # Remove original file
+            remove(original_file)
+            # Move new file
+            move(abs_path, original_file)
+
+        def skip_units(line):
+            return (
+                ("generating PyTorch" in line and not output_pytorch)
+                or ("generating TensorFlow" in line and not output_tensorflow)
+                or ("generating Flax" in line and not output_flax)
+            )
+
+        def replace_in_files(path_to_datafile):
+            with open(path_to_datafile) as datafile:
+                lines_to_copy = []
+                skip_file = False
+                skip_snippet = False
+                for line in datafile:
+                    if "# To replace in: " in line and "##" not in line:
+                        file_to_replace_in = line.split('"')[1]
+                        skip_file = skip_units(line)
+                    elif "# Below: " in line and "##" not in line:
+                        line_to_copy_below = line.split('"')[1]
+                        skip_snippet = skip_units(line)
+                    elif "# End." in line and "##" not in line:
+                        if not skip_file and not skip_snippet:
+                            replace(file_to_replace_in, line_to_copy_below, lines_to_copy)
+
+                        lines_to_copy = []
+                    elif "# Replace with" in line and "##" not in line:
+                        lines_to_copy = []
+                    elif "##" not in line:
+                        lines_to_copy.append(line)
+
+            remove(path_to_datafile)
+
+        replace_in_files(f"{directory}/to_replace_{lowercase_model_name}.py")
+        os.rmdir(directory)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/add_new_model_like.py

@@ -0,0 +1,1710 @@
+# Copyright 2021 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import difflib
+import json
+import os
+import re
+from argparse import ArgumentParser, Namespace
+from dataclasses import dataclass
+from datetime import date
+from itertools import chain
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Pattern, Tuple, Union
+
+from ..models import auto as auto_module
+from ..models.auto.configuration_auto import model_type_to_module_name
+from ..utils import is_flax_available, is_tf_available, is_torch_available, logging
+from . import BaseTransformersCLICommand
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+CURRENT_YEAR = date.today().year
+TRANSFORMERS_PATH = Path(__file__).parent.parent
+REPO_PATH = TRANSFORMERS_PATH.parent.parent
+
+
+@dataclass
+class ModelPatterns:
+    """
+    Holds the basic information about a new model for the add-new-model-like command.
+
+    Args:
+        model_name (`str`): The model name.
+        checkpoint (`str`): The checkpoint to use for doc examples.
+        model_type (`str`, *optional*):
+            The model type, the identifier used internally in the library like `bert` or `xlm-roberta`. Will default to
+            `model_name` lowercased with spaces replaced with minuses (-).
+        model_lower_cased (`str`, *optional*):
+            The lowercased version of the model name, to use for the module name or function names. Will default to
+            `model_name` lowercased with spaces and minuses replaced with underscores.
+        model_camel_cased (`str`, *optional*):
+            The camel-cased version of the model name, to use for the class names. Will default to `model_name`
+            camel-cased (with spaces and minuses both considered as word separators.
+        model_upper_cased (`str`, *optional*):
+            The uppercased version of the model name, to use for the constant names. Will default to `model_name`
+            uppercased with spaces and minuses replaced with underscores.
+        config_class (`str`, *optional*):
+            The tokenizer class associated with this model. Will default to `"{model_camel_cased}Config"`.
+        tokenizer_class (`str`, *optional*):
+            The tokenizer class associated with this model (leave to `None` for models that don't use a tokenizer).
+        image_processor_class (`str`, *optional*):
+            The image processor class associated with this model (leave to `None` for models that don't use an image
+            processor).
+        feature_extractor_class (`str`, *optional*):
+            The feature extractor class associated with this model (leave to `None` for models that don't use a feature
+            extractor).
+        processor_class (`str`, *optional*):
+            The processor class associated with this model (leave to `None` for models that don't use a processor).
+    """
+
+    model_name: str
+    checkpoint: str
+    model_type: Optional[str] = None
+    model_lower_cased: Optional[str] = None
+    model_camel_cased: Optional[str] = None
+    model_upper_cased: Optional[str] = None
+    config_class: Optional[str] = None
+    tokenizer_class: Optional[str] = None
+    image_processor_class: Optional[str] = None
+    feature_extractor_class: Optional[str] = None
+    processor_class: Optional[str] = None
+
+    def __post_init__(self):
+        if self.model_type is None:
+            self.model_type = self.model_name.lower().replace(" ", "-")
+        if self.model_lower_cased is None:
+            self.model_lower_cased = self.model_name.lower().replace(" ", "_").replace("-", "_")
+        if self.model_camel_cased is None:
+            # Split the model name on - and space
+            words = self.model_name.split(" ")
+            words = list(chain(*[w.split("-") for w in words]))
+            # Make sure each word is capitalized
+            words = [w[0].upper() + w[1:] for w in words]
+            self.model_camel_cased = "".join(words)
+        if self.model_upper_cased is None:
+            self.model_upper_cased = self.model_name.upper().replace(" ", "_").replace("-", "_")
+        if self.config_class is None:
+            self.config_class = f"{self.model_camel_cased}Config"
+
+
+ATTRIBUTE_TO_PLACEHOLDER = {
+    "config_class": "[CONFIG_CLASS]",
+    "tokenizer_class": "[TOKENIZER_CLASS]",
+    "image_processor_class": "[IMAGE_PROCESSOR_CLASS]",
+    "feature_extractor_class": "[FEATURE_EXTRACTOR_CLASS]",
+    "processor_class": "[PROCESSOR_CLASS]",
+    "checkpoint": "[CHECKPOINT]",
+    "model_type": "[MODEL_TYPE]",
+    "model_upper_cased": "[MODEL_UPPER_CASED]",
+    "model_camel_cased": "[MODEL_CAMELCASED]",
+    "model_lower_cased": "[MODEL_LOWER_CASED]",
+    "model_name": "[MODEL_NAME]",
+}
+
+
+def is_empty_line(line: str) -> bool:
+    """
+    Determines whether a line is empty or not.
+    """
+    return len(line) == 0 or line.isspace()
+
+
+def find_indent(line: str) -> int:
+    """
+    Returns the number of spaces that start a line indent.
+    """
+    search = re.search(r"^(\s*)(?:\S|$)", line)
+    if search is None:
+        return 0
+    return len(search.groups()[0])
+
+
+def parse_module_content(content: str) -> List[str]:
+    """
+    Parse the content of a module in the list of objects it defines.
+
+    Args:
+        content (`str`): The content to parse
+
+    Returns:
+        `List[str]`: The list of objects defined in the module.
+    """
+    objects = []
+    current_object = []
+    lines = content.split("\n")
+    # Doc-styler takes everything between two triple quotes in docstrings, so we need a fake """ here to go with this.
+    end_markers = [")", "]", "}", '"""']
+
+    for line in lines:
+        # End of an object
+        is_valid_object = len(current_object) > 0
+        if is_valid_object and len(current_object) == 1:
+            is_valid_object = not current_object[0].startswith("# Copied from")
+        if not is_empty_line(line) and find_indent(line) == 0 and is_valid_object:
+            # Closing parts should be included in current object
+            if line in end_markers:
+                current_object.append(line)
+                objects.append("\n".join(current_object))
+                current_object = []
+            else:
+                objects.append("\n".join(current_object))
+                current_object = [line]
+        else:
+            current_object.append(line)
+
+    # Add last object
+    if len(current_object) > 0:
+        objects.append("\n".join(current_object))
+
+    return objects
+
+
+def extract_block(content: str, indent_level: int = 0) -> str:
+    """Return the first block in `content` with the indent level `indent_level`.
+
+    The first line in `content` should be indented at `indent_level` level, otherwise an error will be thrown.
+
+    This method will immediately stop the search when a (non-empty) line with indent level less than `indent_level` is
+    encountered.
+
+    Args:
+        content (`str`): The content to parse
+        indent_level (`int`, *optional*, default to 0): The indent level of the blocks to search for
+
+    Returns:
+        `str`: The first block in `content` with the indent level `indent_level`.
+    """
+    current_object = []
+    lines = content.split("\n")
+    # Doc-styler takes everything between two triple quotes in docstrings, so we need a fake """ here to go with this.
+    end_markers = [")", "]", "}", '"""']
+
+    for idx, line in enumerate(lines):
+        if idx == 0 and indent_level > 0 and not is_empty_line(line) and find_indent(line) != indent_level:
+            raise ValueError(
+                f"When `indent_level > 0`, the first line in `content` should have indent level {indent_level}. Got "
+                f"{find_indent(line)} instead."
+            )
+
+        if find_indent(line) < indent_level and not is_empty_line(line):
+            break
+
+        # End of an object
+        is_valid_object = len(current_object) > 0
+        if (
+            not is_empty_line(line)
+            and not line.endswith(":")
+            and find_indent(line) == indent_level
+            and is_valid_object
+        ):
+            # Closing parts should be included in current object
+            if line.lstrip() in end_markers:
+                current_object.append(line)
+            return "\n".join(current_object)
+        else:
+            current_object.append(line)
+
+    # Add last object
+    if len(current_object) > 0:
+        return "\n".join(current_object)
+
+
+def add_content_to_text(
+    text: str,
+    content: str,
+    add_after: Optional[Union[str, Pattern]] = None,
+    add_before: Optional[Union[str, Pattern]] = None,
+    exact_match: bool = False,
+) -> str:
+    """
+    A utility to add some content inside a given text.
+
+    Args:
+       text (`str`): The text in which we want to insert some content.
+       content (`str`): The content to add.
+       add_after (`str` or `Pattern`):
+           The pattern to test on a line of `text`, the new content is added after the first instance matching it.
+       add_before (`str` or `Pattern`):
+           The pattern to test on a line of `text`, the new content is added before the first instance matching it.
+       exact_match (`bool`, *optional*, defaults to `False`):
+           A line is considered a match with `add_after` or `add_before` if it matches exactly when `exact_match=True`,
+           otherwise, if `add_after`/`add_before` is present in the line.
+
+    <Tip warning={true}>
+
+    The arguments `add_after` and `add_before` are mutually exclusive, and one exactly needs to be provided.
+
+    </Tip>
+
+    Returns:
+        `str`: The text with the new content added if a match was found.
+    """
+    if add_after is None and add_before is None:
+        raise ValueError("You need to pass either `add_after` or `add_before`")
+    if add_after is not None and add_before is not None:
+        raise ValueError("You can't pass both `add_after` or `add_before`")
+    pattern = add_after if add_before is None else add_before
+
+    def this_is_the_line(line):
+        if isinstance(pattern, Pattern):
+            return pattern.search(line) is not None
+        elif exact_match:
+            return pattern == line
+        else:
+            return pattern in line
+
+    new_lines = []
+    for line in text.split("\n"):
+        if this_is_the_line(line):
+            if add_before is not None:
+                new_lines.append(content)
+            new_lines.append(line)
+            if add_after is not None:
+                new_lines.append(content)
+        else:
+            new_lines.append(line)
+
+    return "\n".join(new_lines)
+
+
+def add_content_to_file(
+    file_name: Union[str, os.PathLike],
+    content: str,
+    add_after: Optional[Union[str, Pattern]] = None,
+    add_before: Optional[Union[str, Pattern]] = None,
+    exact_match: bool = False,
+):
+    """
+    A utility to add some content inside a given file.
+
+    Args:
+       file_name (`str` or `os.PathLike`): The name of the file in which we want to insert some content.
+       content (`str`): The content to add.
+       add_after (`str` or `Pattern`):
+           The pattern to test on a line of `text`, the new content is added after the first instance matching it.
+       add_before (`str` or `Pattern`):
+           The pattern to test on a line of `text`, the new content is added before the first instance matching it.
+       exact_match (`bool`, *optional*, defaults to `False`):
+           A line is considered a match with `add_after` or `add_before` if it matches exactly when `exact_match=True`,
+           otherwise, if `add_after`/`add_before` is present in the line.
+
+    <Tip warning={true}>
+
+    The arguments `add_after` and `add_before` are mutually exclusive, and one exactly needs to be provided.
+
+    </Tip>
+    """
+    with open(file_name, "r", encoding="utf-8") as f:
+        old_content = f.read()
+
+    new_content = add_content_to_text(
+        old_content, content, add_after=add_after, add_before=add_before, exact_match=exact_match
+    )
+
+    with open(file_name, "w", encoding="utf-8") as f:
+        f.write(new_content)
+
+
+def replace_model_patterns(
+    text: str, old_model_patterns: ModelPatterns, new_model_patterns: ModelPatterns
+) -> Tuple[str, str]:
+    """
+    Replace all patterns present in a given text.
+
+    Args:
+        text (`str`): The text to treat.
+        old_model_patterns (`ModelPatterns`): The patterns for the old model.
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+
+    Returns:
+        `Tuple(str, str)`: A tuple of with the treated text and the replacement actually done in it.
+    """
+    # The order is crucially important as we will check and replace in that order. For instance the config probably
+    # contains the camel-cased named, but will be treated before.
+    attributes_to_check = ["config_class"]
+    # Add relevant preprocessing classes
+    for attr in ["tokenizer_class", "image_processor_class", "feature_extractor_class", "processor_class"]:
+        if getattr(old_model_patterns, attr) is not None and getattr(new_model_patterns, attr) is not None:
+            attributes_to_check.append(attr)
+
+    # Special cases for checkpoint and model_type
+    if old_model_patterns.checkpoint not in [old_model_patterns.model_type, old_model_patterns.model_lower_cased]:
+        attributes_to_check.append("checkpoint")
+    if old_model_patterns.model_type != old_model_patterns.model_lower_cased:
+        attributes_to_check.append("model_type")
+    else:
+        text = re.sub(
+            rf'(\s*)model_type = "{old_model_patterns.model_type}"',
+            r'\1model_type = "[MODEL_TYPE]"',
+            text,
+        )
+
+    # Special case when the model camel cased and upper cased names are the same for the old model (like for GPT2) but
+    # not the new one. We can't just do a replace in all the text and will need a special regex
+    if old_model_patterns.model_upper_cased == old_model_patterns.model_camel_cased:
+        old_model_value = old_model_patterns.model_upper_cased
+        if re.search(rf"{old_model_value}_[A-Z_]*[^A-Z_]", text) is not None:
+            text = re.sub(rf"{old_model_value}([A-Z_]*)([^a-zA-Z_])", r"[MODEL_UPPER_CASED]\1\2", text)
+    else:
+        attributes_to_check.append("model_upper_cased")
+
+    attributes_to_check.extend(["model_camel_cased", "model_lower_cased", "model_name"])
+
+    # Now let's replace every other attribute by their placeholder
+    for attr in attributes_to_check:
+        text = text.replace(getattr(old_model_patterns, attr), ATTRIBUTE_TO_PLACEHOLDER[attr])
+
+    # Finally we can replace the placeholder byt the new values.
+    replacements = []
+    for attr, placeholder in ATTRIBUTE_TO_PLACEHOLDER.items():
+        if placeholder in text:
+            replacements.append((getattr(old_model_patterns, attr), getattr(new_model_patterns, attr)))
+            text = text.replace(placeholder, getattr(new_model_patterns, attr))
+
+    # If we have two inconsistent replacements, we don't return anything (ex: GPT2->GPT_NEW and GPT2->GPTNew)
+    old_replacement_values = [old for old, new in replacements]
+    if len(set(old_replacement_values)) != len(old_replacement_values):
+        return text, ""
+
+    replacements = simplify_replacements(replacements)
+    replacements = [f"{old}->{new}" for old, new in replacements]
+    return text, ",".join(replacements)
+
+
+def simplify_replacements(replacements):
+    """
+    Simplify a list of replacement patterns to make sure there are no needless ones.
+
+    For instance in the sequence "Bert->BertNew, BertConfig->BertNewConfig, bert->bert_new", the replacement
+    "BertConfig->BertNewConfig" is implied by "Bert->BertNew" so not needed.
+
+    Args:
+        replacements (`List[Tuple[str, str]]`): List of patterns (old, new)
+
+    Returns:
+        `List[Tuple[str, str]]`: The list of patterns simplified.
+    """
+    if len(replacements) <= 1:
+        # Nothing to simplify
+        return replacements
+
+    # Next let's sort replacements by length as a replacement can only "imply" another replacement if it's shorter.
+    replacements.sort(key=lambda x: len(x[0]))
+
+    idx = 0
+    while idx < len(replacements):
+        old, new = replacements[idx]
+        # Loop through all replacements after
+        j = idx + 1
+        while j < len(replacements):
+            old_2, new_2 = replacements[j]
+            # If the replacement is implied by the current one, we can drop it.
+            if old_2.replace(old, new) == new_2:
+                replacements.pop(j)
+            else:
+                j += 1
+        idx += 1
+
+    return replacements
+
+
+def get_module_from_file(module_file: Union[str, os.PathLike]) -> str:
+    """
+    Returns the module name corresponding to a module file.
+    """
+    full_module_path = Path(module_file).absolute()
+    module_parts = full_module_path.with_suffix("").parts
+
+    # Find the first part named transformers, starting from the end.
+    idx = len(module_parts) - 1
+    while idx >= 0 and module_parts[idx] != "transformers":
+        idx -= 1
+    if idx < 0:
+        raise ValueError(f"{module_file} is not a transformers module.")
+
+    return ".".join(module_parts[idx:])
+
+
+SPECIAL_PATTERNS = {
+    "_CHECKPOINT_FOR_DOC =": "checkpoint",
+    "_CONFIG_FOR_DOC =": "config_class",
+    "_TOKENIZER_FOR_DOC =": "tokenizer_class",
+    "_IMAGE_PROCESSOR_FOR_DOC =": "image_processor_class",
+    "_FEAT_EXTRACTOR_FOR_DOC =": "feature_extractor_class",
+    "_PROCESSOR_FOR_DOC =": "processor_class",
+}
+
+
+_re_class_func = re.compile(r"^(?:class|def)\s+([^\s:\(]+)\s*(?:\(|\:)", flags=re.MULTILINE)
+
+
+def remove_attributes(obj, target_attr):
+    """Remove `target_attr` in `obj`."""
+    lines = obj.split(os.linesep)
+
+    target_idx = None
+    for idx, line in enumerate(lines):
+        # search for assignment
+        if line.lstrip().startswith(f"{target_attr} = "):
+            target_idx = idx
+            break
+        # search for function/method definition
+        elif line.lstrip().startswith(f"def {target_attr}("):
+            target_idx = idx
+            break
+
+    # target not found
+    if target_idx is None:
+        return obj
+
+    line = lines[target_idx]
+    indent_level = find_indent(line)
+    # forward pass to find the ending of the block (including empty lines)
+    parsed = extract_block("\n".join(lines[target_idx:]), indent_level)
+    num_lines = len(parsed.split("\n"))
+    for idx in range(num_lines):
+        lines[target_idx + idx] = None
+
+    # backward pass to find comments or decorator
+    for idx in range(target_idx - 1, -1, -1):
+        line = lines[idx]
+        if (line.lstrip().startswith("#") or line.lstrip().startswith("@")) and find_indent(line) == indent_level:
+            lines[idx] = None
+        else:
+            break
+
+    new_obj = os.linesep.join([x for x in lines if x is not None])
+
+    return new_obj
+
+
+def duplicate_module(
+    module_file: Union[str, os.PathLike],
+    old_model_patterns: ModelPatterns,
+    new_model_patterns: ModelPatterns,
+    dest_file: Optional[str] = None,
+    add_copied_from: bool = True,
+    attrs_to_remove: List[str] = None,
+):
+    """
+    Create a new module from an existing one and adapting all function and classes names from old patterns to new ones.
+
+    Args:
+        module_file (`str` or `os.PathLike`): Path to the module to duplicate.
+        old_model_patterns (`ModelPatterns`): The patterns for the old model.
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+        dest_file (`str` or `os.PathLike`, *optional*): Path to the new module.
+        add_copied_from (`bool`, *optional*, defaults to `True`):
+            Whether or not to add `# Copied from` statements in the duplicated module.
+    """
+    if dest_file is None:
+        dest_file = str(module_file).replace(
+            old_model_patterns.model_lower_cased, new_model_patterns.model_lower_cased
+        )
+
+    with open(module_file, "r", encoding="utf-8") as f:
+        content = f.read()
+
+    content = re.sub(r"# Copyright (\d+)\s", f"# Copyright {CURRENT_YEAR} ", content)
+    objects = parse_module_content(content)
+
+    # Loop and treat all objects
+    new_objects = []
+    for obj in objects:
+        # Special cases
+        if "PRETRAINED_CONFIG_ARCHIVE_MAP = {" in obj:
+            # docstyle-ignore
+            obj = (
+                f"{new_model_patterns.model_upper_cased}_PRETRAINED_CONFIG_ARCHIVE_MAP = "
+                + "{"
+                + f"""
+    "{new_model_patterns.checkpoint}": "https://huggingface.co/{new_model_patterns.checkpoint}/resolve/main/config.json",
+"""
+                + "}\n"
+            )
+            new_objects.append(obj)
+            continue
+        elif "PRETRAINED_MODEL_ARCHIVE_LIST = [" in obj:
+            if obj.startswith("TF_"):
+                prefix = "TF_"
+            elif obj.startswith("FLAX_"):
+                prefix = "FLAX_"
+            else:
+                prefix = ""
+            # docstyle-ignore
+            obj = f"""{prefix}{new_model_patterns.model_upper_cased}_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "{new_model_patterns.checkpoint}",
+    # See all {new_model_patterns.model_name} models at https://huggingface.co/models?filter={new_model_patterns.model_type}
+]
+"""
+            new_objects.append(obj)
+            continue
+
+        special_pattern = False
+        for pattern, attr in SPECIAL_PATTERNS.items():
+            if pattern in obj:
+                obj = obj.replace(getattr(old_model_patterns, attr), getattr(new_model_patterns, attr))
+                new_objects.append(obj)
+                special_pattern = True
+                break
+
+        if special_pattern:
+            continue
+
+        # Regular classes functions
+        old_obj = obj
+        obj, replacement = replace_model_patterns(obj, old_model_patterns, new_model_patterns)
+        has_copied_from = re.search(r"^#\s+Copied from", obj, flags=re.MULTILINE) is not None
+        if add_copied_from and not has_copied_from and _re_class_func.search(obj) is not None and len(replacement) > 0:
+            # Copied from statement must be added just before the class/function definition, which may not be the
+            # first line because of decorators.
+            module_name = get_module_from_file(module_file)
+            old_object_name = _re_class_func.search(old_obj).groups()[0]
+            obj = add_content_to_text(
+                obj, f"# Copied from {module_name}.{old_object_name} with {replacement}", add_before=_re_class_func
+            )
+        # In all cases, we remove Copied from statement with indent on methods.
+        obj = re.sub("\n[ ]+# Copied from [^\n]*\n", "\n", obj)
+
+        new_objects.append(obj)
+
+    content = "\n".join(new_objects)
+    # Remove some attributes that we don't want to copy to the new file(s)
+    if attrs_to_remove is not None:
+        for attr in attrs_to_remove:
+            content = remove_attributes(content, target_attr=attr)
+
+    with open(dest_file, "w", encoding="utf-8") as f:
+        f.write(content)
+
+
+def filter_framework_files(
+    files: List[Union[str, os.PathLike]], frameworks: Optional[List[str]] = None
+) -> List[Union[str, os.PathLike]]:
+    """
+    Filter a list of files to only keep the ones corresponding to a list of frameworks.
+
+    Args:
+        files (`List[Union[str, os.PathLike]]`): The list of files to filter.
+        frameworks (`List[str]`, *optional*): The list of allowed frameworks.
+
+    Returns:
+        `List[Union[str, os.PathLike]]`: The list of filtered files.
+    """
+    if frameworks is None:
+        frameworks = get_default_frameworks()
+
+    framework_to_file = {}
+    others = []
+    for f in files:
+        parts = Path(f).name.split("_")
+        if "modeling" not in parts:
+            others.append(f)
+            continue
+        if "tf" in parts:
+            framework_to_file["tf"] = f
+        elif "flax" in parts:
+            framework_to_file["flax"] = f
+        else:
+            framework_to_file["pt"] = f
+
+    return [framework_to_file[f] for f in frameworks if f in framework_to_file] + others
+
+
+def get_model_files(model_type: str, frameworks: Optional[List[str]] = None) -> Dict[str, Union[Path, List[Path]]]:
+    """
+    Retrieves all the files associated to a model.
+
+    Args:
+        model_type (`str`): A valid model type (like "bert" or "gpt2")
+        frameworks (`List[str]`, *optional*):
+            If passed, will only keep the model files corresponding to the passed frameworks.
+
+    Returns:
+        `Dict[str, Union[Path, List[Path]]]`: A dictionary with the following keys:
+        - **doc_file** -- The documentation file for the model.
+        - **model_files** -- All the files in the model module.
+        - **test_files** -- The test files for the model.
+    """
+    module_name = model_type_to_module_name(model_type)
+
+    model_module = TRANSFORMERS_PATH / "models" / module_name
+    model_files = list(model_module.glob("*.py"))
+    model_files = filter_framework_files(model_files, frameworks=frameworks)
+
+    doc_file = REPO_PATH / "docs" / "source" / "en" / "model_doc" / f"{model_type}.mdx"
+
+    # Basic pattern for test files
+    test_files = [
+        f"test_modeling_{module_name}.py",
+        f"test_modeling_tf_{module_name}.py",
+        f"test_modeling_flax_{module_name}.py",
+        f"test_tokenization_{module_name}.py",
+        f"test_image_processing_{module_name}.py",
+        f"test_feature_extraction_{module_name}.py",
+        f"test_processor_{module_name}.py",
+    ]
+    test_files = filter_framework_files(test_files, frameworks=frameworks)
+    # Add the test directory
+    test_files = [REPO_PATH / "tests" / "models" / module_name / f for f in test_files]
+    # Filter by existing files
+    test_files = [f for f in test_files if f.exists()]
+
+    return {"doc_file": doc_file, "model_files": model_files, "module_name": module_name, "test_files": test_files}
+
+
+_re_checkpoint_for_doc = re.compile(r"^_CHECKPOINT_FOR_DOC\s+=\s+(\S*)\s*$", flags=re.MULTILINE)
+
+
+def find_base_model_checkpoint(
+    model_type: str, model_files: Optional[Dict[str, Union[Path, List[Path]]]] = None
+) -> str:
+    """
+    Finds the model checkpoint used in the docstrings for a given model.
+
+    Args:
+        model_type (`str`): A valid model type (like "bert" or "gpt2")
+        model_files (`Dict[str, Union[Path, List[Path]]`, *optional*):
+            The files associated to `model_type`. Can be passed to speed up the function, otherwise will be computed.
+
+    Returns:
+        `str`: The checkpoint used.
+    """
+    if model_files is None:
+        model_files = get_model_files(model_type)
+    module_files = model_files["model_files"]
+    for fname in module_files:
+        if "modeling" not in str(fname):
+            continue
+
+        with open(fname, "r", encoding="utf-8") as f:
+            content = f.read()
+            if _re_checkpoint_for_doc.search(content) is not None:
+                checkpoint = _re_checkpoint_for_doc.search(content).groups()[0]
+                # Remove quotes
+                checkpoint = checkpoint.replace('"', "")
+                checkpoint = checkpoint.replace("'", "")
+                return checkpoint
+
+    # TODO: Find some kind of fallback if there is no _CHECKPOINT_FOR_DOC in any of the modeling file.
+    return ""
+
+
+def get_default_frameworks():
+    """
+    Returns the list of frameworks (PyTorch, TensorFlow, Flax) that are installed in the environment.
+    """
+    frameworks = []
+    if is_torch_available():
+        frameworks.append("pt")
+    if is_tf_available():
+        frameworks.append("tf")
+    if is_flax_available():
+        frameworks.append("flax")
+    return frameworks
+
+
+_re_model_mapping = re.compile("MODEL_([A-Z_]*)MAPPING_NAMES")
+
+
+def retrieve_model_classes(model_type: str, frameworks: Optional[List[str]] = None) -> Dict[str, List[str]]:
+    """
+    Retrieve the model classes associated to a given model.
+
+    Args:
+        model_type (`str`): A valid model type (like "bert" or "gpt2")
+        frameworks (`List[str]`, *optional*):
+            The frameworks to look for. Will default to `["pt", "tf", "flax"]`, passing a smaller list will restrict
+            the classes returned.
+
+    Returns:
+        `Dict[str, List[str]]`: A dictionary with one key per framework and the list of model classes associated to
+        that framework as values.
+    """
+    if frameworks is None:
+        frameworks = get_default_frameworks()
+
+    modules = {
+        "pt": auto_module.modeling_auto if is_torch_available() else None,
+        "tf": auto_module.modeling_tf_auto if is_tf_available() else None,
+        "flax": auto_module.modeling_flax_auto if is_flax_available() else None,
+    }
+
+    model_classes = {}
+    for framework in frameworks:
+        new_model_classes = []
+        if modules[framework] is None:
+            raise ValueError(f"You selected {framework} in the frameworks, but it is not installed.")
+        model_mappings = [attr for attr in dir(modules[framework]) if _re_model_mapping.search(attr) is not None]
+        for model_mapping_name in model_mappings:
+            model_mapping = getattr(modules[framework], model_mapping_name)
+            if model_type in model_mapping:
+                new_model_classes.append(model_mapping[model_type])
+
+        if len(new_model_classes) > 0:
+            # Remove duplicates
+            model_classes[framework] = list(set(new_model_classes))
+
+    return model_classes
+
+
+def retrieve_info_for_model(model_type, frameworks: Optional[List[str]] = None):
+    """
+    Retrieves all the information from a given model_type.
+
+    Args:
+        model_type (`str`): A valid model type (like "bert" or "gpt2")
+        frameworks (`List[str]`, *optional*):
+            If passed, will only keep the info corresponding to the passed frameworks.
+
+    Returns:
+        `Dict`: A dictionary with the following keys:
+        - **frameworks** (`List[str]`): The list of frameworks that back this model type.
+        - **model_classes** (`Dict[str, List[str]]`): The model classes implemented for that model type.
+        - **model_files** (`Dict[str, Union[Path, List[Path]]]`): The files associated with that model type.
+        - **model_patterns** (`ModelPatterns`): The various patterns for the model.
+    """
+    if model_type not in auto_module.MODEL_NAMES_MAPPING:
+        raise ValueError(f"{model_type} is not a valid model type.")
+
+    model_name = auto_module.MODEL_NAMES_MAPPING[model_type]
+    config_class = auto_module.configuration_auto.CONFIG_MAPPING_NAMES[model_type]
+    archive_map = auto_module.configuration_auto.CONFIG_ARCHIVE_MAP_MAPPING_NAMES.get(model_type, None)
+    if model_type in auto_module.tokenization_auto.TOKENIZER_MAPPING_NAMES:
+        tokenizer_classes = auto_module.tokenization_auto.TOKENIZER_MAPPING_NAMES[model_type]
+        tokenizer_class = tokenizer_classes[0] if tokenizer_classes[0] is not None else tokenizer_classes[1]
+    else:
+        tokenizer_class = None
+    image_processor_class = auto_module.image_processing_auto.IMAGE_PROCESSOR_MAPPING_NAMES.get(model_type, None)
+    feature_extractor_class = auto_module.feature_extraction_auto.FEATURE_EXTRACTOR_MAPPING_NAMES.get(model_type, None)
+    processor_class = auto_module.processing_auto.PROCESSOR_MAPPING_NAMES.get(model_type, None)
+
+    model_files = get_model_files(model_type, frameworks=frameworks)
+    model_camel_cased = config_class.replace("Config", "")
+
+    available_frameworks = []
+    for fname in model_files["model_files"]:
+        if "modeling_tf" in str(fname):
+            available_frameworks.append("tf")
+        elif "modeling_flax" in str(fname):
+            available_frameworks.append("flax")
+        elif "modeling" in str(fname):
+            available_frameworks.append("pt")
+
+    if frameworks is None:
+        frameworks = get_default_frameworks()
+
+    frameworks = [f for f in frameworks if f in available_frameworks]
+
+    model_classes = retrieve_model_classes(model_type, frameworks=frameworks)
+
+    # Retrieve model upper-cased name from the constant name of the pretrained archive map.
+    if archive_map is None:
+        model_upper_cased = model_camel_cased.upper()
+    else:
+        parts = archive_map.split("_")
+        idx = 0
+        while idx < len(parts) and parts[idx] != "PRETRAINED":
+            idx += 1
+        if idx < len(parts):
+            model_upper_cased = "_".join(parts[:idx])
+        else:
+            model_upper_cased = model_camel_cased.upper()
+
+    model_patterns = ModelPatterns(
+        model_name,
+        checkpoint=find_base_model_checkpoint(model_type, model_files=model_files),
+        model_type=model_type,
+        model_camel_cased=model_camel_cased,
+        model_lower_cased=model_files["module_name"],
+        model_upper_cased=model_upper_cased,
+        config_class=config_class,
+        tokenizer_class=tokenizer_class,
+        image_processor_class=image_processor_class,
+        feature_extractor_class=feature_extractor_class,
+        processor_class=processor_class,
+    )
+
+    return {
+        "frameworks": frameworks,
+        "model_classes": model_classes,
+        "model_files": model_files,
+        "model_patterns": model_patterns,
+    }
+
+
+def clean_frameworks_in_init(
+    init_file: Union[str, os.PathLike], frameworks: Optional[List[str]] = None, keep_processing: bool = True
+):
+    """
+    Removes all the import lines that don't belong to a given list of frameworks or concern tokenizers/feature
+    extractors/image processors/processors in an init.
+
+    Args:
+        init_file (`str` or `os.PathLike`): The path to the init to treat.
+        frameworks (`List[str]`, *optional*):
+           If passed, this will remove all imports that are subject to a framework not in frameworks
+        keep_processing (`bool`, *optional*, defaults to `True`):
+            Whether or not to keep the preprocessing (tokenizer, feature extractor, image processor, processor) imports
+            in the init.
+    """
+    if frameworks is None:
+        frameworks = get_default_frameworks()
+
+    names = {"pt": "torch"}
+    to_remove = [names.get(f, f) for f in ["pt", "tf", "flax"] if f not in frameworks]
+    if not keep_processing:
+        to_remove.extend(["sentencepiece", "tokenizers", "vision"])
+
+    if len(to_remove) == 0:
+        # Nothing to do
+        return
+
+    remove_pattern = "|".join(to_remove)
+    re_conditional_imports = re.compile(rf"^\s*if not is_({remove_pattern})_available\(\):\s*$")
+    re_try = re.compile(r"\s*try:")
+    re_else = re.compile(r"\s*else:")
+    re_is_xxx_available = re.compile(rf"is_({remove_pattern})_available")
+
+    with open(init_file, "r", encoding="utf-8") as f:
+        content = f.read()
+
+    lines = content.split("\n")
+    new_lines = []
+    idx = 0
+    while idx < len(lines):
+        # Conditional imports in try-except-else blocks
+        if (re_conditional_imports.search(lines[idx]) is not None) and (re_try.search(lines[idx - 1]) is not None):
+            # Remove the preceding `try:`
+            new_lines.pop()
+            idx += 1
+            # Iterate until `else:`
+            while is_empty_line(lines[idx]) or re_else.search(lines[idx]) is None:
+                idx += 1
+            idx += 1
+            indent = find_indent(lines[idx])
+            while find_indent(lines[idx]) >= indent or is_empty_line(lines[idx]):
+                idx += 1
+        # Remove the import from utils
+        elif re_is_xxx_available.search(lines[idx]) is not None:
+            line = lines[idx]
+            for framework in to_remove:
+                line = line.replace(f", is_{framework}_available", "")
+                line = line.replace(f"is_{framework}_available, ", "")
+                line = line.replace(f"is_{framework}_available,", "")
+                line = line.replace(f"is_{framework}_available", "")
+
+            if len(line.strip()) > 0:
+                new_lines.append(line)
+            idx += 1
+        # Otherwise we keep the line, except if it's a tokenizer import and we don't want to keep it.
+        elif keep_processing or (
+            re.search(r'^\s*"(tokenization|processing|feature_extraction|image_processing)', lines[idx]) is None
+            and re.search(r"^\s*from .(tokenization|processing|feature_extraction|image_processing)", lines[idx])
+            is None
+        ):
+            new_lines.append(lines[idx])
+            idx += 1
+        else:
+            idx += 1
+
+    with open(init_file, "w", encoding="utf-8") as f:
+        f.write("\n".join(new_lines))
+
+
+def add_model_to_main_init(
+    old_model_patterns: ModelPatterns,
+    new_model_patterns: ModelPatterns,
+    frameworks: Optional[List[str]] = None,
+    with_processing: bool = True,
+):
+    """
+    Add a model to the main init of Transformers.
+
+    Args:
+        old_model_patterns (`ModelPatterns`): The patterns for the old model.
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+        frameworks (`List[str]`, *optional*):
+            If specified, only the models implemented in those frameworks will be added.
+        with_processsing (`bool`, *optional*, defaults to `True`):
+            Whether the tokenizer/feature extractor/processor of the model should also be added to the init or not.
+    """
+    with open(TRANSFORMERS_PATH / "__init__.py", "r", encoding="utf-8") as f:
+        content = f.read()
+
+    lines = content.split("\n")
+    idx = 0
+    new_lines = []
+    framework = None
+    while idx < len(lines):
+        new_framework = False
+        if not is_empty_line(lines[idx]) and find_indent(lines[idx]) == 0:
+            framework = None
+        elif lines[idx].lstrip().startswith("if not is_torch_available"):
+            framework = "pt"
+            new_framework = True
+        elif lines[idx].lstrip().startswith("if not is_tf_available"):
+            framework = "tf"
+            new_framework = True
+        elif lines[idx].lstrip().startswith("if not is_flax_available"):
+            framework = "flax"
+            new_framework = True
+
+        if new_framework:
+            # For a new framework, we need to skip until the else: block to get where the imports are.
+            while lines[idx].strip() != "else:":
+                new_lines.append(lines[idx])
+                idx += 1
+
+        # Skip if we are in a framework not wanted.
+        if framework is not None and frameworks is not None and framework not in frameworks:
+            new_lines.append(lines[idx])
+            idx += 1
+        elif re.search(rf'models.{old_model_patterns.model_lower_cased}( |")', lines[idx]) is not None:
+            block = [lines[idx]]
+            indent = find_indent(lines[idx])
+            idx += 1
+            while find_indent(lines[idx]) > indent:
+                block.append(lines[idx])
+                idx += 1
+            if lines[idx].strip() in [")", "]", "],"]:
+                block.append(lines[idx])
+                idx += 1
+            block = "\n".join(block)
+            new_lines.append(block)
+
+            add_block = True
+            if not with_processing:
+                processing_classes = [
+                    old_model_patterns.tokenizer_class,
+                    old_model_patterns.image_processor_class,
+                    old_model_patterns.feature_extractor_class,
+                    old_model_patterns.processor_class,
+                ]
+                # Only keep the ones that are not None
+                processing_classes = [c for c in processing_classes if c is not None]
+                for processing_class in processing_classes:
+                    block = block.replace(f' "{processing_class}",', "")
+                    block = block.replace(f', "{processing_class}"', "")
+                    block = block.replace(f" {processing_class},", "")
+                    block = block.replace(f", {processing_class}", "")
+
+                    if processing_class in block:
+                        add_block = False
+            if add_block:
+                new_lines.append(replace_model_patterns(block, old_model_patterns, new_model_patterns)[0])
+        else:
+            new_lines.append(lines[idx])
+            idx += 1
+
+    with open(TRANSFORMERS_PATH / "__init__.py", "w", encoding="utf-8") as f:
+        f.write("\n".join(new_lines))
+
+
+def insert_tokenizer_in_auto_module(old_model_patterns: ModelPatterns, new_model_patterns: ModelPatterns):
+    """
+    Add a tokenizer to the relevant mappings in the auto module.
+
+    Args:
+        old_model_patterns (`ModelPatterns`): The patterns for the old model.
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+    """
+    if old_model_patterns.tokenizer_class is None or new_model_patterns.tokenizer_class is None:
+        return
+
+    with open(TRANSFORMERS_PATH / "models" / "auto" / "tokenization_auto.py", "r", encoding="utf-8") as f:
+        content = f.read()
+
+    lines = content.split("\n")
+    idx = 0
+    # First we get to the TOKENIZER_MAPPING_NAMES block.
+    while not lines[idx].startswith("    TOKENIZER_MAPPING_NAMES = OrderedDict("):
+        idx += 1
+    idx += 1
+
+    # That block will end at this prompt:
+    while not lines[idx].startswith("TOKENIZER_MAPPING = _LazyAutoMapping"):
+        # Either all the tokenizer block is defined on one line, in which case, it ends with "),"
+        if lines[idx].endswith(","):
+            block = lines[idx]
+        # Otherwise it takes several lines until we get to a "),"
+        else:
+            block = []
+            while not lines[idx].startswith("            ),"):
+                block.append(lines[idx])
+                idx += 1
+            block = "\n".join(block)
+        idx += 1
+
+        # If we find the model type and tokenizer class in that block, we have the old model tokenizer block
+        if f'"{old_model_patterns.model_type}"' in block and old_model_patterns.tokenizer_class in block:
+            break
+
+    new_block = block.replace(old_model_patterns.model_type, new_model_patterns.model_type)
+    new_block = new_block.replace(old_model_patterns.tokenizer_class, new_model_patterns.tokenizer_class)
+
+    new_lines = lines[:idx] + [new_block] + lines[idx:]
+    with open(TRANSFORMERS_PATH / "models" / "auto" / "tokenization_auto.py", "w", encoding="utf-8") as f:
+        f.write("\n".join(new_lines))
+
+
+AUTO_CLASSES_PATTERNS = {
+    "configuration_auto.py": [
+        '        ("{model_type}", "{model_name}"),',
+        '        ("{model_type}", "{config_class}"),',
+        '        ("{model_type}", "{pretrained_archive_map}"),',
+    ],
+    "feature_extraction_auto.py": ['        ("{model_type}", "{feature_extractor_class}"),'],
+    "image_processing_auto.py": ['        ("{model_type}", "{image_processor_class}"),'],
+    "modeling_auto.py": ['        ("{model_type}", "{any_pt_class}"),'],
+    "modeling_tf_auto.py": ['        ("{model_type}", "{any_tf_class}"),'],
+    "modeling_flax_auto.py": ['        ("{model_type}", "{any_flax_class}"),'],
+    "processing_auto.py": ['        ("{model_type}", "{processor_class}"),'],
+}
+
+
+def add_model_to_auto_classes(
+    old_model_patterns: ModelPatterns, new_model_patterns: ModelPatterns, model_classes: Dict[str, List[str]]
+):
+    """
+    Add a model to the relevant mappings in the auto module.
+
+    Args:
+        old_model_patterns (`ModelPatterns`): The patterns for the old model.
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+        model_classes (`Dict[str, List[str]]`): A dictionary framework to list of model classes implemented.
+    """
+    for filename in AUTO_CLASSES_PATTERNS:
+        # Extend patterns with all model classes if necessary
+        new_patterns = []
+        for pattern in AUTO_CLASSES_PATTERNS[filename]:
+            if re.search("any_([a-z]*)_class", pattern) is not None:
+                framework = re.search("any_([a-z]*)_class", pattern).groups()[0]
+                if framework in model_classes:
+                    new_patterns.extend(
+                        [
+                            pattern.replace("{" + f"any_{framework}_class" + "}", cls)
+                            for cls in model_classes[framework]
+                        ]
+                    )
+            elif "{config_class}" in pattern:
+                new_patterns.append(pattern.replace("{config_class}", old_model_patterns.config_class))
+            elif "{image_processor_class}" in pattern:
+                if (
+                    old_model_patterns.image_processor_class is not None
+                    and new_model_patterns.image_processor_class is not None
+                ):
+                    new_patterns.append(
+                        pattern.replace("{image_processor_class}", old_model_patterns.image_processor_class)
+                    )
+            elif "{feature_extractor_class}" in pattern:
+                if (
+                    old_model_patterns.feature_extractor_class is not None
+                    and new_model_patterns.feature_extractor_class is not None
+                ):
+                    new_patterns.append(
+                        pattern.replace("{feature_extractor_class}", old_model_patterns.feature_extractor_class)
+                    )
+            elif "{processor_class}" in pattern:
+                if old_model_patterns.processor_class is not None and new_model_patterns.processor_class is not None:
+                    new_patterns.append(pattern.replace("{processor_class}", old_model_patterns.processor_class))
+            else:
+                new_patterns.append(pattern)
+
+        # Loop through all patterns.
+        for pattern in new_patterns:
+            full_name = TRANSFORMERS_PATH / "models" / "auto" / filename
+            old_model_line = pattern
+            new_model_line = pattern
+            for attr in ["model_type", "model_name"]:
+                old_model_line = old_model_line.replace("{" + attr + "}", getattr(old_model_patterns, attr))
+                new_model_line = new_model_line.replace("{" + attr + "}", getattr(new_model_patterns, attr))
+            if "pretrained_archive_map" in pattern:
+                old_model_line = old_model_line.replace(
+                    "{pretrained_archive_map}", f"{old_model_patterns.model_upper_cased}_PRETRAINED_CONFIG_ARCHIVE_MAP"
+                )
+                new_model_line = new_model_line.replace(
+                    "{pretrained_archive_map}", f"{new_model_patterns.model_upper_cased}_PRETRAINED_CONFIG_ARCHIVE_MAP"
+                )
+
+            new_model_line = new_model_line.replace(
+                old_model_patterns.model_camel_cased, new_model_patterns.model_camel_cased
+            )
+
+            add_content_to_file(full_name, new_model_line, add_after=old_model_line)
+
+    # Tokenizers require special handling
+    insert_tokenizer_in_auto_module(old_model_patterns, new_model_patterns)
+
+
+DOC_OVERVIEW_TEMPLATE = """## Overview
+
+The {model_name} model was proposed in [<INSERT PAPER NAME HERE>](<INSERT PAPER LINK HERE>) by <INSERT AUTHORS HERE>.
+<INSERT SHORT SUMMARY HERE>
+
+The abstract from the paper is the following:
+
+*<INSERT PAPER ABSTRACT HERE>*
+
+Tips:
+
+<INSERT TIPS ABOUT MODEL HERE>
+
+This model was contributed by [INSERT YOUR HF USERNAME HERE](https://huggingface.co/<INSERT YOUR HF USERNAME HERE>).
+The original code can be found [here](<INSERT LINK TO GITHUB REPO HERE>).
+
+"""
+
+
+def duplicate_doc_file(
+    doc_file: Union[str, os.PathLike],
+    old_model_patterns: ModelPatterns,
+    new_model_patterns: ModelPatterns,
+    dest_file: Optional[Union[str, os.PathLike]] = None,
+    frameworks: Optional[List[str]] = None,
+):
+    """
+    Duplicate a documentation file and adapts it for a new model.
+
+    Args:
+        module_file (`str` or `os.PathLike`): Path to the doc file to duplicate.
+        old_model_patterns (`ModelPatterns`): The patterns for the old model.
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+        dest_file (`str` or `os.PathLike`, *optional*): Path to the new doc file.
+            Will default to the a file named `{new_model_patterns.model_type}.mdx` in the same folder as `module_file`.
+        frameworks (`List[str]`, *optional*):
+            If passed, will only keep the model classes corresponding to this list of frameworks in the new doc file.
+    """
+    with open(doc_file, "r", encoding="utf-8") as f:
+        content = f.read()
+
+    content = re.sub(r"<!--\s*Copyright (\d+)\s", f"<!--Copyright {CURRENT_YEAR} ", content)
+    if frameworks is None:
+        frameworks = get_default_frameworks()
+    if dest_file is None:
+        dest_file = Path(doc_file).parent / f"{new_model_patterns.model_type}.mdx"
+
+    # Parse the doc file in blocks. One block per section/header
+    lines = content.split("\n")
+    blocks = []
+    current_block = []
+
+    for line in lines:
+        if line.startswith("#"):
+            blocks.append("\n".join(current_block))
+            current_block = [line]
+        else:
+            current_block.append(line)
+    blocks.append("\n".join(current_block))
+
+    new_blocks = []
+    in_classes = False
+    for block in blocks:
+        # Copyright
+        if not block.startswith("#"):
+            new_blocks.append(block)
+        # Main title
+        elif re.search(r"^#\s+\S+", block) is not None:
+            new_blocks.append(f"# {new_model_patterns.model_name}\n")
+        # The config starts the part of the doc with the classes.
+        elif not in_classes and old_model_patterns.config_class in block.split("\n")[0]:
+            in_classes = True
+            new_blocks.append(DOC_OVERVIEW_TEMPLATE.format(model_name=new_model_patterns.model_name))
+            new_block, _ = replace_model_patterns(block, old_model_patterns, new_model_patterns)
+            new_blocks.append(new_block)
+        # In classes
+        elif in_classes:
+            in_classes = True
+            block_title = block.split("\n")[0]
+            block_class = re.search(r"^#+\s+(\S.*)$", block_title).groups()[0]
+            new_block, _ = replace_model_patterns(block, old_model_patterns, new_model_patterns)
+
+            if "Tokenizer" in block_class:
+                # We only add the tokenizer if necessary
+                if old_model_patterns.tokenizer_class != new_model_patterns.tokenizer_class:
+                    new_blocks.append(new_block)
+            elif "ImageProcessor" in block_class:
+                # We only add the image processor if necessary
+                if old_model_patterns.image_processor_class != new_model_patterns.image_processor_class:
+                    new_blocks.append(new_block)
+            elif "FeatureExtractor" in block_class:
+                # We only add the feature extractor if necessary
+                if old_model_patterns.feature_extractor_class != new_model_patterns.feature_extractor_class:
+                    new_blocks.append(new_block)
+            elif "Processor" in block_class:
+                # We only add the processor if necessary
+                if old_model_patterns.processor_class != new_model_patterns.processor_class:
+                    new_blocks.append(new_block)
+            elif block_class.startswith("Flax"):
+                # We only add Flax models if in the selected frameworks
+                if "flax" in frameworks:
+                    new_blocks.append(new_block)
+            elif block_class.startswith("TF"):
+                # We only add TF models if in the selected frameworks
+                if "tf" in frameworks:
+                    new_blocks.append(new_block)
+            elif len(block_class.split(" ")) == 1:
+                # We only add PyTorch models if in the selected frameworks
+                if "pt" in frameworks:
+                    new_blocks.append(new_block)
+            else:
+                new_blocks.append(new_block)
+
+    with open(dest_file, "w", encoding="utf-8") as f:
+        f.write("\n".join(new_blocks))
+
+
+def create_new_model_like(
+    model_type: str,
+    new_model_patterns: ModelPatterns,
+    add_copied_from: bool = True,
+    frameworks: Optional[List[str]] = None,
+    old_checkpoint: Optional[str] = None,
+):
+    """
+    Creates a new model module like a given model of the Transformers library.
+
+    Args:
+        model_type (`str`): The model type to duplicate (like "bert" or "gpt2")
+        new_model_patterns (`ModelPatterns`): The patterns for the new model.
+        add_copied_from (`bool`, *optional*, defaults to `True`):
+            Whether or not to add "Copied from" statements to all classes in the new model modeling files.
+        frameworks (`List[str]`, *optional*):
+            If passed, will limit the duplicate to the frameworks specified.
+        old_checkpoint (`str`, *optional*):
+            The name of the base checkpoint for the old model. Should be passed along when it can't be automatically
+            recovered from the `model_type`.
+    """
+    # Retrieve all the old model info.
+    model_info = retrieve_info_for_model(model_type, frameworks=frameworks)
+    model_files = model_info["model_files"]
+    old_model_patterns = model_info["model_patterns"]
+    if old_checkpoint is not None:
+        old_model_patterns.checkpoint = old_checkpoint
+    if len(old_model_patterns.checkpoint) == 0:
+        raise ValueError(
+            "The old model checkpoint could not be recovered from the model type. Please pass it to the "
+            "`old_checkpoint` argument."
+        )
+
+    keep_old_processing = True
+    for processing_attr in ["image_processor_class", "feature_extractor_class", "processor_class", "tokenizer_class"]:
+        if getattr(old_model_patterns, processing_attr) != getattr(new_model_patterns, processing_attr):
+            keep_old_processing = False
+
+    model_classes = model_info["model_classes"]
+
+    # 1. We create the module for our new model.
+    old_module_name = model_files["module_name"]
+    module_folder = TRANSFORMERS_PATH / "models" / new_model_patterns.model_lower_cased
+    os.makedirs(module_folder, exist_ok=True)
+
+    files_to_adapt = model_files["model_files"]
+    if keep_old_processing:
+        files_to_adapt = [
+            f
+            for f in files_to_adapt
+            if "tokenization" not in str(f)
+            and "processing" not in str(f)
+            and "feature_extraction" not in str(f)
+            and "image_processing" not in str(f)
+        ]
+
+    os.makedirs(module_folder, exist_ok=True)
+    for module_file in files_to_adapt:
+        new_module_name = module_file.name.replace(
+            old_model_patterns.model_lower_cased, new_model_patterns.model_lower_cased
+        )
+        dest_file = module_folder / new_module_name
+        duplicate_module(
+            module_file,
+            old_model_patterns,
+            new_model_patterns,
+            dest_file=dest_file,
+            add_copied_from=add_copied_from and "modeling" in new_module_name,
+        )
+
+    clean_frameworks_in_init(
+        module_folder / "__init__.py", frameworks=frameworks, keep_processing=not keep_old_processing
+    )
+
+    # 2. We add our new model to the models init and the main init
+    add_content_to_file(
+        TRANSFORMERS_PATH / "models" / "__init__.py",
+        f"    {new_model_patterns.model_lower_cased},",
+        add_after=f"    {old_module_name},",
+        exact_match=True,
+    )
+    add_model_to_main_init(
+        old_model_patterns, new_model_patterns, frameworks=frameworks, with_processing=not keep_old_processing
+    )
+
+    # 3. Add test files
+    files_to_adapt = model_files["test_files"]
+    if keep_old_processing:
+        files_to_adapt = [
+            f
+            for f in files_to_adapt
+            if "tokenization" not in str(f)
+            and "processor" not in str(f)
+            and "feature_extraction" not in str(f)
+            and "image_processing" not in str(f)
+        ]
+
+    def disable_fx_test(filename: Path) -> bool:
+        with open(filename) as fp:
+            content = fp.read()
+        new_content = re.sub(r"fx_compatible\s*=\s*True", "fx_compatible = False", content)
+        with open(filename, "w") as fp:
+            fp.write(new_content)
+        return content != new_content
+
+    disabled_fx_test = False
+
+    tests_folder = REPO_PATH / "tests" / "models" / new_model_patterns.model_lower_cased
+    os.makedirs(tests_folder, exist_ok=True)
+    with open(tests_folder / "__init__.py", "w"):
+        pass
+
+    for test_file in files_to_adapt:
+        new_test_file_name = test_file.name.replace(
+            old_model_patterns.model_lower_cased, new_model_patterns.model_lower_cased
+        )
+        dest_file = test_file.parent.parent / new_model_patterns.model_lower_cased / new_test_file_name
+        duplicate_module(
+            test_file,
+            old_model_patterns,
+            new_model_patterns,
+            dest_file=dest_file,
+            add_copied_from=False,
+            attrs_to_remove=["pipeline_model_mapping", "is_pipeline_test_to_skip"],
+        )
+        disabled_fx_test = disabled_fx_test | disable_fx_test(dest_file)
+
+    if disabled_fx_test:
+        print(
+            "The tests for symbolic tracing with torch.fx were disabled, you can add those once symbolic tracing works"
+            " for your new model."
+        )
+
+    # 4. Add model to auto classes
+    add_model_to_auto_classes(old_model_patterns, new_model_patterns, model_classes)
+
+    # 5. Add doc file
+    doc_file = REPO_PATH / "docs" / "source" / "en" / "model_doc" / f"{old_model_patterns.model_type}.mdx"
+    duplicate_doc_file(doc_file, old_model_patterns, new_model_patterns, frameworks=frameworks)
+
+    # 6. Warn the user for duplicate patterns
+    if old_model_patterns.model_type == old_model_patterns.checkpoint:
+        print(
+            "The model you picked has the same name for the model type and the checkpoint name "
+            f"({old_model_patterns.model_type}). As a result, it's possible some places where the new checkpoint "
+            f"should be, you have {new_model_patterns.model_type} instead. You should search for all instances of "
+            f"{new_model_patterns.model_type} in the new files and check they're not badly used as checkpoints."
+        )
+    elif old_model_patterns.model_lower_cased == old_model_patterns.checkpoint:
+        print(
+            "The model you picked has the same name for the model type and the checkpoint name "
+            f"({old_model_patterns.model_lower_cased}). As a result, it's possible some places where the new "
+            f"checkpoint should be, you have {new_model_patterns.model_lower_cased} instead. You should search for "
+            f"all instances of {new_model_patterns.model_lower_cased} in the new files and check they're not badly "
+            "used as checkpoints."
+        )
+    if (
+        old_model_patterns.model_type == old_model_patterns.model_lower_cased
+        and new_model_patterns.model_type != new_model_patterns.model_lower_cased
+    ):
+        print(
+            "The model you picked has the same name for the model type and the lowercased model name "
+            f"({old_model_patterns.model_lower_cased}). As a result, it's possible some places where the new "
+            f"model type should be, you have {new_model_patterns.model_lower_cased} instead. You should search for "
+            f"all instances of {new_model_patterns.model_lower_cased} in the new files and check they're not badly "
+            "used as the model type."
+        )
+
+    if not keep_old_processing and old_model_patterns.tokenizer_class is not None:
+        print(
+            "The constants at the start of the new tokenizer file created needs to be manually fixed. If your new "
+            "model has a tokenizer fast, you will also need to manually add the converter in the "
+            "`SLOW_TO_FAST_CONVERTERS` constant of `convert_slow_tokenizer.py`."
+        )
+
+
+def add_new_model_like_command_factory(args: Namespace):
+    return AddNewModelLikeCommand(config_file=args.config_file, path_to_repo=args.path_to_repo)
+
+
+class AddNewModelLikeCommand(BaseTransformersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        add_new_model_like_parser = parser.add_parser("add-new-model-like")
+        add_new_model_like_parser.add_argument(
+            "--config_file", type=str, help="A file with all the information for this model creation."
+        )
+        add_new_model_like_parser.add_argument(
+            "--path_to_repo", type=str, help="When not using an editable install, the path to the Transformers repo."
+        )
+        add_new_model_like_parser.set_defaults(func=add_new_model_like_command_factory)
+
+    def __init__(self, config_file=None, path_to_repo=None, *args):
+        if config_file is not None:
+            with open(config_file, "r", encoding="utf-8") as f:
+                config = json.load(f)
+            self.old_model_type = config["old_model_type"]
+            self.model_patterns = ModelPatterns(**config["new_model_patterns"])
+            self.add_copied_from = config.get("add_copied_from", True)
+            self.frameworks = config.get("frameworks", get_default_frameworks())
+            self.old_checkpoint = config.get("old_checkpoint", None)
+        else:
+            (
+                self.old_model_type,
+                self.model_patterns,
+                self.add_copied_from,
+                self.frameworks,
+                self.old_checkpoint,
+            ) = get_user_input()
+
+        self.path_to_repo = path_to_repo
+
+    def run(self):
+        if self.path_to_repo is not None:
+            # Adapt constants
+            global TRANSFORMERS_PATH
+            global REPO_PATH
+
+            REPO_PATH = Path(self.path_to_repo)
+            TRANSFORMERS_PATH = REPO_PATH / "src" / "transformers"
+
+        create_new_model_like(
+            model_type=self.old_model_type,
+            new_model_patterns=self.model_patterns,
+            add_copied_from=self.add_copied_from,
+            frameworks=self.frameworks,
+            old_checkpoint=self.old_checkpoint,
+        )
+
+
+def get_user_field(
+    question: str,
+    default_value: Optional[str] = None,
+    is_valid_answer: Optional[Callable] = None,
+    convert_to: Optional[Callable] = None,
+    fallback_message: Optional[str] = None,
+) -> Any:
+    """
+    A utility function that asks a question to the user to get an answer, potentially looping until it gets a valid
+    answer.
+
+    Args:
+        question (`str`): The question to ask the user.
+        default_value (`str`, *optional*): A potential default value that will be used when the answer is empty.
+        is_valid_answer (`Callable`, *optional*):
+            If set, the question will be asked until this function returns `True` on the provided answer.
+        convert_to (`Callable`, *optional*):
+            If set, the answer will be passed to this function. If this function raises an error on the procided
+            answer, the question will be asked again.
+        fallback_message (`str`, *optional*):
+            A message that will be displayed each time the question is asked again to the user.
+
+    Returns:
+        `Any`: The answer provided by the user (or the default), passed through the potential conversion function.
+    """
+    if not question.endswith(" "):
+        question = question + " "
+    if default_value is not None:
+        question = f"{question} [{default_value}] "
+
+    valid_answer = False
+    while not valid_answer:
+        answer = input(question)
+        if default_value is not None and len(answer) == 0:
+            answer = default_value
+        if is_valid_answer is not None:
+            valid_answer = is_valid_answer(answer)
+        elif convert_to is not None:
+            try:
+                answer = convert_to(answer)
+                valid_answer = True
+            except Exception:
+                valid_answer = False
+        else:
+            valid_answer = True
+
+        if not valid_answer:
+            print(fallback_message)
+
+    return answer
+
+
+def convert_to_bool(x: str) -> bool:
+    """
+    Converts a string to a bool.
+    """
+    if x.lower() in ["1", "y", "yes", "true"]:
+        return True
+    if x.lower() in ["0", "n", "no", "false"]:
+        return False
+    raise ValueError(f"{x} is not a value that can be converted to a bool.")
+
+
+def get_user_input():
+    """
+    Ask the user for the necessary inputs to add the new model.
+    """
+    model_types = list(auto_module.configuration_auto.MODEL_NAMES_MAPPING.keys())
+
+    # Get old model type
+    valid_model_type = False
+    while not valid_model_type:
+        old_model_type = input(
+            "What is the model you would like to duplicate? Please provide the lowercase `model_type` (e.g. roberta): "
+        )
+        if old_model_type in model_types:
+            valid_model_type = True
+        else:
+            print(f"{old_model_type} is not a valid model type.")
+            near_choices = difflib.get_close_matches(old_model_type, model_types)
+            if len(near_choices) >= 1:
+                if len(near_choices) > 1:
+                    near_choices = " or ".join(near_choices)
+                print(f"Did you mean {near_choices}?")
+
+    old_model_info = retrieve_info_for_model(old_model_type)
+    old_tokenizer_class = old_model_info["model_patterns"].tokenizer_class
+    old_image_processor_class = old_model_info["model_patterns"].image_processor_class
+    old_feature_extractor_class = old_model_info["model_patterns"].feature_extractor_class
+    old_processor_class = old_model_info["model_patterns"].processor_class
+    old_frameworks = old_model_info["frameworks"]
+
+    old_checkpoint = None
+    if len(old_model_info["model_patterns"].checkpoint) == 0:
+        old_checkpoint = get_user_field(
+            "We couldn't find the name of the base checkpoint for that model, please enter it here."
+        )
+
+    model_name = get_user_field(
+        "What is the name (with no special casing) for your new model in the paper (e.g. RoBERTa)? "
+    )
+    default_patterns = ModelPatterns(model_name, model_name)
+
+    model_type = get_user_field(
+        "What identifier would you like to use for the `model_type` of this model? ",
+        default_value=default_patterns.model_type,
+    )
+    model_lower_cased = get_user_field(
+        "What lowercase name would you like to use for the module (folder) of this model? ",
+        default_value=default_patterns.model_lower_cased,
+    )
+    model_camel_cased = get_user_field(
+        "What prefix (camel-cased) would you like to use for the model classes of this model (e.g. Roberta)? ",
+        default_value=default_patterns.model_camel_cased,
+    )
+    model_upper_cased = get_user_field(
+        "What prefix (upper-cased) would you like to use for the constants relative to this model? ",
+        default_value=default_patterns.model_upper_cased,
+    )
+    config_class = get_user_field(
+        "What will be the name of the config class for this model? ", default_value=f"{model_camel_cased}Config"
+    )
+    checkpoint = get_user_field(
+        "Please give a checkpoint identifier (on the model Hub) for this new model (e.g. facebook/roberta-base): "
+    )
+
+    old_processing_classes = [
+        c
+        for c in [old_image_processor_class, old_feature_extractor_class, old_tokenizer_class, old_processor_class]
+        if c is not None
+    ]
+    old_processing_classes = ", ".join(old_processing_classes)
+    keep_processing = get_user_field(
+        f"Will your new model use the same processing class as {old_model_type} ({old_processing_classes}) (yes/no)? ",
+        convert_to=convert_to_bool,
+        fallback_message="Please answer yes/no, y/n, true/false or 1/0. ",
+    )
+    if keep_processing:
+        image_processor_class = old_image_processor_class
+        feature_extractor_class = old_feature_extractor_class
+        processor_class = old_processor_class
+        tokenizer_class = old_tokenizer_class
+    else:
+        if old_tokenizer_class is not None:
+            tokenizer_class = get_user_field(
+                "What will be the name of the tokenizer class for this model? ",
+                default_value=f"{model_camel_cased}Tokenizer",
+            )
+        else:
+            tokenizer_class = None
+        if old_image_processor_class is not None:
+            image_processor_class = get_user_field(
+                "What will be the name of the image processor class for this model? ",
+                default_value=f"{model_camel_cased}ImageProcessor",
+            )
+        else:
+            image_processor_class = None
+        if old_feature_extractor_class is not None:
+            feature_extractor_class = get_user_field(
+                "What will be the name of the feature extractor class for this model? ",
+                default_value=f"{model_camel_cased}FeatureExtractor",
+            )
+        else:
+            feature_extractor_class = None
+        if old_processor_class is not None:
+            processor_class = get_user_field(
+                "What will be the name of the processor class for this model? ",
+                default_value=f"{model_camel_cased}Processor",
+            )
+        else:
+            processor_class = None
+
+    model_patterns = ModelPatterns(
+        model_name,
+        checkpoint,
+        model_type=model_type,
+        model_lower_cased=model_lower_cased,
+        model_camel_cased=model_camel_cased,
+        model_upper_cased=model_upper_cased,
+        config_class=config_class,
+        tokenizer_class=tokenizer_class,
+        image_processor_class=image_processor_class,
+        feature_extractor_class=feature_extractor_class,
+        processor_class=processor_class,
+    )
+
+    add_copied_from = get_user_field(
+        "Should we add # Copied from statements when creating the new modeling file (yes/no)? ",
+        convert_to=convert_to_bool,
+        default_value="yes",
+        fallback_message="Please answer yes/no, y/n, true/false or 1/0.",
+    )
+
+    all_frameworks = get_user_field(
+        "Should we add a version of your new model in all the frameworks implemented by"
+        f" {old_model_type} ({old_frameworks}) (yes/no)? ",
+        convert_to=convert_to_bool,
+        default_value="yes",
+        fallback_message="Please answer yes/no, y/n, true/false or 1/0.",
+    )
+    if all_frameworks:
+        frameworks = None
+    else:
+        frameworks = get_user_field(
+            "Please enter the list of framworks you want (pt, tf, flax) separated by spaces",
+            is_valid_answer=lambda x: all(p in ["pt", "tf", "flax"] for p in x.split(" ")),
+        )
+        frameworks = list(set(frameworks.split(" ")))
+
+    return (old_model_type, model_patterns, add_copied_from, frameworks, old_checkpoint)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/download.py

@@ -0,0 +1,46 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import ArgumentParser
+
+from . import BaseTransformersCLICommand
+
+
+def download_command_factory(args):
+    return DownloadCommand(args.model, args.cache_dir, args.force)
+
+
+class DownloadCommand(BaseTransformersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        download_parser = parser.add_parser("download")
+        download_parser.add_argument(
+            "--cache-dir", type=str, default=None, help="Path to location to store the models"
+        )
+        download_parser.add_argument(
+            "--force", action="store_true", help="Force the model to be download even if already in cache-dir"
+        )
+        download_parser.add_argument("model", type=str, help="Name of the model to download")
+        download_parser.set_defaults(func=download_command_factory)
+
+    def __init__(self, model: str, cache: str, force: bool):
+        self._model = model
+        self._cache = cache
+        self._force = force
+
+    def run(self):
+        from ..models.auto import AutoModel, AutoTokenizer
+
+        AutoModel.from_pretrained(self._model, cache_dir=self._cache, force_download=self._force)
+        AutoTokenizer.from_pretrained(self._model, cache_dir=self._cache, force_download=self._force)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/env.py

@@ -0,0 +1,104 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib.util
+import platform
+from argparse import ArgumentParser
+
+import huggingface_hub
+
+from .. import __version__ as version
+from ..utils import is_flax_available, is_safetensors_available, is_tf_available, is_torch_available
+from . import BaseTransformersCLICommand
+
+
+def info_command_factory(_):
+    return EnvironmentCommand()
+
+
+class EnvironmentCommand(BaseTransformersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        download_parser = parser.add_parser("env")
+        download_parser.set_defaults(func=info_command_factory)
+
+    def run(self):
+        safetensors_version = "not installed"
+        if is_safetensors_available():
+            import safetensors
+
+            safetensors_version = safetensors.__version__
+        elif importlib.util.find_spec("safetensors") is not None:
+            import safetensors
+
+            safetensors_version = f"{safetensors.__version__} but is ignored because of PyTorch version too old."
+
+        pt_version = "not installed"
+        pt_cuda_available = "NA"
+        if is_torch_available():
+            import torch
+
+            pt_version = torch.__version__
+            pt_cuda_available = torch.cuda.is_available()
+
+        tf_version = "not installed"
+        tf_cuda_available = "NA"
+        if is_tf_available():
+            import tensorflow as tf
+
+            tf_version = tf.__version__
+            try:
+                # deprecated in v2.1
+                tf_cuda_available = tf.test.is_gpu_available()
+            except AttributeError:
+                # returns list of devices, convert to bool
+                tf_cuda_available = bool(tf.config.list_physical_devices("GPU"))
+
+        flax_version = "not installed"
+        jax_version = "not installed"
+        jaxlib_version = "not installed"
+        jax_backend = "NA"
+        if is_flax_available():
+            import flax
+            import jax
+            import jaxlib
+
+            flax_version = flax.__version__
+            jax_version = jax.__version__
+            jaxlib_version = jaxlib.__version__
+            jax_backend = jax.lib.xla_bridge.get_backend().platform
+
+        info = {
+            "`transformers` version": version,
+            "Platform": platform.platform(),
+            "Python version": platform.python_version(),
+            "Huggingface_hub version": huggingface_hub.__version__,
+            "Safetensors version": f"{safetensors_version}",
+            "PyTorch version (GPU?)": f"{pt_version} ({pt_cuda_available})",
+            "Tensorflow version (GPU?)": f"{tf_version} ({tf_cuda_available})",
+            "Flax version (CPU?/GPU?/TPU?)": f"{flax_version} ({jax_backend})",
+            "Jax version": f"{jax_version}",
+            "JaxLib version": f"{jaxlib_version}",
+            "Using GPU in script?": "<fill in>",
+            "Using distributed or parallel set-up in script?": "<fill in>",
+        }
+
+        print("\nCopy-and-paste the text below in your GitHub issue and FILL OUT the two last points.\n")
+        print(self.format_dict(info))
+
+        return info
+
+    @staticmethod
+    def format_dict(d):
+        return "\n".join([f"- {prop}: {val}" for prop, val in d.items()]) + "\n"

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/lfs.py

@@ -0,0 +1,226 @@
+"""
+Implementation of a custom transfer agent for the transfer type "multipart" for git-lfs.
+
+Inspired by: github.com/cbartz/git-lfs-swift-transfer-agent/blob/master/git_lfs_swift_transfer.py
+
+Spec is: github.com/git-lfs/git-lfs/blob/master/docs/custom-transfers.md
+
+
+To launch debugger while developing:
+
+``` [lfs "customtransfer.multipart"]
+path = /path/to/transformers/.env/bin/python args = -m debugpy --listen 5678 --wait-for-client
+/path/to/transformers/src/transformers/commands/transformers_cli.py lfs-multipart-upload ```"""
+
+import json
+import os
+import subprocess
+import sys
+import warnings
+from argparse import ArgumentParser
+from contextlib import AbstractContextManager
+from typing import Dict, List, Optional
+
+import requests
+
+from ..utils import logging
+from . import BaseTransformersCLICommand
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+LFS_MULTIPART_UPLOAD_COMMAND = "lfs-multipart-upload"
+
+
+class LfsCommands(BaseTransformersCLICommand):
+    """
+    Implementation of a custom transfer agent for the transfer type "multipart" for git-lfs. This lets users upload
+    large files >5GB 🔥. Spec for LFS custom transfer agent is:
+    https://github.com/git-lfs/git-lfs/blob/master/docs/custom-transfers.md
+
+    This introduces two commands to the CLI:
+
+    1. $ transformers-cli lfs-enable-largefiles
+
+    This should be executed once for each model repo that contains a model file >5GB. It's documented in the error
+    message you get if you just try to git push a 5GB file without having enabled it before.
+
+    2. $ transformers-cli lfs-multipart-upload
+
+    This command is called by lfs directly and is not meant to be called by the user.
+    """
+
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        enable_parser = parser.add_parser(
+            "lfs-enable-largefiles",
+            help=(
+                "Deprecated: use `huggingface-cli` instead. Configure your repository to enable upload of files > 5GB."
+            ),
+        )
+        enable_parser.add_argument("path", type=str, help="Local path to repository you want to configure.")
+        enable_parser.set_defaults(func=lambda args: LfsEnableCommand(args))
+
+        upload_parser = parser.add_parser(
+            LFS_MULTIPART_UPLOAD_COMMAND,
+            help=(
+                "Deprecated: use `huggingface-cli` instead. "
+                "Command will get called by git-lfs, do not call it directly."
+            ),
+        )
+        upload_parser.set_defaults(func=lambda args: LfsUploadCommand(args))
+
+
+class LfsEnableCommand:
+    def __init__(self, args):
+        self.args = args
+
+    def run(self):
+        warnings.warn(
+            "Managing repositories through transformers-cli is deprecated. Please use `huggingface-cli` instead."
+        )
+        local_path = os.path.abspath(self.args.path)
+        if not os.path.isdir(local_path):
+            print("This does not look like a valid git repo.")
+            exit(1)
+        subprocess.run(
+            "git config lfs.customtransfer.multipart.path transformers-cli".split(), check=True, cwd=local_path
+        )
+        subprocess.run(
+            f"git config lfs.customtransfer.multipart.args {LFS_MULTIPART_UPLOAD_COMMAND}".split(),
+            check=True,
+            cwd=local_path,
+        )
+        print("Local repo set up for largefiles")
+
+
+def write_msg(msg: Dict):
+    """Write out the message in Line delimited JSON."""
+    msg = json.dumps(msg) + "\n"
+    sys.stdout.write(msg)
+    sys.stdout.flush()
+
+
+def read_msg() -> Optional[Dict]:
+    """Read Line delimited JSON from stdin."""
+    msg = json.loads(sys.stdin.readline().strip())
+
+    if "terminate" in (msg.get("type"), msg.get("event")):
+        # terminate message received
+        return None
+
+    if msg.get("event") not in ("download", "upload"):
+        logger.critical("Received unexpected message")
+        sys.exit(1)
+
+    return msg
+
+
+class FileSlice(AbstractContextManager):
+    """
+    File-like object that only reads a slice of a file
+
+    Inspired by stackoverflow.com/a/29838711/593036
+    """
+
+    def __init__(self, filepath: str, seek_from: int, read_limit: int):
+        self.filepath = filepath
+        self.seek_from = seek_from
+        self.read_limit = read_limit
+        self.n_seen = 0
+
+    def __enter__(self):
+        self.f = open(self.filepath, "rb")
+        self.f.seek(self.seek_from)
+        return self
+
+    def __len__(self):
+        total_length = os.fstat(self.f.fileno()).st_size
+        return min(self.read_limit, total_length - self.seek_from)
+
+    def read(self, n=-1):
+        if self.n_seen >= self.read_limit:
+            return b""
+        remaining_amount = self.read_limit - self.n_seen
+        data = self.f.read(remaining_amount if n < 0 else min(n, remaining_amount))
+        self.n_seen += len(data)
+        return data
+
+    def __iter__(self):
+        yield self.read(n=4 * 1024 * 1024)
+
+    def __exit__(self, *args):
+        self.f.close()
+
+
+class LfsUploadCommand:
+    def __init__(self, args):
+        self.args = args
+
+    def run(self):
+        # Immediately after invoking a custom transfer process, git-lfs
+        # sends initiation data to the process over stdin.
+        # This tells the process useful information about the configuration.
+        init_msg = json.loads(sys.stdin.readline().strip())
+        if not (init_msg.get("event") == "init" and init_msg.get("operation") == "upload"):
+            write_msg({"error": {"code": 32, "message": "Wrong lfs init operation"}})
+            sys.exit(1)
+
+        # The transfer process should use the information it needs from the
+        # initiation structure, and also perform any one-off setup tasks it
+        # needs to do. It should then respond on stdout with a simple empty
+        # confirmation structure, as follows:
+        write_msg({})
+
+        # After the initiation exchange, git-lfs will send any number of
+        # transfer requests to the stdin of the transfer process, in a serial sequence.
+        while True:
+            msg = read_msg()
+            if msg is None:
+                # When all transfers have been processed, git-lfs will send
+                # a terminate event to the stdin of the transfer process.
+                # On receiving this message the transfer process should
+                # clean up and terminate. No response is expected.
+                sys.exit(0)
+
+            oid = msg["oid"]
+            filepath = msg["path"]
+            completion_url = msg["action"]["href"]
+            header = msg["action"]["header"]
+            chunk_size = int(header.pop("chunk_size"))
+            presigned_urls: List[str] = list(header.values())
+
+            parts = []
+            for i, presigned_url in enumerate(presigned_urls):
+                with FileSlice(filepath, seek_from=i * chunk_size, read_limit=chunk_size) as data:
+                    r = requests.put(presigned_url, data=data)
+                    r.raise_for_status()
+                    parts.append(
+                        {
+                            "etag": r.headers.get("etag"),
+                            "partNumber": i + 1,
+                        }
+                    )
+                    # In order to support progress reporting while data is uploading / downloading,
+                    # the transfer process should post messages to stdout
+                    write_msg(
+                        {
+                            "event": "progress",
+                            "oid": oid,
+                            "bytesSoFar": (i + 1) * chunk_size,
+                            "bytesSinceLast": chunk_size,
+                        }
+                    )
+                    # Not precise but that's ok.
+
+            r = requests.post(
+                completion_url,
+                json={
+                    "oid": oid,
+                    "parts": parts,
+                },
+            )
+            r.raise_for_status()
+
+            write_msg({"event": "complete", "oid": oid})

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/run.py

@@ -0,0 +1,110 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import ArgumentParser
+
+from ..pipelines import Pipeline, PipelineDataFormat, get_supported_tasks, pipeline
+from ..utils import logging
+from . import BaseTransformersCLICommand
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def try_infer_format_from_ext(path: str):
+    if not path:
+        return "pipe"
+
+    for ext in PipelineDataFormat.SUPPORTED_FORMATS:
+        if path.endswith(ext):
+            return ext
+
+    raise Exception(
+        f"Unable to determine file format from file extension {path}. "
+        f"Please provide the format through --format {PipelineDataFormat.SUPPORTED_FORMATS}"
+    )
+
+
+def run_command_factory(args):
+    nlp = pipeline(
+        task=args.task,
+        model=args.model if args.model else None,
+        config=args.config,
+        tokenizer=args.tokenizer,
+        device=args.device,
+    )
+    format = try_infer_format_from_ext(args.input) if args.format == "infer" else args.format
+    reader = PipelineDataFormat.from_str(
+        format=format,
+        output_path=args.output,
+        input_path=args.input,
+        column=args.column if args.column else nlp.default_input_names,
+        overwrite=args.overwrite,
+    )
+    return RunCommand(nlp, reader)
+
+
+class RunCommand(BaseTransformersCLICommand):
+    def __init__(self, nlp: Pipeline, reader: PipelineDataFormat):
+        self._nlp = nlp
+        self._reader = reader
+
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        run_parser = parser.add_parser("run", help="Run a pipeline through the CLI")
+        run_parser.add_argument("--task", choices=get_supported_tasks(), help="Task to run")
+        run_parser.add_argument("--input", type=str, help="Path to the file to use for inference")
+        run_parser.add_argument("--output", type=str, help="Path to the file that will be used post to write results.")
+        run_parser.add_argument("--model", type=str, help="Name or path to the model to instantiate.")
+        run_parser.add_argument("--config", type=str, help="Name or path to the model's config to instantiate.")
+        run_parser.add_argument(
+            "--tokenizer", type=str, help="Name of the tokenizer to use. (default: same as the model name)"
+        )
+        run_parser.add_argument(
+            "--column",
+            type=str,
+            help="Name of the column to use as input. (For multi columns input as QA use column1,columns2)",
+        )
+        run_parser.add_argument(
+            "--format",
+            type=str,
+            default="infer",
+            choices=PipelineDataFormat.SUPPORTED_FORMATS,
+            help="Input format to read from",
+        )
+        run_parser.add_argument(
+            "--device",
+            type=int,
+            default=-1,
+            help="Indicate the device to run onto, -1 indicates CPU, >= 0 indicates GPU (default: -1)",
+        )
+        run_parser.add_argument("--overwrite", action="store_true", help="Allow overwriting the output file.")
+        run_parser.set_defaults(func=run_command_factory)
+
+    def run(self):
+        nlp, outputs = self._nlp, []
+
+        for entry in self._reader:
+            output = nlp(**entry) if self._reader.is_multi_columns else nlp(entry)
+            if isinstance(output, dict):
+                outputs.append(output)
+            else:
+                outputs += output
+
+        # Saving data
+        if self._nlp.binary_output:
+            binary_path = self._reader.save_binary(outputs)
+            logger.warning(f"Current pipeline requires output to be in binary format, saving at {binary_path}")
+        else:
+            self._reader.save(outputs)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/serving.py

@@ -0,0 +1,228 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import ArgumentParser, Namespace
+from typing import Any, List, Optional
+
+from ..pipelines import Pipeline, get_supported_tasks, pipeline
+from ..utils import logging
+from . import BaseTransformersCLICommand
+
+
+try:
+    from fastapi import Body, FastAPI, HTTPException
+    from fastapi.routing import APIRoute
+    from pydantic import BaseModel
+    from starlette.responses import JSONResponse
+    from uvicorn import run
+
+    _serve_dependencies_installed = True
+except (ImportError, AttributeError):
+    BaseModel = object
+
+    def Body(*x, **y):
+        pass
+
+    _serve_dependencies_installed = False
+
+
+logger = logging.get_logger("transformers-cli/serving")
+
+
+def serve_command_factory(args: Namespace):
+    """
+    Factory function used to instantiate serving server from provided command line arguments.
+
+    Returns: ServeCommand
+    """
+    nlp = pipeline(
+        task=args.task,
+        model=args.model if args.model else None,
+        config=args.config,
+        tokenizer=args.tokenizer,
+        device=args.device,
+    )
+    return ServeCommand(nlp, args.host, args.port, args.workers)
+
+
+class ServeModelInfoResult(BaseModel):
+    """
+    Expose model information
+    """
+
+    infos: dict
+
+
+class ServeTokenizeResult(BaseModel):
+    """
+    Tokenize result model
+    """
+
+    tokens: List[str]
+    tokens_ids: Optional[List[int]]
+
+
+class ServeDeTokenizeResult(BaseModel):
+    """
+    DeTokenize result model
+    """
+
+    text: str
+
+
+class ServeForwardResult(BaseModel):
+    """
+    Forward result model
+    """
+
+    output: Any
+
+
+class ServeCommand(BaseTransformersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        """
+        Register this command to argparse so it's available for the transformer-cli
+
+        Args:
+            parser: Root parser to register command-specific arguments
+        """
+        serve_parser = parser.add_parser(
+            "serve", help="CLI tool to run inference requests through REST and GraphQL endpoints."
+        )
+        serve_parser.add_argument(
+            "--task",
+            type=str,
+            choices=get_supported_tasks(),
+            help="The task to run the pipeline on",
+        )
+        serve_parser.add_argument("--host", type=str, default="localhost", help="Interface the server will listen on.")
+        serve_parser.add_argument("--port", type=int, default=8888, help="Port the serving will listen to.")
+        serve_parser.add_argument("--workers", type=int, default=1, help="Number of http workers")
+        serve_parser.add_argument("--model", type=str, help="Model's name or path to stored model.")
+        serve_parser.add_argument("--config", type=str, help="Model's config name or path to stored model.")
+        serve_parser.add_argument("--tokenizer", type=str, help="Tokenizer name to use.")
+        serve_parser.add_argument(
+            "--device",
+            type=int,
+            default=-1,
+            help="Indicate the device to run onto, -1 indicates CPU, >= 0 indicates GPU (default: -1)",
+        )
+        serve_parser.set_defaults(func=serve_command_factory)
+
+    def __init__(self, pipeline: Pipeline, host: str, port: int, workers: int):
+        self._pipeline = pipeline
+
+        self.host = host
+        self.port = port
+        self.workers = workers
+
+        if not _serve_dependencies_installed:
+            raise RuntimeError(
+                "Using serve command requires FastAPI and uvicorn. "
+                'Please install transformers with [serving]: pip install "transformers[serving]".'
+                "Or install FastAPI and uvicorn separately."
+            )
+        else:
+            logger.info(f"Serving model over {host}:{port}")
+            self._app = FastAPI(
+                routes=[
+                    APIRoute(
+                        "/",
+                        self.model_info,
+                        response_model=ServeModelInfoResult,
+                        response_class=JSONResponse,
+                        methods=["GET"],
+                    ),
+                    APIRoute(
+                        "/tokenize",
+                        self.tokenize,
+                        response_model=ServeTokenizeResult,
+                        response_class=JSONResponse,
+                        methods=["POST"],
+                    ),
+                    APIRoute(
+                        "/detokenize",
+                        self.detokenize,
+                        response_model=ServeDeTokenizeResult,
+                        response_class=JSONResponse,
+                        methods=["POST"],
+                    ),
+                    APIRoute(
+                        "/forward",
+                        self.forward,
+                        response_model=ServeForwardResult,
+                        response_class=JSONResponse,
+                        methods=["POST"],
+                    ),
+                ],
+                timeout=600,
+            )
+
+    def run(self):
+        run(self._app, host=self.host, port=self.port, workers=self.workers)
+
+    def model_info(self):
+        return ServeModelInfoResult(infos=vars(self._pipeline.model.config))
+
+    def tokenize(self, text_input: str = Body(None, embed=True), return_ids: bool = Body(False, embed=True)):
+        """
+        Tokenize the provided input and eventually returns corresponding tokens id: - **text_input**: String to
+        tokenize - **return_ids**: Boolean flags indicating if the tokens have to be converted to their integer
+        mapping.
+        """
+        try:
+            tokens_txt = self._pipeline.tokenizer.tokenize(text_input)
+
+            if return_ids:
+                tokens_ids = self._pipeline.tokenizer.convert_tokens_to_ids(tokens_txt)
+                return ServeTokenizeResult(tokens=tokens_txt, tokens_ids=tokens_ids)
+            else:
+                return ServeTokenizeResult(tokens=tokens_txt)
+
+        except Exception as e:
+            raise HTTPException(status_code=500, detail={"model": "", "error": str(e)})
+
+    def detokenize(
+        self,
+        tokens_ids: List[int] = Body(None, embed=True),
+        skip_special_tokens: bool = Body(False, embed=True),
+        cleanup_tokenization_spaces: bool = Body(True, embed=True),
+    ):
+        """
+        Detokenize the provided tokens ids to readable text: - **tokens_ids**: List of tokens ids -
+        **skip_special_tokens**: Flag indicating to not try to decode special tokens - **cleanup_tokenization_spaces**:
+        Flag indicating to remove all leading/trailing spaces and intermediate ones.
+        """
+        try:
+            decoded_str = self._pipeline.tokenizer.decode(tokens_ids, skip_special_tokens, cleanup_tokenization_spaces)
+            return ServeDeTokenizeResult(model="", text=decoded_str)
+        except Exception as e:
+            raise HTTPException(status_code=500, detail={"model": "", "error": str(e)})
+
+    async def forward(self, inputs=Body(None, embed=True)):
+        """
+        **inputs**: **attention_mask**: **tokens_type_ids**:
+        """
+
+        # Check we don't have empty string
+        if len(inputs) == 0:
+            return ServeForwardResult(output=[], attention=[])
+
+        try:
+            # Forward through the model
+            output = self._pipeline(inputs)
+            return ServeForwardResult(output=output)
+        except Exception as e:
+            raise HTTPException(500, {"error": str(e)})

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/transformers_cli.py

@@ -0,0 +1,59 @@
+#!/usr/bin/env python
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import ArgumentParser
+
+from .add_new_model import AddNewModelCommand
+from .add_new_model_like import AddNewModelLikeCommand
+from .convert import ConvertCommand
+from .download import DownloadCommand
+from .env import EnvironmentCommand
+from .lfs import LfsCommands
+from .pt_to_tf import PTtoTFCommand
+from .run import RunCommand
+from .serving import ServeCommand
+from .user import UserCommands
+
+
+def main():
+    parser = ArgumentParser("Transformers CLI tool", usage="transformers-cli <command> [<args>]")
+    commands_parser = parser.add_subparsers(help="transformers-cli command helpers")
+
+    # Register commands
+    ConvertCommand.register_subcommand(commands_parser)
+    DownloadCommand.register_subcommand(commands_parser)
+    EnvironmentCommand.register_subcommand(commands_parser)
+    RunCommand.register_subcommand(commands_parser)
+    ServeCommand.register_subcommand(commands_parser)
+    UserCommands.register_subcommand(commands_parser)
+    AddNewModelCommand.register_subcommand(commands_parser)
+    AddNewModelLikeCommand.register_subcommand(commands_parser)
+    LfsCommands.register_subcommand(commands_parser)
+    PTtoTFCommand.register_subcommand(commands_parser)
+
+    # Let's go
+    args = parser.parse_args()
+
+    if not hasattr(args, "func"):
+        parser.print_help()
+        exit(1)
+
+    # Run
+    service = args.func(args)
+    service.run()
+
+
+if __name__ == "__main__":
+    main()

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/commands/user.py

@@ -0,0 +1,197 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import subprocess
+from argparse import ArgumentParser
+from typing import List, Union
+
+from huggingface_hub.hf_api import HfFolder, create_repo, whoami
+from requests.exceptions import HTTPError
+
+from . import BaseTransformersCLICommand
+
+
+class UserCommands(BaseTransformersCLICommand):
+    @staticmethod
+    def register_subcommand(parser: ArgumentParser):
+        login_parser = parser.add_parser("login", help="Log in using the same credentials as on huggingface.co")
+        login_parser.set_defaults(func=lambda args: LoginCommand(args))
+        whoami_parser = parser.add_parser("whoami", help="Find out which huggingface.co account you are logged in as.")
+        whoami_parser.set_defaults(func=lambda args: WhoamiCommand(args))
+        logout_parser = parser.add_parser("logout", help="Log out")
+        logout_parser.set_defaults(func=lambda args: LogoutCommand(args))
+
+        # new system: git-based repo system
+        repo_parser = parser.add_parser(
+            "repo",
+            help="Deprecated: use `huggingface-cli` instead. Commands to interact with your huggingface.co repos.",
+        )
+        repo_subparsers = repo_parser.add_subparsers(
+            help="Deprecated: use `huggingface-cli` instead. huggingface.co repos related commands"
+        )
+        repo_create_parser = repo_subparsers.add_parser(
+            "create", help="Deprecated: use `huggingface-cli` instead. Create a new repo on huggingface.co"
+        )
+        repo_create_parser.add_argument(
+            "name",
+            type=str,
+            help="Name for your model's repo. Will be namespaced under your username to build the model id.",
+        )
+        repo_create_parser.add_argument("--organization", type=str, help="Optional: organization namespace.")
+        repo_create_parser.add_argument("-y", "--yes", action="store_true", help="Optional: answer Yes to the prompt")
+        repo_create_parser.set_defaults(func=lambda args: RepoCreateCommand(args))
+
+
+class ANSI:
+    """
+    Helper for en.wikipedia.org/wiki/ANSI_escape_code
+    """
+
+    _bold = "\u001b[1m"
+    _red = "\u001b[31m"
+    _gray = "\u001b[90m"
+    _reset = "\u001b[0m"
+
+    @classmethod
+    def bold(cls, s):
+        return f"{cls._bold}{s}{cls._reset}"
+
+    @classmethod
+    def red(cls, s):
+        return f"{cls._bold}{cls._red}{s}{cls._reset}"
+
+    @classmethod
+    def gray(cls, s):
+        return f"{cls._gray}{s}{cls._reset}"
+
+
+def tabulate(rows: List[List[Union[str, int]]], headers: List[str]) -> str:
+    """
+    Inspired by:
+
+    - stackoverflow.com/a/8356620/593036
+    - stackoverflow.com/questions/9535954/printing-lists-as-tabular-data
+    """
+    col_widths = [max(len(str(x)) for x in col) for col in zip(*rows, headers)]
+    row_format = ("{{:{}}} " * len(headers)).format(*col_widths)
+    lines = []
+    lines.append(row_format.format(*headers))
+    lines.append(row_format.format(*["-" * w for w in col_widths]))
+    for row in rows:
+        lines.append(row_format.format(*row))
+    return "\n".join(lines)
+
+
+class BaseUserCommand:
+    def __init__(self, args):
+        self.args = args
+
+
+class LoginCommand(BaseUserCommand):
+    def run(self):
+        print(
+            ANSI.red(
+                "ERROR! `huggingface-cli login` uses an outdated login mechanism "
+                "that is not compatible with the Hugging Face Hub backend anymore. "
+                "Please use `huggingface-cli login instead."
+            )
+        )
+
+
+class WhoamiCommand(BaseUserCommand):
+    def run(self):
+        print(
+            ANSI.red(
+                "WARNING! `transformers-cli whoami` is deprecated and will be removed in v5. Please use "
+                "`huggingface-cli whoami` instead."
+            )
+        )
+        token = HfFolder.get_token()
+        if token is None:
+            print("Not logged in")
+            exit()
+        try:
+            user, orgs = whoami(token)
+            print(user)
+            if orgs:
+                print(ANSI.bold("orgs: "), ",".join(orgs))
+        except HTTPError as e:
+            print(e)
+            print(ANSI.red(e.response.text))
+            exit(1)
+
+
+class LogoutCommand(BaseUserCommand):
+    def run(self):
+        print(
+            ANSI.red(
+                "ERROR! `transformers-cli logout` uses an outdated logout mechanism "
+                "that is not compatible with the Hugging Face Hub backend anymore. "
+                "Please use `huggingface-cli logout instead."
+            )
+        )
+
+
+class RepoCreateCommand(BaseUserCommand):
+    def run(self):
+        print(
+            ANSI.red(
+                "WARNING! Managing repositories through transformers-cli is deprecated. "
+                "Please use `huggingface-cli` instead."
+            )
+        )
+        token = HfFolder.get_token()
+        if token is None:
+            print("Not logged in")
+            exit(1)
+        try:
+            stdout = subprocess.check_output(["git", "--version"]).decode("utf-8")
+            print(ANSI.gray(stdout.strip()))
+        except FileNotFoundError:
+            print("Looks like you do not have git installed, please install.")
+
+        try:
+            stdout = subprocess.check_output(["git-lfs", "--version"]).decode("utf-8")
+            print(ANSI.gray(stdout.strip()))
+        except FileNotFoundError:
+            print(
+                ANSI.red(
+                    "Looks like you do not have git-lfs installed, please install."
+                    " You can install from https://git-lfs.github.com/."
+                    " Then run `git lfs install` (you only have to do this once)."
+                )
+            )
+        print("")
+
+        user, _ = whoami(token)
+        namespace = self.args.organization if self.args.organization is not None else user
+        full_name = f"{namespace}/{self.args.name}"
+        print(f"You are about to create {ANSI.bold(full_name)}")
+
+        if not self.args.yes:
+            choice = input("Proceed? [Y/n] ").lower()
+            if not (choice == "" or choice == "y" or choice == "yes"):
+                print("Abort")
+                exit()
+        try:
+            url = create_repo(token, name=self.args.name, organization=self.args.organization)
+        except HTTPError as e:
+            print(e)
+            print(ANSI.red(e.response.text))
+            exit(1)
+        print("\nYour repo now lives at:")
+        print(f"  {ANSI.bold(url)}")
+        print("\nYou can clone it locally with the command below, and commit/push as usual.")
+        print(f"\n  git clone {url}")
+        print("")

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/data/test_generation_utils.py

@@ -0,0 +1,99 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+import unittest
+
+import timeout_decorator
+
+from ..testing_utils import require_torch
+from ..utils import cached_property, is_torch_available
+
+
+if is_torch_available():
+    import torch
+
+    from ..models.marian import MarianConfig, MarianMTModel
+
+
+@require_torch
+class GenerationUtilsTest(unittest.TestCase):
+    @cached_property
+    def config(self):
+        config = MarianConfig.from_pretrained("sshleifer/tiny-marian-en-de")
+        return config
+
+    @cached_property
+    def model(self):
+        return MarianMTModel(self.config)
+
+    def test_postprocess_next_token_scores(self):
+        config = self.config
+        model = self.model
+        # Initialize an input id tensor with batch size 8 and sequence length 12
+        input_ids = torch.arange(0, 96, 1).view((8, 12))
+        eos = config.eos_token_id
+        bad_words_ids_test_cases = [[[299]], [[23, 24], [54]], [[config.eos_token_id]], []]
+        masked_scores = [
+            [(0, 299), (1, 299), (2, 299), (3, 299), (4, 299), (5, 299), (6, 299), (7, 299)],
+            [(1, 24), (0, 54), (1, 54), (2, 54), (3, 54), (4, 54), (5, 54), (6, 54), (7, 54)],
+            [(0, eos), (1, eos), (2, eos), (3, eos), (4, eos), (5, eos), (6, eos), (7, eos)],
+            [],
+        ]
+
+        for test_case_index, bad_words_ids in enumerate(bad_words_ids_test_cases):
+            # Initialize a scores tensor with batch size 8 and vocabulary size 300
+            scores = torch.rand((8, 300))
+            output = model.postprocess_next_token_scores(
+                scores,
+                input_ids,
+                0,
+                bad_words_ids,
+                13,
+                15,
+                config.max_length,
+                config.eos_token_id,
+                config.repetition_penalty,
+                32,
+                5,
+            )
+            for masked_score in masked_scores[test_case_index]:
+                self.assertTrue(output[masked_score[0], masked_score[1]] == -float("inf"))
+
+    @timeout_decorator.timeout(10)
+    def test_postprocess_next_token_scores_large_bad_words_list(self):
+        config = self.config
+        model = self.model
+        # Initialize an input id tensor with batch size 8 and sequence length 12
+        input_ids = torch.arange(0, 96, 1).view((8, 12))
+
+        bad_words_ids = []
+        for _ in range(100):
+            length_bad_word = random.randint(1, 4)
+            bad_words_ids.append(random.sample(range(1, 300), length_bad_word))
+
+        scores = torch.rand((8, 300))
+        _ = model.postprocess_next_token_scores(
+            scores,
+            input_ids,
+            0,
+            bad_words_ids,
+            13,
+            15,
+            config.max_length,
+            config.eos_token_id,
+            config.repetition_penalty,
+            32,
+            5,
+        )

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/generation/__init__.py

@@ -0,0 +1,272 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ..utils import OptionalDependencyNotAvailable, _LazyModule, is_flax_available, is_tf_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_utils": ["GenerationConfig"],
+    "streamers": ["TextIteratorStreamer", "TextStreamer"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["beam_constraints"] = [
+        "Constraint",
+        "ConstraintListState",
+        "DisjunctiveConstraint",
+        "PhrasalConstraint",
+    ]
+    _import_structure["beam_search"] = [
+        "BeamHypotheses",
+        "BeamScorer",
+        "BeamSearchScorer",
+        "ConstrainedBeamSearchScorer",
+    ]
+    _import_structure["logits_process"] = [
+        "EpsilonLogitsWarper",
+        "EtaLogitsWarper",
+        "ForcedBOSTokenLogitsProcessor",
+        "ForcedEOSTokenLogitsProcessor",
+        "HammingDiversityLogitsProcessor",
+        "InfNanRemoveLogitsProcessor",
+        "LogitsProcessor",
+        "LogitsProcessorList",
+        "LogitsWarper",
+        "MinLengthLogitsProcessor",
+        "MinNewTokensLengthLogitsProcessor",
+        "NoBadWordsLogitsProcessor",
+        "NoRepeatNGramLogitsProcessor",
+        "PrefixConstrainedLogitsProcessor",
+        "RepetitionPenaltyLogitsProcessor",
+        "EncoderRepetitionPenaltyLogitsProcessor",
+        "TemperatureLogitsWarper",
+        "TopKLogitsWarper",
+        "TopPLogitsWarper",
+        "TypicalLogitsWarper",
+        "EncoderNoRepeatNGramLogitsProcessor",
+        "ExponentialDecayLengthPenalty",
+        "LogitNormalization",
+    ]
+    _import_structure["stopping_criteria"] = [
+        "MaxNewTokensCriteria",
+        "MaxLengthCriteria",
+        "MaxTimeCriteria",
+        "StoppingCriteria",
+        "StoppingCriteriaList",
+        "validate_stopping_criteria",
+    ]
+    _import_structure["utils"] = [
+        "GenerationMixin",
+        "top_k_top_p_filtering",
+        "GreedySearchEncoderDecoderOutput",
+        "GreedySearchDecoderOnlyOutput",
+        "SampleEncoderDecoderOutput",
+        "SampleDecoderOnlyOutput",
+        "BeamSearchEncoderDecoderOutput",
+        "BeamSearchDecoderOnlyOutput",
+        "BeamSampleEncoderDecoderOutput",
+        "BeamSampleDecoderOnlyOutput",
+        "ContrastiveSearchEncoderDecoderOutput",
+        "ContrastiveSearchDecoderOnlyOutput",
+    ]
+
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tf_logits_process"] = [
+        "TFForcedBOSTokenLogitsProcessor",
+        "TFForcedEOSTokenLogitsProcessor",
+        "TFLogitsProcessor",
+        "TFLogitsProcessorList",
+        "TFLogitsWarper",
+        "TFMinLengthLogitsProcessor",
+        "TFNoBadWordsLogitsProcessor",
+        "TFNoRepeatNGramLogitsProcessor",
+        "TFRepetitionPenaltyLogitsProcessor",
+        "TFTemperatureLogitsWarper",
+        "TFTopKLogitsWarper",
+        "TFTopPLogitsWarper",
+        "TFForceTokensLogitsProcessor",
+        "TFSuppressTokensAtBeginLogitsProcessor",
+        "TFSuppressTokensLogitsProcessor",
+    ]
+    _import_structure["tf_utils"] = [
+        "TFGenerationMixin",
+        "tf_top_k_top_p_filtering",
+        "TFGreedySearchDecoderOnlyOutput",
+        "TFGreedySearchEncoderDecoderOutput",
+        "TFSampleEncoderDecoderOutput",
+        "TFSampleDecoderOnlyOutput",
+        "TFBeamSearchEncoderDecoderOutput",
+        "TFBeamSearchDecoderOnlyOutput",
+        "TFBeamSampleEncoderDecoderOutput",
+        "TFBeamSampleDecoderOnlyOutput",
+        "TFContrastiveSearchEncoderDecoderOutput",
+        "TFContrastiveSearchDecoderOnlyOutput",
+    ]
+
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["flax_logits_process"] = [
+        "FlaxForcedBOSTokenLogitsProcessor",
+        "FlaxForcedEOSTokenLogitsProcessor",
+        "FlaxLogitsProcessor",
+        "FlaxLogitsProcessorList",
+        "FlaxLogitsWarper",
+        "FlaxMinLengthLogitsProcessor",
+        "FlaxTemperatureLogitsWarper",
+        "FlaxTopKLogitsWarper",
+        "FlaxTopPLogitsWarper",
+    ]
+    _import_structure["flax_utils"] = [
+        "FlaxGenerationMixin",
+        "FlaxGreedySearchOutput",
+        "FlaxSampleOutput",
+        "FlaxBeamSearchOutput",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_utils import GenerationConfig
+    from .streamers import TextIteratorStreamer, TextStreamer
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .beam_constraints import Constraint, ConstraintListState, DisjunctiveConstraint, PhrasalConstraint
+        from .beam_search import BeamHypotheses, BeamScorer, BeamSearchScorer, ConstrainedBeamSearchScorer
+        from .logits_process import (
+            EncoderNoRepeatNGramLogitsProcessor,
+            EncoderRepetitionPenaltyLogitsProcessor,
+            EpsilonLogitsWarper,
+            EtaLogitsWarper,
+            ExponentialDecayLengthPenalty,
+            ForcedBOSTokenLogitsProcessor,
+            ForcedEOSTokenLogitsProcessor,
+            HammingDiversityLogitsProcessor,
+            InfNanRemoveLogitsProcessor,
+            LogitNormalization,
+            LogitsProcessor,
+            LogitsProcessorList,
+            LogitsWarper,
+            MinLengthLogitsProcessor,
+            MinNewTokensLengthLogitsProcessor,
+            NoBadWordsLogitsProcessor,
+            NoRepeatNGramLogitsProcessor,
+            PrefixConstrainedLogitsProcessor,
+            RepetitionPenaltyLogitsProcessor,
+            TemperatureLogitsWarper,
+            TopKLogitsWarper,
+            TopPLogitsWarper,
+            TypicalLogitsWarper,
+        )
+        from .stopping_criteria import (
+            MaxLengthCriteria,
+            MaxNewTokensCriteria,
+            MaxTimeCriteria,
+            StoppingCriteria,
+            StoppingCriteriaList,
+            validate_stopping_criteria,
+        )
+        from .utils import (
+            BeamSampleDecoderOnlyOutput,
+            BeamSampleEncoderDecoderOutput,
+            BeamSearchDecoderOnlyOutput,
+            BeamSearchEncoderDecoderOutput,
+            ContrastiveSearchDecoderOnlyOutput,
+            ContrastiveSearchEncoderDecoderOutput,
+            GenerationMixin,
+            GreedySearchDecoderOnlyOutput,
+            GreedySearchEncoderDecoderOutput,
+            SampleDecoderOnlyOutput,
+            SampleEncoderDecoderOutput,
+            top_k_top_p_filtering,
+        )
+
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tf_logits_process import (
+            TFForcedBOSTokenLogitsProcessor,
+            TFForcedEOSTokenLogitsProcessor,
+            TFForceTokensLogitsProcessor,
+            TFLogitsProcessor,
+            TFLogitsProcessorList,
+            TFLogitsWarper,
+            TFMinLengthLogitsProcessor,
+            TFNoBadWordsLogitsProcessor,
+            TFNoRepeatNGramLogitsProcessor,
+            TFRepetitionPenaltyLogitsProcessor,
+            TFSuppressTokensAtBeginLogitsProcessor,
+            TFSuppressTokensLogitsProcessor,
+            TFTemperatureLogitsWarper,
+            TFTopKLogitsWarper,
+            TFTopPLogitsWarper,
+        )
+        from .tf_utils import (
+            TFBeamSampleDecoderOnlyOutput,
+            TFBeamSampleEncoderDecoderOutput,
+            TFBeamSearchDecoderOnlyOutput,
+            TFBeamSearchEncoderDecoderOutput,
+            TFContrastiveSearchDecoderOnlyOutput,
+            TFContrastiveSearchEncoderDecoderOutput,
+            TFGenerationMixin,
+            TFGreedySearchDecoderOnlyOutput,
+            TFGreedySearchEncoderDecoderOutput,
+            TFSampleDecoderOnlyOutput,
+            TFSampleEncoderDecoderOutput,
+            tf_top_k_top_p_filtering,
+        )
+
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .flax_logits_process import (
+            FlaxForcedBOSTokenLogitsProcessor,
+            FlaxForcedEOSTokenLogitsProcessor,
+            FlaxLogitsProcessor,
+            FlaxLogitsProcessorList,
+            FlaxLogitsWarper,
+            FlaxMinLengthLogitsProcessor,
+            FlaxTemperatureLogitsWarper,
+            FlaxTopKLogitsWarper,
+            FlaxTopPLogitsWarper,
+        )
+        from .flax_utils import FlaxBeamSearchOutput, FlaxGenerationMixin, FlaxGreedySearchOutput, FlaxSampleOutput
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/kernels/rwkv/wkv_op.cpp

@@ -0,0 +1,66 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+typedef at::BFloat16 bf16;
+
+void cuda_forward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y);
+void cuda_forward_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y);
+void cuda_forward_with_state(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *s);
+void cuda_forward_with_state_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, float *s);
+void cuda_backward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *gy, float *gw, float *gu, float *gk, float *gv);
+void cuda_backward_bf16(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, bf16 *gy, bf16 *gw, bf16 *gu, bf16 *gk, bf16 *gv);
+
+void forward(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
+    const int B = k.size(0);
+    const int T = k.size(1);
+    const int C = k.size(2);
+    cuda_forward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>());
+}
+void forward_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
+    const int B = k.size(0);
+    const int T = k.size(1);
+    const int C = k.size(2);
+    cuda_forward_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>());
+}
+void forward_with_state(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &s) {
+    const int B = k.size(0);
+    const int T = k.size(1);
+    const int C = k.size(2);
+    cuda_forward_with_state(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), s.data_ptr<float>());
+}
+void forward_with_state_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &s) {
+    const int B = k.size(0);
+    const int T = k.size(1);
+    const int C = k.size(2);
+    cuda_forward_with_state_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(), s.data_ptr<float>());
+}
+void backward(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
+    const int B = k.size(0);
+    const int T = k.size(1);
+    const int C = k.size(2);
+    cuda_backward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), gy.data_ptr<float>(), gw.data_ptr<float>(), gu.data_ptr<float>(), gk.data_ptr<float>(), gv.data_ptr<float>());
+}
+void backward_bf16(torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
+    const int B = k.size(0);
+    const int T = k.size(1);
+    const int C = k.size(2);
+    cuda_backward_bf16(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(),
+        gy.data_ptr<bf16>(), gw.data_ptr<bf16>(), gu.data_ptr<bf16>(), gk.data_ptr<bf16>(), gv.data_ptr<bf16>());
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward", &forward, "wkv forward");
+    m.def("forward_bf16", &forward_bf16, "wkv forward bf16");
+    m.def("forward_with_state", &forward_with_state, "wkv forward with state");
+    m.def("forward_with_state_bf16", &forward_with_state_bf16, "wkv forward with state bf16");
+    m.def("backward", &backward, "wkv backward");
+    m.def("backward_bf16", &backward_bf16, "wkv backward bf16");
+}
+
+TORCH_LIBRARY(wkv, m) {
+    m.def("forward", forward);
+    m.def("forward_bf16", forward_bf16);
+    m.def("forward_with_state", forward_with_state);
+    m.def("forward_with_state_bf16", forward_with_state_bf16);
+    m.def("backward", backward);
+    m.def("backward_bf16", backward_bf16);
+}

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/modelcard.py

@@ -0,0 +1,903 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+
+import copy
+import json
+import os
+import warnings
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Union
+
+import requests
+import yaml
+from huggingface_hub import model_info
+from huggingface_hub.utils import HFValidationError
+
+from . import __version__
+from .models.auto.modeling_auto import (
+    MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
+    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
+    MODEL_FOR_CTC_MAPPING_NAMES,
+    MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
+    MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES,
+    MODEL_FOR_MASKED_LM_MAPPING_NAMES,
+    MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES,
+    MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES,
+    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
+    MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
+    MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES,
+    MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES,
+    MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES,
+    MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES,
+)
+from .training_args import ParallelMode
+from .utils import (
+    MODEL_CARD_NAME,
+    cached_file,
+    is_datasets_available,
+    is_offline_mode,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+    logging,
+)
+
+
+TASK_MAPPING = {
+    "text-generation": MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
+    "image-classification": MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
+    "image-segmentation": MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES,
+    "fill-mask": MODEL_FOR_MASKED_LM_MAPPING_NAMES,
+    "object-detection": MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES,
+    "question-answering": MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES,
+    "text2text-generation": MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
+    "text-classification": MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
+    "table-question-answering": MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES,
+    "token-classification": MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES,
+    "audio-classification": MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
+    "automatic-speech-recognition": {**MODEL_FOR_CTC_MAPPING_NAMES, **MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES},
+    "zero-shot-image-classification": MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES,
+}
+
+logger = logging.get_logger(__name__)
+
+
+class ModelCard:
+    r"""
+    Structured Model Card class. Store model card as well as methods for loading/downloading/saving model cards.
+
+    Please read the following paper for details and explanation on the sections: "Model Cards for Model Reporting" by
+    Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer,
+    Inioluwa Deborah Raji and Timnit Gebru for the proposal behind model cards. Link: https://arxiv.org/abs/1810.03993
+
+    Note: A model card can be loaded and saved to disk.
+    """
+
+    def __init__(self, **kwargs):
+        warnings.warn(
+            "The class `ModelCard` is deprecated and will be removed in version 5 of Transformers", FutureWarning
+        )
+        # Recommended attributes from https://arxiv.org/abs/1810.03993 (see papers)
+        self.model_details = kwargs.pop("model_details", {})
+        self.intended_use = kwargs.pop("intended_use", {})
+        self.factors = kwargs.pop("factors", {})
+        self.metrics = kwargs.pop("metrics", {})
+        self.evaluation_data = kwargs.pop("evaluation_data", {})
+        self.training_data = kwargs.pop("training_data", {})
+        self.quantitative_analyses = kwargs.pop("quantitative_analyses", {})
+        self.ethical_considerations = kwargs.pop("ethical_considerations", {})
+        self.caveats_and_recommendations = kwargs.pop("caveats_and_recommendations", {})
+
+        # Open additional attributes
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error(f"Can't set {key} with value {value} for {self}")
+                raise err
+
+    def save_pretrained(self, save_directory_or_file):
+        """Save a model card object to the directory or file `save_directory_or_file`."""
+        if os.path.isdir(save_directory_or_file):
+            # If we save using the predefined names, we can load using `from_pretrained`
+            output_model_card_file = os.path.join(save_directory_or_file, MODEL_CARD_NAME)
+        else:
+            output_model_card_file = save_directory_or_file
+
+        self.to_json_file(output_model_card_file)
+        logger.info(f"Model card saved in {output_model_card_file}")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r"""
+        Instantiate a [`ModelCard`] from a pre-trained model model card.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string, the *model id* of a pretrained model card hosted inside a model repo on huggingface.co.
+                  Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                  user or organization name, like `dbmdz/bert-base-german-cased`.
+                - a path to a *directory* containing a model card file saved using the [`~ModelCard.save_pretrained`]
+                  method, e.g.: `./my_model_directory/`.
+                - a path or url to a saved model card JSON *file*, e.g.: `./my_model_directory/modelcard.json`.
+
+            cache_dir: (*optional*) string:
+                Path to a directory in which a downloaded pre-trained model card should be cached if the standard cache
+                should not be used.
+
+            kwargs: (*optional*) dict: key/value pairs with which to update the ModelCard object after loading.
+
+                - The values in kwargs of any keys which are model card attributes will be used to override the loaded
+                  values.
+                - Behavior concerning key/value pairs whose keys are *not* model card attributes is controlled by the
+                  *return_unused_kwargs* keyword parameter.
+
+            proxies: (*optional*) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}. The proxies are used on each request.
+
+            return_unused_kwargs: (*optional*) bool:
+
+                - If False, then this function returns just the final model card object.
+                - If True, then this functions returns a tuple *(model card, unused_kwargs)* where *unused_kwargs* is a
+                  dictionary consisting of the key/value pairs whose keys are not model card attributes: ie the part of
+                  kwargs which has not been used to update *ModelCard* and is otherwise ignored.
+
+        Examples:
+
+        ```python
+        # Download model card from huggingface.co and cache.
+        modelcard = ModelCard.from_pretrained("bert-base-uncased")
+        # Model card was saved using *save_pretrained('./test/saved_model/')*
+        modelcard = ModelCard.from_pretrained("./test/saved_model/")
+        modelcard = ModelCard.from_pretrained("./test/saved_model/modelcard.json")
+        modelcard = ModelCard.from_pretrained("bert-base-uncased", output_attentions=True, foo=False)
+        ```"""
+        cache_dir = kwargs.pop("cache_dir", None)
+        proxies = kwargs.pop("proxies", None)
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+        from_pipeline = kwargs.pop("_from_pipeline", None)
+
+        user_agent = {"file_type": "model_card"}
+        if from_pipeline is not None:
+            user_agent["using_pipeline"] = from_pipeline
+
+        is_local = os.path.isdir(pretrained_model_name_or_path)
+        if os.path.isfile(pretrained_model_name_or_path):
+            resolved_model_card_file = pretrained_model_name_or_path
+            is_local = True
+        else:
+            try:
+                # Load from URL or cache if already cached
+                resolved_model_card_file = cached_file(
+                    pretrained_model_name_or_path,
+                    filename=MODEL_CARD_NAME,
+                    cache_dir=cache_dir,
+                    proxies=proxies,
+                    user_agent=user_agent,
+                )
+                if is_local:
+                    logger.info(f"loading model card file {resolved_model_card_file}")
+                else:
+                    logger.info(f"loading model card file {MODEL_CARD_NAME} from cache at {resolved_model_card_file}")
+                # Load model card
+                modelcard = cls.from_json_file(resolved_model_card_file)
+
+            except (EnvironmentError, json.JSONDecodeError):
+                # We fall back on creating an empty model card
+                modelcard = cls()
+
+        # Update model card with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(modelcard, key):
+                setattr(modelcard, key, value)
+                to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info(f"Model card: {modelcard}")
+        if return_unused_kwargs:
+            return modelcard, kwargs
+        else:
+            return modelcard
+
+    @classmethod
+    def from_dict(cls, json_object):
+        """Constructs a `ModelCard` from a Python dictionary of parameters."""
+        return cls(**json_object)
+
+    @classmethod
+    def from_json_file(cls, json_file):
+        """Constructs a `ModelCard` from a json file of parameters."""
+        with open(json_file, "r", encoding="utf-8") as reader:
+            text = reader.read()
+        dict_obj = json.loads(text)
+        return cls(**dict_obj)
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return str(self.to_json_string())
+
+    def to_dict(self):
+        """Serializes this instance to a Python dictionary."""
+        output = copy.deepcopy(self.__dict__)
+        return output
+
+    def to_json_string(self):
+        """Serializes this instance to a JSON string."""
+        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path):
+        """Save this instance to a json file."""
+        with open(json_file_path, "w", encoding="utf-8") as writer:
+            writer.write(self.to_json_string())
+
+
+AUTOGENERATED_TRAINER_COMMENT = """
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+"""
+
+AUTOGENERATED_KERAS_COMMENT = """
+<!-- This model card has been generated automatically according to the information Keras had access to. You should
+probably proofread and complete it, then remove this comment. -->
+"""
+
+
+TASK_TAG_TO_NAME_MAPPING = {
+    "fill-mask": "Masked Language Modeling",
+    "image-classification": "Image Classification",
+    "image-segmentation": "Image Segmentation",
+    "multiple-choice": "Multiple Choice",
+    "object-detection": "Object Detection",
+    "question-answering": "Question Answering",
+    "summarization": "Summarization",
+    "table-question-answering": "Table Question Answering",
+    "text-classification": "Text Classification",
+    "text-generation": "Causal Language Modeling",
+    "text2text-generation": "Sequence-to-sequence Language Modeling",
+    "token-classification": "Token Classification",
+    "translation": "Translation",
+    "zero-shot-classification": "Zero Shot Classification",
+    "automatic-speech-recognition": "Automatic Speech Recognition",
+}
+
+
+METRIC_TAGS = [
+    "accuracy",
+    "bleu",
+    "f1",
+    "matthews_correlation",
+    "pearsonr",
+    "precision",
+    "recall",
+    "rouge",
+    "sacrebleu",
+    "spearmanr",
+    "wer",
+]
+
+
+def _listify(obj):
+    if obj is None:
+        return []
+    elif isinstance(obj, str):
+        return [obj]
+    else:
+        return obj
+
+
+def _insert_values_as_list(metadata, name, values):
+    if values is None:
+        return metadata
+    if isinstance(values, str):
+        values = [values]
+    values = [v for v in values if v is not None]
+    if len(values) == 0:
+        return metadata
+    metadata[name] = values
+    return metadata
+
+
+def infer_metric_tags_from_eval_results(eval_results):
+    if eval_results is None:
+        return {}
+    result = {}
+    for key in eval_results.keys():
+        if key.lower().replace(" ", "_") in METRIC_TAGS:
+            result[key.lower().replace(" ", "_")] = key
+        elif key.lower() == "rouge1":
+            result["rouge"] = key
+    return result
+
+
+def _insert_value(metadata, name, value):
+    if value is None:
+        return metadata
+    metadata[name] = value
+    return metadata
+
+
+def is_hf_dataset(dataset):
+    if not is_datasets_available():
+        return False
+
+    from datasets import Dataset, IterableDataset
+
+    return isinstance(dataset, (Dataset, IterableDataset))
+
+
+def _get_mapping_values(mapping):
+    result = []
+    for v in mapping.values():
+        if isinstance(v, (tuple, list)):
+            result += list(v)
+        else:
+            result.append(v)
+    return result
+
+
+@dataclass
+class TrainingSummary:
+    model_name: str
+    language: Optional[Union[str, List[str]]] = None
+    license: Optional[str] = None
+    tags: Optional[Union[str, List[str]]] = None
+    finetuned_from: Optional[str] = None
+    tasks: Optional[Union[str, List[str]]] = None
+    dataset: Optional[Union[str, List[str]]] = None
+    dataset_tags: Optional[Union[str, List[str]]] = None
+    dataset_args: Optional[Union[str, List[str]]] = None
+    dataset_metadata: Optional[Dict[str, Any]] = None
+    eval_results: Optional[Dict[str, float]] = None
+    eval_lines: Optional[List[str]] = None
+    hyperparameters: Optional[Dict[str, Any]] = None
+    source: Optional[str] = "trainer"
+
+    def __post_init__(self):
+        # Infer default license from the checkpoint used, if possible.
+        if (
+            self.license is None
+            and not is_offline_mode()
+            and self.finetuned_from is not None
+            and len(self.finetuned_from) > 0
+        ):
+            try:
+                info = model_info(self.finetuned_from)
+                for tag in info.tags:
+                    if tag.startswith("license:"):
+                        self.license = tag[8:]
+            except (requests.exceptions.HTTPError, HFValidationError):
+                pass
+
+    def create_model_index(self, metric_mapping):
+        model_index = {"name": self.model_name}
+
+        # Dataset mapping tag -> name
+        dataset_names = _listify(self.dataset)
+        dataset_tags = _listify(self.dataset_tags)
+        dataset_args = _listify(self.dataset_args)
+        dataset_metadata = _listify(self.dataset_metadata)
+        if len(dataset_args) < len(dataset_tags):
+            dataset_args = dataset_args + [None] * (len(dataset_tags) - len(dataset_args))
+        dataset_mapping = dict(zip(dataset_tags, dataset_names))
+        dataset_arg_mapping = dict(zip(dataset_tags, dataset_args))
+        dataset_metadata_mapping = dict(zip(dataset_tags, dataset_metadata))
+
+        task_mapping = {
+            task: TASK_TAG_TO_NAME_MAPPING[task] for task in _listify(self.tasks) if task in TASK_TAG_TO_NAME_MAPPING
+        }
+
+        model_index["results"] = []
+
+        if len(task_mapping) == 0 and len(dataset_mapping) == 0:
+            return [model_index]
+        if len(task_mapping) == 0:
+            task_mapping = {None: None}
+        if len(dataset_mapping) == 0:
+            dataset_mapping = {None: None}
+
+        # One entry per dataset and per task
+        all_possibilities = [(task_tag, ds_tag) for task_tag in task_mapping for ds_tag in dataset_mapping]
+        for task_tag, ds_tag in all_possibilities:
+            result = {}
+            if task_tag is not None:
+                result["task"] = {"name": task_mapping[task_tag], "type": task_tag}
+
+            if ds_tag is not None:
+                metadata = dataset_metadata_mapping.get(ds_tag, {})
+                result["dataset"] = {
+                    "name": dataset_mapping[ds_tag],
+                    "type": ds_tag,
+                    **metadata,
+                }
+                if dataset_arg_mapping[ds_tag] is not None:
+                    result["dataset"]["args"] = dataset_arg_mapping[ds_tag]
+
+            if len(metric_mapping) > 0:
+                result["metrics"] = []
+                for metric_tag, metric_name in metric_mapping.items():
+                    result["metrics"].append(
+                        {
+                            "name": metric_name,
+                            "type": metric_tag,
+                            "value": self.eval_results[metric_name],
+                        }
+                    )
+
+            # Remove partial results to avoid the model card being rejected.
+            if "task" in result and "dataset" in result and "metrics" in result:
+                model_index["results"].append(result)
+            else:
+                logger.info(f"Dropping the following result as it does not have all the necessary fields:\n{result}")
+
+        return [model_index]
+
+    def create_metadata(self):
+        metric_mapping = infer_metric_tags_from_eval_results(self.eval_results)
+
+        metadata = {}
+        metadata = _insert_values_as_list(metadata, "language", self.language)
+        metadata = _insert_value(metadata, "license", self.license)
+        metadata = _insert_values_as_list(metadata, "tags", self.tags)
+        metadata = _insert_values_as_list(metadata, "datasets", self.dataset_tags)
+        metadata = _insert_values_as_list(metadata, "metrics", list(metric_mapping.keys()))
+        metadata["model-index"] = self.create_model_index(metric_mapping)
+
+        return metadata
+
+    def to_model_card(self):
+        model_card = ""
+
+        metadata = yaml.dump(self.create_metadata(), sort_keys=False)
+        if len(metadata) > 0:
+            model_card = f"---\n{metadata}---\n"
+
+        # Now the model card for realsies.
+        if self.source == "trainer":
+            model_card += AUTOGENERATED_TRAINER_COMMENT
+        else:
+            model_card += AUTOGENERATED_KERAS_COMMENT
+
+        model_card += f"\n# {self.model_name}\n\n"
+
+        if self.finetuned_from is None:
+            model_card += "This model was trained from scratch on "
+        else:
+            model_card += (
+                "This model is a fine-tuned version of"
+                f" [{self.finetuned_from}](https://huggingface.co/{self.finetuned_from}) on "
+            )
+
+        if self.dataset is None:
+            model_card += "an unknown dataset."
+        else:
+            if isinstance(self.dataset, str):
+                model_card += f"the {self.dataset} dataset."
+            elif isinstance(self.dataset, (tuple, list)) and len(self.dataset) == 1:
+                model_card += f"the {self.dataset[0]} dataset."
+            else:
+                model_card += (
+                    ", ".join([f"the {ds}" for ds in self.dataset[:-1]]) + f" and the {self.dataset[-1]} datasets."
+                )
+
+        if self.eval_results is not None:
+            model_card += "\nIt achieves the following results on the evaluation set:\n"
+            model_card += "\n".join([f"- {name}: {_maybe_round(value)}" for name, value in self.eval_results.items()])
+        model_card += "\n"
+
+        model_card += "\n## Model description\n\nMore information needed\n"
+        model_card += "\n## Intended uses & limitations\n\nMore information needed\n"
+        model_card += "\n## Training and evaluation data\n\nMore information needed\n"
+
+        model_card += "\n## Training procedure\n"
+        model_card += "\n### Training hyperparameters\n"
+        if self.hyperparameters is not None:
+            model_card += "\nThe following hyperparameters were used during training:\n"
+            model_card += "\n".join([f"- {name}: {value}" for name, value in self.hyperparameters.items()])
+            model_card += "\n"
+        else:
+            model_card += "\nMore information needed\n"
+
+        if self.eval_lines is not None:
+            model_card += "\n### Training results\n\n"
+            model_card += make_markdown_table(self.eval_lines)
+            model_card += "\n"
+
+        model_card += "\n### Framework versions\n\n"
+        model_card += f"- Transformers {__version__}\n"
+
+        if self.source == "trainer" and is_torch_available():
+            import torch
+
+            model_card += f"- Pytorch {torch.__version__}\n"
+        elif self.source == "keras" and is_tf_available():
+            import tensorflow as tf
+
+            model_card += f"- TensorFlow {tf.__version__}\n"
+        if is_datasets_available():
+            import datasets
+
+            model_card += f"- Datasets {datasets.__version__}\n"
+        if is_tokenizers_available():
+            import tokenizers
+
+            model_card += f"- Tokenizers {tokenizers.__version__}\n"
+
+        return model_card
+
+    @classmethod
+    def from_trainer(
+        cls,
+        trainer,
+        language=None,
+        license=None,
+        tags=None,
+        model_name=None,
+        finetuned_from=None,
+        tasks=None,
+        dataset_tags=None,
+        dataset_metadata=None,
+        dataset=None,
+        dataset_args=None,
+    ):
+        # Infer default from dataset
+        one_dataset = trainer.eval_dataset if trainer.eval_dataset is not None else trainer.train_dataset
+        if is_hf_dataset(one_dataset) and (dataset_tags is None or dataset_args is None or dataset_metadata is None):
+            default_tag = one_dataset.builder_name
+            # Those are not real datasets from the Hub so we exclude them.
+            if default_tag not in ["csv", "json", "pandas", "parquet", "text"]:
+                if dataset_metadata is None:
+                    dataset_metadata = [{"config": one_dataset.config_name, "split": str(one_dataset.split)}]
+                if dataset_tags is None:
+                    dataset_tags = [default_tag]
+                if dataset_args is None:
+                    dataset_args = [one_dataset.config_name]
+
+        if dataset is None and dataset_tags is not None:
+            dataset = dataset_tags
+
+        # Infer default finetuned_from
+        if (
+            finetuned_from is None
+            and hasattr(trainer.model.config, "_name_or_path")
+            and not os.path.isdir(trainer.model.config._name_or_path)
+        ):
+            finetuned_from = trainer.model.config._name_or_path
+
+        # Infer default task tag:
+        if tasks is None:
+            model_class_name = trainer.model.__class__.__name__
+            for task, mapping in TASK_MAPPING.items():
+                if model_class_name in _get_mapping_values(mapping):
+                    tasks = task
+
+        if model_name is None:
+            model_name = Path(trainer.args.output_dir).name
+        if len(model_name) == 0:
+            model_name = finetuned_from
+
+        # Add `generated_from_trainer` to the tags
+        if tags is None:
+            tags = ["generated_from_trainer"]
+        elif isinstance(tags, str) and tags != "generated_from_trainer":
+            tags = [tags, "generated_from_trainer"]
+        elif "generated_from_trainer" not in tags:
+            tags.append("generated_from_trainer")
+
+        _, eval_lines, eval_results = parse_log_history(trainer.state.log_history)
+        hyperparameters = extract_hyperparameters_from_trainer(trainer)
+
+        return cls(
+            language=language,
+            license=license,
+            tags=tags,
+            model_name=model_name,
+            finetuned_from=finetuned_from,
+            tasks=tasks,
+            dataset=dataset,
+            dataset_tags=dataset_tags,
+            dataset_args=dataset_args,
+            dataset_metadata=dataset_metadata,
+            eval_results=eval_results,
+            eval_lines=eval_lines,
+            hyperparameters=hyperparameters,
+        )
+
+    @classmethod
+    def from_keras(
+        cls,
+        model,
+        model_name,
+        keras_history=None,
+        language=None,
+        license=None,
+        tags=None,
+        finetuned_from=None,
+        tasks=None,
+        dataset_tags=None,
+        dataset=None,
+        dataset_args=None,
+    ):
+        # Infer default from dataset
+        if dataset is not None:
+            if is_hf_dataset(dataset) and (dataset_tags is None or dataset_args is None):
+                default_tag = dataset.builder_name
+                # Those are not real datasets from the Hub so we exclude them.
+                if default_tag not in ["csv", "json", "pandas", "parquet", "text"]:
+                    if dataset_tags is None:
+                        dataset_tags = [default_tag]
+                    if dataset_args is None:
+                        dataset_args = [dataset.config_name]
+
+        if dataset is None and dataset_tags is not None:
+            dataset = dataset_tags
+
+        # Infer default finetuned_from
+        if (
+            finetuned_from is None
+            and hasattr(model.config, "_name_or_path")
+            and not os.path.isdir(model.config._name_or_path)
+        ):
+            finetuned_from = model.config._name_or_path
+
+        # Infer default task tag:
+        if tasks is None:
+            model_class_name = model.__class__.__name__
+            for task, mapping in TASK_MAPPING.items():
+                if model_class_name in _get_mapping_values(mapping):
+                    tasks = task
+
+        # Add `generated_from_keras_callback` to the tags
+        if tags is None:
+            tags = ["generated_from_keras_callback"]
+        elif isinstance(tags, str) and tags != "generated_from_keras_callback":
+            tags = [tags, "generated_from_keras_callback"]
+        elif "generated_from_keras_callback" not in tags:
+            tags.append("generated_from_keras_callback")
+
+        if keras_history is not None:
+            _, eval_lines, eval_results = parse_keras_history(keras_history)
+        else:
+            eval_lines = []
+            eval_results = {}
+        hyperparameters = extract_hyperparameters_from_keras(model)
+
+        return cls(
+            language=language,
+            license=license,
+            tags=tags,
+            model_name=model_name,
+            finetuned_from=finetuned_from,
+            tasks=tasks,
+            dataset_tags=dataset_tags,
+            dataset=dataset,
+            dataset_args=dataset_args,
+            eval_results=eval_results,
+            eval_lines=eval_lines,
+            hyperparameters=hyperparameters,
+            source="keras",
+        )
+
+
+def parse_keras_history(logs):
+    """
+    Parse the `logs` of either a `tf.keras.History` object returned by `model.fit()` or an accumulated logs `dict`
+    passed to the `PushToHubCallback`. Returns lines and logs compatible with those returned by `parse_log_history`.
+    """
+    if hasattr(logs, "history"):
+        # This looks like a `History` object
+        if not hasattr(logs, "epoch"):
+            # This history looks empty, return empty results
+            return None, [], {}
+        logs.history["epoch"] = logs.epoch
+        logs = logs.history
+    else:
+        # Training logs is a list of dicts, let's invert it to a dict of lists to match a History object
+        logs = {log_key: [single_dict[log_key] for single_dict in logs] for log_key in logs[0]}
+
+    lines = []
+    for i in range(len(logs["epoch"])):
+        epoch_dict = {log_key: log_value_list[i] for log_key, log_value_list in logs.items()}
+        values = {}
+        for k, v in epoch_dict.items():
+            if k.startswith("val_"):
+                k = "validation_" + k[4:]
+            elif k != "epoch":
+                k = "train_" + k
+            splits = k.split("_")
+            name = " ".join([part.capitalize() for part in splits])
+            values[name] = v
+        lines.append(values)
+
+    eval_results = lines[-1]
+
+    return logs, lines, eval_results
+
+
+def parse_log_history(log_history):
+    """
+    Parse the `log_history` of a Trainer to get the intermediate and final evaluation results.
+    """
+    idx = 0
+    while idx < len(log_history) and "train_runtime" not in log_history[idx]:
+        idx += 1
+
+    # If there are no training logs
+    if idx == len(log_history):
+        idx -= 1
+        while idx >= 0 and "eval_loss" not in log_history[idx]:
+            idx -= 1
+
+        if idx >= 0:
+            return None, None, log_history[idx]
+        else:
+            return None, None, None
+
+    # From now one we can assume we have training logs:
+    train_log = log_history[idx]
+    lines = []
+    training_loss = "No log"
+    for i in range(idx):
+        if "loss" in log_history[i]:
+            training_loss = log_history[i]["loss"]
+        if "eval_loss" in log_history[i]:
+            metrics = log_history[i].copy()
+            _ = metrics.pop("total_flos", None)
+            epoch = metrics.pop("epoch", None)
+            step = metrics.pop("step", None)
+            _ = metrics.pop("eval_runtime", None)
+            _ = metrics.pop("eval_samples_per_second", None)
+            _ = metrics.pop("eval_steps_per_second", None)
+            _ = metrics.pop("eval_jit_compilation_time", None)
+            values = {"Training Loss": training_loss, "Epoch": epoch, "Step": step}
+            for k, v in metrics.items():
+                if k == "eval_loss":
+                    values["Validation Loss"] = v
+                else:
+                    splits = k.split("_")
+                    name = " ".join([part.capitalize() for part in splits[1:]])
+                    values[name] = v
+            lines.append(values)
+
+    idx = len(log_history) - 1
+    while idx >= 0 and "eval_loss" not in log_history[idx]:
+        idx -= 1
+
+    if idx > 0:
+        eval_results = {}
+        for key, value in log_history[idx].items():
+            if key.startswith("eval_"):
+                key = key[5:]
+            if key not in ["runtime", "samples_per_second", "steps_per_second", "epoch", "step"]:
+                camel_cased_key = " ".join([part.capitalize() for part in key.split("_")])
+                eval_results[camel_cased_key] = value
+        return train_log, lines, eval_results
+    else:
+        return train_log, lines, None
+
+
+def extract_hyperparameters_from_keras(model):
+    import tensorflow as tf
+
+    hyperparameters = {}
+    if hasattr(model, "optimizer") and model.optimizer is not None:
+        hyperparameters["optimizer"] = model.optimizer.get_config()
+    else:
+        hyperparameters["optimizer"] = None
+    hyperparameters["training_precision"] = tf.keras.mixed_precision.global_policy().name
+
+    return hyperparameters
+
+
+def _maybe_round(v, decimals=4):
+    if isinstance(v, float) and len(str(v).split(".")) > 1 and len(str(v).split(".")[1]) > decimals:
+        return f"{v:.{decimals}f}"
+    return str(v)
+
+
+def _regular_table_line(values, col_widths):
+    values_with_space = [f"| {v}" + " " * (w - len(v) + 1) for v, w in zip(values, col_widths)]
+    return "".join(values_with_space) + "|\n"
+
+
+def _second_table_line(col_widths):
+    values = ["|:" + "-" * w + ":" for w in col_widths]
+    return "".join(values) + "|\n"
+
+
+def make_markdown_table(lines):
+    """
+    Create a nice Markdown table from the results in `lines`.
+    """
+    if lines is None or len(lines) == 0:
+        return ""
+    col_widths = {key: len(str(key)) for key in lines[0].keys()}
+    for line in lines:
+        for key, value in line.items():
+            if col_widths[key] < len(_maybe_round(value)):
+                col_widths[key] = len(_maybe_round(value))
+
+    table = _regular_table_line(list(lines[0].keys()), list(col_widths.values()))
+    table += _second_table_line(list(col_widths.values()))
+    for line in lines:
+        table += _regular_table_line([_maybe_round(v) for v in line.values()], list(col_widths.values()))
+    return table
+
+
+_TRAINING_ARGS_KEYS = [
+    "learning_rate",
+    "train_batch_size",
+    "eval_batch_size",
+    "seed",
+]
+
+
+def extract_hyperparameters_from_trainer(trainer):
+    hyperparameters = {k: getattr(trainer.args, k) for k in _TRAINING_ARGS_KEYS}
+
+    if trainer.args.parallel_mode not in [ParallelMode.NOT_PARALLEL, ParallelMode.NOT_DISTRIBUTED]:
+        hyperparameters["distributed_type"] = (
+            "multi-GPU" if trainer.args.parallel_mode == ParallelMode.DISTRIBUTED else trainer.args.parallel_mode.value
+        )
+    if trainer.args.world_size > 1:
+        hyperparameters["num_devices"] = trainer.args.world_size
+    if trainer.args.gradient_accumulation_steps > 1:
+        hyperparameters["gradient_accumulation_steps"] = trainer.args.gradient_accumulation_steps
+
+    total_train_batch_size = (
+        trainer.args.train_batch_size * trainer.args.world_size * trainer.args.gradient_accumulation_steps
+    )
+    if total_train_batch_size != hyperparameters["train_batch_size"]:
+        hyperparameters["total_train_batch_size"] = total_train_batch_size
+    total_eval_batch_size = trainer.args.eval_batch_size * trainer.args.world_size
+    if total_eval_batch_size != hyperparameters["eval_batch_size"]:
+        hyperparameters["total_eval_batch_size"] = total_eval_batch_size
+
+    if trainer.args.adafactor:
+        hyperparameters["optimizer"] = "Adafactor"
+    else:
+        hyperparameters["optimizer"] = (
+            f"Adam with betas=({trainer.args.adam_beta1},{trainer.args.adam_beta2}) and"
+            f" epsilon={trainer.args.adam_epsilon}"
+        )
+
+    hyperparameters["lr_scheduler_type"] = trainer.args.lr_scheduler_type.value
+    if trainer.args.warmup_ratio != 0.0:
+        hyperparameters["lr_scheduler_warmup_ratio"] = trainer.args.warmup_ratio
+    if trainer.args.warmup_steps != 0.0:
+        hyperparameters["lr_scheduler_warmup_steps"] = trainer.args.warmup_steps
+    if trainer.args.max_steps != -1:
+        hyperparameters["training_steps"] = trainer.args.max_steps
+    else:
+        hyperparameters["num_epochs"] = trainer.args.num_train_epochs
+
+    if trainer.args.fp16:
+        if trainer.use_cuda_amp:
+            hyperparameters["mixed_precision_training"] = "Native AMP"
+        elif trainer.use_apex:
+            hyperparameters["mixed_precision_training"] = f"Apex, opt level {trainer.args.fp16_opt_level}"
+
+    if trainer.args.label_smoothing_factor != 0.0:
+        hyperparameters["label_smoothing_factor"] = trainer.args.label_smoothing_factor
+
+    return hyperparameters

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/modeling_flax_pytorch_utils.py

@@ -0,0 +1,421 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch - Flax general utilities."""
+
+
+import os
+from pickle import UnpicklingError
+from typing import Dict, Tuple
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.serialization import from_bytes
+from flax.traverse_util import flatten_dict, unflatten_dict
+
+import transformers
+
+from .utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+#####################
+# PyTorch => Flax #
+#####################
+
+
+def load_pytorch_checkpoint_in_flax_state_dict(
+    flax_model, pytorch_checkpoint_path, is_sharded, allow_missing_keys=False
+):
+    """Load pytorch checkpoints in a flax model"""
+    try:
+        import torch  # noqa: F401
+    except ImportError:
+        logger.error(
+            "Loading a PyTorch model in Flax, requires both PyTorch and Flax to be installed. Please see"
+            " https://pytorch.org/ and https://flax.readthedocs.io/en/latest/installation.html for installation"
+            " instructions."
+        )
+        raise
+
+    if not is_sharded:
+        pt_path = os.path.abspath(pytorch_checkpoint_path)
+        logger.info(f"Loading PyTorch weights from {pt_path}")
+
+        pt_state_dict = torch.load(pt_path, map_location="cpu")
+        logger.info(f"PyTorch checkpoint contains {sum(t.numel() for t in pt_state_dict.values()):,} parameters.")
+
+        flax_state_dict = convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model)
+    else:
+        # model is sharded and pytorch_checkpoint_path already contains the list of .pt shard files
+        flax_state_dict = convert_pytorch_sharded_state_dict_to_flax(pytorch_checkpoint_path, flax_model)
+    return flax_state_dict
+
+
+def rename_key_and_reshape_tensor(
+    pt_tuple_key: Tuple[str],
+    pt_tensor: np.ndarray,
+    random_flax_state_dict: Dict[str, jnp.ndarray],
+    model_prefix: str,
+) -> (Tuple[str], np.ndarray):
+    """Rename PT weight names to corresponding Flax weight names and reshape tensor if necessary"""
+
+    def is_key_or_prefix_key_in_dict(key: Tuple[str]) -> bool:
+        """Checks if `key` of `(prefix,) + key` is in random_flax_state_dict"""
+        return len(set(random_flax_state_dict) & {key, (model_prefix,) + key}) > 0
+
+    # layer norm
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
+    if pt_tuple_key[-1] in ["weight", "gamma"] and is_key_or_prefix_key_in_dict(renamed_pt_tuple_key):
+        return renamed_pt_tuple_key, pt_tensor
+
+    # batch norm layer mean
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("mean",)
+    if pt_tuple_key[-1] == "running_mean" and not is_key_or_prefix_key_in_dict(pt_tuple_key):
+        return renamed_pt_tuple_key, pt_tensor
+
+    # batch norm layer var
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("var",)
+    if pt_tuple_key[-1] == "running_var" and not is_key_or_prefix_key_in_dict(pt_tuple_key):
+        return renamed_pt_tuple_key, pt_tensor
+
+    # embedding
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("embedding",)
+    if pt_tuple_key[-1] == "weight" and is_key_or_prefix_key_in_dict(renamed_pt_tuple_key):
+        return renamed_pt_tuple_key, pt_tensor
+
+    # conv layer
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)
+    if pt_tuple_key[-1] == "weight" and pt_tensor.ndim == 4 and not is_key_or_prefix_key_in_dict(pt_tuple_key):
+        pt_tensor = pt_tensor.transpose(2, 3, 1, 0)
+        return renamed_pt_tuple_key, pt_tensor
+
+    # linear layer
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)
+    if pt_tuple_key[-1] == "weight" and not is_key_or_prefix_key_in_dict(pt_tuple_key):
+        pt_tensor = pt_tensor.T
+        return renamed_pt_tuple_key, pt_tensor
+
+    # old PyTorch layer norm weight
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("weight",)
+    if pt_tuple_key[-1] == "gamma":
+        return renamed_pt_tuple_key, pt_tensor
+
+    # old PyTorch layer norm bias
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("bias",)
+    if pt_tuple_key[-1] == "beta":
+        return renamed_pt_tuple_key, pt_tensor
+
+    return pt_tuple_key, pt_tensor
+
+
+def convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model):
+    # convert pytorch tensor to numpy
+    pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
+
+    model_prefix = flax_model.base_model_prefix
+
+    # use params dict if the model contains batch norm layers
+    if "params" in flax_model.params:
+        flax_model_params = flax_model.params["params"]
+    else:
+        flax_model_params = flax_model.params
+    random_flax_state_dict = flatten_dict(flax_model_params)
+
+    # add batch_stats keys,values to dict
+    if "batch_stats" in flax_model.params:
+        flax_batch_stats = flatten_dict(flax_model.params["batch_stats"])
+        random_flax_state_dict.update(flax_batch_stats)
+
+    flax_state_dict = {}
+
+    load_model_with_head_into_base_model = (model_prefix not in flax_model_params) and (
+        model_prefix in {k.split(".")[0] for k in pt_state_dict.keys()}
+    )
+    load_base_model_into_model_with_head = (model_prefix in flax_model_params) and (
+        model_prefix not in {k.split(".")[0] for k in pt_state_dict.keys()}
+    )
+
+    # Need to change some parameters name to match Flax names
+    for pt_key, pt_tensor in pt_state_dict.items():
+        pt_tuple_key = tuple(pt_key.split("."))
+
+        # remove base model prefix if necessary
+        has_base_model_prefix = pt_tuple_key[0] == model_prefix
+        if load_model_with_head_into_base_model and has_base_model_prefix:
+            pt_tuple_key = pt_tuple_key[1:]
+
+        # Correctly rename weight parameters
+        flax_key, flax_tensor = rename_key_and_reshape_tensor(
+            pt_tuple_key, pt_tensor, random_flax_state_dict, model_prefix
+        )
+
+        # add model prefix if necessary
+        require_base_model_prefix = (model_prefix,) + flax_key in random_flax_state_dict
+        if load_base_model_into_model_with_head and require_base_model_prefix:
+            flax_key = (model_prefix,) + flax_key
+
+        if flax_key in random_flax_state_dict:
+            if flax_tensor.shape != random_flax_state_dict[flax_key].shape:
+                raise ValueError(
+                    f"PyTorch checkpoint seems to be incorrect. Weight {pt_key} was expected to be of shape "
+                    f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."
+                )
+
+        # add batch stats if the model contains batchnorm layers
+        if "batch_stats" in flax_model.params:
+            if "mean" in flax_key[-1] or "var" in flax_key[-1]:
+                flax_state_dict[("batch_stats",) + flax_key] = jnp.asarray(flax_tensor)
+                continue
+            # remove num_batches_tracked key
+            if "num_batches_tracked" in flax_key[-1]:
+                flax_state_dict.pop(flax_key, None)
+                continue
+
+            # also add unexpected weight so that warning is thrown
+            flax_state_dict[("params",) + flax_key] = jnp.asarray(flax_tensor)
+
+        else:
+            # also add unexpected weight so that warning is thrown
+            flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
+
+    return unflatten_dict(flax_state_dict)
+
+
+############################
+# Sharded Pytorch => Flax #
+############################
+
+
+def convert_pytorch_sharded_state_dict_to_flax(shard_filenames, flax_model):
+    import torch
+
+    # Load the index
+    flax_state_dict = {}
+    for shard_file in shard_filenames:
+        # load using msgpack utils
+        pt_state_dict = torch.load(shard_file)
+        pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
+
+        model_prefix = flax_model.base_model_prefix
+
+        # use params dict if the model contains batch norm layers and then add batch_stats keys,values to dict
+        if "batch_stats" in flax_model.params:
+            flax_model_params = flax_model.params["params"]
+
+            random_flax_state_dict = flatten_dict(flax_model_params)
+            random_flax_state_dict.update(flatten_dict(flax_model.params["batch_stats"]))
+        else:
+            flax_model_params = flax_model.params
+            random_flax_state_dict = flatten_dict(flax_model_params)
+
+        load_model_with_head_into_base_model = (model_prefix not in flax_model_params) and (
+            model_prefix in {k.split(".")[0] for k in pt_state_dict.keys()}
+        )
+        load_base_model_into_model_with_head = (model_prefix in flax_model_params) and (
+            model_prefix not in {k.split(".")[0] for k in pt_state_dict.keys()}
+        )
+        # Need to change some parameters name to match Flax names
+        for pt_key, pt_tensor in pt_state_dict.items():
+            pt_tuple_key = tuple(pt_key.split("."))
+
+            # remove base model prefix if necessary
+            has_base_model_prefix = pt_tuple_key[0] == model_prefix
+            if load_model_with_head_into_base_model and has_base_model_prefix:
+                pt_tuple_key = pt_tuple_key[1:]
+
+            # Correctly rename weight parameters
+            flax_key, flax_tensor = rename_key_and_reshape_tensor(
+                pt_tuple_key, pt_tensor, random_flax_state_dict, model_prefix
+            )
+            # add model prefix if necessary
+            require_base_model_prefix = (model_prefix,) + flax_key in random_flax_state_dict
+            if load_base_model_into_model_with_head and require_base_model_prefix:
+                flax_key = (model_prefix,) + flax_key
+
+            if flax_key in random_flax_state_dict:
+                if flax_tensor.shape != random_flax_state_dict[flax_key].shape:
+                    raise ValueError(
+                        f"PyTorch checkpoint seems to be incorrect. Weight {pt_key} was expected to be of shape "
+                        f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."
+                    )
+
+            # add batch stats if the model contains batchnorm layers
+            if "batch_stats" in flax_model.params:
+                if "mean" in flax_key[-1]:
+                    flax_state_dict[("batch_stats",) + flax_key] = jnp.asarray(flax_tensor)
+                    continue
+                if "var" in flax_key[-1]:
+                    flax_state_dict[("batch_stats",) + flax_key] = jnp.asarray(flax_tensor)
+                    continue
+                # remove num_batches_tracked key
+                if "num_batches_tracked" in flax_key[-1]:
+                    flax_state_dict.pop(flax_key, None)
+                    continue
+
+                # also add unexpected weight so that warning is thrown
+                flax_state_dict[("params",) + flax_key] = jnp.asarray(flax_tensor)
+
+            else:
+                # also add unexpected weight so that warning is thrown
+                flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
+    return unflatten_dict(flax_state_dict)
+
+
+#####################
+# Flax => PyTorch #
+#####################
+
+
+def load_flax_checkpoint_in_pytorch_model(model, flax_checkpoint_path):
+    """Load flax checkpoints in a PyTorch model"""
+    flax_checkpoint_path = os.path.abspath(flax_checkpoint_path)
+    logger.info(f"Loading Flax weights from {flax_checkpoint_path}")
+
+    # import correct flax class
+    flax_cls = getattr(transformers, "Flax" + model.__class__.__name__)
+
+    # load flax weight dict
+    with open(flax_checkpoint_path, "rb") as state_f:
+        try:
+            flax_state_dict = from_bytes(flax_cls, state_f.read())
+        except UnpicklingError:
+            raise EnvironmentError(f"Unable to convert {flax_checkpoint_path} to Flax deserializable object. ")
+
+    return load_flax_weights_in_pytorch_model(model, flax_state_dict)
+
+
+def load_flax_weights_in_pytorch_model(pt_model, flax_state):
+    """Load flax checkpoints in a PyTorch model"""
+
+    try:
+        import torch  # noqa: F401
+    except ImportError:
+        logger.error(
+            "Loading a Flax weights in PyTorch, requires both PyTorch and Flax to be installed. Please see"
+            " https://pytorch.org/ and https://flax.readthedocs.io/en/latest/installation.html for installation"
+            " instructions."
+        )
+        raise
+
+    # check if we have bf16 weights
+    is_type_bf16 = flatten_dict(jax.tree_util.tree_map(lambda x: x.dtype == jnp.bfloat16, flax_state)).values()
+    if any(is_type_bf16):
+        # convert all weights to fp32 if the are bf16 since torch.from_numpy can-not handle bf16
+        # and bf16 is not fully supported in PT yet.
+        logger.warning(
+            "Found ``bfloat16`` weights in Flax model. Casting all ``bfloat16`` weights to ``float32`` "
+            "before loading those in PyTorch model."
+        )
+        flax_state = jax.tree_util.tree_map(
+            lambda params: params.astype(np.float32) if params.dtype == jnp.bfloat16 else params, flax_state
+        )
+
+    flax_state_dict = flatten_dict(flax_state)
+    pt_model_dict = pt_model.state_dict()
+
+    load_model_with_head_into_base_model = (pt_model.base_model_prefix in flax_state) and (
+        pt_model.base_model_prefix not in {k.split(".")[0] for k in pt_model_dict.keys()}
+    )
+    load_base_model_into_model_with_head = (pt_model.base_model_prefix not in flax_state) and (
+        pt_model.base_model_prefix in {k.split(".")[0] for k in pt_model_dict.keys()}
+    )
+
+    # keep track of unexpected & missing keys
+    unexpected_keys = []
+    missing_keys = set(pt_model_dict.keys())
+
+    for flax_key_tuple, flax_tensor in flax_state_dict.items():
+        has_base_model_prefix = flax_key_tuple[0] == pt_model.base_model_prefix
+        require_base_model_prefix = ".".join((pt_model.base_model_prefix,) + flax_key_tuple) in pt_model_dict
+
+        # adapt flax_key to prepare for loading from/to base model only
+        if load_model_with_head_into_base_model and has_base_model_prefix:
+            flax_key_tuple = flax_key_tuple[1:]
+        elif load_base_model_into_model_with_head and require_base_model_prefix:
+            flax_key_tuple = (pt_model.base_model_prefix,) + flax_key_tuple
+
+        # rename flax weights to PyTorch format
+        if flax_key_tuple[-1] == "kernel" and flax_tensor.ndim == 4 and ".".join(flax_key_tuple) not in pt_model_dict:
+            # conv layer
+            flax_key_tuple = flax_key_tuple[:-1] + ("weight",)
+            flax_tensor = jnp.transpose(flax_tensor, (3, 2, 0, 1))
+        elif flax_key_tuple[-1] == "kernel" and ".".join(flax_key_tuple) not in pt_model_dict:
+            # linear layer
+            flax_key_tuple = flax_key_tuple[:-1] + ("weight",)
+            flax_tensor = flax_tensor.T
+        elif flax_key_tuple[-1] in ["scale", "embedding"]:
+            flax_key_tuple = flax_key_tuple[:-1] + ("weight",)
+
+        # adding batch stats from flax batch norm to pt
+        elif "mean" in flax_key_tuple[-1]:
+            flax_key_tuple = flax_key_tuple[:-1] + ("running_mean",)
+        elif "var" in flax_key_tuple[-1]:
+            flax_key_tuple = flax_key_tuple[:-1] + ("running_var",)
+
+        if "batch_stats" in flax_state:
+            flax_key = ".".join(flax_key_tuple[1:])  # Remove the params/batch_stats header
+        else:
+            flax_key = ".".join(flax_key_tuple)
+
+        if flax_key in pt_model_dict:
+            if flax_tensor.shape != pt_model_dict[flax_key].shape:
+                raise ValueError(
+                    f"Flax checkpoint seems to be incorrect. Weight {flax_key_tuple} was expected "
+                    f"to be of shape {pt_model_dict[flax_key].shape}, but is {flax_tensor.shape}."
+                )
+            else:
+                # add weight to pytorch dict
+                flax_tensor = np.asarray(flax_tensor) if not isinstance(flax_tensor, np.ndarray) else flax_tensor
+                pt_model_dict[flax_key] = torch.from_numpy(flax_tensor)
+                # remove from missing keys
+                missing_keys.remove(flax_key)
+        else:
+            # weight is not expected by PyTorch model
+            unexpected_keys.append(flax_key)
+
+    pt_model.load_state_dict(pt_model_dict)
+
+    # re-transform missing_keys to list
+    missing_keys = list(missing_keys)
+
+    if len(unexpected_keys) > 0:
+        logger.warning(
+            "Some weights of the Flax model were not used when initializing the PyTorch model"
+            f" {pt_model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are initializing"
+            f" {pt_model.__class__.__name__} from a Flax model trained on another task or with another architecture"
+            " (e.g. initializing a BertForSequenceClassification model from a FlaxBertForPreTraining model).\n- This"
+            f" IS NOT expected if you are initializing {pt_model.__class__.__name__} from a Flax model that you expect"
+            " to be exactly identical (e.g. initializing a BertForSequenceClassification model from a"
+            " FlaxBertForSequenceClassification model)."
+        )
+    else:
+        logger.warning(f"All Flax model weights were used when initializing {pt_model.__class__.__name__}.\n")
+    if len(missing_keys) > 0:
+        logger.warning(
+            f"Some weights of {pt_model.__class__.__name__} were not initialized from the Flax model and are newly"
+            f" initialized: {missing_keys}\nYou should probably TRAIN this model on a down-stream task to be able to"
+            " use it for predictions and inference."
+        )
+    else:
+        logger.warning(
+            f"All the weights of {pt_model.__class__.__name__} were initialized from the Flax model.\n"
+            "If your task is similar to the task the model of the checkpoint was trained on, "
+            f"you can already use {pt_model.__class__.__name__} for predictions without further training."
+        )
+
+    return pt_model

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/modeling_flax_utils.py

@@ -0,0 +1,1170 @@
+# coding=utf-8
+# Copyright 2021 The Google Flax Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import gc
+import json
+import os
+import re
+from functools import partial
+from pickle import UnpicklingError
+from typing import Any, Dict, Set, Tuple, Union
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+import msgpack.exceptions
+from flax.core.frozen_dict import FrozenDict, unfreeze
+from flax.serialization import from_bytes, to_bytes
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax.random import PRNGKey
+
+from .configuration_utils import PretrainedConfig
+from .dynamic_module_utils import custom_object_save
+from .generation import FlaxGenerationMixin, GenerationConfig
+from .modeling_flax_pytorch_utils import load_pytorch_checkpoint_in_flax_state_dict
+from .utils import (
+    FLAX_WEIGHTS_INDEX_NAME,
+    FLAX_WEIGHTS_NAME,
+    WEIGHTS_INDEX_NAME,
+    WEIGHTS_NAME,
+    PushToHubMixin,
+    add_code_sample_docstrings,
+    add_start_docstrings_to_model_forward,
+    cached_file,
+    copy_func,
+    download_url,
+    has_file,
+    is_offline_mode,
+    is_remote_url,
+    logging,
+    replace_return_docstrings,
+)
+from .utils.hub import convert_file_size_to_int, get_checkpoint_shard_files
+
+
+logger = logging.get_logger(__name__)
+
+
+def quick_gelu(x):
+    return x * jax.nn.sigmoid(1.702 * x)
+
+
+ACT2FN = {
+    "gelu": partial(nn.gelu, approximate=False),
+    "relu": nn.relu,
+    "silu": nn.swish,
+    "swish": nn.swish,
+    "gelu_new": partial(nn.gelu, approximate=True),
+    "quick_gelu": quick_gelu,
+}
+
+
+def dtype_byte_size(dtype):
+    """
+    Returns the size (in bytes) occupied by one parameter of type `dtype`. Example:
+    ```py
+    >>> dtype_byte_size(np.float32)
+    4
+    ```
+    """
+    if dtype == bool:
+        return 1 / 8
+    bit_search = re.search(r"[^\d](\d+)$", dtype.name)
+    if bit_search is None:
+        raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
+    bit_size = int(bit_search.groups()[0])
+    return bit_size // 8
+
+
+def flax_shard_checkpoint(params, max_shard_size="10GB"):
+    """
+    Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
+    given size. The sub-checkpoints are determined by iterating through the `state_dict` in the order of its keys, so
+    there is no optimization made to make each sub-checkpoint as close as possible to the maximum size passed. For
+    example, if the limit is 10GB and we have weights of sizes [6GB, 6GB, 2GB, 6GB, 2GB, 2GB] they will get sharded as
+    [6GB], [6+2GB], [6+2+2GB] and not [6+2+2GB], [6+2GB], [6GB].
+
+    <Tip warning={true}>
+
+    If one of the model's weight is bigger that `max_shard_size`, it will end up in its own sub-checkpoint which will
+    have a size greater than `max_shard_size`.
+
+    </Tip>
+
+    Args:
+        params (`Union[Dict, FrozenDict]`): A `PyTree` of model parameters.
+        max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
+            The maximum size of each sub-checkpoint. If expressed as a string, needs to be digits followed by a unit
+            (like `"5MB"`).
+    """
+    max_shard_size = convert_file_size_to_int(max_shard_size)
+
+    sharded_state_dicts = []
+    current_block = {}
+    current_block_size = 0
+    total_size = 0
+
+    # flatten the weights to chunk
+    weights = flatten_dict(params, sep="/")
+    for item in weights:
+        weight_size = weights[item].size * dtype_byte_size(weights[item].dtype)
+
+        # If this weight is going to tip up over the maximal size, we split.
+        if current_block_size + weight_size > max_shard_size:
+            sharded_state_dicts.append(current_block)
+            current_block = {}
+            current_block_size = 0
+
+        current_block[item] = weights[item]
+        current_block_size += weight_size
+        total_size += weight_size
+
+    # Add the last block
+    sharded_state_dicts.append(current_block)
+
+    # If we only have one shard, we return it
+    if len(sharded_state_dicts) == 1:
+        return {FLAX_WEIGHTS_NAME: sharded_state_dicts[0]}, None
+
+    # Otherwise, let's build the index
+    weight_map = {}
+    shards = {}
+    for idx, shard in enumerate(sharded_state_dicts):
+        shard_file = FLAX_WEIGHTS_NAME.replace(".msgpack", f"-{idx+1:05d}-of-{len(sharded_state_dicts):05d}.msgpack")
+        shards[shard_file] = shard
+        for weight_name in shard.keys():
+            weight_map[weight_name] = shard_file
+
+    # Add the metadata
+    metadata = {"total_size": total_size}
+    index = {"metadata": metadata, "weight_map": weight_map}
+    return shards, index
+
+
+class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
+    r"""
+    Base class for all models.
+
+    [`FlaxPreTrainedModel`] takes care of storing the configuration of the models and handles methods for loading,
+    downloading and saving models.
+
+    Class attributes (overridden by derived classes):
+
+        - **config_class** ([`PretrainedConfig`]) -- A subclass of [`PretrainedConfig`] to use as configuration class
+          for this model architecture.
+        - **base_model_prefix** (`str`) -- A string indicating the attribute associated to the base model in derived
+          classes of the same architecture adding modules on top of the base model.
+        - **main_input_name** (`str`) -- The name of the principal input to the model (often `input_ids` for NLP
+          models, `pixel_values` for vision models and `input_values` for speech models).
+    """
+    config_class = None
+    base_model_prefix = ""
+    main_input_name = "input_ids"
+    _auto_class = None
+    _missing_keys = set()
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        module: nn.Module,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+    ):
+        if config is None:
+            raise ValueError("config cannot be None")
+
+        if module is None:
+            raise ValueError("module cannot be None")
+
+        # Those are private to be exposed as typed property on derived classes.
+        self._config = config
+        self._module = module
+
+        # Those are public as their type is generic to every derived classes.
+        self.key = PRNGKey(seed)
+        self.dtype = dtype
+        self.input_shape = input_shape
+        self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
+
+        # To check if the model was intialized automatically.
+        self._is_initialized = _do_init
+
+        if _do_init:
+            # randomly initialized parameters
+            random_params = self.init_weights(self.key, input_shape)
+            params_shape_tree = jax.eval_shape(lambda params: params, random_params)
+        else:
+            init_fn = partial(self.init_weights, input_shape=input_shape)
+            params_shape_tree = jax.eval_shape(init_fn, self.key)
+
+            logger.info(
+                "Model weights are not initialized as `_do_init` is set to `False`. "
+                f"Make sure to call `{self.__class__.__name__}.init_weights` manually to initialize the weights."
+            )
+
+        # get the shape of the parameters
+        self._params_shape_tree = params_shape_tree
+
+        # save required_params as set
+        self._required_params = set(flatten_dict(unfreeze(params_shape_tree)).keys())
+
+        # initialize the parameters
+        if _do_init:
+            self.params = random_params
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> Dict:
+        raise NotImplementedError(f"init method has to be implemented for {self}")
+
+    def enable_gradient_checkpointing(self):
+        raise NotImplementedError(f"gradient checkpointing method has to be implemented for {self}")
+
+    @classmethod
+    def _from_config(cls, config, **kwargs):
+        """
+        All context managers that the model should be initialized under go here.
+        """
+        return cls(config, **kwargs)
+
+    @property
+    def framework(self) -> str:
+        """
+        :str: Identifies that this is a Flax model.
+        """
+        return "flax"
+
+    @property
+    def config(self) -> PretrainedConfig:
+        return self._config
+
+    @property
+    def module(self) -> nn.Module:
+        return self._module
+
+    @property
+    def params(self) -> Union[Dict, FrozenDict]:
+        if not self._is_initialized:
+            raise ValueError(
+                "`params` cannot be accessed from model when the model is created with `_do_init=False`. "
+                "You must call `init_weights` manually and store the params outside of the model and "
+                "pass it explicitly where needed."
+            )
+        return self._params
+
+    @property
+    def required_params(self) -> Set:
+        return self._required_params
+
+    @property
+    def params_shape_tree(self) -> Dict:
+        return self._params_shape_tree
+
+    @params.setter
+    def params(self, params: Union[Dict, FrozenDict]):
+        # don't set params if the model is not initialized
+        if not self._is_initialized:
+            raise ValueError(
+                "`params` cannot be set from model when the model is created with `_do_init=False`. "
+                "You store the params outside of the model."
+            )
+
+        if isinstance(params, FrozenDict):
+            params = unfreeze(params)
+        param_keys = set(flatten_dict(params).keys())
+        if len(self.required_params - param_keys) > 0:
+            raise ValueError(
+                "Some parameters are missing. Make sure that `params` include the following "
+                f"parameters {self.required_params - param_keys}"
+            )
+        self._params = params
+
+    def _cast_floating_to(self, params: Union[Dict, FrozenDict], dtype: jnp.dtype, mask: Any = None) -> Any:
+        """
+        Helper method to cast floating-point values of given parameter `PyTree` to given `dtype`.
+        """
+
+        # taken from https://github.com/deepmind/jmp/blob/3a8318abc3292be38582794dbf7b094e6583b192/jmp/_src/policy.py#L27
+        def conditional_cast(param):
+            if isinstance(param, jnp.ndarray) and jnp.issubdtype(param.dtype, jnp.floating):
+                param = param.astype(dtype)
+            return param
+
+        if mask is None:
+            return jax.tree_util.tree_map(conditional_cast, params)
+
+        flat_params = flatten_dict(params)
+        flat_mask, _ = jax.tree_util.tree_flatten(mask)
+
+        for masked, key in zip(flat_mask, flat_params.keys()):
+            if masked:
+                param = flat_params[key]
+                flat_params[key] = conditional_cast(param)
+
+        return unflatten_dict(flat_params)
+
+    def to_bf16(self, params: Union[Dict, FrozenDict], mask: Any = None):
+        r"""
+        Cast the floating-point `params` to `jax.numpy.bfloat16`. This returns a new `params` tree and does not cast
+        the `params` in place.
+
+        This method can be used on TPU to explicitly convert the model parameters to bfloat16 precision to do full
+        half-precision training or to save weights in bfloat16 for inference in order to save memory and improve speed.
+
+        Arguments:
+            params (`Union[Dict, FrozenDict]`):
+                A `PyTree` of model parameters.
+            mask (`Union[Dict, FrozenDict]`):
+                A `PyTree` with same structure as the `params` tree. The leaves should be booleans, `True` for params
+                you want to cast, and should be `False` for those you want to skip.
+
+        Examples:
+
+        ```python
+        >>> from transformers import FlaxBertModel
+
+        >>> # load model
+        >>> model = FlaxBertModel.from_pretrained("bert-base-cased")
+        >>> # By default, the model parameters will be in fp32 precision, to cast these to bfloat16 precision
+        >>> model.params = model.to_bf16(model.params)
+        >>> # If you want don't want to cast certain parameters (for example layer norm bias and scale)
+        >>> # then pass the mask as follows
+        >>> from flax import traverse_util
+
+        >>> model = FlaxBertModel.from_pretrained("bert-base-cased")
+        >>> flat_params = traverse_util.flatten_dict(model.params)
+        >>> mask = {
+        ...     path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale"))
+        ...     for path in flat_params
+        ... }
+        >>> mask = traverse_util.unflatten_dict(mask)
+        >>> model.params = model.to_bf16(model.params, mask)
+        ```"""
+        return self._cast_floating_to(params, jnp.bfloat16, mask)
+
+    def to_fp32(self, params: Union[Dict, FrozenDict], mask: Any = None):
+        r"""
+        Cast the floating-point `parmas` to `jax.numpy.float32`. This method can be used to explicitly convert the
+        model parameters to fp32 precision. This returns a new `params` tree and does not cast the `params` in place.
+
+        Arguments:
+            params (`Union[Dict, FrozenDict]`):
+                A `PyTree` of model parameters.
+            mask (`Union[Dict, FrozenDict]`):
+                A `PyTree` with same structure as the `params` tree. The leaves should be booleans, `True` for params
+                you want to cast, and should be `False` for those you want to skip
+
+        Examples:
+
+        ```python
+        >>> from transformers import FlaxBertModel
+
+        >>> # Download model and configuration from huggingface.co
+        >>> model = FlaxBertModel.from_pretrained("bert-base-cased")
+        >>> # By default, the model params will be in fp32, to illustrate the use of this method,
+        >>> # we'll first cast to fp16 and back to fp32
+        >>> model.params = model.to_f16(model.params)
+        >>> # now cast back to fp32
+        >>> model.params = model.to_fp32(model.params)
+        ```"""
+        return self._cast_floating_to(params, jnp.float32, mask)
+
+    def to_fp16(self, params: Union[Dict, FrozenDict], mask: Any = None):
+        r"""
+        Cast the floating-point `parmas` to `jax.numpy.float16`. This returns a new `params` tree and does not cast the
+        `params` in place.
+
+        This method can be used on GPU to explicitly convert the model parameters to float16 precision to do full
+        half-precision training or to save weights in float16 for inference in order to save memory and improve speed.
+
+        Arguments:
+            params (`Union[Dict, FrozenDict]`):
+                A `PyTree` of model parameters.
+            mask (`Union[Dict, FrozenDict]`):
+                A `PyTree` with same structure as the `params` tree. The leaves should be booleans, `True` for params
+                you want to cast, and should be `False` for those you want to skip
+
+        Examples:
+
+        ```python
+        >>> from transformers import FlaxBertModel
+
+        >>> # load model
+        >>> model = FlaxBertModel.from_pretrained("bert-base-cased")
+        >>> # By default, the model params will be in fp32, to cast these to float16
+        >>> model.params = model.to_fp16(model.params)
+        >>> # If you want don't want to cast certain parameters (for example layer norm bias and scale)
+        >>> # then pass the mask as follows
+        >>> from flax import traverse_util
+
+        >>> model = FlaxBertModel.from_pretrained("bert-base-cased")
+        >>> flat_params = traverse_util.flatten_dict(model.params)
+        >>> mask = {
+        ...     path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale"))
+        ...     for path in flat_params
+        ... }
+        >>> mask = traverse_util.unflatten_dict(mask)
+        >>> model.params = model.to_fp16(model.params, mask)
+        ```"""
+        return self._cast_floating_to(params, jnp.float16, mask)
+
+    @classmethod
+    def load_flax_sharded_weights(cls, shard_files):
+        """
+        This is the same as [`flax.serialization.from_bytes`]
+        (https:lax.readthedocs.io/en/latest/_modules/flax/serialization.html#from_bytes) but for a sharded checkpoint.
+
+        This load is performed efficiently: each checkpoint shard is loaded one by one in RAM and deleted after being
+        loaded in the model.
+
+        Args:
+            shard_files (`List[str]`:
+                The list of shard files to load.
+
+        Returns:
+            `Dict`: A nested dictionary of the model parameters, in the expected format for flax models : `{'model':
+            {'params': {'...'}}}`.
+        """
+
+        # Load the index
+        state_sharded_dict = {}
+
+        for shard_file in shard_files:
+            # load using msgpack utils
+            try:
+                with open(shard_file, "rb") as state_f:
+                    state = from_bytes(cls, state_f.read())
+            except (UnpicklingError, msgpack.exceptions.ExtraData) as e:
+                with open(shard_file) as f:
+                    if f.read().startswith("version"):
+                        raise OSError(
+                            "You seem to have cloned a repository without having git-lfs installed. Please"
+                            " install git-lfs and run `git lfs install` followed by `git lfs pull` in the"
+                            " folder you cloned."
+                        )
+                    else:
+                        raise ValueError from e
+            except (UnicodeDecodeError, ValueError):
+                raise EnvironmentError(f"Unable to convert {shard_file} to Flax deserializable object. ")
+
+            state = flatten_dict(state, sep="/")
+            state_sharded_dict.update(state)
+            del state
+            gc.collect()
+
+        # the state dict is unflattened to the match the format of model.params
+        return unflatten_dict(state_sharded_dict, sep="/")
+
+    def can_generate(self) -> bool:
+        """
+        Returns whether this model can generate sequences with `.generate()`. Returns:
+            `bool`: Whether this model can generate sequences with `.generate()`.
+        """
+        # Detects whether `prepare_inputs_for_generation` has been overwritten, which is a requirement for generation
+        if "GenerationMixin" in str(self.prepare_inputs_for_generation):
+            return False
+        return True
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, os.PathLike],
+        dtype: jnp.dtype = jnp.float32,
+        *model_args,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a pretrained flax model from a pre-trained model configuration.
+
+        The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
+        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+        task.
+
+        The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
+        weights are discarded.
+
+        Parameters:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                Can be either:
+
+                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                      user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing model weights saved using
+                      [`~FlaxPreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
+                    - A path or url to a *pt index checkpoint file* (e.g, `./tf_model/model.ckpt.index`). In this case,
+                      `from_pt` should be set to `True`.
+            dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+                The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
+                `jax.numpy.bfloat16` (on TPUs).
+
+                This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
+                specified all the computation will be performed with the given `dtype`.
+
+                **Note that this only specifies the dtype of the computation and does not influence the dtype of model
+                parameters.**
+
+                If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`] and
+                [`~FlaxPreTrainedModel.to_bf16`].
+            model_args (sequence of positional arguments, *optional*):
+                All remaining positional arguments will be passed to the underlying model's `__init__` method.
+            config (`Union[PretrainedConfig, str, os.PathLike]`, *optional*):
+                Can be either:
+
+                    - an instance of a class derived from [`PretrainedConfig`],
+                    - a string or path valid as input to [`~PretrainedConfig.from_pretrained`].
+
+                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
+                be automatically loaded when:
+
+                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
+                      model).
+                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
+                      save directory.
+                    - The model is loaded by supplying a local directory as `pretrained_model_name_or_path` and a
+                      configuration JSON file named *config.json* is found in the directory.
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            from_pt (`bool`, *optional*, defaults to `False`):
+                Load the model weights from a PyTorch checkpoint save file (see docstring of
+                `pretrained_model_name_or_path` argument).
+            ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`):
+                Whether or not to raise an error if some of the weights from the checkpoint do not have the same size
+                as the weights of the model (if for instance, you are instantiating a model with 10 labels from a
+                checkpoint with 3 labels).
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only(`bool`, *optional*, defaults to `False`):
+                Whether or not to only look at local files (i.e., do not try to download the model).
+            use_auth_token (`str` or `bool`, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, or not specified, will use
+                the token generated when running `huggingface-cli login` (stored in `~/.huggingface`).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+
+
+                <Tip>
+
+                To test a pull request you made on the Hub, you can pass `revision="refs/pr/<pr_number>".
+
+                </Tip>
+
+            subfolder (`str`, *optional*, defaults to `""`):
+                In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can
+                specify the folder name here.
+            kwargs (remaining dictionary of keyword arguments, *optional*):
+                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
+                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
+                automatically loaded:
+
+                    - If a configuration is provided with `config`, `**kwargs` will be directly passed to the
+                      underlying model's `__init__` method (we assume all relevant updates to the configuration have
+                      already been done)
+                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
+                      initialization function ([`~PretrainedConfig.from_pretrained`]). Each key of `kwargs` that
+                      corresponds to a configuration attribute will be used to override said attribute with the
+                      supplied `kwargs` value. Remaining keys that do not correspond to any configuration attribute
+                      will be passed to the underlying model's `__init__` function.
+
+        Examples:
+
+        ```python
+        >>> from transformers import BertConfig, FlaxBertModel
+
+        >>> # Download model and configuration from huggingface.co and cache.
+        >>> model = FlaxBertModel.from_pretrained("bert-base-cased")
+        >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
+        >>> model = FlaxBertModel.from_pretrained("./test/saved_model/")
+        >>> # Loading from a PyTorch checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
+        >>> config = BertConfig.from_json_file("./pt_model/config.json")
+        >>> model = FlaxBertModel.from_pretrained("./pt_model/pytorch_model.bin", from_pt=True, config=config)
+        ```"""
+        config = kwargs.pop("config", None)
+        cache_dir = kwargs.pop("cache_dir", None)
+        from_pt = kwargs.pop("from_pt", False)
+        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", False)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", None)
+        from_pipeline = kwargs.pop("_from_pipeline", None)
+        from_auto_class = kwargs.pop("_from_auto", False)
+        _do_init = kwargs.pop("_do_init", True)
+        subfolder = kwargs.pop("subfolder", "")
+        commit_hash = kwargs.pop("_commit_hash", None)
+
+        if trust_remote_code is True:
+            logger.warning(
+                "The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is"
+                " ignored."
+            )
+
+        user_agent = {"file_type": "model", "framework": "flax", "from_auto_class": from_auto_class}
+        if from_pipeline is not None:
+            user_agent["using_pipeline"] = from_pipeline
+
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
+        # Load config if we don't provide a configuration
+        if not isinstance(config, PretrainedConfig):
+            config_path = config if config is not None else pretrained_model_name_or_path
+            config, model_kwargs = cls.config_class.from_pretrained(
+                config_path,
+                cache_dir=cache_dir,
+                return_unused_kwargs=True,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                subfolder=subfolder,
+                _from_auto=from_auto_class,
+                _from_pipeline=from_pipeline,
+                _commit_hash=commit_hash,
+                **kwargs,
+            )
+        else:
+            model_kwargs = kwargs.copy()
+
+        if commit_hash is None:
+            commit_hash = getattr(config, "_commit_hash", None)
+
+        # Add the dtype to model_kwargs
+        model_kwargs["dtype"] = dtype
+
+        # This variable will flag if we're loading a sharded checkpoint. In this case the archive file is just the
+        # index of the files.
+        is_sharded = False
+
+        # Load model
+        if pretrained_model_name_or_path is not None:
+            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+            is_local = os.path.isdir(pretrained_model_name_or_path)
+            if os.path.isdir(pretrained_model_name_or_path):
+                if from_pt and os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_NAME)):
+                    # Load from a PyTorch checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_NAME)
+                elif from_pt and os.path.isfile(
+                    os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_INDEX_NAME)
+                ):
+                    # Load from a sharded pytorch checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_INDEX_NAME)
+                    is_sharded = True
+                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)):
+                    # Load from a Flax checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
+                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_INDEX_NAME)):
+                    # Load from a sharded Flax checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_INDEX_NAME)
+                    is_sharded = True
+                # At this stage we don't have a weight file so we will raise an error.
+                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_NAME)):
+                    raise EnvironmentError(
+                        f"Error no file named {FLAX_WEIGHTS_NAME} found in directory {pretrained_model_name_or_path} "
+                        "but there is a file for PyTorch weights. Use `from_pt=True` to load this model from those "
+                        "weights."
+                    )
+                else:
+                    raise EnvironmentError(
+                        f"Error no file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME} found in directory "
+                        f"{pretrained_model_name_or_path}."
+                    )
+            elif os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):
+                archive_file = pretrained_model_name_or_path
+                is_local = True
+            elif is_remote_url(pretrained_model_name_or_path):
+                filename = pretrained_model_name_or_path
+                resolved_archive_file = download_url(pretrained_model_name_or_path)
+            else:
+                filename = WEIGHTS_NAME if from_pt else FLAX_WEIGHTS_NAME
+                try:
+                    # Load from URL or cache if already cached
+                    cached_file_kwargs = {
+                        "cache_dir": cache_dir,
+                        "force_download": force_download,
+                        "proxies": proxies,
+                        "resume_download": resume_download,
+                        "local_files_only": local_files_only,
+                        "use_auth_token": use_auth_token,
+                        "user_agent": user_agent,
+                        "revision": revision,
+                        "subfolder": subfolder,
+                        "_raise_exceptions_for_missing_entries": False,
+                        "_commit_hash": commit_hash,
+                    }
+                    resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
+
+                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an expection but a None
+                    # result when internet is up, the repo and revision exist, but the file does not.
+                    if resolved_archive_file is None and filename == FLAX_WEIGHTS_NAME:
+                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
+                        resolved_archive_file = cached_file(
+                            pretrained_model_name_or_path, FLAX_WEIGHTS_INDEX_NAME, **cached_file_kwargs
+                        )
+                        if resolved_archive_file is not None:
+                            is_sharded = True
+                    # Maybe the checkpoint is pytorch sharded, we try to grab the pytorch index name in this case.
+                    elif resolved_archive_file is None and from_pt:
+                        resolved_archive_file = cached_file(
+                            pretrained_model_name_or_path, WEIGHTS_INDEX_NAME, **cached_file_kwargs
+                        )
+                        if resolved_archive_file is not None:
+                            is_sharded = True
+                    if resolved_archive_file is None:
+                        # Otherwise, maybe there is a TF or Flax model file.  We try those to give a helpful error
+                        # message.
+                        has_file_kwargs = {
+                            "revision": revision,
+                            "proxies": proxies,
+                            "use_auth_token": use_auth_token,
+                        }
+                        if has_file(pretrained_model_name_or_path, WEIGHTS_NAME, **has_file_kwargs):
+                            raise EnvironmentError(
+                                f"{pretrained_model_name_or_path} does not appear to have a file named"
+                                f" {FLAX_WEIGHTS_NAME} but there is a file for PyTorch weights. Use `from_pt=True` to"
+                                " load this model from those weights."
+                            )
+                        elif has_file(pretrained_model_name_or_path, WEIGHTS_INDEX_NAME, **has_file_kwargs):
+                            raise EnvironmentError(
+                                f"{pretrained_model_name_or_path} does not appear to have a file named"
+                                f" {FLAX_WEIGHTS_INDEX_NAME} but there is a sharded file for PyTorch weights. Use"
+                                " `from_pt=True` to load this model from those weights."
+                            )
+                        else:
+                            raise EnvironmentError(
+                                f"{pretrained_model_name_or_path} does not appear to have a file named"
+                                f" {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}."
+                            )
+                except EnvironmentError:
+                    # Raise any environment error raise by `cached_file`. It will have a helpful error message adapted
+                    # to the original exception.
+                    raise
+                except Exception:
+                    # For any other exception, we throw a generic error.
+                    raise EnvironmentError(
+                        f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it"
+                        " from 'https://huggingface.co/models', make sure you don't have a local directory with the"
+                        f" same name. Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a"
+                        f" directory containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}."
+                    )
+
+            if is_local:
+                logger.info(f"loading weights file {archive_file}")
+                resolved_archive_file = archive_file
+            else:
+                logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
+        else:
+            resolved_archive_file = None
+
+        # We'll need to download and cache each checkpoint shard if the checkpoint is sharded.
+        if is_sharded:
+            # resolved_archive_file becomes a list of files that point to the different checkpoint shards in this case.
+            resolved_archive_file, _ = get_checkpoint_shard_files(
+                pretrained_model_name_or_path,
+                resolved_archive_file,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                user_agent=user_agent,
+                revision=revision,
+                subfolder=subfolder,
+                _commit_hash=commit_hash,
+            )
+
+        # init random models
+        model = cls(config, *model_args, _do_init=_do_init, **model_kwargs)
+
+        if from_pt:
+            state = load_pytorch_checkpoint_in_flax_state_dict(model, resolved_archive_file, is_sharded)
+        else:
+            if is_sharded:
+                state = cls.load_flax_sharded_weights(resolved_archive_file)
+            else:
+                try:
+                    with open(resolved_archive_file, "rb") as state_f:
+                        state = from_bytes(cls, state_f.read())
+                except (UnpicklingError, msgpack.exceptions.ExtraData) as e:
+                    try:
+                        with open(resolved_archive_file) as f:
+                            if f.read().startswith("version"):
+                                raise OSError(
+                                    "You seem to have cloned a repository without having git-lfs installed. Please"
+                                    " install git-lfs and run `git lfs install` followed by `git lfs pull` in the"
+                                    " folder you cloned."
+                                )
+                            else:
+                                raise ValueError from e
+                    except (UnicodeDecodeError, ValueError):
+                        raise EnvironmentError(f"Unable to convert {archive_file} to Flax deserializable object. ")
+            # make sure all arrays are stored as jnp.arrays
+            # NOTE: This is to prevent a bug this will be fixed in Flax >= v0.3.4:
+            # https://github.com/google/flax/issues/1261
+            if _do_init:
+                state = jax.tree_util.tree_map(jnp.array, state)
+            else:
+                # keep the params on CPU if we don't want to initialize
+                state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.devices("cpu")[0]), state)
+
+        if "batch_stats" in state:  # if flax model contains batch norm layers
+            # if model is base model only use model_prefix key
+            if (
+                cls.base_model_prefix not in dict(model.params_shape_tree["params"])
+                and cls.base_model_prefix in state["params"]
+            ):
+                state["params"] = state["params"][cls.base_model_prefix]
+                state["batch_stats"] = state["batch_stats"][cls.base_model_prefix]
+
+            # if model is head model and we are loading weights from base model
+            # we initialize new params dict with base_model_prefix
+            if (
+                cls.base_model_prefix in dict(model.params_shape_tree["params"])
+                and cls.base_model_prefix not in state["params"]
+            ):
+                state = {
+                    "params": {cls.base_model_prefix: state["params"]},
+                    "batch_stats": {cls.base_model_prefix: state["batch_stats"]},
+                }
+
+        else:
+            # if model is base model only use model_prefix key
+            if cls.base_model_prefix not in dict(model.params_shape_tree) and cls.base_model_prefix in state:
+                state = state[cls.base_model_prefix]
+
+            # if model is head model and we are loading weights from base model
+            # we initialize new params dict with base_model_prefix
+            if cls.base_model_prefix in dict(model.params_shape_tree) and cls.base_model_prefix not in state:
+                state = {cls.base_model_prefix: state}
+
+        # flatten dicts
+        state = flatten_dict(state)
+
+        random_state = flatten_dict(unfreeze(model.params if _do_init else model.params_shape_tree))
+
+        missing_keys = model.required_params - set(state.keys())
+        unexpected_keys = set(state.keys()) - model.required_params
+
+        # Disabling warning when porting pytorch weights to flax, flax does not uses num_batches_tracked
+        for unexpected_key in unexpected_keys.copy():
+            if "num_batches_tracked" in unexpected_key[-1]:
+                unexpected_keys.remove(unexpected_key)
+
+        if missing_keys and not _do_init:
+            logger.warning(
+                f"The checkpoint {pretrained_model_name_or_path} is missing required keys: {missing_keys}. "
+                "Make sure to call model.init_weights to initialize the missing weights."
+            )
+            cls._missing_keys = missing_keys
+
+        # Mistmatched keys contains tuples key/shape1/shape2 of weights in the checkpoint that have a shape not
+        # matching the weights in the model.
+        mismatched_keys = []
+        for key in state.keys():
+            if key in random_state and state[key].shape != random_state[key].shape:
+                if ignore_mismatched_sizes:
+                    mismatched_keys.append((key, state[key].shape, random_state[key].shape))
+                    state[key] = random_state[key]
+                else:
+                    raise ValueError(
+                        f"Trying to load the pretrained weight for {key} failed: checkpoint has shape "
+                        f"{state[key].shape} which is incompatible with the model shape {random_state[key].shape}. "
+                        "Using `ignore_mismatched_sizes=True` if you really want to load this checkpoint inside this "
+                        "model."
+                    )
+
+        # add missing keys as random parameters if we are initializing
+        if missing_keys and _do_init:
+            for missing_key in missing_keys:
+                state[missing_key] = random_state[missing_key]
+
+        # remove unexpected keys to not be saved again
+        for unexpected_key in unexpected_keys:
+            del state[unexpected_key]
+
+        if len(unexpected_keys) > 0:
+            logger.warning(
+                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
+                f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
+                f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task or"
+                " with another architecture (e.g. initializing a BertForSequenceClassification model from a"
+                " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
+                f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly identical"
+                " (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
+            )
+        else:
+            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
+
+        if len(missing_keys) > 0:
+            logger.warning(
+                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+                f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
+                " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
+            )
+        elif len(mismatched_keys) == 0:
+            logger.info(
+                f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
+                f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the checkpoint"
+                f" was trained on, you can already use {model.__class__.__name__} for predictions without further"
+                " training."
+            )
+        if len(mismatched_keys) > 0:
+            mismatched_warning = "\n".join(
+                [
+                    f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
+                    for key, shape1, shape2 in mismatched_keys
+                ]
+            )
+            logger.warning(
+                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+                f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
+                f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be able"
+                " to use it for predictions and inference."
+            )
+
+        # dictionary of key: dtypes for the model params
+        param_dtypes = jax.tree_util.tree_map(lambda x: x.dtype, state)
+        # extract keys of parameters not in jnp.float32
+        fp16_params = [k for k in param_dtypes if param_dtypes[k] == jnp.float16]
+        bf16_params = [k for k in param_dtypes if param_dtypes[k] == jnp.bfloat16]
+
+        # raise a warning if any of the parameters are not in jnp.float32
+        if len(fp16_params) > 0:
+            logger.warning(
+                f"Some of the weights of {model.__class__.__name__} were initialized in float16 precision from "
+                f"the model checkpoint at {pretrained_model_name_or_path}:\n{fp16_params}\n"
+                "You should probably UPCAST the model weights to float32 if this was not intended. "
+                "See [`~FlaxPreTrainedModel.to_fp32`] for further information on how to do this."
+            )
+
+        if len(bf16_params) > 0:
+            logger.warning(
+                f"Some of the weights of {model.__class__.__name__} were initialized in bfloat16 precision from "
+                f"the model checkpoint at {pretrained_model_name_or_path}:\n{bf16_params}\n"
+                "You should probably UPCAST the model weights to float32 if this was not intended. "
+                "See [`~FlaxPreTrainedModel.to_fp32`] for further information on how to do this."
+            )
+
+        # If it is a model with generation capabilities, attempt to load the generation config
+        if model.can_generate():
+            try:
+                model.generation_config = GenerationConfig.from_pretrained(
+                    pretrained_model_name_or_path,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    _from_auto=from_auto_class,
+                    _from_pipeline=from_pipeline,
+                    **kwargs,
+                )
+            except OSError:
+                logger.info(
+                    "Generation config file not found, using a generation config created from the model config."
+                )
+                pass
+
+        if _do_init:
+            # set correct parameters
+            model.params = unflatten_dict(state)
+            return model
+        else:
+            return model, unflatten_dict(state)
+
+    def save_pretrained(
+        self, save_directory: Union[str, os.PathLike], params=None, push_to_hub=False, max_shard_size="10GB", **kwargs
+    ):
+        """
+        Save a model and its configuration file to a directory, so that it can be re-loaded using the
+        `[`~FlaxPreTrainedModel.from_pretrained`]` class method
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to which to save. Will be created if it doesn't exist.
+            push_to_hub (`bool`, *optional*, defaults to `False`):
+                Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the
+                repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+                namespace).
+            max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
+                The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size
+                lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5MB"`).
+
+                <Tip warning={true}>
+
+                If a single weight of the model is bigger than `max_shard_size`, it will be in its own checkpoint shard
+                which will be bigger than `max_shard_size`.
+
+                </Tip>
+
+            kwargs:
+                Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+        """
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+
+        os.makedirs(save_directory, exist_ok=True)
+
+        if push_to_hub:
+            commit_message = kwargs.pop("commit_message", None)
+            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+            repo_id = self._create_repo(repo_id, **kwargs)
+            files_timestamps = self._get_files_timestamps(save_directory)
+
+        # get abs dir
+        save_directory = os.path.abspath(save_directory)
+        # save config as well
+        self.config.architectures = [self.__class__.__name__[4:]]
+
+        # If we have a custom model, we copy the file defining it in the folder and set the attributes so it can be
+        # loaded from the Hub.
+        if self._auto_class is not None:
+            custom_object_save(self, save_directory, config=self.config)
+
+        self.config.save_pretrained(save_directory)
+        if self.can_generate():
+            self.generation_config.save_pretrained(save_directory)
+
+        # save model
+        output_model_file = os.path.join(save_directory, FLAX_WEIGHTS_NAME)
+
+        shards, index = flax_shard_checkpoint(params if params is not None else self.params, max_shard_size)
+        # Clean the folder from a previous save
+        for filename in os.listdir(save_directory):
+            full_filename = os.path.join(save_directory, filename)
+            if (
+                filename.startswith(FLAX_WEIGHTS_NAME[:-4])
+                and os.path.isfile(full_filename)
+                and filename not in shards.keys()
+            ):
+                os.remove(full_filename)
+
+        if index is None:
+            with open(output_model_file, "wb") as f:
+                params = params if params is not None else self.params
+                model_bytes = to_bytes(params)
+                f.write(model_bytes)
+
+        else:
+            save_index_file = os.path.join(save_directory, FLAX_WEIGHTS_INDEX_NAME)
+            # Save the index as well
+            with open(save_index_file, "w", encoding="utf-8") as f:
+                content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+                f.write(content)
+            logger.info(
+                f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "
+                f"split in {len(shards)} checkpoint shards. You can find where each parameters has been saved in the "
+                f"index located at {save_index_file}."
+            )
+            for shard_file, shard in shards.items():
+                # the shard item are unflattened, to save them we need to flatten them again
+                with open(os.path.join(save_directory, shard_file), mode="wb") as f:
+                    params = unflatten_dict(shard, sep="/")
+                    shard_bytes = to_bytes(params)
+                    f.write(shard_bytes)
+
+        logger.info(f"Model weights saved in {output_model_file}")
+
+        if push_to_hub:
+            self._upload_modified_files(
+                save_directory,
+                repo_id,
+                files_timestamps,
+                commit_message=commit_message,
+                token=kwargs.get("use_auth_token"),
+            )
+
+    @classmethod
+    def register_for_auto_class(cls, auto_class="FlaxAutoModel"):
+        """
+        Register this class with a given auto class. This should only be used for custom models as the ones in the
+        library are already mapped with an auto class.
+
+        <Tip warning={true}>
+
+        This API is experimental and may have some slight breaking changes in the next releases.
+
+        </Tip>
+
+        Args:
+            auto_class (`str` or `type`, *optional*, defaults to `"FlaxAutoModel"`):
+                The auto class to register this new model with.
+        """
+        if not isinstance(auto_class, str):
+            auto_class = auto_class.__name__
+
+        import transformers.models.auto as auto_module
+
+        if not hasattr(auto_module, auto_class):
+            raise ValueError(f"{auto_class} is not a valid auto class.")
+
+        cls._auto_class = auto_class
+
+
+# To update the docstring, we need to copy the method, otherwise we change the original docstring.
+FlaxPreTrainedModel.push_to_hub = copy_func(FlaxPreTrainedModel.push_to_hub)
+if FlaxPreTrainedModel.push_to_hub.__doc__ is not None:
+    FlaxPreTrainedModel.push_to_hub.__doc__ = FlaxPreTrainedModel.push_to_hub.__doc__.format(
+        object="model", object_class="FlaxAutoModel", object_files="model checkpoint"
+    )
+
+
+def overwrite_call_docstring(model_class, docstring):
+    # copy __call__ function to be sure docstring is changed only for this function
+    model_class.__call__ = copy_func(model_class.__call__)
+    # delete existing docstring
+    model_class.__call__.__doc__ = None
+    # set correct docstring
+    model_class.__call__ = add_start_docstrings_to_model_forward(docstring)(model_class.__call__)
+
+
+def append_call_sample_docstring(model_class, checkpoint, output_type, config_class, mask=None):
+    model_class.__call__ = copy_func(model_class.__call__)
+    model_class.__call__ = add_code_sample_docstrings(
+        checkpoint=checkpoint,
+        output_type=output_type,
+        config_class=config_class,
+        model_cls=model_class.__name__,
+    )(model_class.__call__)
+
+
+def append_replace_return_docstrings(model_class, output_type, config_class):
+    model_class.__call__ = copy_func(model_class.__call__)
+    model_class.__call__ = replace_return_docstrings(
+        output_type=output_type,
+        config_class=config_class,
+    )(model_class.__call__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/modeling_tf_pytorch_utils.py

@@ -0,0 +1,578 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch - TF 2.0 general utilities."""
+
+
+import os
+import re
+
+import numpy
+
+from .utils import ExplicitEnum, expand_dims, is_numpy_array, is_torch_tensor, logging, reshape, squeeze, tensor_size
+from .utils import transpose as transpose_func
+
+
+logger = logging.get_logger(__name__)
+
+
+class TransposeType(ExplicitEnum):
+    """
+    Possible ...
+    """
+
+    NO = "no"
+    SIMPLE = "simple"
+    CONV1D = "conv1d"
+    CONV2D = "conv2d"
+
+
+def convert_tf_weight_name_to_pt_weight_name(
+    tf_name, start_prefix_to_remove="", tf_weight_shape=None, name_scope=None
+):
+    """
+    Convert a TF 2.0 model variable name in a pytorch model weight name.
+
+    Conventions for TF2.0 scopes -> PyTorch attribute names conversions:
+
+        - '$1___$2' is replaced by $2 (can be used to duplicate or remove layers in TF2.0 vs PyTorch)
+        - '_._' is replaced by a new level separation (can be used to convert TF2.0 lists in PyTorch nn.ModulesList)
+
+    return tuple with:
+
+        - pytorch model weight name
+        - transpose: `TransposeType` member indicating whether and how TF2.0 and PyTorch weights matrices should be
+          transposed with regards to each other
+    """
+    if name_scope is not None:
+        if not tf_name.startswith(name_scope):
+            raise ValueError(
+                f"Weight name {tf_name} does not start with name_scope {name_scope}. This is an internal error "
+                "in Transformers, so (unless you were doing something really evil) please open an issue to report it!"
+            )
+        tf_name = tf_name[len(name_scope) :]
+        tf_name = tf_name.lstrip("/")
+    tf_name = tf_name.replace(":0", "")  # device ids
+    tf_name = re.sub(
+        r"/[^/]*___([^/]*)/", r"/\1/", tf_name
+    )  # '$1___$2' is replaced by $2 (can be used to duplicate or remove layers in TF2.0 vs PyTorch)
+    tf_name = tf_name.replace(
+        "_._", "/"
+    )  # '_._' is replaced by a level separation (can be used to convert TF2.0 lists in PyTorch nn.ModulesList)
+    tf_name = re.sub(r"//+", "/", tf_name)  # Remove empty levels at the end
+    tf_name = tf_name.split("/")  # Convert from TF2.0 '/' separators to PyTorch '.' separators
+    # Some weights have a single name without "/" such as final_logits_bias in BART
+    if len(tf_name) > 1:
+        tf_name = tf_name[1:]  # Remove level zero
+
+    tf_weight_shape = list(tf_weight_shape)
+
+    # When should we transpose the weights
+    if tf_name[-1] == "kernel" and tf_weight_shape is not None and len(tf_weight_shape) == 4:
+        transpose = TransposeType.CONV2D
+    elif tf_name[-1] == "kernel" and tf_weight_shape is not None and len(tf_weight_shape) == 3:
+        transpose = TransposeType.CONV1D
+    elif bool(
+        tf_name[-1] in ["kernel", "pointwise_kernel", "depthwise_kernel"]
+        or "emb_projs" in tf_name
+        or "out_projs" in tf_name
+    ):
+        transpose = TransposeType.SIMPLE
+    else:
+        transpose = TransposeType.NO
+
+    # Convert standard TF2.0 names in PyTorch names
+    if tf_name[-1] == "kernel" or tf_name[-1] == "embeddings" or tf_name[-1] == "gamma":
+        tf_name[-1] = "weight"
+    if tf_name[-1] == "beta":
+        tf_name[-1] = "bias"
+
+    # The SeparableConv1D TF layer contains two weights that are translated to PyTorch Conv1D here
+    if tf_name[-1] == "pointwise_kernel" or tf_name[-1] == "depthwise_kernel":
+        tf_name[-1] = tf_name[-1].replace("_kernel", ".weight")
+
+    # Remove prefix if needed
+    tf_name = ".".join(tf_name)
+    if start_prefix_to_remove:
+        tf_name = tf_name.replace(start_prefix_to_remove, "", 1)
+
+    return tf_name, transpose
+
+
+def apply_transpose(transpose: TransposeType, weight, match_shape=None, pt_to_tf=True):
+    """
+    Apply a transpose to some weight then tries to reshape the weight to the same shape as a given shape, all in a
+    framework agnostic way.
+    """
+    if transpose is TransposeType.CONV2D:
+        # Conv2D weight:
+        #    PT: (num_out_channel, num_in_channel, kernel[0], kernel[1])
+        # -> TF: (kernel[0], kernel[1], num_in_channel, num_out_channel)
+        axes = (2, 3, 1, 0) if pt_to_tf else (3, 2, 0, 1)
+        weight = transpose_func(weight, axes=axes)
+    elif transpose is TransposeType.CONV1D:
+        # Conv1D weight:
+        #    PT: (num_out_channel, num_in_channel, kernel)
+        # -> TF: (kernel, num_in_channel, num_out_channel)
+        weight = transpose_func(weight, axes=(2, 1, 0))
+    elif transpose is TransposeType.SIMPLE:
+        weight = transpose_func(weight)
+
+    if match_shape is None:
+        return weight
+
+    if len(match_shape) < len(weight.shape):
+        weight = squeeze(weight)
+    elif len(match_shape) > len(weight.shape):
+        weight = expand_dims(weight, axis=0)
+
+    if list(match_shape) != list(weight.shape):
+        try:
+            weight = reshape(weight, match_shape)
+        except AssertionError as e:
+            e.args += (match_shape, match_shape)
+            raise e
+
+    return weight
+
+
+#####################
+# PyTorch => TF 2.0 #
+#####################
+
+
+def load_pytorch_checkpoint_in_tf2_model(
+    tf_model,
+    pytorch_checkpoint_path,
+    tf_inputs=None,
+    allow_missing_keys=False,
+    output_loading_info=False,
+    _prefix=None,
+    tf_to_pt_weight_rename=None,
+):
+    """Load pytorch checkpoints in a TF 2.0 model"""
+    try:
+        import tensorflow as tf  # noqa: F401
+        import torch  # noqa: F401
+    except ImportError:
+        logger.error(
+            "Loading a PyTorch model in TensorFlow, requires both PyTorch and TensorFlow to be installed. Please see "
+            "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+
+    # Treats a single file as a collection of shards with 1 shard.
+    if isinstance(pytorch_checkpoint_path, str):
+        pytorch_checkpoint_path = [pytorch_checkpoint_path]
+
+    # Loads all shards into a single state dictionary
+    pt_state_dict = {}
+    for path in pytorch_checkpoint_path:
+        pt_path = os.path.abspath(path)
+        logger.info(f"Loading PyTorch weights from {pt_path}")
+        pt_state_dict.update(torch.load(pt_path, map_location="cpu"))
+
+    logger.info(f"PyTorch checkpoint contains {sum(t.numel() for t in pt_state_dict.values()):,} parameters")
+
+    return load_pytorch_weights_in_tf2_model(
+        tf_model,
+        pt_state_dict,
+        tf_inputs=tf_inputs,
+        allow_missing_keys=allow_missing_keys,
+        output_loading_info=output_loading_info,
+        _prefix=_prefix,
+        tf_to_pt_weight_rename=tf_to_pt_weight_rename,
+    )
+
+
+def load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=None, allow_missing_keys=False):
+    """Load pytorch checkpoints in a TF 2.0 model"""
+    pt_state_dict = pt_model.state_dict()
+
+    return load_pytorch_weights_in_tf2_model(
+        tf_model, pt_state_dict, tf_inputs=tf_inputs, allow_missing_keys=allow_missing_keys
+    )
+
+
+def load_pytorch_weights_in_tf2_model(
+    tf_model,
+    pt_state_dict,
+    tf_inputs=None,
+    allow_missing_keys=False,
+    output_loading_info=False,
+    _prefix=None,
+    tf_to_pt_weight_rename=None,
+):
+    """Load pytorch state_dict in a TF 2.0 model."""
+    try:
+        import tensorflow as tf  # noqa: F401
+        import torch  # noqa: F401
+    except ImportError:
+        logger.error(
+            "Loading a PyTorch model in TensorFlow, requires both PyTorch and TensorFlow to be installed. Please see "
+            "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+
+    pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
+    return load_pytorch_state_dict_in_tf2_model(
+        tf_model,
+        pt_state_dict,
+        tf_inputs=tf_inputs,
+        allow_missing_keys=allow_missing_keys,
+        output_loading_info=output_loading_info,
+        _prefix=_prefix,
+        tf_to_pt_weight_rename=tf_to_pt_weight_rename,
+    )
+
+
+def load_pytorch_state_dict_in_tf2_model(
+    tf_model,
+    pt_state_dict,
+    tf_inputs=None,
+    allow_missing_keys=False,
+    output_loading_info=False,
+    _prefix=None,
+    tf_to_pt_weight_rename=None,
+    ignore_mismatched_sizes=False,
+):
+    """Load a pytorch state_dict in a TF 2.0 model."""
+    import tensorflow as tf
+    from packaging.version import parse
+
+    if parse(tf.__version__) >= parse("2.11.0"):
+        from keras import backend as K
+    else:
+        from tensorflow.python.keras import backend as K
+
+    if tf_inputs is None:
+        tf_inputs = tf_model.dummy_inputs
+
+    if _prefix is None:
+        _prefix = ""
+    if tf_inputs is not None:
+        with tf.name_scope(_prefix):
+            tf_model(tf_inputs, training=False)  # Make sure model is built
+    # Adapt state dict - TODO remove this and update the AWS weights files instead
+    # Convert old format to new format if needed from a PyTorch state_dict
+    old_keys = []
+    new_keys = []
+    for key in pt_state_dict.keys():
+        new_key = None
+        if "gamma" in key:
+            new_key = key.replace("gamma", "weight")
+        if "beta" in key:
+            new_key = key.replace("beta", "bias")
+        if "running_var" in key:
+            new_key = key.replace("running_var", "moving_variance")
+        if "running_mean" in key:
+            new_key = key.replace("running_mean", "moving_mean")
+        if new_key:
+            old_keys.append(key)
+            new_keys.append(new_key)
+    for old_key, new_key in zip(old_keys, new_keys):
+        pt_state_dict[new_key] = pt_state_dict.pop(old_key)
+
+    # Matt: All TF models store the actual model stem in a MainLayer class, including the base model.
+    # In PT, the derived models (with heads) use the base model class as the stem instead, and the base model
+    # just contains the stem itself, and there is no MainLayer class. This means that TF base classes have one
+    # extra layer in their weight names, corresponding to the MainLayer class. This code block compensates for that.
+    start_prefix_to_remove = ""
+    if not any(s.startswith(tf_model.base_model_prefix) for s in pt_state_dict.keys()):
+        start_prefix_to_remove = tf_model.base_model_prefix + "."
+
+    symbolic_weights = tf_model.trainable_weights + tf_model.non_trainable_weights
+    tf_loaded_numel = 0
+    weight_value_tuples = []
+    all_pytorch_weights = set(pt_state_dict.keys())
+    missing_keys = []
+    mismatched_keys = []
+    for symbolic_weight in symbolic_weights:
+        sw_name = symbolic_weight.name
+        name, transpose = convert_tf_weight_name_to_pt_weight_name(
+            sw_name,
+            start_prefix_to_remove=start_prefix_to_remove,
+            tf_weight_shape=symbolic_weight.shape,
+            name_scope=_prefix,
+        )
+        if tf_to_pt_weight_rename is not None:
+            name = tf_to_pt_weight_rename(name)
+
+        # Find associated numpy array in pytorch model state dict
+        if name not in pt_state_dict:
+            if allow_missing_keys:
+                missing_keys.append(name)
+                continue
+            elif tf_model._keys_to_ignore_on_load_missing is not None:
+                # authorized missing keys don't have to be loaded
+                if any(re.search(pat, name) is not None for pat in tf_model._keys_to_ignore_on_load_missing):
+                    continue
+            raise AttributeError(f"{name} not found in PyTorch model")
+
+        try:
+            array = apply_transpose(transpose, pt_state_dict[name], symbolic_weight.shape)
+        except tf.errors.InvalidArgumentError as e:
+            if not ignore_mismatched_sizes:
+                error_msg = str(e)
+                error_msg += (
+                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
+                )
+                raise tf.errors.InvalidArgumentError(error_msg)
+            else:
+                mismatched_keys.append((name, pt_state_dict[name].shape, symbolic_weight.shape))
+                continue
+
+        tf_loaded_numel += tensor_size(array)
+
+        weight_value_tuples.append((symbolic_weight, array))
+        all_pytorch_weights.discard(name)
+
+    K.batch_set_value(weight_value_tuples)
+
+    if tf_inputs is not None:
+        tf_model(tf_inputs, training=False)  # Make sure restore ops are run
+
+    logger.info(f"Loaded {tf_loaded_numel:,} parameters in the TF 2.0 model.")
+
+    unexpected_keys = list(all_pytorch_weights)
+
+    if tf_model._keys_to_ignore_on_load_missing is not None:
+        for pat in tf_model._keys_to_ignore_on_load_missing:
+            missing_keys = [k for k in missing_keys if re.search(pat, k) is None]
+    if tf_model._keys_to_ignore_on_load_unexpected is not None:
+        for pat in tf_model._keys_to_ignore_on_load_unexpected:
+            unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+    if len(unexpected_keys) > 0:
+        logger.warning(
+            "Some weights of the PyTorch model were not used when initializing the TF 2.0 model"
+            f" {tf_model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are initializing"
+            f" {tf_model.__class__.__name__} from a PyTorch model trained on another task or with another architecture"
+            " (e.g. initializing a TFBertForSequenceClassification model from a BertForPreTraining model).\n- This IS"
+            f" NOT expected if you are initializing {tf_model.__class__.__name__} from a PyTorch model that you expect"
+            " to be exactly identical (e.g. initializing a TFBertForSequenceClassification model from a"
+            " BertForSequenceClassification model)."
+        )
+    else:
+        logger.warning(f"All PyTorch model weights were used when initializing {tf_model.__class__.__name__}.\n")
+    if len(missing_keys) > 0:
+        logger.warning(
+            f"Some weights or buffers of the TF 2.0 model {tf_model.__class__.__name__} were not initialized from the"
+            f" PyTorch model and are newly initialized: {missing_keys}\nYou should probably TRAIN this model on a"
+            " down-stream task to be able to use it for predictions and inference."
+        )
+    else:
+        logger.warning(
+            f"All the weights of {tf_model.__class__.__name__} were initialized from the PyTorch model.\n"
+            "If your task is similar to the task the model of the checkpoint was trained on, "
+            f"you can already use {tf_model.__class__.__name__} for predictions without further training."
+        )
+
+    if len(mismatched_keys) > 0:
+        mismatched_warning = "\n".join(
+            [
+                f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
+                for key, shape1, shape2 in mismatched_keys
+            ]
+        )
+        logger.warning(
+            f"Some weights of {tf_model.__class__.__name__} were not initialized from the model checkpoint"
+            f" are newly initialized because the shapes did not"
+            f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be able"
+            " to use it for predictions and inference."
+        )
+
+    if output_loading_info:
+        loading_info = {
+            "missing_keys": missing_keys,
+            "unexpected_keys": unexpected_keys,
+            "mismatched_keys": mismatched_keys,
+        }
+        return tf_model, loading_info
+
+    return tf_model
+
+
+#####################
+# TF 2.0 => PyTorch #
+#####################
+
+
+def load_tf2_checkpoint_in_pytorch_model(
+    pt_model, tf_checkpoint_path, tf_inputs=None, allow_missing_keys=False, output_loading_info=False
+):
+    """
+    Load TF 2.0 HDF5 checkpoint in a PyTorch model We use HDF5 to easily do transfer learning (see
+    https://github.com/tensorflow/tensorflow/blob/ee16fcac960ae660e0e4496658a366e2f745e1f0/tensorflow/python/keras/engine/network.py#L1352-L1357).
+    """
+    try:
+        import tensorflow as tf  # noqa: F401
+        import torch  # noqa: F401
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires both PyTorch and TensorFlow to be installed. Please see "
+            "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+
+    import transformers
+
+    from .modeling_tf_utils import load_tf_weights
+
+    logger.info(f"Loading TensorFlow weights from {tf_checkpoint_path}")
+
+    # Instantiate and load the associated TF 2.0 model
+    tf_model_class_name = "TF" + pt_model.__class__.__name__  # Add "TF" at the beginning
+    tf_model_class = getattr(transformers, tf_model_class_name)
+    tf_model = tf_model_class(pt_model.config)
+
+    if tf_inputs is None:
+        tf_inputs = tf_model.dummy_inputs
+
+    if tf_inputs is not None:
+        tf_model(tf_inputs, training=False)  # Make sure model is built
+
+    load_tf_weights(tf_model, tf_checkpoint_path)
+
+    return load_tf2_model_in_pytorch_model(
+        pt_model, tf_model, allow_missing_keys=allow_missing_keys, output_loading_info=output_loading_info
+    )
+
+
+def load_tf2_model_in_pytorch_model(pt_model, tf_model, allow_missing_keys=False, output_loading_info=False):
+    """Load TF 2.0 model in a pytorch model"""
+    weights = tf_model.weights
+
+    return load_tf2_weights_in_pytorch_model(
+        pt_model, weights, allow_missing_keys=allow_missing_keys, output_loading_info=output_loading_info
+    )
+
+
+def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=False, output_loading_info=False):
+    """Load TF2.0 symbolic weights in a PyTorch model"""
+    try:
+        import tensorflow as tf  # noqa: F401
+        import torch  # noqa: F401
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires both PyTorch and TensorFlow to be installed. Please see "
+            "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+
+    tf_state_dict = {tf_weight.name: tf_weight.numpy() for tf_weight in tf_weights}
+    return load_tf2_state_dict_in_pytorch_model(
+        pt_model, tf_state_dict, allow_missing_keys=allow_missing_keys, output_loading_info=output_loading_info
+    )
+
+
+def load_tf2_state_dict_in_pytorch_model(pt_model, tf_state_dict, allow_missing_keys=False, output_loading_info=False):
+    import torch
+
+    new_pt_params_dict = {}
+    current_pt_params_dict = dict(pt_model.named_parameters())
+
+    # Make sure we are able to load PyTorch base models as well as derived models (with heads)
+    # TF models always have a prefix, some of PyTorch models (base ones) don't
+    start_prefix_to_remove = ""
+    if not any(s.startswith(pt_model.base_model_prefix) for s in current_pt_params_dict.keys()):
+        start_prefix_to_remove = pt_model.base_model_prefix + "."
+
+    # Build a map from potential PyTorch weight names to TF 2.0 Variables
+    tf_weights_map = {}
+    for name, tf_weight in tf_state_dict.items():
+        pt_name, transpose = convert_tf_weight_name_to_pt_weight_name(
+            name, start_prefix_to_remove=start_prefix_to_remove, tf_weight_shape=tf_weight.shape
+        )
+        tf_weights_map[pt_name] = (tf_weight, transpose)
+
+    all_tf_weights = set(tf_weights_map.keys())
+    loaded_pt_weights_data_ptr = {}
+    missing_keys_pt = []
+    for pt_weight_name, pt_weight in current_pt_params_dict.items():
+        # Handle PyTorch shared weight ()not duplicated in TF 2.0
+        if pt_weight.data_ptr() in loaded_pt_weights_data_ptr:
+            new_pt_params_dict[pt_weight_name] = loaded_pt_weights_data_ptr[pt_weight.data_ptr()]
+            continue
+
+        # Find associated numpy array in pytorch model state dict
+        if pt_weight_name not in tf_weights_map:
+            if allow_missing_keys:
+                missing_keys_pt.append(pt_weight_name)
+                continue
+
+            raise AttributeError(f"{pt_weight_name} not found in TF 2.0 model")
+
+        array, transpose = tf_weights_map[pt_weight_name]
+
+        array = apply_transpose(transpose, array, pt_weight.shape, pt_to_tf=False)
+
+        if numpy.isscalar(array):
+            array = numpy.array(array)
+        if not is_torch_tensor(array) and not is_numpy_array(array):
+            array = array.numpy()
+        if is_numpy_array(array):
+            # Convert to torch tensor
+            array = torch.from_numpy(array)
+
+        new_pt_params_dict[pt_weight_name] = array
+        loaded_pt_weights_data_ptr[pt_weight.data_ptr()] = array
+        all_tf_weights.discard(pt_weight_name)
+
+    missing_keys, unexpected_keys = pt_model.load_state_dict(new_pt_params_dict, strict=False)
+    missing_keys += missing_keys_pt
+
+    # Some models may have keys that are not in the state by design, removing them before needlessly warning
+    # the user.
+    if pt_model._keys_to_ignore_on_load_missing is not None:
+        for pat in pt_model._keys_to_ignore_on_load_missing:
+            missing_keys = [k for k in missing_keys if re.search(pat, k) is None]
+
+    if pt_model._keys_to_ignore_on_load_unexpected is not None:
+        for pat in pt_model._keys_to_ignore_on_load_unexpected:
+            unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+    if len(unexpected_keys) > 0:
+        logger.warning(
+            "Some weights of the TF 2.0 model were not used when initializing the PyTorch model"
+            f" {pt_model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are initializing"
+            f" {pt_model.__class__.__name__} from a TF 2.0 model trained on another task or with another architecture"
+            " (e.g. initializing a BertForSequenceClassification model from a TFBertForPreTraining model).\n- This IS"
+            f" NOT expected if you are initializing {pt_model.__class__.__name__} from a TF 2.0 model that you expect"
+            " to be exactly identical (e.g. initializing a BertForSequenceClassification model from a"
+            " TFBertForSequenceClassification model)."
+        )
+    else:
+        logger.warning(f"All TF 2.0 model weights were used when initializing {pt_model.__class__.__name__}.\n")
+    if len(missing_keys) > 0:
+        logger.warning(
+            f"Some weights of {pt_model.__class__.__name__} were not initialized from the TF 2.0 model and are newly"
+            f" initialized: {missing_keys}\nYou should probably TRAIN this model on a down-stream task to be able to"
+            " use it for predictions and inference."
+        )
+    else:
+        logger.warning(
+            f"All the weights of {pt_model.__class__.__name__} were initialized from the TF 2.0 model.\n"
+            "If your task is similar to the task the model of the checkpoint was trained on, "
+            f"you can already use {pt_model.__class__.__name__} for predictions without further training."
+        )
+
+    logger.info(f"Weights or buffers not loaded from TF 2.0 model: {all_tf_weights}")
+
+    if output_loading_info:
+        loading_info = {"missing_keys": missing_keys, "unexpected_keys": unexpected_keys}
+        return pt_model, loading_info
+
+    return pt_model

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/__init__.py

@@ -0,0 +1,179 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_sentencepiece_available,
+    is_tf_available,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_albert": ["ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "AlbertConfig", "AlbertOnnxConfig"],
+}
+
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_albert"] = ["AlbertTokenizer"]
+
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_albert_fast"] = ["AlbertTokenizerFast"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_albert"] = [
+        "ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AlbertForMaskedLM",
+        "AlbertForMultipleChoice",
+        "AlbertForPreTraining",
+        "AlbertForQuestionAnswering",
+        "AlbertForSequenceClassification",
+        "AlbertForTokenClassification",
+        "AlbertModel",
+        "AlbertPreTrainedModel",
+        "load_tf_weights_in_albert",
+    ]
+
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_tf_albert"] = [
+        "TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFAlbertForMaskedLM",
+        "TFAlbertForMultipleChoice",
+        "TFAlbertForPreTraining",
+        "TFAlbertForQuestionAnswering",
+        "TFAlbertForSequenceClassification",
+        "TFAlbertForTokenClassification",
+        "TFAlbertMainLayer",
+        "TFAlbertModel",
+        "TFAlbertPreTrainedModel",
+    ]
+
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_flax_albert"] = [
+        "FlaxAlbertForMaskedLM",
+        "FlaxAlbertForMultipleChoice",
+        "FlaxAlbertForPreTraining",
+        "FlaxAlbertForQuestionAnswering",
+        "FlaxAlbertForSequenceClassification",
+        "FlaxAlbertForTokenClassification",
+        "FlaxAlbertModel",
+        "FlaxAlbertPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig, AlbertOnnxConfig
+
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_albert import AlbertTokenizer
+
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_albert_fast import AlbertTokenizerFast
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_albert import (
+            ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AlbertForMaskedLM,
+            AlbertForMultipleChoice,
+            AlbertForPreTraining,
+            AlbertForQuestionAnswering,
+            AlbertForSequenceClassification,
+            AlbertForTokenClassification,
+            AlbertModel,
+            AlbertPreTrainedModel,
+            load_tf_weights_in_albert,
+        )
+
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_tf_albert import (
+            TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TFAlbertForMaskedLM,
+            TFAlbertForMultipleChoice,
+            TFAlbertForPreTraining,
+            TFAlbertForQuestionAnswering,
+            TFAlbertForSequenceClassification,
+            TFAlbertForTokenClassification,
+            TFAlbertMainLayer,
+            TFAlbertModel,
+            TFAlbertPreTrainedModel,
+        )
+
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_flax_albert import (
+            FlaxAlbertForMaskedLM,
+            FlaxAlbertForMultipleChoice,
+            FlaxAlbertForPreTraining,
+            FlaxAlbertForQuestionAnswering,
+            FlaxAlbertForSequenceClassification,
+            FlaxAlbertForTokenClassification,
+            FlaxAlbertModel,
+            FlaxAlbertPreTrainedModel,
+        )
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/configuration_albert.py

@@ -0,0 +1,172 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" ALBERT model configuration"""
+from collections import OrderedDict
+from typing import Mapping
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+
+
+ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "albert-base-v1": "https://huggingface.co/albert-base-v1/resolve/main/config.json",
+    "albert-large-v1": "https://huggingface.co/albert-large-v1/resolve/main/config.json",
+    "albert-xlarge-v1": "https://huggingface.co/albert-xlarge-v1/resolve/main/config.json",
+    "albert-xxlarge-v1": "https://huggingface.co/albert-xxlarge-v1/resolve/main/config.json",
+    "albert-base-v2": "https://huggingface.co/albert-base-v2/resolve/main/config.json",
+    "albert-large-v2": "https://huggingface.co/albert-large-v2/resolve/main/config.json",
+    "albert-xlarge-v2": "https://huggingface.co/albert-xlarge-v2/resolve/main/config.json",
+    "albert-xxlarge-v2": "https://huggingface.co/albert-xxlarge-v2/resolve/main/config.json",
+}
+
+
+class AlbertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlbertModel`] or a [`TFAlbertModel`]. It is used
+    to instantiate an ALBERT model according to the specified arguments, defining the model architecture. Instantiating
+    a configuration with the defaults will yield a similar configuration to that of the ALBERT
+    [albert-xxlarge-v2](https://huggingface.co/albert-xxlarge-v2) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30000):
+            Vocabulary size of the ALBERT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`AlbertModel`] or [`TFAlbertModel`].
+        embedding_size (`int`, *optional*, defaults to 128):
+            Dimensionality of vocabulary embeddings.
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_hidden_groups (`int`, *optional*, defaults to 1):
+            Number of groups for the hidden layers, parameters in the same group are shared.
+        num_attention_heads (`int`, *optional*, defaults to 64):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 16384):
+            The dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        inner_group_num (`int`, *optional*, defaults to 1):
+            The number of inner repetition of attention and ffn.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu_new"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`AlbertModel`] or [`TFAlbertModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        classifier_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for attached classifiers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+
+    Examples:
+
+    ```python
+    >>> from transformers import AlbertConfig, AlbertModel
+
+    >>> # Initializing an ALBERT-xxlarge style configuration
+    >>> albert_xxlarge_configuration = AlbertConfig()
+
+    >>> # Initializing an ALBERT-base style configuration
+    >>> albert_base_configuration = AlbertConfig(
+    ...     hidden_size=768,
+    ...     num_attention_heads=12,
+    ...     intermediate_size=3072,
+    ... )
+
+    >>> # Initializing a model (with random weights) from the ALBERT-base style configuration
+    >>> model = AlbertModel(albert_xxlarge_configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "albert"
+
+    def __init__(
+        self,
+        vocab_size=30000,
+        embedding_size=128,
+        hidden_size=4096,
+        num_hidden_layers=12,
+        num_hidden_groups=1,
+        num_attention_heads=64,
+        intermediate_size=16384,
+        inner_group_num=1,
+        hidden_act="gelu_new",
+        hidden_dropout_prob=0,
+        attention_probs_dropout_prob=0,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        classifier_dropout_prob=0.1,
+        position_embedding_type="absolute",
+        pad_token_id=0,
+        bos_token_id=2,
+        eos_token_id=3,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.embedding_size = embedding_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_hidden_groups = num_hidden_groups
+        self.num_attention_heads = num_attention_heads
+        self.inner_group_num = inner_group_num
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.classifier_dropout_prob = classifier_dropout_prob
+        self.position_embedding_type = position_embedding_type
+
+
+# Copied from transformers.models.bert.configuration_bert.BertOnnxConfig with Roberta->Albert
+class AlbertOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task == "multiple-choice":
+            dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
+        else:
+            dynamic_axis = {0: "batch", 1: "sequence"}
+        return OrderedDict(
+            [
+                ("input_ids", dynamic_axis),
+                ("attention_mask", dynamic_axis),
+                ("token_type_ids", dynamic_axis),
+            ]
+        )

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/convert_albert_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,63 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ALBERT checkpoint."""
+
+
+import argparse
+
+import torch
+
+from ...utils import logging
+from . import AlbertConfig, AlbertForPreTraining, load_tf_weights_in_albert
+
+
+logging.set_verbosity_info()
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, albert_config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = AlbertConfig.from_json_file(albert_config_file)
+    print(f"Building PyTorch model from configuration: {config}")
+    model = AlbertForPreTraining(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_albert(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print(f"Save PyTorch model to {pytorch_dump_path}")
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
+    )
+    parser.add_argument(
+        "--albert_config_file",
+        default=None,
+        type=str,
+        required=True,
+        help=(
+            "The config json file corresponding to the pre-trained ALBERT model. \n"
+            "This specifies the model architecture."
+        ),
+    )
+    parser.add_argument(
+        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.albert_config_file, args.pytorch_dump_path)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/modeling_albert.py

@@ -0,0 +1,1403 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch ALBERT model."""
+
+import math
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPooling,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_albert import AlbertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+
+
+ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "albert-base-v1",
+    "albert-large-v1",
+    "albert-xlarge-v1",
+    "albert-xxlarge-v1",
+    "albert-base-v2",
+    "albert-large-v2",
+    "albert-xlarge-v2",
+    "albert-xxlarge-v2",
+    # See all ALBERT models at https://huggingface.co/models?filter=albert
+]
+
+
+def load_tf_weights_in_albert(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        print(name)
+
+    for name, array in zip(names, arrays):
+        original_name = name
+
+        # If saved from the TF HUB module
+        name = name.replace("module/", "")
+
+        # Renaming and simplifying
+        name = name.replace("ffn_1", "ffn")
+        name = name.replace("bert/", "albert/")
+        name = name.replace("attention_1", "attention")
+        name = name.replace("transform/", "")
+        name = name.replace("LayerNorm_1", "full_layer_layer_norm")
+        name = name.replace("LayerNorm", "attention/LayerNorm")
+        name = name.replace("transformer/", "")
+
+        # The feed forward layer had an 'intermediate' step which has been abstracted away
+        name = name.replace("intermediate/dense/", "")
+        name = name.replace("ffn/intermediate/output/dense/", "ffn_output/")
+
+        # ALBERT attention was split between self and output which have been abstracted away
+        name = name.replace("/output/", "/")
+        name = name.replace("/self/", "/")
+
+        # The pooler is a linear layer
+        name = name.replace("pooler/dense", "pooler")
+
+        # The classifier was simplified to predictions from cls/predictions
+        name = name.replace("cls/predictions", "predictions")
+        name = name.replace("predictions/attention", "predictions")
+
+        # Naming was changed to be more explicit
+        name = name.replace("embeddings/attention", "embeddings")
+        name = name.replace("inner_group_", "albert_layers/")
+        name = name.replace("group_", "albert_layer_groups/")
+
+        # Classifier
+        if len(name.split("/")) == 1 and ("output_bias" in name or "output_weights" in name):
+            name = "classifier/" + name
+
+        # No ALBERT model currently handles the next sentence prediction task
+        if "seq_relationship" in name:
+            name = name.replace("seq_relationship/output_", "sop_classifier/classifier/")
+            name = name.replace("weights", "weight")
+
+        name = name.split("/")
+
+        # Ignore the gradients applied by the LAMB/ADAM optimizers.
+        if (
+            "adam_m" in name
+            or "adam_v" in name
+            or "AdamWeightDecayOptimizer" in name
+            or "AdamWeightDecayOptimizer_1" in name
+            or "global_step" in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            if pointer.shape != array.shape:
+                raise ValueError(f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched")
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        print(f"Initialize PyTorch weight {name} from {original_name}")
+        pointer.data = torch.from_numpy(array)
+
+    return model
+
+
+class AlbertEmbeddings(nn.Module):
+    """
+    Construct the embeddings from word, position and token_type embeddings.
+    """
+
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
+        )
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values_length: int = 0,
+    ) -> torch.Tensor:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class AlbertAttention(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads}"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.attention_head_size = config.hidden_size // config.num_attention_heads
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.attention_dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.output_dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.pruned_heads = set()
+
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+    # Copied from transformers.models.bert.modeling_bert.BertSelfAttention.transpose_for_scores
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def prune_heads(self, heads: List[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.num_attention_heads, self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.query = prune_linear_layer(self.query, index)
+        self.key = prune_linear_layer(self.key, index)
+        self.value = prune_linear_layer(self.value, index)
+        self.dense = prune_linear_layer(self.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.num_attention_heads = self.num_attention_heads - len(heads)
+        self.all_head_size = self.attention_head_size * self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.attention_dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = context_layer.transpose(2, 1).flatten(2)
+
+        projected_context_layer = self.dense(context_layer)
+        projected_context_layer_dropout = self.output_dropout(projected_context_layer)
+        layernormed_context_layer = self.LayerNorm(hidden_states + projected_context_layer_dropout)
+        return (layernormed_context_layer, attention_probs) if output_attentions else (layernormed_context_layer,)
+
+
+class AlbertLayer(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.full_layer_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = AlbertAttention(config)
+        self.ffn = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.ffn_output = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.activation = ACT2FN[config.hidden_act]
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        attention_output = self.attention(hidden_states, attention_mask, head_mask, output_attentions)
+
+        ffn_output = apply_chunking_to_forward(
+            self.ff_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output[0],
+        )
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output[0])
+
+        return (hidden_states,) + attention_output[1:]  # add attentions if we output them
+
+    def ff_chunk(self, attention_output: torch.Tensor) -> torch.Tensor:
+        ffn_output = self.ffn(attention_output)
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+        return ffn_output
+
+
+class AlbertLayerGroup(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            layer_output = albert_layer(hidden_states, attention_mask, head_mask[layer_index], output_attentions)
+            hidden_states = layer_output[0]
+
+            if output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
+
+
+class AlbertTransformer(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.config = config
+        self.embedding_hidden_mapping_in = nn.Linear(config.embedding_size, config.hidden_size)
+        self.albert_layer_groups = nn.ModuleList([AlbertLayerGroup(config) for _ in range(config.num_hidden_groups)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[BaseModelOutput, Tuple]:
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        head_mask = [None] * self.config.num_hidden_layers if head_mask is None else head_mask
+
+        for i in range(self.config.num_hidden_layers):
+            # Number of layers in a hidden group
+            layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+            # Index of the hidden group
+            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+            layer_group_output = self.albert_layer_groups[group_idx](
+                hidden_states,
+                attention_mask,
+                head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
+                output_attentions,
+                output_hidden_states,
+            )
+            hidden_states = layer_group_output[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+class AlbertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlbertConfig
+    load_tf_weights = load_tf_weights_in_albert
+    base_model_prefix = "albert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+@dataclass
+class AlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of [`AlbertForPreTraining`].
+
+    Args:
+        loss (*optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction
+            (classification) loss.
+        prediction_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (`torch.FloatTensor` of shape `(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    prediction_logits: torch.FloatTensor = None
+    sop_logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+ALBERT_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Args:
+        config ([`AlbertConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
+            [`PreTrainedTokenizer.encode`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ALBERT Model transformer outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class AlbertModel(AlbertPreTrainedModel):
+    config_class = AlbertConfig
+    base_model_prefix = "albert"
+
+    def __init__(self, config: AlbertConfig, add_pooling_layer: bool = True):
+        super().__init__(config)
+
+        self.config = config
+        self.embeddings = AlbertEmbeddings(config)
+        self.encoder = AlbertTransformer(config)
+        if add_pooling_layer:
+            self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
+            self.pooler_activation = nn.Tanh()
+        else:
+            self.pooler = None
+            self.pooler_activation = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value: nn.Embedding) -> None:
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} ALBERT has
+        a different architecture in that its layers are shared across groups, which then has inner groups. If an ALBERT
+        model has 12 hidden layers and 2 hidden groups, with two inner groups, there is a total of 4 different layers.
+
+        These layers are flattened: the indices [0,1] correspond to the two inner groups of the first hidden layer,
+        while [2,3] correspond to the two inner groups of the second hidden layer.
+
+        Any layer with in index other than [0,1,2,3] will result in an error. See base class PreTrainedModel for more
+        information about head pruning
+        """
+        for layer, heads in heads_to_prune.items():
+            group_idx = int(layer / self.config.inner_group_num)
+            inner_group_idx = int(layer - group_idx * self.config.inner_group_num)
+            self.encoder.albert_layer_groups[group_idx].albert_layers[inner_group_idx].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[None] = None,
+        output_hidden_states: Optional[None] = None,
+        return_dict: Optional[None] = None,
+    ) -> Union[BaseModelOutputWithPooling, Tuple]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(self.dtype).min
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = encoder_outputs[0]
+
+        pooled_output = self.pooler_activation(self.pooler(sequence_output[:, 0])) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a
+    `sentence order prediction (classification)` head.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForPreTraining(AlbertPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [
+        "predictions.decoder.weight",
+        "predictions.decoder.bias",
+        "embeddings.position_ids",
+    ]
+
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.predictions = AlbertMLMHead(config)
+        self.sop_classifier = AlbertSOPHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.predictions.decoder = new_embeddings
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.albert.embeddings.word_embeddings
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=AlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        sentence_order_label: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[AlbertForPreTrainingOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        sentence_order_label (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
+            (see `input_ids` docstring) Indices should be in `[0, 1]`. `0` indicates original order (sequence A, then
+            sequence B), `1` indicates switched order (sequence B, then sequence A).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, AlbertForPreTraining
+        >>> import torch
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+        >>> model = AlbertForPreTraining.from_pretrained("albert-base-v2")
+
+        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)
+        >>> # Batch size 1
+        >>> outputs = model(input_ids)
+
+        >>> prediction_logits = outputs.prediction_logits
+        >>> sop_logits = outputs.sop_logits
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output, pooled_output = outputs[:2]
+
+        prediction_scores = self.predictions(sequence_output)
+        sop_scores = self.sop_classifier(pooled_output)
+
+        total_loss = None
+        if labels is not None and sentence_order_label is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            sentence_order_loss = loss_fct(sop_scores.view(-1, 2), sentence_order_label.view(-1))
+            total_loss = masked_lm_loss + sentence_order_loss
+
+        if not return_dict:
+            output = (prediction_scores, sop_scores) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return AlbertForPreTrainingOutput(
+            loss=total_loss,
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class AlbertMLMHead(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.dense = nn.Linear(config.hidden_size, config.embedding_size)
+        self.decoder = nn.Linear(config.embedding_size, config.vocab_size)
+        self.activation = ACT2FN[config.hidden_act]
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+
+        prediction_scores = hidden_states
+
+        return prediction_scores
+
+    def _tie_weights(self) -> None:
+        # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+        self.bias = self.decoder.bias
+
+
+class AlbertSOPHead(nn.Module):
+    def __init__(self, config: AlbertConfig):
+        super().__init__()
+
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, pooled_output: torch.Tensor) -> torch.Tensor:
+        dropout_pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(dropout_pooled_output)
+        return logits
+
+
+@add_start_docstrings(
+    "Albert Model with a `language modeling` head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForMaskedLM(AlbertPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [
+        "predictions.decoder.weight",
+        "predictions.decoder.bias",
+        "embeddings.position_ids",
+    ]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        self.predictions = AlbertMLMHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.predictions.decoder = new_embeddings
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.albert.embeddings.word_embeddings
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[MaskedLMOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoTokenizer, AlbertForMaskedLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+        >>> model = AlbertForMaskedLM.from_pretrained("albert-base-v2")
+
+        >>> # add mask_token
+        >>> inputs = tokenizer("The capital of [MASK] is Paris.", return_tensors="pt")
+        >>> with torch.no_grad():
+        ...     logits = model(**inputs).logits
+
+        >>> # retrieve index of [MASK]
+        >>> mask_token_index = (inputs.input_ids == tokenizer.mask_token_id)[0].nonzero(as_tuple=True)[0]
+        >>> predicted_token_id = logits[0, mask_token_index].argmax(axis=-1)
+        >>> tokenizer.decode(predicted_token_id)
+        'france'
+        ```
+
+        ```python
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="pt")["input_ids"]
+        >>> labels = torch.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100)
+        >>> outputs = model(**inputs, labels=labels)
+        >>> round(outputs.loss.item(), 2)
+        0.81
+        ```
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_outputs = outputs[0]
+
+        prediction_scores = self.predictions(sequence_outputs)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForSequenceClassification(AlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="textattack/albert-base-v2-imdb",
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="'LABEL_1'",
+        expected_loss=0.12,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[SequenceClassifierOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForTokenClassification(AlbertPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        classifier_dropout_prob = (
+            config.classifier_dropout_prob
+            if config.classifier_dropout_prob is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[TokenClassifierOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForQuestionAnswering(AlbertPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config, add_pooling_layer=False)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="twmkn9/albert-base-v2-squad2",
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        qa_target_start_index=12,
+        qa_target_end_index=13,
+        expected_output="'a nice puppet'",
+        expected_loss=7.36,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[AlbertForPreTrainingOutput, Tuple]:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits: torch.Tensor = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForMultipleChoice(AlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig):
+        super().__init__(config)
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[AlbertForPreTrainingOutput, Tuple]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where *num_choices* is the size of the second dimension of the input tensors. (see
+            *input_ids* above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits: torch.Tensor = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/modeling_flax_albert.py

@@ -0,0 +1,1119 @@
+# coding=utf-8
+# Copyright 2021 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Tuple
+
+import flax
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen.attention import dot_product_attention_weights
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax import lax
+
+from ...modeling_flax_outputs import (
+    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPooling,
+    FlaxMaskedLMOutput,
+    FlaxMultipleChoiceModelOutput,
+    FlaxQuestionAnsweringModelOutput,
+    FlaxSequenceClassifierOutput,
+    FlaxTokenClassifierOutput,
+)
+from ...modeling_flax_utils import (
+    ACT2FN,
+    FlaxPreTrainedModel,
+    append_call_sample_docstring,
+    append_replace_return_docstrings,
+    overwrite_call_docstring,
+)
+from ...utils import ModelOutput, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_albert import AlbertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+
+
+@flax.struct.dataclass
+class FlaxAlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of [`FlaxAlbertForPreTraining`].
+
+    Args:
+        prediction_logits (`jnp.ndarray` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (`jnp.ndarray` of shape `(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings + one for the output of each layer) of shape
+            `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(jnp.ndarray)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `jnp.ndarray` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    prediction_logits: jnp.ndarray = None
+    sop_logits: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray]] = None
+    attentions: Optional[Tuple[jnp.ndarray]] = None
+
+
+ALBERT_START_DOCSTRING = r"""
+
+    This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading, saving and converting weights from PyTorch models)
+
+    This model is also a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
+    subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
+    general usage and behavior.
+
+    Finally, this model supports inherent JAX features such as:
+
+    - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+    - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+    - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+    - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+    Parameters:
+        config ([`AlbertConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~FlaxPreTrainedModel.from_pretrained`] method to load the model weights.
+        dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+            The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
+            `jax.numpy.bfloat16` (on TPUs).
+
+            This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
+            specified all the computation will be performed with the given `dtype`.
+
+            **Note that this only specifies the dtype of the computation and does not influence the dtype of model
+            parameters.**
+
+            If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`] and
+            [`~FlaxPreTrainedModel.to_bf16`].
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`numpy.ndarray` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`numpy.ndarray` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`numpy.ndarray` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`numpy.ndarray` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+
+"""
+
+
+class FlaxAlbertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.word_embeddings = nn.Embed(
+            self.config.vocab_size,
+            self.config.embedding_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+        self.position_embeddings = nn.Embed(
+            self.config.max_position_embeddings,
+            self.config.embedding_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+        self.token_type_embeddings = nn.Embed(
+            self.config.type_vocab_size,
+            self.config.embedding_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+        self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertEmbeddings.__call__
+    def __call__(self, input_ids, token_type_ids, position_ids, deterministic: bool = True):
+        # Embed
+        inputs_embeds = self.word_embeddings(input_ids.astype("i4"))
+        position_embeds = self.position_embeddings(position_ids.astype("i4"))
+        token_type_embeddings = self.token_type_embeddings(token_type_ids.astype("i4"))
+
+        # Sum all embeddings
+        hidden_states = inputs_embeds + token_type_embeddings + position_embeds
+
+        # Layer Norm
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+        return hidden_states
+
+
+class FlaxAlbertSelfAttention(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        if self.config.hidden_size % self.config.num_attention_heads != 0:
+            raise ValueError(
+                "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads` "
+                "                   : {self.config.num_attention_heads}"
+            )
+
+        self.query = nn.Dense(
+            self.config.hidden_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+        self.key = nn.Dense(
+            self.config.hidden_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+        self.value = nn.Dense(
+            self.config.hidden_size,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+        self.dense = nn.Dense(
+            self.config.hidden_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+
+    def __call__(self, hidden_states, attention_mask, deterministic=True, output_attentions: bool = False):
+        head_dim = self.config.hidden_size // self.config.num_attention_heads
+
+        query_states = self.query(hidden_states).reshape(
+            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        )
+        value_states = self.value(hidden_states).reshape(
+            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        )
+        key_states = self.key(hidden_states).reshape(
+            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        )
+
+        # Convert the boolean attention mask to an attention bias.
+        if attention_mask is not None:
+            # attention mask in the form of attention bias
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+            attention_bias = lax.select(
+                attention_mask > 0,
+                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+                jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
+            )
+        else:
+            attention_bias = None
+
+        dropout_rng = None
+        if not deterministic and self.config.attention_probs_dropout_prob > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        attn_weights = dot_product_attention_weights(
+            query_states,
+            key_states,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.config.attention_probs_dropout_prob,
+            broadcast_dropout=True,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
+        attn_output = attn_output.reshape(attn_output.shape[:2] + (-1,))
+
+        projected_attn_output = self.dense(attn_output)
+        projected_attn_output = self.dropout(projected_attn_output, deterministic=deterministic)
+        layernormed_attn_output = self.LayerNorm(projected_attn_output + hidden_states)
+        outputs = (layernormed_attn_output, attn_weights) if output_attentions else (layernormed_attn_output,)
+        return outputs
+
+
+class FlaxAlbertLayer(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.attention = FlaxAlbertSelfAttention(self.config, dtype=self.dtype)
+        self.ffn = nn.Dense(
+            self.config.intermediate_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.activation = ACT2FN[self.config.hidden_act]
+        self.ffn_output = nn.Dense(
+            self.config.hidden_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.full_layer_layer_norm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+    ):
+        attention_outputs = self.attention(
+            hidden_states, attention_mask, deterministic=deterministic, output_attentions=output_attentions
+        )
+        attention_output = attention_outputs[0]
+        ffn_output = self.ffn(attention_output)
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+        ffn_output = self.dropout(ffn_output, deterministic=deterministic)
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attention_outputs[1],)
+        return outputs
+
+
+class FlaxAlbertLayerCollection(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxAlbertLayer(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.inner_group_num)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ):
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.layers):
+            layer_output = albert_layer(
+                hidden_states,
+                attention_mask,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            hidden_states = layer_output[0]
+
+            if output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
+
+
+class FlaxAlbertLayerCollections(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    layer_index: Optional[str] = None
+
+    def setup(self):
+        self.albert_layers = FlaxAlbertLayerCollection(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+    ):
+        outputs = self.albert_layers(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        return outputs
+
+
+class FlaxAlbertLayerGroups(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.layers = [
+            FlaxAlbertLayerCollections(self.config, name=str(i), layer_index=str(i), dtype=self.dtype)
+            for i in range(self.config.num_hidden_groups)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        all_attentions = () if output_attentions else None
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for i in range(self.config.num_hidden_layers):
+            # Index of the hidden group
+            group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+            layer_group_output = self.layers[group_idx](
+                hidden_states,
+                attention_mask,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+            )
+            hidden_states = layer_group_output[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+class FlaxAlbertEncoder(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+        self.embedding_hidden_mapping_in = nn.Dense(
+            self.config.hidden_size,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.albert_layer_groups = FlaxAlbertLayerGroups(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+        return self.albert_layer_groups(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+
+class FlaxAlbertOnlyMLMHead(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+    bias_init: Callable[..., np.ndarray] = jax.nn.initializers.zeros
+
+    def setup(self):
+        self.dense = nn.Dense(self.config.embedding_size, dtype=self.dtype)
+        self.activation = ACT2FN[self.config.hidden_act]
+        self.LayerNorm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
+        self.decoder = nn.Dense(self.config.vocab_size, dtype=self.dtype, use_bias=False)
+        self.bias = self.param("bias", self.bias_init, (self.config.vocab_size,))
+
+    def __call__(self, hidden_states, shared_embedding=None):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+
+        if shared_embedding is not None:
+            hidden_states = self.decoder.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+        else:
+            hidden_states = self.decoder(hidden_states)
+
+        hidden_states += self.bias
+        return hidden_states
+
+
+class FlaxAlbertSOPHead(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.dropout = nn.Dropout(self.config.classifier_dropout_prob)
+        self.classifier = nn.Dense(2, dtype=self.dtype)
+
+    def __call__(self, pooled_output, deterministic=True):
+        pooled_output = self.dropout(pooled_output, deterministic=deterministic)
+        logits = self.classifier(pooled_output)
+        return logits
+
+
+class FlaxAlbertPreTrainedModel(FlaxPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlbertConfig
+    base_model_prefix = "albert"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: AlbertConfig,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        token_type_ids = jnp.zeros_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        attention_mask = jnp.ones_like(input_ids)
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(
+            rngs, input_ids, attention_mask, token_type_ids, position_ids, return_dict=False
+        )["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        # init input tensors if not passed
+        if token_type_ids is None:
+            token_type_ids = jnp.zeros_like(input_ids)
+
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(token_type_ids, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+        )
+
+
+class FlaxAlbertModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    add_pooling_layer: bool = True
+
+    def setup(self):
+        self.embeddings = FlaxAlbertEmbeddings(self.config, dtype=self.dtype)
+        self.encoder = FlaxAlbertEncoder(self.config, dtype=self.dtype)
+        if self.add_pooling_layer:
+            self.pooler = nn.Dense(
+                self.config.hidden_size,
+                kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+                dtype=self.dtype,
+                name="pooler",
+            )
+            self.pooler_activation = nn.tanh
+        else:
+            self.pooler = None
+            self.pooler_activation = None
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids: Optional[np.ndarray] = None,
+        position_ids: Optional[np.ndarray] = None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # make sure `token_type_ids` is correctly initialized when not passed
+        if token_type_ids is None:
+            token_type_ids = jnp.zeros_like(input_ids)
+
+        # make sure `position_ids` is correctly initialized when not passed
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        hidden_states = self.embeddings(input_ids, token_type_ids, position_ids, deterministic=deterministic)
+
+        outputs = self.encoder(
+            hidden_states,
+            attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = outputs[0]
+        if self.add_pooling_layer:
+            pooled = self.pooler(hidden_states[:, 0])
+            pooled = self.pooler_activation(pooled)
+        else:
+            pooled = None
+
+        if not return_dict:
+            # if pooled is None, don't return it
+            if pooled is None:
+                return (hidden_states,) + outputs[1:]
+            return (hidden_states, pooled) + outputs[1:]
+
+        return FlaxBaseModelOutputWithPooling(
+            last_hidden_state=hidden_states,
+            pooler_output=pooled,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Albert Model transformer outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertModel(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertModule
+
+
+append_call_sample_docstring(FlaxAlbertModel, _CHECKPOINT_FOR_DOC, FlaxBaseModelOutputWithPooling, _CONFIG_FOR_DOC)
+
+
+class FlaxAlbertForPreTrainingModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype)
+        self.predictions = FlaxAlbertOnlyMLMHead(config=self.config, dtype=self.dtype)
+        self.sop_classifier = FlaxAlbertSOPHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.albert.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        hidden_states = outputs[0]
+        pooled_output = outputs[1]
+
+        prediction_scores = self.predictions(hidden_states, shared_embedding=shared_embedding)
+        sop_scores = self.sop_classifier(pooled_output, deterministic=deterministic)
+
+        if not return_dict:
+            return (prediction_scores, sop_scores) + outputs[2:]
+
+        return FlaxAlbertForPreTrainingOutput(
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a
+    `sentence order prediction (classification)` head.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForPreTraining(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForPreTrainingModule
+
+
+FLAX_ALBERT_FOR_PRETRAINING_DOCSTRING = """
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, FlaxAlbertForPreTraining
+
+    >>> tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+    >>> model = FlaxAlbertForPreTraining.from_pretrained("albert-base-v2")
+
+    >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="np")
+    >>> outputs = model(**inputs)
+
+    >>> prediction_logits = outputs.prediction_logits
+    >>> seq_relationship_logits = outputs.sop_logits
+    ```
+"""
+
+overwrite_call_docstring(
+    FlaxAlbertForPreTraining,
+    ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length") + FLAX_ALBERT_FOR_PRETRAINING_DOCSTRING,
+)
+append_replace_return_docstrings(
+    FlaxAlbertForPreTraining, output_type=FlaxAlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC
+)
+
+
+class FlaxAlbertForMaskedLMModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, add_pooling_layer=False, dtype=self.dtype)
+        self.predictions = FlaxAlbertOnlyMLMHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.albert.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        # Compute the prediction scores
+        logits = self.predictions(hidden_states, shared_embedding=shared_embedding)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxMaskedLMOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top.""", ALBERT_START_DOCSTRING)
+class FlaxAlbertForMaskedLM(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForMaskedLMModule
+
+
+append_call_sample_docstring(FlaxAlbertForMaskedLM, _CHECKPOINT_FOR_DOC, FlaxMaskedLMOutput, _CONFIG_FOR_DOC)
+
+
+class FlaxAlbertForSequenceClassificationModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype)
+        classifier_dropout = (
+            self.config.classifier_dropout_prob
+            if self.config.classifier_dropout_prob is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
+        self.classifier = nn.Dense(
+            self.config.num_labels,
+            dtype=self.dtype,
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(pooled_output, deterministic=deterministic)
+        logits = self.classifier(pooled_output)
+
+        if not return_dict:
+            return (logits,) + outputs[2:]
+
+        return FlaxSequenceClassifierOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForSequenceClassification(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForSequenceClassificationModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForSequenceClassification,
+    _CHECKPOINT_FOR_DOC,
+    FlaxSequenceClassifierOutput,
+    _CONFIG_FOR_DOC,
+)
+
+
+class FlaxAlbertForMultipleChoiceModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype)
+        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+        self.classifier = nn.Dense(1, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        num_choices = input_ids.shape[1]
+        input_ids = input_ids.reshape(-1, input_ids.shape[-1]) if input_ids is not None else None
+        attention_mask = attention_mask.reshape(-1, attention_mask.shape[-1]) if attention_mask is not None else None
+        token_type_ids = token_type_ids.reshape(-1, token_type_ids.shape[-1]) if token_type_ids is not None else None
+        position_ids = position_ids.reshape(-1, position_ids.shape[-1]) if position_ids is not None else None
+
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(pooled_output, deterministic=deterministic)
+        logits = self.classifier(pooled_output)
+
+        reshaped_logits = logits.reshape(-1, num_choices)
+
+        if not return_dict:
+            return (reshaped_logits,) + outputs[2:]
+
+        return FlaxMultipleChoiceModelOutput(
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForMultipleChoice(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForMultipleChoiceModule
+
+
+overwrite_call_docstring(
+    FlaxAlbertForMultipleChoice, ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
+)
+append_call_sample_docstring(
+    FlaxAlbertForMultipleChoice,
+    _CHECKPOINT_FOR_DOC,
+    FlaxMultipleChoiceModelOutput,
+    _CONFIG_FOR_DOC,
+)
+
+
+class FlaxAlbertForTokenClassificationModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
+        classifier_dropout = (
+            self.config.classifier_dropout_prob
+            if self.config.classifier_dropout_prob is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
+        self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+        logits = self.classifier(hidden_states)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxTokenClassifierOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForTokenClassification(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForTokenClassificationModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForTokenClassification,
+    _CHECKPOINT_FOR_DOC,
+    FlaxTokenClassifierOutput,
+    _CONFIG_FOR_DOC,
+)
+
+
+class FlaxAlbertForQuestionAnsweringModule(nn.Module):
+    config: AlbertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.albert = FlaxAlbertModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
+        self.qa_outputs = nn.Dense(self.config.num_labels, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        token_type_ids,
+        position_ids,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.albert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+
+        logits = self.qa_outputs(hidden_states)
+        start_logits, end_logits = logits.split(self.config.num_labels, axis=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        if not return_dict:
+            return (start_logits, end_logits) + outputs[1:]
+
+        return FlaxQuestionAnsweringModelOutput(
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class FlaxAlbertForQuestionAnswering(FlaxAlbertPreTrainedModel):
+    module_class = FlaxAlbertForQuestionAnsweringModule
+
+
+append_call_sample_docstring(
+    FlaxAlbertForQuestionAnswering,
+    _CHECKPOINT_FOR_DOC,
+    FlaxQuestionAnsweringModelOutput,
+    _CONFIG_FOR_DOC,
+)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/modeling_tf_albert.py

@@ -0,0 +1,1487 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TF 2.0 ALBERT model."""
+
+import math
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import numpy as np
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPooling,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
+from ...modeling_tf_utils import (
+    TFMaskedLanguageModelingLoss,
+    TFModelInputType,
+    TFMultipleChoiceLoss,
+    TFPreTrainedModel,
+    TFQuestionAnsweringLoss,
+    TFSequenceClassificationLoss,
+    TFTokenClassificationLoss,
+    get_initializer,
+    keras_serializable,
+    unpack_inputs,
+)
+from ...tf_utils import shape_list, stable_softmax
+from ...utils import (
+    MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_albert import AlbertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "albert-base-v2"
+_CONFIG_FOR_DOC = "AlbertConfig"
+
+TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "albert-base-v1",
+    "albert-large-v1",
+    "albert-xlarge-v1",
+    "albert-xxlarge-v1",
+    "albert-base-v2",
+    "albert-large-v2",
+    "albert-xlarge-v2",
+    "albert-xxlarge-v2",
+    # See all ALBERT models at https://huggingface.co/models?filter=albert
+]
+
+
+class TFAlbertPreTrainingLoss:
+    """
+    Loss function suitable for ALBERT pretraining, that is, the task of pretraining a language model by combining SOP +
+    MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation.
+    """
+
+    def hf_compute_loss(self, labels: tf.Tensor, logits: tf.Tensor) -> tf.Tensor:
+        loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
+            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
+        )
+        if self.config.tf_legacy_loss:
+            # make sure only labels that are not equal to -100
+            # are taken into account as loss
+            masked_lm_active_loss = tf.not_equal(tf.reshape(tensor=labels["labels"], shape=(-1,)), -100)
+            masked_lm_reduced_logits = tf.boolean_mask(
+                tensor=tf.reshape(tensor=logits[0], shape=(-1, shape_list(logits[0])[2])),
+                mask=masked_lm_active_loss,
+            )
+            masked_lm_labels = tf.boolean_mask(
+                tensor=tf.reshape(tensor=labels["labels"], shape=(-1,)), mask=masked_lm_active_loss
+            )
+            sentence_order_active_loss = tf.not_equal(
+                tf.reshape(tensor=labels["sentence_order_label"], shape=(-1,)), -100
+            )
+            sentence_order_reduced_logits = tf.boolean_mask(
+                tensor=tf.reshape(tensor=logits[1], shape=(-1, 2)), mask=sentence_order_active_loss
+            )
+            sentence_order_label = tf.boolean_mask(
+                tensor=tf.reshape(tensor=labels["sentence_order_label"], shape=(-1,)), mask=sentence_order_active_loss
+            )
+            masked_lm_loss = loss_fn(y_true=masked_lm_labels, y_pred=masked_lm_reduced_logits)
+            sentence_order_loss = loss_fn(y_true=sentence_order_label, y_pred=sentence_order_reduced_logits)
+            masked_lm_loss = tf.reshape(tensor=masked_lm_loss, shape=(-1, shape_list(sentence_order_loss)[0]))
+            masked_lm_loss = tf.reduce_mean(input_tensor=masked_lm_loss, axis=0)
+
+            return masked_lm_loss + sentence_order_loss
+
+        # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway
+        unmasked_lm_losses = loss_fn(y_true=tf.nn.relu(labels["labels"]), y_pred=logits[0])
+        # make sure only labels that are not equal to -100
+        # are taken into account for the loss computation
+        lm_loss_mask = tf.cast(labels["labels"] != -100, dtype=unmasked_lm_losses.dtype)
+        masked_lm_losses = unmasked_lm_losses * lm_loss_mask
+        reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses) / tf.reduce_sum(lm_loss_mask)
+
+        sop_logits = tf.reshape(logits[1], (-1, 2))
+        # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway
+        unmasked_sop_loss = loss_fn(y_true=tf.nn.relu(labels["sentence_order_label"]), y_pred=sop_logits)
+        sop_loss_mask = tf.cast(labels["sentence_order_label"] != -100, dtype=unmasked_sop_loss.dtype)
+
+        masked_sop_loss = unmasked_sop_loss * sop_loss_mask
+        reduced_masked_sop_loss = tf.reduce_sum(masked_sop_loss) / tf.reduce_sum(sop_loss_mask)
+
+        return tf.reshape(reduced_masked_lm_loss + reduced_masked_sop_loss, (1,))
+
+
+class TFAlbertEmbeddings(tf.keras.layers.Layer):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+        self.embedding_size = config.embedding_size
+        self.max_position_embeddings = config.max_position_embeddings
+        self.initializer_range = config.initializer_range
+        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
+        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+
+    def build(self, input_shape: tf.TensorShape):
+        with tf.name_scope("word_embeddings"):
+            self.weight = self.add_weight(
+                name="weight",
+                shape=[self.config.vocab_size, self.embedding_size],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        with tf.name_scope("token_type_embeddings"):
+            self.token_type_embeddings = self.add_weight(
+                name="embeddings",
+                shape=[self.config.type_vocab_size, self.embedding_size],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        with tf.name_scope("position_embeddings"):
+            self.position_embeddings = self.add_weight(
+                name="embeddings",
+                shape=[self.max_position_embeddings, self.embedding_size],
+                initializer=get_initializer(self.initializer_range),
+            )
+
+        super().build(input_shape)
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertEmbeddings.call
+    def call(
+        self,
+        input_ids: tf.Tensor = None,
+        position_ids: tf.Tensor = None,
+        token_type_ids: tf.Tensor = None,
+        inputs_embeds: tf.Tensor = None,
+        past_key_values_length=0,
+        training: bool = False,
+    ) -> tf.Tensor:
+        """
+        Applies embedding based on inputs tensor.
+
+        Returns:
+            final_embeddings (`tf.Tensor`): output embedding tensor.
+        """
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("Need to provide either `input_ids` or `input_embeds`.")
+
+        if input_ids is not None:
+            # Note: tf.gather, on which the embedding layer is based, won't check positive out of bound
+            # indices on GPU, returning zeros instead. This is a dangerous silent behavior.
+            tf.debugging.assert_less(
+                input_ids,
+                tf.cast(self.config.vocab_size, dtype=input_ids.dtype),
+                message=(
+                    "input_ids must be smaller than the embedding layer's input dimension (got"
+                    f" {tf.math.reduce_max(input_ids)} >= {self.config.vocab_size})"
+                ),
+            )
+            inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
+
+        input_shape = shape_list(inputs_embeds)[:-1]
+
+        if token_type_ids is None:
+            token_type_ids = tf.fill(dims=input_shape, value=0)
+
+        if position_ids is None:
+            position_ids = tf.expand_dims(
+                tf.range(start=past_key_values_length, limit=input_shape[1] + past_key_values_length), axis=0
+            )
+
+        position_embeds = tf.gather(params=self.position_embeddings, indices=position_ids)
+        token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids)
+        final_embeddings = inputs_embeds + position_embeds + token_type_embeds
+        final_embeddings = self.LayerNorm(inputs=final_embeddings)
+        final_embeddings = self.dropout(inputs=final_embeddings, training=training)
+
+        return final_embeddings
+
+
+class TFAlbertAttention(tf.keras.layers.Layer):
+    """Contains the complete attention sublayer, including both dropouts and layer norm."""
+
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number "
+                f"of attention heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.sqrt_att_head_size = math.sqrt(self.attention_head_size)
+        self.output_attentions = config.output_attentions
+
+        self.query = tf.keras.layers.Dense(
+            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
+        )
+        self.key = tf.keras.layers.Dense(
+            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
+        )
+        self.value = tf.keras.layers.Dense(
+            units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
+        )
+        self.dense = tf.keras.layers.Dense(
+            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
+        )
+        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
+        # Two different dropout probabilities; see https://github.com/google-research/albert/blob/master/modeling.py#L971-L993
+        self.attention_dropout = tf.keras.layers.Dropout(rate=config.attention_probs_dropout_prob)
+        self.output_dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+
+    def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
+        # Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
+        tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))
+
+        # Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
+        return tf.transpose(tensor, perm=[0, 2, 1, 3])
+
+    def call(
+        self,
+        input_tensor: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        training: bool = False,
+    ) -> Tuple[tf.Tensor]:
+        batch_size = shape_list(input_tensor)[0]
+        mixed_query_layer = self.query(inputs=input_tensor)
+        mixed_key_layer = self.key(inputs=input_tensor)
+        mixed_value_layer = self.value(inputs=input_tensor)
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
+        attention_scores = tf.divide(attention_scores, dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFAlbertModel call() function)
+            attention_scores = tf.add(attention_scores, attention_mask)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = stable_softmax(logits=attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.attention_dropout(inputs=attention_probs, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = tf.multiply(attention_probs, head_mask)
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+
+        # (batch_size, seq_len_q, all_head_size)
+        context_layer = tf.reshape(tensor=context_layer, shape=(batch_size, -1, self.all_head_size))
+        self_outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+        hidden_states = self_outputs[0]
+        hidden_states = self.dense(inputs=hidden_states)
+        hidden_states = self.output_dropout(inputs=hidden_states, training=training)
+        attention_output = self.LayerNorm(inputs=hidden_states + input_tensor)
+
+        # add attentions if we output them
+        outputs = (attention_output,) + self_outputs[1:]
+
+        return outputs
+
+
+class TFAlbertLayer(tf.keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.attention = TFAlbertAttention(config, name="attention")
+        self.ffn = tf.keras.layers.Dense(
+            units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="ffn"
+        )
+
+        if isinstance(config.hidden_act, str):
+            self.activation = get_tf_activation(config.hidden_act)
+        else:
+            self.activation = config.hidden_act
+
+        self.ffn_output = tf.keras.layers.Dense(
+            units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="ffn_output"
+        )
+        self.full_layer_layer_norm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name="full_layer_layer_norm"
+        )
+        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        training: bool = False,
+    ) -> Tuple[tf.Tensor]:
+        attention_outputs = self.attention(
+            input_tensor=hidden_states,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            training=training,
+        )
+        ffn_output = self.ffn(inputs=attention_outputs[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(inputs=ffn_output)
+        ffn_output = self.dropout(inputs=ffn_output, training=training)
+        hidden_states = self.full_layer_layer_norm(inputs=ffn_output + attention_outputs[0])
+
+        # add attentions if we output them
+        outputs = (hidden_states,) + attention_outputs[1:]
+
+        return outputs
+
+
+class TFAlbertLayerGroup(tf.keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.albert_layers = [
+            TFAlbertLayer(config, name=f"albert_layers_._{i}") for i in range(config.inner_group_num)
+        ]
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
+        layer_hidden_states = () if output_hidden_states else None
+        layer_attentions = () if output_attentions else None
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            if output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+            layer_output = albert_layer(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask[layer_index],
+                output_attentions=output_attentions,
+                training=training,
+            )
+            hidden_states = layer_output[0]
+
+            if output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+        # Add last layer
+        if output_hidden_states:
+            layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        return tuple(v for v in [hidden_states, layer_hidden_states, layer_attentions] if v is not None)
+
+
+class TFAlbertTransformer(tf.keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.num_hidden_layers = config.num_hidden_layers
+        self.num_hidden_groups = config.num_hidden_groups
+        # Number of layers in a hidden group
+        self.layers_per_group = int(config.num_hidden_layers / config.num_hidden_groups)
+        self.embedding_hidden_mapping_in = tf.keras.layers.Dense(
+            units=config.hidden_size,
+            kernel_initializer=get_initializer(config.initializer_range),
+            name="embedding_hidden_mapping_in",
+        )
+        self.albert_layer_groups = [
+            TFAlbertLayerGroup(config, name=f"albert_layer_groups_._{i}") for i in range(config.num_hidden_groups)
+        ]
+
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        attention_mask: tf.Tensor,
+        head_mask: tf.Tensor,
+        output_attentions: bool,
+        output_hidden_states: bool,
+        return_dict: bool,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
+        hidden_states = self.embedding_hidden_mapping_in(inputs=hidden_states)
+        all_attentions = () if output_attentions else None
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for i in range(self.num_hidden_layers):
+            # Index of the hidden group
+            group_idx = int(i / (self.num_hidden_layers / self.num_hidden_groups))
+            layer_group_output = self.albert_layer_groups[group_idx](
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask[group_idx * self.layers_per_group : (group_idx + 1) * self.layers_per_group],
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                training=training,
+            )
+            hidden_states = layer_group_output[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+class TFAlbertPreTrainedModel(TFPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlbertConfig
+    base_model_prefix = "albert"
+
+
+class TFAlbertMLMHead(tf.keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, input_embeddings: tf.keras.layers.Layer, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+        self.embedding_size = config.embedding_size
+        self.dense = tf.keras.layers.Dense(
+            config.embedding_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
+        )
+        if isinstance(config.hidden_act, str):
+            self.activation = get_tf_activation(config.hidden_act)
+        else:
+            self.activation = config.hidden_act
+
+        self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = input_embeddings
+
+    def build(self, input_shape: tf.TensorShape):
+        self.bias = self.add_weight(shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="bias")
+        self.decoder_bias = self.add_weight(
+            shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="decoder/bias"
+        )
+
+        super().build(input_shape)
+
+    def get_output_embeddings(self) -> tf.keras.layers.Layer:
+        return self.decoder
+
+    def set_output_embeddings(self, value: tf.Variable):
+        self.decoder.weight = value
+        self.decoder.vocab_size = shape_list(value)[0]
+
+    def get_bias(self) -> Dict[str, tf.Variable]:
+        return {"bias": self.bias, "decoder_bias": self.decoder_bias}
+
+    def set_bias(self, value: tf.Variable):
+        self.bias = value["bias"]
+        self.decoder_bias = value["decoder_bias"]
+        self.config.vocab_size = shape_list(value["bias"])[0]
+
+    def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
+        hidden_states = self.dense(inputs=hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(inputs=hidden_states)
+        seq_length = shape_list(tensor=hidden_states)[1]
+        hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.embedding_size])
+        hidden_states = tf.matmul(a=hidden_states, b=self.decoder.weight, transpose_b=True)
+        hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, seq_length, self.config.vocab_size])
+        hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.decoder_bias)
+
+        return hidden_states
+
+
+@keras_serializable
+class TFAlbertMainLayer(tf.keras.layers.Layer):
+    config_class = AlbertConfig
+
+    def __init__(self, config: AlbertConfig, add_pooling_layer: bool = True, **kwargs):
+        super().__init__(**kwargs)
+
+        self.config = config
+
+        self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
+        self.encoder = TFAlbertTransformer(config, name="encoder")
+        self.pooler = (
+            tf.keras.layers.Dense(
+                units=config.hidden_size,
+                kernel_initializer=get_initializer(config.initializer_range),
+                activation="tanh",
+                name="pooler",
+            )
+            if add_pooling_layer
+            else None
+        )
+
+    def get_input_embeddings(self) -> tf.keras.layers.Layer:
+        return self.embeddings
+
+    def set_input_embeddings(self, value: tf.Variable):
+        self.embeddings.weight = value
+        self.embeddings.vocab_size = shape_list(value)[0]
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        raise NotImplementedError
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: bool = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if attention_mask is None:
+            attention_mask = tf.fill(dims=input_shape, value=1)
+
+        if token_type_ids is None:
+            token_type_ids = tf.fill(dims=input_shape, value=0)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            training=training,
+        )
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = tf.reshape(attention_mask, (input_shape[0], 1, 1, input_shape[1]))
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)
+        one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
+        ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
+        extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        encoder_outputs = self.encoder(
+            hidden_states=embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(inputs=sequence_output[:, 0]) if self.pooler is not None else None
+
+        if not return_dict:
+            return (
+                sequence_output,
+                pooled_output,
+            ) + encoder_outputs[1:]
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@dataclass
+class TFAlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of [`TFAlbertForPreTraining`].
+
+    Args:
+        prediction_logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (`tf.Tensor` of shape `(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape
+            `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: tf.Tensor = None
+    prediction_logits: tf.Tensor = None
+    sop_logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+ALBERT_START_DOCSTRING = r"""
+
+    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a [tf.keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
+    as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
+    behavior.
+
+    <Tip>
+
+    TensorFlow models and layers in `transformers` accept two formats as input:
+
+    - having all inputs as keyword arguments (like PyTorch models), or
+    - having all inputs as a list, tuple or dict in the first positional argument.
+
+    The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
+    and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
+    pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second
+    format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with
+    the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first
+    positional argument:
+
+    - a single Tensor with `input_ids` only and nothing else: `model(input_ids)`
+    - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+    `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+    - a dictionary with one or several input Tensors associated to the input names given in the docstring:
+    `model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
+
+    Note that when creating models and layers with
+    [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry
+    about any of this, as you can just pass inputs like you would to any other Python function!
+
+    </Tip>
+
+    Args:
+        config ([`AlbertConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`Numpy array` or `tf.Tensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
+            [`PreTrainedTokenizer.encode`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`Numpy array` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`tf.Tensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the
+            config will be used instead.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail. This argument can be used only in eager mode, in graph mode the value in the config will be
+            used instead.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in
+            eager mode, in graph mode the value will always be set to True.
+        training (`bool`, *optional*, defaults to `False`):
+            Whether or not to use the model in training mode (some modules like dropout modules have different
+            behaviors between training and evaluation).
+"""
+
+
+@add_start_docstrings(
+    "The bare Albert Model transformer outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertModel(TFAlbertPreTrainedModel):
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        return outputs
+
+    def serving_output(self, output: TFBaseModelOutputWithPooling) -> TFBaseModelOutputWithPooling:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=output.last_hidden_state,
+            pooler_output=output.pooler_output,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with two heads on top for pretraining: a `masked language modeling` head and a `sentence order
+    prediction` (classification) head.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForPreTraining(TFAlbertPreTrainedModel, TFAlbertPreTrainingLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"predictions.decoder.weight"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.predictions = TFAlbertMLMHead(config, input_embeddings=self.albert.embeddings, name="predictions")
+        self.sop_classifier = TFAlbertSOPHead(config, name="sop_classifier")
+
+    def get_lm_head(self) -> tf.keras.layers.Layer:
+        return self.predictions
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TFAlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        sentence_order_label: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFAlbertForPreTrainingOutput, Tuple[tf.Tensor]]:
+        r"""
+        Return:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import AutoTokenizer, TFAlbertForPreTraining
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+        >>> model = TFAlbertForPreTraining.from_pretrained("albert-base-v2")
+
+        >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]
+        >>> # Batch size 1
+        >>> outputs = model(input_ids)
+
+        >>> prediction_logits = outputs.prediction_logits
+        >>> sop_logits = outputs.sop_logits
+        ```"""
+
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output, pooled_output = outputs[:2]
+        prediction_scores = self.predictions(hidden_states=sequence_output)
+        sop_scores = self.sop_classifier(pooled_output=pooled_output, training=training)
+        total_loss = None
+
+        if labels is not None and sentence_order_label is not None:
+            d_labels = {"labels": labels}
+            d_labels["sentence_order_label"] = sentence_order_label
+            total_loss = self.hf_compute_loss(labels=d_labels, logits=(prediction_scores, sop_scores))
+
+        if not return_dict:
+            output = (prediction_scores, sop_scores) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return TFAlbertForPreTrainingOutput(
+            loss=total_loss,
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def serving_output(self, output: TFAlbertForPreTrainingOutput) -> TFAlbertForPreTrainingOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFAlbertForPreTrainingOutput(
+            prediction_logits=output.prediction_logits,
+            sop_logits=output.sop_logits,
+            hidden_states=hs,
+            attentions=attns,
+        )
+
+
+class TFAlbertSOPHead(tf.keras.layers.Layer):
+    def __init__(self, config: AlbertConfig, **kwargs):
+        super().__init__(**kwargs)
+
+        self.dropout = tf.keras.layers.Dropout(rate=config.classifier_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(
+            units=config.num_labels,
+            kernel_initializer=get_initializer(config.initializer_range),
+            name="classifier",
+        )
+
+    def call(self, pooled_output: tf.Tensor, training: bool) -> tf.Tensor:
+        dropout_pooled_output = self.dropout(inputs=pooled_output, training=training)
+        logits = self.classifier(inputs=dropout_pooled_output)
+
+        return logits
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top.""", ALBERT_START_DOCSTRING)
+class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions.decoder.weight"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
+        self.predictions = TFAlbertMLMHead(config, input_embeddings=self.albert.embeddings, name="predictions")
+
+    def get_lm_head(self) -> tf.keras.layers.Layer:
+        return self.predictions
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=TFMaskedLMOutput, config_class=_CONFIG_FOR_DOC)
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import AutoTokenizer, TFAlbertForMaskedLM
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+        >>> model = TFAlbertForMaskedLM.from_pretrained("albert-base-v2")
+
+        >>> # add mask_token
+        >>> inputs = tokenizer(f"The capital of [MASK] is Paris.", return_tensors="tf")
+        >>> logits = model(**inputs).logits
+
+        >>> # retrieve index of [MASK]
+        >>> mask_token_index = tf.where(inputs.input_ids == tokenizer.mask_token_id)[0][1]
+        >>> predicted_token_id = tf.math.argmax(logits[0, mask_token_index], axis=-1)
+        >>> tokenizer.decode(predicted_token_id)
+        'france'
+        ```
+
+        ```python
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"]
+        >>> labels = tf.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100)
+        >>> outputs = model(**inputs, labels=labels)
+        >>> round(float(outputs.loss), 2)
+        0.81
+        ```
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.predictions(hidden_states=sequence_output, training=training)
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=prediction_scores)
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMaskedLMOutput(
+            loss=loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
+    def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
+
+
+@add_start_docstrings(
+    """
+    Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClassificationLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"predictions"]
+    _keys_to_ignore_on_load_missing = [r"dropout"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.dropout = tf.keras.layers.Dropout(rate=config.classifier_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(
+            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="vumichien/albert-base-v2-imdb",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output="'LABEL_1'",
+        expected_loss=0.12,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(inputs=pooled_output, training=training)
+        logits = self.classifier(inputs=pooled_output)
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
+    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificationLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions"]
+    _keys_to_ignore_on_load_missing = [r"dropout"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
+        classifier_dropout_prob = (
+            config.classifier_dropout_prob
+            if config.classifier_dropout_prob is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(rate=classifier_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(
+            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(inputs=sequence_output, training=training)
+        logits = self.classifier(inputs=sequence_output)
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFTokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
+    def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
+        self.qa_outputs = tf.keras.layers.Dense(
+            units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
+        )
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint="vumichien/albert-base-v2-squad2",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        qa_target_start_index=12,
+        qa_target_end_index=13,
+        expected_output="'a nice puppet'",
+        expected_loss=7.36,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
+        r"""
+        start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+        outputs = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = outputs[0]
+        logits = self.qa_outputs(inputs=sequence_output)
+        start_logits, end_logits = tf.split(value=logits, num_or_size_splits=2, axis=-1)
+        start_logits = tf.squeeze(input=start_logits, axis=-1)
+        end_logits = tf.squeeze(input=end_logits, axis=-1)
+        loss = None
+
+        if start_positions is not None and end_positions is not None:
+            labels = {"start_position": start_positions}
+            labels["end_position"] = end_positions
+            loss = self.hf_compute_loss(labels=labels, logits=(start_logits, end_logits))
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
+    def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFQuestionAnsweringModelOutput(
+            start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
+        )
+
+
+@add_start_docstrings(
+    """
+    Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    ALBERT_START_DOCSTRING,
+)
+class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
+    # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
+    _keys_to_ignore_on_load_unexpected = [r"pooler", r"predictions"]
+    _keys_to_ignore_on_load_missing = [r"dropout"]
+
+    def __init__(self, config: AlbertConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertMainLayer(config, name="albert")
+        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(
+            units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
+        )
+
+    @property
+    def dummy_inputs(self):
+        """
+        Dummy inputs to build the network.
+
+        Returns:
+            tf.Tensor with dummy inputs
+        """
+        return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS, dtype=tf.int32)}
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: Optional[bool] = False,
+    ) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]`
+            where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above)
+        """
+
+        if input_ids is not None:
+            num_choices = shape_list(input_ids)[1]
+            seq_length = shape_list(input_ids)[2]
+        else:
+            num_choices = shape_list(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[2]
+
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
+        flat_attention_mask = (
+            tf.reshape(tensor=attention_mask, shape=(-1, seq_length)) if attention_mask is not None else None
+        )
+        flat_token_type_ids = (
+            tf.reshape(tensor=token_type_ids, shape=(-1, seq_length)) if token_type_ids is not None else None
+        )
+        flat_position_ids = (
+            tf.reshape(tensor=position_ids, shape=(-1, seq_length)) if position_ids is not None else None
+        )
+        flat_inputs_embeds = (
+            tf.reshape(tensor=inputs_embeds, shape=(-1, seq_length, shape_list(inputs_embeds)[3]))
+            if inputs_embeds is not None
+            else None
+        )
+        outputs = self.albert(
+            input_ids=flat_input_ids,
+            attention_mask=flat_attention_mask,
+            token_type_ids=flat_token_type_ids,
+            position_ids=flat_position_ids,
+            head_mask=head_mask,
+            inputs_embeds=flat_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        pooled_output = outputs[1]
+        pooled_output = self.dropout(inputs=pooled_output, training=training)
+        logits = self.classifier(inputs=pooled_output)
+        reshaped_logits = tf.reshape(tensor=logits, shape=(-1, num_choices))
+        loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=reshaped_logits)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    @tf.function(
+        input_signature=[
+            {
+                "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
+                "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
+            }
+        ]
+    )
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
+    def serving(self, inputs: Dict[str, tf.Tensor]) -> TFMultipleChoiceModelOutput:
+        output = self.call(input_ids=inputs)
+
+        return self.serving_output(output)
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
+    def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/tokenization_albert.py

@@ -0,0 +1,369 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Tokenization classes for ALBERT model."""
+
+
+import os
+import unicodedata
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "albert-base-v1": "https://huggingface.co/albert-base-v1/resolve/main/spiece.model",
+        "albert-large-v1": "https://huggingface.co/albert-large-v1/resolve/main/spiece.model",
+        "albert-xlarge-v1": "https://huggingface.co/albert-xlarge-v1/resolve/main/spiece.model",
+        "albert-xxlarge-v1": "https://huggingface.co/albert-xxlarge-v1/resolve/main/spiece.model",
+        "albert-base-v2": "https://huggingface.co/albert-base-v2/resolve/main/spiece.model",
+        "albert-large-v2": "https://huggingface.co/albert-large-v2/resolve/main/spiece.model",
+        "albert-xlarge-v2": "https://huggingface.co/albert-xlarge-v2/resolve/main/spiece.model",
+        "albert-xxlarge-v2": "https://huggingface.co/albert-xxlarge-v2/resolve/main/spiece.model",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "albert-base-v1": 512,
+    "albert-large-v1": 512,
+    "albert-xlarge-v1": 512,
+    "albert-xxlarge-v1": 512,
+    "albert-base-v2": 512,
+    "albert-large-v2": 512,
+    "albert-xlarge-v2": 512,
+    "albert-xxlarge-v2": 512,
+}
+
+SPIECE_UNDERLINE = "▁"
+
+
+class AlbertTokenizer(PreTrainedTokenizer):
+    """
+    Construct an ALBERT tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        remove_space (`bool`, *optional*, defaults to `True`):
+            Whether or not to strip the text when tokenizing (removing excess spaces before and after the string).
+        keep_accents (`bool`, *optional*, defaults to `False`):
+            Whether or not to keep accents when tokenizing.
+        bos_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        eos_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The end of sequence token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the end of sequence.
+            The token used is the `sep_token`.
+
+            </Tip>
+
+        unk_token (`str`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        sep_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        mask_token (`str`, *optional*, defaults to `"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        sp_model_kwargs (`dict`, *optional*):
+            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
+            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
+            to set:
+
+            - `enable_sampling`: Enable subword regularization.
+            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - `nbest_size = {0,1}`: No sampling is performed.
+              - `nbest_size > 1`: samples from the nbest_size results.
+              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+
+    Attributes:
+        sp_model (`SentencePieceProcessor`):
+            The *SentencePiece* processor that is used for every conversion (string, tokens and IDs).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(
+        self,
+        vocab_file,
+        do_lower_case=True,
+        remove_space=True,
+        keep_accents=False,
+        bos_token="[CLS]",
+        eos_token="[SEP]",
+        unk_token="<unk>",
+        sep_token="[SEP]",
+        pad_token="<pad>",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ) -> None:
+        # Mask token behave like a normal word, i.e. include the space before it and
+        # is included in the raw text, there should be a match in a non-normalized sentence.
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
+
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
+        super().__init__(
+            do_lower_case=do_lower_case,
+            remove_space=remove_space,
+            keep_accents=keep_accents,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            **kwargs,
+        )
+
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.keep_accents = keep_accents
+        self.vocab_file = vocab_file
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+
+    @property
+    def vocab_size(self):
+        return len(self.sp_model)
+
+    def get_vocab(self):
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    def preprocess_text(self, inputs):
+        if self.remove_space:
+            outputs = " ".join(inputs.strip().split())
+        else:
+            outputs = inputs
+        outputs = outputs.replace("``", '"').replace("''", '"')
+
+        if not self.keep_accents:
+            outputs = unicodedata.normalize("NFKD", outputs)
+            outputs = "".join([c for c in outputs if not unicodedata.combining(c)])
+        if self.do_lower_case:
+            outputs = outputs.lower()
+
+        return outputs
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Tokenize a string."""
+        text = self.preprocess_text(text)
+        pieces = self.sp_model.encode(text, out_type=str)
+        new_pieces = []
+        for piece in pieces:
+            if len(piece) > 1 and piece[-1] == str(",") and piece[-2].isdigit():
+                cur_pieces = self.sp_model.EncodeAsPieces(piece[:-1].replace(SPIECE_UNDERLINE, ""))
+                if piece[0] != SPIECE_UNDERLINE and cur_pieces[0][0] == SPIECE_UNDERLINE:
+                    if len(cur_pieces[0]) == 1:
+                        cur_pieces = cur_pieces[1:]
+                    else:
+                        cur_pieces[0] = cur_pieces[0][1:]
+                cur_pieces.append(piece[-1])
+                new_pieces.extend(cur_pieces)
+            else:
+                new_pieces.append(piece)
+
+        return new_pieces
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.PieceToId(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.sp_model.IdToPiece(index)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string.strip()
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. An ALBERT sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return cls + token_ids_0 + sep
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/albert/tokenization_albert_fast.py

@@ -0,0 +1,242 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Tokenization classes for ALBERT model."""
+
+
+import os
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+from ...tokenization_utils import AddedToken
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import is_sentencepiece_available, logging
+
+
+if is_sentencepiece_available():
+    from .tokenization_albert import AlbertTokenizer
+else:
+    AlbertTokenizer = None
+
+logger = logging.get_logger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "spiece.model", "tokenizer_file": "tokenizer.json"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "albert-base-v1": "https://huggingface.co/albert-base-v1/resolve/main/spiece.model",
+        "albert-large-v1": "https://huggingface.co/albert-large-v1/resolve/main/spiece.model",
+        "albert-xlarge-v1": "https://huggingface.co/albert-xlarge-v1/resolve/main/spiece.model",
+        "albert-xxlarge-v1": "https://huggingface.co/albert-xxlarge-v1/resolve/main/spiece.model",
+        "albert-base-v2": "https://huggingface.co/albert-base-v2/resolve/main/spiece.model",
+        "albert-large-v2": "https://huggingface.co/albert-large-v2/resolve/main/spiece.model",
+        "albert-xlarge-v2": "https://huggingface.co/albert-xlarge-v2/resolve/main/spiece.model",
+        "albert-xxlarge-v2": "https://huggingface.co/albert-xxlarge-v2/resolve/main/spiece.model",
+    },
+    "tokenizer_file": {
+        "albert-base-v1": "https://huggingface.co/albert-base-v1/resolve/main/tokenizer.json",
+        "albert-large-v1": "https://huggingface.co/albert-large-v1/resolve/main/tokenizer.json",
+        "albert-xlarge-v1": "https://huggingface.co/albert-xlarge-v1/resolve/main/tokenizer.json",
+        "albert-xxlarge-v1": "https://huggingface.co/albert-xxlarge-v1/resolve/main/tokenizer.json",
+        "albert-base-v2": "https://huggingface.co/albert-base-v2/resolve/main/tokenizer.json",
+        "albert-large-v2": "https://huggingface.co/albert-large-v2/resolve/main/tokenizer.json",
+        "albert-xlarge-v2": "https://huggingface.co/albert-xlarge-v2/resolve/main/tokenizer.json",
+        "albert-xxlarge-v2": "https://huggingface.co/albert-xxlarge-v2/resolve/main/tokenizer.json",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "albert-base-v1": 512,
+    "albert-large-v1": 512,
+    "albert-xlarge-v1": 512,
+    "albert-xxlarge-v1": 512,
+    "albert-base-v2": 512,
+    "albert-large-v2": 512,
+    "albert-xlarge-v2": 512,
+    "albert-xxlarge-v2": 512,
+}
+
+SPIECE_UNDERLINE = "▁"
+
+
+class AlbertTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" ALBERT tokenizer (backed by HuggingFace's *tokenizers* library). Based on
+    [Unigram](https://huggingface.co/docs/tokenizers/python/latest/components.html?highlight=unigram#models). This
+    tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods
+
+    Args:
+        vocab_file (`str`):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        remove_space (`bool`, *optional*, defaults to `True`):
+            Whether or not to strip the text when tokenizing (removing excess spaces before and after the string).
+        keep_accents (`bool`, *optional*, defaults to `False`):
+            Whether or not to keep accents when tokenizing.
+        bos_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        eos_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The end of sequence token. .. note:: When building a sequence using special tokens, this is not the token
+            that is used for the end of sequence. The token used is the `sep_token`.
+        unk_token (`str`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        sep_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        mask_token (`str`, *optional*, defaults to `"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    slow_tokenizer_class = AlbertTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        tokenizer_file=None,
+        do_lower_case=True,
+        remove_space=True,
+        keep_accents=False,
+        bos_token="[CLS]",
+        eos_token="[SEP]",
+        unk_token="<unk>",
+        sep_token="[SEP]",
+        pad_token="<pad>",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        **kwargs,
+    ):
+        # Mask token behave like a normal word, i.e. include the space before it and
+        # is included in the raw text, there should be a match in a non-normalized sentence.
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
+
+        super().__init__(
+            vocab_file,
+            tokenizer_file=tokenizer_file,
+            do_lower_case=do_lower_case,
+            remove_space=remove_space,
+            keep_accents=keep_accents,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            **kwargs,
+        )
+
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.keep_accents = keep_accents
+        self.vocab_file = vocab_file
+        self.can_save_slow_tokenizer = False if not self.vocab_file else True
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. An ALBERT sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: list of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return cls + token_ids_0 + sep
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        if token_ids_1 is None, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not self.can_save_slow_tokenizer:
+            raise ValueError(
+                "Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
+                "tokenizer."
+            )
+
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/align/__init__.py

@@ -0,0 +1,73 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_align": [
+        "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "AlignConfig",
+        "AlignTextConfig",
+        "AlignVisionConfig",
+    ],
+    "processing_align": ["AlignProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_align"] = [
+        "ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AlignModel",
+        "AlignPreTrainedModel",
+        "AlignTextModel",
+        "AlignVisionModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_align import (
+        ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        AlignConfig,
+        AlignTextConfig,
+        AlignVisionConfig,
+    )
+    from .processing_align import AlignProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_align import (
+            ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AlignModel,
+            AlignPreTrainedModel,
+            AlignTextModel,
+            AlignVisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/align/configuration_align.py

@@ -0,0 +1,398 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" ALIGN model configuration"""
+
+import copy
+import os
+from typing import TYPE_CHECKING, List, Union
+
+
+if TYPE_CHECKING:
+    pass
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "kakaobrain/align-base": "https://huggingface.co/kakaobrain/align-base/resolve/main/config.json",
+}
+
+
+class AlignTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlignTextModel`]. It is used to instantiate a
+    ALIGN text encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the text encoder of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values here are
+    copied from BERT.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the Align Text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`AlignTextModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`AlignTextModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*, defaults to 0)
+            Padding token id.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignTextConfig, AlignTextModel
+
+    >>> # Initializing a AlignTextConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignTextConfig()
+
+    >>> # Initializing a AlignTextModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "align_text_model"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        use_cache=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.pad_token_id = pad_token_id
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from AlignConfig
+        if config_dict.get("model_type") == "align":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class AlignVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlignVisionModel`]. It is used to instantiate a
+    ALIGN vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values are copied
+    from EfficientNet (efficientnet-b7)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 600):
+            The input image size.
+        width_coefficient (`float`, *optional*, defaults to 2.0):
+            Scaling coefficient for network width at each stage.
+        depth_coefficient (`float`, *optional*, defaults to 3.1):
+            Scaling coefficient for network depth at each stage.
+        depth_divisor `int`, *optional*, defaults to 8):
+            A unit of network width.
+        kernel_sizes (`List[int]`, *optional*, defaults to `[3, 3, 5, 3, 5, 5, 3]`):
+            List of kernel sizes to be used in each block.
+        in_channels (`List[int]`, *optional*, defaults to `[32, 16, 24, 40, 80, 112, 192]`):
+            List of input channel sizes to be used in each block for convolutional layers.
+        out_channels (`List[int]`, *optional*, defaults to `[16, 24, 40, 80, 112, 192, 320]`):
+            List of output channel sizes to be used in each block for convolutional layers.
+        depthwise_padding (`List[int]`, *optional*, defaults to `[]`):
+            List of block indices with square padding.
+        strides: (`List[int]`, *optional*, defaults to `[1, 2, 2, 2, 1, 2, 1]`):
+            List of stride sizes to be used in each block for convolutional layers.
+        num_block_repeats (`List[int]`, *optional*, defaults to `[1, 2, 2, 3, 3, 4, 1]`):
+            List of the number of times each block is to repeated.
+        expand_ratios (`List[int]`, *optional*, defaults to `[1, 6, 6, 6, 6, 6, 6]`):
+            List of scaling coefficient of each block.
+        squeeze_expansion_ratio (`float`, *optional*, defaults to 0.25):
+            Squeeze expansion ratio.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in each block. If string, `"gelu"`, `"relu"`,
+            `"selu", `"gelu_new"`, `"silu"` and `"mish"` are supported.
+        hiddem_dim (`int`, *optional*, defaults to 1280):
+            The hidden dimension of the layer before the classification head.
+        pooling_type (`str` or `function`, *optional*, defaults to `"mean"`):
+            Type of final pooling to be applied before the dense classification head. Available options are [`"mean"`,
+            `"max"`]
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        batch_norm_eps (`float`, *optional*, defaults to 1e-3):
+            The epsilon used by the batch normalization layers.
+        batch_norm_momentum (`float`, *optional*, defaults to 0.99):
+            The momentum used by the batch normalization layers.
+        dropout_rate (`float`, *optional*, defaults to 0.5):
+            The dropout rate to be applied before final classifier layer.
+        drop_connect_rate (`float`, *optional*, defaults to 0.2):
+            The drop rate for skip connections.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignVisionConfig, AlignVisionModel
+
+    >>> # Initializing a AlignVisionConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignVisionConfig()
+
+    >>> # Initializing a AlignVisionModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "align_vision_model"
+
+    def __init__(
+        self,
+        num_channels: int = 3,
+        image_size: int = 600,
+        width_coefficient: float = 2.0,
+        depth_coefficient: float = 3.1,
+        depth_divisor: int = 8,
+        kernel_sizes: List[int] = [3, 3, 5, 3, 5, 5, 3],
+        in_channels: List[int] = [32, 16, 24, 40, 80, 112, 192],
+        out_channels: List[int] = [16, 24, 40, 80, 112, 192, 320],
+        depthwise_padding: List[int] = [],
+        strides: List[int] = [1, 2, 2, 2, 1, 2, 1],
+        num_block_repeats: List[int] = [1, 2, 2, 3, 3, 4, 1],
+        expand_ratios: List[int] = [1, 6, 6, 6, 6, 6, 6],
+        squeeze_expansion_ratio: float = 0.25,
+        hidden_act: str = "swish",
+        hidden_dim: int = 2560,
+        pooling_type: str = "mean",
+        initializer_range: float = 0.02,
+        batch_norm_eps: float = 0.001,
+        batch_norm_momentum: float = 0.99,
+        dropout_rate: float = 0.5,
+        drop_connect_rate: float = 0.2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.width_coefficient = width_coefficient
+        self.depth_coefficient = depth_coefficient
+        self.depth_divisor = depth_divisor
+        self.kernel_sizes = kernel_sizes
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.depthwise_padding = depthwise_padding
+        self.strides = strides
+        self.num_block_repeats = num_block_repeats
+        self.expand_ratios = expand_ratios
+        self.squeeze_expansion_ratio = squeeze_expansion_ratio
+        self.hidden_act = hidden_act
+        self.hidden_dim = hidden_dim
+        self.pooling_type = pooling_type
+        self.initializer_range = initializer_range
+        self.batch_norm_eps = batch_norm_eps
+        self.batch_norm_momentum = batch_norm_momentum
+        self.dropout_rate = dropout_rate
+        self.drop_connect_rate = drop_connect_rate
+        self.num_hidden_layers = sum(num_block_repeats) * 4
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from AlignConfig
+        if config_dict.get("model_type") == "align":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class AlignConfig(PretrainedConfig):
+    r"""
+    [`AlignConfig`] is the configuration class to store the configuration of a [`AlignModel`]. It is used to
+    instantiate a ALIGN model according to the specified arguments, defining the text model and vision model configs.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AlignTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AlignVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 640):
+            Dimentionality of text and vision projection layers.
+        temperature_init_value (`float`, *optional*, defaults to 1.0):
+            The inital value of the *temperature* paramter. Default is used as per the original ALIGN implementation.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignConfig, AlignModel
+
+    >>> # Initializing a AlignConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignConfig()
+
+    >>> # Initializing a AlignModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a AlignConfig from a AlignTextConfig and a AlignVisionConfig
+    >>> from transformers import AlignTextConfig, AlignVisionConfig
+
+    >>> # Initializing ALIGN Text and Vision configurations
+    >>> config_text = AlignTextConfig()
+    >>> config_vision = AlignVisionConfig()
+
+    >>> config = AlignConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "align"
+    is_composition = True
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        projection_dim=640,
+        temperature_init_value=1.0,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the AlignTextConfig with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. Initializing the AlignVisionConfig with default values.")
+
+        self.text_config = AlignTextConfig(**text_config)
+        self.vision_config = AlignVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.temperature_init_value = temperature_init_value
+        self.initializer_range = initializer_range
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: AlignTextConfig, vision_config: AlignVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`AlignConfig`] (or a derived class) from align text model configuration and align vision model
+        configuration.
+
+        Returns:
+            [`AlignConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["text_config"] = self.text_config.to_dict()
+        output["vision_config"] = self.vision_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/align/convert_align_tf_to_hf.py

@@ -0,0 +1,387 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ALIGN checkpoints from the original repository."""
+
+import argparse
+import os
+
+import align
+import numpy as np
+import requests
+import tensorflow as tf
+import torch
+from PIL import Image
+from tokenizer import Tokenizer
+
+from transformers import (
+    AlignConfig,
+    AlignModel,
+    AlignProcessor,
+    BertConfig,
+    BertTokenizer,
+    EfficientNetConfig,
+    EfficientNetImageProcessor,
+)
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def preprocess(image):
+    image = tf.image.resize(image, (346, 346))
+    image = tf.image.crop_to_bounding_box(image, (346 - 289) // 2, (346 - 289) // 2, 289, 289)
+    return image
+
+
+def get_align_config():
+    vision_config = EfficientNetConfig.from_pretrained("google/efficientnet-b7")
+    vision_config.image_size = 289
+    vision_config.hidden_dim = 640
+    vision_config.id2label = {"0": "LABEL_0", "1": "LABEL_1"}
+    vision_config.label2id = {"LABEL_0": 0, "LABEL_1": 1}
+    vision_config.depthwise_padding = []
+
+    text_config = BertConfig()
+    config = AlignConfig.from_text_vision_configs(
+        text_config=text_config, vision_config=vision_config, projection_dim=640
+    )
+    return config
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+def get_processor():
+    image_processor = EfficientNetImageProcessor(
+        do_center_crop=True,
+        rescale_factor=1 / 127.5,
+        rescale_offset=True,
+        do_normalize=False,
+        include_top=False,
+        resample=Image.BILINEAR,
+    )
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+    tokenizer.model_max_length = 64
+    processor = AlignProcessor(image_processor=image_processor, tokenizer=tokenizer)
+    return processor
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def rename_keys(original_param_names):
+    # EfficientNet image encoder
+    block_names = [v.split("_")[0].split("block")[1] for v in original_param_names if v.startswith("block")]
+    block_names = list(set(block_names))
+    block_names = sorted(block_names)
+    num_blocks = len(block_names)
+    block_name_mapping = {b: str(i) for b, i in zip(block_names, range(num_blocks))}
+
+    rename_keys = []
+    rename_keys.append(("stem_conv/kernel:0", "embeddings.convolution.weight"))
+    rename_keys.append(("stem_bn/gamma:0", "embeddings.batchnorm.weight"))
+    rename_keys.append(("stem_bn/beta:0", "embeddings.batchnorm.bias"))
+    rename_keys.append(("stem_bn/moving_mean:0", "embeddings.batchnorm.running_mean"))
+    rename_keys.append(("stem_bn/moving_variance:0", "embeddings.batchnorm.running_var"))
+
+    for b in block_names:
+        hf_b = block_name_mapping[b]
+        rename_keys.append((f"block{b}_expand_conv/kernel:0", f"encoder.blocks.{hf_b}.expansion.expand_conv.weight"))
+        rename_keys.append((f"block{b}_expand_bn/gamma:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.weight"))
+        rename_keys.append((f"block{b}_expand_bn/beta:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.bias"))
+        rename_keys.append(
+            (f"block{b}_expand_bn/moving_mean:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_expand_bn/moving_variance:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_var")
+        )
+        rename_keys.append(
+            (f"block{b}_dwconv/depthwise_kernel:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_conv.weight")
+        )
+        rename_keys.append((f"block{b}_bn/gamma:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.weight"))
+        rename_keys.append((f"block{b}_bn/beta:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.bias"))
+        rename_keys.append(
+            (f"block{b}_bn/moving_mean:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_bn/moving_variance:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_var")
+        )
+
+        rename_keys.append((f"block{b}_se_reduce/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.weight"))
+        rename_keys.append((f"block{b}_se_reduce/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.bias"))
+        rename_keys.append((f"block{b}_se_expand/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.weight"))
+        rename_keys.append((f"block{b}_se_expand/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.bias"))
+        rename_keys.append(
+            (f"block{b}_project_conv/kernel:0", f"encoder.blocks.{hf_b}.projection.project_conv.weight")
+        )
+        rename_keys.append((f"block{b}_project_bn/gamma:0", f"encoder.blocks.{hf_b}.projection.project_bn.weight"))
+        rename_keys.append((f"block{b}_project_bn/beta:0", f"encoder.blocks.{hf_b}.projection.project_bn.bias"))
+        rename_keys.append(
+            (f"block{b}_project_bn/moving_mean:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_project_bn/moving_variance:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_var")
+        )
+
+    key_mapping = {}
+    for item in rename_keys:
+        if item[0] in original_param_names:
+            key_mapping[item[0]] = "vision_model." + item[1]
+
+    # BERT text encoder
+    rename_keys = []
+    old = "tf_bert_model/bert"
+    new = "text_model"
+    for i in range(12):
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/query/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.query.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/query/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.query.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/key/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.key.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/key/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.key.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/value/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.value.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/value/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.value.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/dense/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.output.dense.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/dense/bias:0",
+                f"{new}.encoder.layer.{i}.attention.output.dense.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/gamma:0",
+                f"{new}.encoder.layer.{i}.attention.output.LayerNorm.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/beta:0",
+                f"{new}.encoder.layer.{i}.attention.output.LayerNorm.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/intermediate/dense/kernel:0",
+                f"{new}.encoder.layer.{i}.intermediate.dense.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/intermediate/dense/bias:0",
+                f"{new}.encoder.layer.{i}.intermediate.dense.bias",
+            )
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/dense/kernel:0", f"{new}.encoder.layer.{i}.output.dense.weight")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/dense/bias:0", f"{new}.encoder.layer.{i}.output.dense.bias")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/LayerNorm/gamma:0", f"{new}.encoder.layer.{i}.output.LayerNorm.weight")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/LayerNorm/beta:0", f"{new}.encoder.layer.{i}.output.LayerNorm.bias")
+        )
+
+    rename_keys.append((f"{old}/embeddings/word_embeddings/weight:0", f"{new}.embeddings.word_embeddings.weight"))
+    rename_keys.append(
+        (f"{old}/embeddings/position_embeddings/embeddings:0", f"{new}.embeddings.position_embeddings.weight")
+    )
+    rename_keys.append(
+        (f"{old}/embeddings/token_type_embeddings/embeddings:0", f"{new}.embeddings.token_type_embeddings.weight")
+    )
+    rename_keys.append((f"{old}/embeddings/LayerNorm/gamma:0", f"{new}.embeddings.LayerNorm.weight"))
+    rename_keys.append((f"{old}/embeddings/LayerNorm/beta:0", f"{new}.embeddings.LayerNorm.bias"))
+
+    rename_keys.append((f"{old}/pooler/dense/kernel:0", f"{new}.pooler.dense.weight"))
+    rename_keys.append((f"{old}/pooler/dense/bias:0", f"{new}.pooler.dense.bias"))
+    rename_keys.append(("dense/kernel:0", "text_projection.weight"))
+    rename_keys.append(("dense/bias:0", "text_projection.bias"))
+    rename_keys.append(("dense/bias:0", "text_projection.bias"))
+    rename_keys.append(("temperature:0", "temperature"))
+
+    for item in rename_keys:
+        if item[0] in original_param_names:
+            key_mapping[item[0]] = item[1]
+    return key_mapping
+
+
+def replace_params(hf_params, tf_params, key_mapping):
+    list(hf_params.keys())
+
+    for key, value in tf_params.items():
+        if key not in key_mapping:
+            continue
+
+        hf_key = key_mapping[key]
+        if "_conv" in key and "kernel" in key:
+            new_hf_value = torch.from_numpy(value).permute(3, 2, 0, 1)
+        elif "embeddings" in key:
+            new_hf_value = torch.from_numpy(value)
+        elif "depthwise_kernel" in key:
+            new_hf_value = torch.from_numpy(value).permute(2, 3, 0, 1)
+        elif "kernel" in key:
+            new_hf_value = torch.from_numpy(np.transpose(value))
+        elif "temperature" in key:
+            new_hf_value = value
+        elif "bn/gamma" or "bn/beta" in key:
+            new_hf_value = torch.from_numpy(np.transpose(value)).squeeze()
+        else:
+            new_hf_value = torch.from_numpy(value)
+
+        # Replace HF parameters with original TF model parameters
+        hf_params[hf_key].copy_(new_hf_value)
+
+
+@torch.no_grad()
+def convert_align_checkpoint(checkpoint_path, pytorch_dump_folder_path, save_model, push_to_hub):
+    """
+    Copy/paste/tweak model's weights to our ALIGN structure.
+    """
+    # Load original model
+    seq_length = 64
+    tok = Tokenizer(seq_length)
+    original_model = align.Align("efficientnet-b7", "bert-base", 640, seq_length, tok.get_vocab_size())
+    original_model.compile()
+    original_model.load_weights(checkpoint_path)
+
+    tf_params = original_model.trainable_variables
+    tf_non_train_params = original_model.non_trainable_variables
+    tf_params = {param.name: param.numpy() for param in tf_params}
+    for param in tf_non_train_params:
+        tf_params[param.name] = param.numpy()
+    tf_param_names = list(tf_params.keys())
+
+    # Load HuggingFace model
+    config = get_align_config()
+    hf_model = AlignModel(config).eval()
+    hf_params = hf_model.state_dict()
+
+    # Create src-to-dst parameter name mapping dictionary
+    print("Converting parameters...")
+    key_mapping = rename_keys(tf_param_names)
+    replace_params(hf_params, tf_params, key_mapping)
+
+    # Initialize processor
+    processor = get_processor()
+    inputs = processor(
+        images=prepare_img(), text="A picture of a cat", padding="max_length", max_length=64, return_tensors="pt"
+    )
+
+    # HF model inference
+    hf_model.eval()
+    with torch.no_grad():
+        outputs = hf_model(**inputs)
+
+    hf_image_features = outputs.image_embeds.detach().numpy()
+    hf_text_features = outputs.text_embeds.detach().numpy()
+
+    # Original model inference
+    original_model.trainable = False
+    tf_image_processor = EfficientNetImageProcessor(
+        do_center_crop=True,
+        do_rescale=False,
+        do_normalize=False,
+        include_top=False,
+        resample=Image.BILINEAR,
+    )
+    image = tf_image_processor(images=prepare_img(), return_tensors="tf", data_format="channels_last")["pixel_values"]
+    text = tok(tf.constant(["A picture of a cat"]))
+
+    image_features = original_model.image_encoder(image, training=False)
+    text_features = original_model.text_encoder(text, training=False)
+
+    image_features = tf.nn.l2_normalize(image_features, axis=-1)
+    text_features = tf.nn.l2_normalize(text_features, axis=-1)
+
+    # Check whether original and HF model outputs match  -> np.allclose
+    assert np.allclose(image_features, hf_image_features, atol=1e-3), "The predicted image features are not the same."
+    assert np.allclose(text_features, hf_text_features, atol=1e-3), "The predicted text features are not the same."
+    print("Model outputs match!")
+
+    if save_model:
+        # Create folder to save model
+        if not os.path.isdir(pytorch_dump_folder_path):
+            os.mkdir(pytorch_dump_folder_path)
+        # Save converted model and feature extractor
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        # Push model and feature extractor to hub
+        print("Pushing converted ALIGN to the hub...")
+        processor.push_to_hub("align-base")
+        hf_model.push_to_hub("align-base")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--checkpoint_path",
+        default="./weights/model-weights",
+        type=str,
+        help="Path to the pretrained TF ALIGN checkpoint.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default="hf_model",
+        type=str,
+        help="Path to the output PyTorch model directory.",
+    )
+    parser.add_argument("--save_model", action="store_true", help="Save model to local")
+    parser.add_argument("--push_to_hub", action="store_true", help="Push model and feature extractor to the hub")
+
+    args = parser.parse_args()
+    convert_align_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.save_model, args.push_to_hub)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/align/modeling_align.py

@@ -0,0 +1,1642 @@
+# coding=utf-8
+# Copyright 2023 The Google Research Team Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch ALIGN model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutputWithNoAttention,
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    BaseModelOutputWithPoolingAndNoAttention,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_align import AlignConfig, AlignTextConfig, AlignVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "kakaobrain/align-base"
+_CONFIG_FOR_DOC = "AlignConfig"
+
+
+ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "kakaobrain/align-base",
+    # See all ALIGN models at https://huggingface.co/models?filter=align
+]
+
+
+ALIGN_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`AlignConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ALIGN_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+ALIGN_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`EfficientNetImageProcessor.__call__`] for details.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+ALIGN_INPUTS_DOCSTRING = r"""
+    Args:
+       input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`EfficientNetImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@dataclass
+class AlignVisionModelOutput(ModelOutput):
+    """
+    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
+
+    Args:
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class AlignTextModelOutput(ModelOutput):
+    """
+    Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The text embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    text_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class AlignOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image:(`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`AlignTextModel`].
+        image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The output of [`AlignVisionModel`].
+        text_model_output(`BaseModelOutputWithPoolingAndCrossAttentions`):
+            The output of the [`AlignTextModel`].
+        vision_model_output(`BaseModelOutputWithPoolingAndNoAttention`):
+            The output of the [`AlignVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    image_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPoolingAndCrossAttentions = None
+    vision_model_output: BaseModelOutputWithPoolingAndNoAttention = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device), label_smoothing=0.1)
+
+
+def align_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.round_filters with EfficientNet -> AlignVision
+def round_filters(config: AlignVisionConfig, num_channels: int):
+    r"""
+    Round number of filters based on depth multiplier.
+    """
+    divisor = config.depth_divisor
+    num_channels *= config.width_coefficient
+    new_dim = max(divisor, int(num_channels + divisor / 2) // divisor * divisor)
+
+    # Make sure that round down does not go down by more than 10%.
+    if new_dim < 0.9 * num_channels:
+        new_dim += divisor
+
+    return int(new_dim)
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.correct_pad
+def correct_pad(kernel_size: Union[int, Tuple], adjust: bool = True):
+    r"""
+    Utility function to get the tuple padding value for the depthwise convolution.
+
+    Args:
+        kernel_size (`int` or `tuple`):
+            Kernel size of the convolution layers.
+        adjust (`bool`, *optional*, defaults to `True`):
+            Adjusts padding value to apply to right and bottom sides of the input.
+    """
+    if isinstance(kernel_size, int):
+        kernel_size = (kernel_size, kernel_size)
+
+    correct = (kernel_size[0] // 2, kernel_size[1] // 2)
+    if adjust:
+        return (correct[1] - 1, correct[1], correct[0] - 1, correct[0])
+    else:
+        return (correct[1], correct[1], correct[0], correct[0])
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetEmbeddings with EfficientNet->AlignVision
+class AlignVisionEmbeddings(nn.Module):
+    r"""
+    A module that corresponds to the stem module of the original work.
+    """
+
+    def __init__(self, config: AlignVisionConfig):
+        super().__init__()
+
+        self.out_dim = round_filters(config, 32)
+        self.padding = nn.ZeroPad2d(padding=(0, 1, 0, 1))
+        self.convolution = nn.Conv2d(
+            config.num_channels, self.out_dim, kernel_size=3, stride=2, padding="valid", bias=False
+        )
+        self.batchnorm = nn.BatchNorm2d(self.out_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum)
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        features = self.padding(pixel_values)
+        features = self.convolution(features)
+        features = self.batchnorm(features)
+        features = self.activation(features)
+
+        return features
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetDepthwiseConv2d with EfficientNet->AlignVision
+class AlignVisionDepthwiseConv2d(nn.Conv2d):
+    def __init__(
+        self,
+        in_channels,
+        depth_multiplier=1,
+        kernel_size=3,
+        stride=1,
+        padding=0,
+        dilation=1,
+        bias=True,
+        padding_mode="zeros",
+    ):
+        out_channels = in_channels * depth_multiplier
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=in_channels,
+            bias=bias,
+            padding_mode=padding_mode,
+        )
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetExpansionLayer with EfficientNet->AlignVision
+class AlignVisionExpansionLayer(nn.Module):
+    r"""
+    This corresponds to the expansion phase of each block in the original implementation.
+    """
+
+    def __init__(self, config: AlignVisionConfig, in_dim: int, out_dim: int, stride: int):
+        super().__init__()
+        self.expand_conv = nn.Conv2d(
+            in_channels=in_dim,
+            out_channels=out_dim,
+            kernel_size=1,
+            padding="same",
+            bias=False,
+        )
+        self.expand_bn = nn.BatchNorm2d(num_features=out_dim, eps=config.batch_norm_eps)
+        self.expand_act = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        # Expand phase
+        hidden_states = self.expand_conv(hidden_states)
+        hidden_states = self.expand_bn(hidden_states)
+        hidden_states = self.expand_act(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetDepthwiseLayer with with EfficientNet->AlignVision
+class AlignVisionDepthwiseLayer(nn.Module):
+    r"""
+    This corresponds to the depthwise convolution phase of each block in the original implementation.
+    """
+
+    def __init__(
+        self,
+        config: AlignVisionConfig,
+        in_dim: int,
+        stride: int,
+        kernel_size: int,
+        adjust_padding: bool,
+    ):
+        super().__init__()
+        self.stride = stride
+        conv_pad = "valid" if self.stride == 2 else "same"
+        padding = correct_pad(kernel_size, adjust=adjust_padding)
+
+        self.depthwise_conv_pad = nn.ZeroPad2d(padding=padding)
+        self.depthwise_conv = AlignVisionDepthwiseConv2d(
+            in_dim, kernel_size=kernel_size, stride=stride, padding=conv_pad, bias=False
+        )
+        self.depthwise_norm = nn.BatchNorm2d(
+            num_features=in_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum
+        )
+        self.depthwise_act = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        # Depthwise convolution
+        if self.stride == 2:
+            hidden_states = self.depthwise_conv_pad(hidden_states)
+
+        hidden_states = self.depthwise_conv(hidden_states)
+        hidden_states = self.depthwise_norm(hidden_states)
+        hidden_states = self.depthwise_act(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.efficientnet.modeling_efficientnet.EfficientNetSqueezeExciteLayer with with EfficientNet->AlignVision
+class AlignVisionSqueezeExciteLayer(nn.Module):
+    r"""
+    This corresponds to the Squeeze and Excitement phase of each block in the original implementation.
+    """
+
+    def __init__(self, config: AlignVisionConfig, in_dim: int, expand_dim: int, expand: bool = False):
+        super().__init__()
+        self.dim = expand_dim if expand else in_dim
+        self.dim_se = max(1, int(in_dim * config.squeeze_expansion_ratio))
+
+        self.squeeze = nn.AdaptiveAvgPool2d(output_size=1)
+        self.reduce = nn.Conv2d(
+            in_channels=self.dim,
+            out_channels=self.dim_se,
+            kernel_size=1,
+            padding="same",
+        )
+        self.expand = nn.Conv2d(
+            in_channels=self.dim_se,
+            out_channels=self.dim,
+            kernel_size=1,
+            padding="same",
+        )
+        self.act_reduce = ACT2FN[config.hidden_act]
+        self.act_expand = nn.Sigmoid()
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        inputs = hidden_states
+        hidden_states = self.squeeze(hidden_states)
+        hidden_states = self.reduce(hidden_states)
+        hidden_states = self.act_reduce(hidden_states)
+
+        hidden_states = self.expand(hidden_states)
+        hidden_states = self.act_expand(hidden_states)
+        hidden_states = torch.mul(inputs, hidden_states)
+
+        return hidden_states
+
+
+class AlignVisionFinalBlockLayer(nn.Module):
+    r"""
+    This corresponds to the final phase of each block in the original implementation.
+    """
+
+    def __init__(
+        self, config: AlignVisionConfig, in_dim: int, out_dim: int, stride: int, drop_rate: float, id_skip: bool
+    ):
+        super().__init__()
+        self.apply_dropout = stride == 1 and not id_skip
+        self.project_conv = nn.Conv2d(
+            in_channels=in_dim,
+            out_channels=out_dim,
+            kernel_size=1,
+            padding="same",
+            bias=False,
+        )
+        self.project_bn = nn.BatchNorm2d(
+            num_features=out_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum
+        )
+        self.dropout = nn.Dropout(p=drop_rate)
+
+    def forward(self, embeddings: torch.FloatTensor, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        hidden_states = self.project_conv(hidden_states)
+        hidden_states = self.project_bn(hidden_states)
+
+        if self.apply_dropout:
+            hidden_states = self.dropout(hidden_states)
+            hidden_states = hidden_states + embeddings
+
+        return hidden_states
+
+
+class AlignVisionBlock(nn.Module):
+    r"""
+    This corresponds to the block module of original the EfficientNet vision encoder implementation.
+
+    Args:
+        config ([`AlignVisionConfig`]):
+            Model configuration class.
+        in_dim (`int`):
+            Number of input channels.
+        out_dim (`int`):
+            Number of output channels.
+        stride (`int`):
+            Stride size to be used in convolution layers.
+        expand_ratio (`int`):
+            Expand ratio to set the output dimensions for the expansion and squeeze-excite layers.
+        kernel_size (`int`):
+            Kernel size for the depthwise convolution layer.
+        drop_rate (`float`):
+            Dropout rate to be used in the final phase of each block.
+        id_skip (`bool`):
+            Whether to apply dropout and sum the final hidden states with the input embeddings during the final phase
+            of each block. Set to `True` for the first block of each stage.
+        adjust_padding (`bool`):
+            Whether to apply padding to only right and bottom side of the input kernel before the depthwise convolution
+            operation, set to `True` for inputs with odd input sizes.
+    """
+
+    def __init__(
+        self,
+        config: AlignVisionConfig,
+        in_dim: int,
+        out_dim: int,
+        stride: int,
+        expand_ratio: int,
+        kernel_size: int,
+        drop_rate: float,
+        id_skip: bool,
+        adjust_padding: bool,
+    ):
+        super().__init__()
+        self.expand_ratio = expand_ratio
+        self.expand = True if self.expand_ratio != 1 else False
+        expand_in_dim = in_dim * expand_ratio
+
+        if self.expand:
+            self.expansion = AlignVisionExpansionLayer(
+                config=config, in_dim=in_dim, out_dim=expand_in_dim, stride=stride
+            )
+
+        self.depthwise_conv = AlignVisionDepthwiseLayer(
+            config=config,
+            in_dim=expand_in_dim if self.expand else in_dim,
+            stride=stride,
+            kernel_size=kernel_size,
+            adjust_padding=adjust_padding,
+        )
+        self.squeeze_excite = AlignVisionSqueezeExciteLayer(
+            config=config, in_dim=in_dim, expand_dim=expand_in_dim, expand=self.expand
+        )
+        self.projection = AlignVisionFinalBlockLayer(
+            config=config,
+            in_dim=expand_in_dim if self.expand else in_dim,
+            out_dim=out_dim,
+            stride=stride,
+            drop_rate=drop_rate,
+            id_skip=id_skip,
+        )
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
+        embeddings = hidden_states
+        # Expansion and depthwise convolution phase
+        if self.expand_ratio != 1:
+            hidden_states = self.expansion(hidden_states)
+        hidden_states = self.depthwise_conv(hidden_states)
+
+        # Squeeze and excite phase
+        hidden_states = self.squeeze_excite(hidden_states)
+        hidden_states = self.projection(embeddings, hidden_states)
+        return hidden_states
+
+
+class AlignVisionEncoder(nn.Module):
+    r"""
+    Forward propogates the embeddings through each vision encoder (EfficientNet) block.
+
+    Args:
+        config ([`AlignVisionConfig`]):
+            Model configuration class.
+    """
+
+    def __init__(self, config: AlignVisionConfig):
+        super().__init__()
+        self.depth_coefficient = config.depth_coefficient
+
+        def round_repeats(repeats):
+            # Round number of block repeats based on depth multiplier.
+            return int(math.ceil(self.depth_coefficient * repeats))
+
+        num_base_blocks = len(config.in_channels)
+        num_blocks = sum(round_repeats(n) for n in config.num_block_repeats)
+
+        curr_block_num = 0
+        blocks = []
+        for i in range(num_base_blocks):
+            in_dim = round_filters(config, config.in_channels[i])
+            out_dim = round_filters(config, config.out_channels[i])
+            stride = config.strides[i]
+            kernel_size = config.kernel_sizes[i]
+            expand_ratio = config.expand_ratios[i]
+
+            for j in range(round_repeats(config.num_block_repeats[i])):
+                id_skip = True if j == 0 else False
+                stride = 1 if j > 0 else stride
+                in_dim = out_dim if j > 0 else in_dim
+                adjust_padding = False if curr_block_num in config.depthwise_padding else True
+                drop_rate = config.drop_connect_rate * curr_block_num / num_blocks
+
+                block = AlignVisionBlock(
+                    config=config,
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    stride=stride,
+                    kernel_size=kernel_size,
+                    expand_ratio=expand_ratio,
+                    drop_rate=drop_rate,
+                    id_skip=id_skip,
+                    adjust_padding=adjust_padding,
+                )
+                blocks.append(block)
+                curr_block_num += 1
+
+        self.blocks = nn.ModuleList(blocks)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> BaseModelOutputWithPoolingAndNoAttention:
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
+
+        for block in self.blocks:
+            hidden_states = block(hidden_states)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states] if v is not None)
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEmbeddings with Bert->AlignText
+class AlignTextEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
+        )
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values_length: int = 0,
+    ) -> torch.Tensor:
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->AlignText
+class AlignTextSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        use_cache = past_key_value is not None
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            query_length, key_length = query_layer.shape[2], key_layer.shape[2]
+            if use_cache:
+                position_ids_l = torch.tensor(key_length - 1, dtype=torch.long, device=hidden_states.device).view(
+                    -1, 1
+                )
+            else:
+                position_ids_l = torch.arange(query_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(key_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in AlignTextModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->AlignText
+class AlignTextSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->AlignText
+class AlignTextAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = AlignTextSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.output = AlignTextSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->AlignText
+class AlignTextIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->AlignText
+class AlignTextOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->AlignText
+class AlignTextLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = AlignTextAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
+            self.crossattention = AlignTextAttention(config, position_embedding_type="absolute")
+        self.intermediate = AlignTextIntermediate(config)
+        self.output = AlignTextOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
+                    " by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->AlignText
+class AlignTextEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([AlignTextLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler with Bert -> AlignText
+class AlignTextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class AlignPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AlignConfig
+    base_model_prefix = "align"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, AlignModel):
+            nn.init.xavier_uniform_(module.text_projection.weight)
+            module.text_projection.bias.data.zero_()
+            module.text_projection._is_hf_initialized = True
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (AlignTextModel, AlignVisionModel)):
+            module.gradient_checkpointing = value
+
+
+@add_start_docstrings(
+    """The text model from ALIGN without any head or projection on top.""",
+    ALIGN_START_DOCSTRING,
+)
+class AlignTextModel(AlignPreTrainedModel):
+    config_class = AlignTextConfig
+
+    def __init__(self, config: AlignTextConfig, add_pooling_layer: bool = True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = AlignTextEmbeddings(config)
+        self.encoder = AlignTextEncoder(config)
+
+        self.pooler = AlignTextPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    @add_start_docstrings_to_model_forward(ALIGN_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndCrossAttentions, config_class=AlignTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, AlignTextModel
+
+        >>> model = AlignTextModel.from_pretrained("kakaobrain/align-base")
+        >>> tokenizer = AutoTokenizer.from_pretrained("kakaobrain/align-base")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """The vision model from ALIGN without any head or projection on top.""",
+    ALIGN_START_DOCSTRING,
+)
+class AlignVisionModel(AlignPreTrainedModel):
+    config_class = AlignVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: AlignVisionConfig):
+        super().__init__(config)
+        self.config = config
+        self.embeddings = AlignVisionEmbeddings(config)
+        self.encoder = AlignVisionEncoder(config)
+
+        # Final pooling layer
+        if config.pooling_type == "mean":
+            self.pooler = nn.AvgPool2d(config.hidden_dim, ceil_mode=True)
+        elif config.pooling_type == "max":
+            self.pooler = nn.MaxPool2d(config.hidden_dim, ceil_mode=True)
+        else:
+            raise ValueError(f"config.pooling must be one of ['mean', 'max'] got {config.pooling}")
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.convolution
+
+    @add_start_docstrings_to_model_forward(ALIGN_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndNoAttention, config_class=AlignVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPoolingAndNoAttention]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AlignVisionModel
+
+        >>> model = AlignVisionModel.from_pretrained("kakaobrain/align-base")
+        >>> processor = AutoProcessor.from_pretrained("kakaobrain/align-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        embedding_output = self.embeddings(pixel_values)
+        encoder_outputs = self.encoder(
+            embedding_output,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        # Apply pooling
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = self.pooler(last_hidden_state)
+        # Reshape (batch_size, projection_dim, 1 , 1) -> (batch_size, projection_dim)
+        pooled_output = pooled_output.reshape(pooled_output.shape[:2])
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndNoAttention(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+        )
+
+
+@add_start_docstrings(ALIGN_START_DOCSTRING)
+class AlignModel(AlignPreTrainedModel):
+    config_class = AlignConfig
+
+    def __init__(self, config: AlignConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, AlignTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type AlignTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, AlignVisionConfig):
+            raise ValueError(
+                "config.vision_config is expected to be of type AlignVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+
+        self.text_model = AlignTextModel(text_config)
+        self.vision_model = AlignVisionModel(vision_config)
+
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim)
+        self.temperature = nn.Parameter(torch.ones([]) * self.config.temperature_init_value)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALIGN_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`AlignTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, AlignModel
+
+        >>> model = AlignModel.from_pretrained("kakaobrain/align-base")
+        >>> tokenizer = AutoTokenizer.from_pretrained("kakaobrain/align-base")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use ALIGN model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = text_outputs[0][:, 0, :]
+        text_features = self.text_projection(last_hidden_state)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(ALIGN_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`AlignVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AlignModel
+
+        >>> model = AlignModel.from_pretrained("kakaobrain/align-base")
+        >>> processor = AutoProcessor.from_pretrained("kakaobrain/align-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        # Use ALIGN model's config for some fields (if specified) instead of those of vision & text components.
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_features = vision_outputs[1]  # pooled_output
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(ALIGN_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=AlignOutput, config_class=AlignConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, AlignOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AlignModel
+
+        >>> model = AlignModel.from_pretrained("kakaobrain/align-base")
+        >>> processor = AutoProcessor.from_pretrained("kakaobrain/align-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use ALIGN model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        text_embeds = text_outputs[0][:, 0, :]
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) / self.temperature
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = align_loss(logits_per_text)
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return AlignOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/align/processing_align.py

@@ -0,0 +1,122 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image/Text processor class for ALIGN
+"""
+
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class AlignProcessor(ProcessorMixin):
+    r"""
+    Constructs an ALIGN processor which wraps [`EfficientNetImageProcessor`] and
+    [`BertTokenizer`]/[`BertTokenizerFast`] into a single processor that interits both the image processor and
+    tokenizer functionalities. See the [`~AlignProcessor.__call__`] and [`~OwlViTProcessor.decode`] for more
+    information.
+
+    Args:
+        image_processor ([`EfficientNetImageProcessor`]):
+            The image processor is a required input.
+        tokenizer ([`BertTokenizer`, `BertTokenizerFast`]):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "EfficientNetImageProcessor"
+    tokenizer_class = ("BertTokenizer", "BertTokenizerFast")
+
+    def __init__(self, image_processor, tokenizer):
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(self, text=None, images=None, padding="max_length", max_length=64, return_tensors=None, **kwargs):
+        """
+        Main method to prepare text(s) and image(s) to be fed as input to the model. This method forwards the `text`
+        and `kwargs` arguments to BertTokenizerFast's [`~BertTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwargs` arguments to
+        EfficientNetImageProcessor's [`~EfficientNetImageProcessor.__call__`] if `images` is not `None`. Please refer
+        to the doctsring of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `max_length`):
+                Activates and controls padding for tokenization of input text. Choose between [`True` or `'longest'`,
+                `'max_length'`, `False` or `'do_not_pad'`]
+            max_length (`int`, *optional*, defaults to `max_length`):
+                Maximum padding value to use to pad the input text during tokenization.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(
+                text, padding=padding, max_length=max_length, return_tensors=return_tensors, **kwargs
+            )
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors, **kwargs)
+
+        if text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/altclip/__init__.py

@@ -0,0 +1,71 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_altclip": [
+        "ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "AltCLIPConfig",
+        "AltCLIPTextConfig",
+        "AltCLIPVisionConfig",
+    ],
+    "processing_altclip": ["AltCLIPProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_altclip"] = [
+        "ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AltCLIPPreTrainedModel",
+        "AltCLIPModel",
+        "AltCLIPTextModel",
+        "AltCLIPVisionModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_altclip import (
+        ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        AltCLIPConfig,
+        AltCLIPTextConfig,
+        AltCLIPVisionConfig,
+    )
+    from .processing_altclip import AltCLIPProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_altclip import (
+            ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AltCLIPModel,
+            AltCLIPPreTrainedModel,
+            AltCLIPTextModel,
+            AltCLIPVisionModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/altclip/configuration_altclip.py

@@ -0,0 +1,405 @@
+# coding=utf-8
+# Copyright 2022 WenXiang ZhongzhiCheng LedellWu LiuGuang BoWenZhang and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" AltCLIP model configuration"""
+import copy
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "BAAI/AltCLIP": "https://huggingface.co/BAAI/AltCLIP/resolve/main/config.json",
+    # See all AltCLIP models at https://huggingface.co/models?filter=altclip
+}
+
+
+class AltCLIPTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AltCLIPTextModel`]. It is used to instantiate a
+    AltCLIP text model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the AltCLIP
+    [BAAI/AltCLIP](https://huggingface.co/BAAI/AltCLIP) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 250002):
+            Vocabulary size of the AltCLIP model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`AltCLIPTextModel`].
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 514):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`AltCLIPTextModel`]
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        project_dim (`int`, *optional*, defaults to 768):
+            The dimentions of the teacher model before the mapping layer.
+
+    Examples:
+
+    ```python
+    >>> from transformers import AltCLIPTextModel, AltCLIPTextConfig
+
+    >>> # Initializing a AltCLIPTextConfig with BAAI/AltCLIP style configuration
+    >>> configuration = AltCLIPTextConfig()
+
+    >>> # Initializing a AltCLIPTextModel (with random weights) from the BAAI/AltCLIP style configuration
+    >>> model = AltCLIPTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "altclip_text_model"
+
+    def __init__(
+        self,
+        vocab_size=250002,
+        hidden_size=1024,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        intermediate_size=4096,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=514,
+        type_vocab_size=1,
+        initializer_range=0.02,
+        initializer_factor=0.02,
+        layer_norm_eps=1e-05,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        position_embedding_type="absolute",
+        use_cache=True,
+        project_dim=768,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.project_dim = project_dim
+
+
+class AltCLIPVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AltCLIPModel`]. It is used to instantiate an
+    AltCLIP model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the AltCLIP
+    [BAAI/AltCLIP](https://huggingface.co/BAAI/AltCLIP) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float``, *optional*, defaults to 1):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import AltCLIPVisionConfig, AltCLIPVisionModel
+
+    >>> # Initializing a AltCLIPVisionConfig with BAAI/AltCLIP style configuration
+    >>> configuration = AltCLIPVisionConfig()
+
+    >>> # Initializing a AltCLIPVisionModel (with random weights) from the BAAI/AltCLIP style configuration
+    >>> model = AltCLIPVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "altclip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from AltCLIPConfig
+        if config_dict.get("model_type") == "altclip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class AltCLIPConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AltCLIPModel`]. It is used to instantiate an
+    AltCLIP model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the AltCLIP
+    [BAAI/AltCLIP](https://huggingface.co/BAAI/AltCLIP) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AltCLIPTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AltCLIPVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The inital value of the *logit_scale* paramter. Default is used as per the original CLIP implementation.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import AltCLIPConfig, AltCLIPModel
+
+    >>> # Initializing a AltCLIPConfig with BAAI/AltCLIP style configuration
+    >>> configuration = AltCLIPConfig()
+
+    >>> # Initializing a AltCLIPModel (with random weights) from the BAAI/AltCLIP style configuration
+    >>> model = AltCLIPModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a AltCLIPConfig from a AltCLIPTextConfig and a AltCLIPVisionConfig
+
+    >>> # Initializing a AltCLIPText and AltCLIPVision configuration
+    >>> config_text = AltCLIPTextConfig()
+    >>> config_vision = AltCLIPVisionConfig()
+
+    >>> config = AltCLIPConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "altclip"
+    is_composition = True
+
+    def __init__(
+        self, text_config=None, vision_config=None, projection_dim=768, logit_scale_init_value=2.6592, **kwargs
+    ):
+        # If `_config_dict` exist, we use them for the backward compatibility.
+        # We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot
+        # of confusion!).
+        text_config_dict = kwargs.pop("text_config_dict", None)
+        vision_config_dict = kwargs.pop("vision_config_dict", None)
+
+        super().__init__(**kwargs)
+
+        # Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in
+        # `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most
+        # cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.
+        if text_config_dict is not None:
+            if text_config is None:
+                text_config = {}
+
+            # This is the complete result when using `text_config_dict`.
+            _text_config_dict = AltCLIPTextConfig(**text_config_dict).to_dict()
+
+            # Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.
+            for key, value in _text_config_dict.items():
+                if key in text_config and value != text_config[key] and key not in ["transformers_version"]:
+                    # If specified in `text_config_dict`
+                    if key in text_config_dict:
+                        message = (
+                            f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "
+                            f'The value `text_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`text_config_dict` is provided which will be used to initialize `AltCLIPTextConfig`. The "
+                            f'value `text_config["{key}"]` will be overriden.'
+                        )
+                    logger.warning(message)
+
+            # Update all values in `text_config` with the ones in `_text_config_dict`.
+            text_config.update(_text_config_dict)
+
+        if vision_config_dict is not None:
+            if vision_config is None:
+                vision_config = {}
+
+            # This is the complete result when using `vision_config_dict`.
+            _vision_config_dict = AltCLIPVisionConfig(**vision_config_dict).to_dict()
+            # convert keys to string instead of integer
+            if "id2label" in _vision_config_dict:
+                _vision_config_dict["id2label"] = {
+                    str(key): value for key, value in _vision_config_dict["id2label"].items()
+                }
+
+            # Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.
+            for key, value in _vision_config_dict.items():
+                if key in vision_config and value != vision_config[key] and key not in ["transformers_version"]:
+                    # If specified in `vision_config_dict`
+                    if key in vision_config_dict:
+                        message = (
+                            f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "
+                            f'values. The value `vision_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`vision_config_dict` is provided which will be used to initialize `AltCLIPVisionConfig`. "
+                            f'The value `vision_config["{key}"]` will be overriden.'
+                        )
+                    logger.warning(message)
+
+            # Update all values in `vision_config` with the ones in `_vision_config_dict`.
+            vision_config.update(_vision_config_dict)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("`text_config` is `None`. Initializing the `AltCLIPTextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `AltCLIPVisionConfig` with default values.")
+
+        self.text_config = AltCLIPTextConfig(**text_config)
+        self.vision_config = AltCLIPVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: AltCLIPTextConfig, vision_config: AltCLIPVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`AltCLIPConfig`] (or a derived class) from altclip text model configuration and altclip vision
+        model configuration.
+
+        Returns:
+            [`AltCLIPConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["text_config"] = self.text_config.to_dict()
+        output["vision_config"] = self.vision_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/altclip/modeling_altclip.py

@@ -0,0 +1,1712 @@
+# coding=utf-8
+# Copyright 2022 The BAAI Teams Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch AltCLIP model."""
+import math
+from dataclasses import dataclass
+from typing import Any, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPooling,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    BaseModelOutputWithPoolingAndProjection,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import ModelOutput, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_altclip import AltCLIPConfig, AltCLIPTextConfig, AltCLIPVisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "BAAI/AltCLIP"
+_CONFIG_FOR_DOC = "AltCLIPConfig"
+
+ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "BAAI/AltCLIP",
+    # See all AltCLIP models at https://huggingface.co/models?filter=altclip
+]
+
+
+ALTCLIP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`CLIPConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+ALTCLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+ALTCLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+ALTCLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
+
+
+def clip_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->AltCLIP
+class AltCLIPOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image:(`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`AltCLIPTextModel`].
+        image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of
+            [`AltCLIPVisionModel`].
+        text_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`AltCLIPTextModel`].
+        vision_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`AltCLIPVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    image_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPooling = None
+    vision_model_output: BaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaEmbeddings with Roberta->AltRoberta
+class AltRobertaEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
+        )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx
+        )
+
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaSelfAttention with Roberta->AltRoberta
+class AltRobertaSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        use_cache = past_key_value is not None
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            query_length, key_length = query_layer.shape[2], key_layer.shape[2]
+            if use_cache:
+                position_ids_l = torch.tensor(key_length - 1, dtype=torch.long, device=hidden_states.device).view(
+                    -1, 1
+                )
+            else:
+                position_ids_l = torch.arange(query_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(key_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in AltRobertaModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaSelfOutput
+class AltRobertaSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaAttention with Roberta->AltRoberta
+class AltRobertaAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = AltRobertaSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.output = AltRobertaSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaIntermediate with Roberta->AltRoberta
+class AltRobertaIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaOutput
+class AltRobertaOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaLayer with Roberta->AltRoberta
+class AltRobertaLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = AltRobertaAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
+            self.crossattention = AltRobertaAttention(config, position_embedding_type="absolute")
+        self.intermediate = AltRobertaIntermediate(config)
+        self.output = AltRobertaOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
+                    " by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaEncoder with Roberta->AltRoberta
+class AltRobertaEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([AltRobertaLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaPooler
+class AltRobertaPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPAttention with CLIP->AltCLIP
+class AltCLIPAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scale
+        key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+        value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.view(*proj_shape)
+        value_states = value_states.view(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        # apply the causal_attention_mask first
+        if causal_attention_mask is not None:
+            if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {causal_attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if output_attentions:
+            # this operation is a bit akward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->AltCLIP
+class AltCLIPMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->AltCLIP
+class AltCLIPEncoderLayer(nn.Module):
+    def __init__(self, config: AltCLIPConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = AltCLIPAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = AltCLIPMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        causal_attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->AltCLIP
+class AltCLIPEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`AltCLIPEncoderLayer`].
+
+    Args:
+        config: AltCLIPConfig
+    """
+
+    def __init__(self, config: AltCLIPConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([AltCLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Causal mask for the text model. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(encoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPVisionEmbeddings with CLIP->AltCLIP
+class AltCLIPVisionEmbeddings(nn.Module):
+    def __init__(self, config: AltCLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            bias=False,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)))
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class AltCLIPPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = AltCLIPConfig
+    base_model_prefix = "altclip"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+        if isinstance(module, AltCLIPVisionEmbeddings):
+            factor = self.config.initializer_factor
+            nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
+            nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
+            nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
+        elif isinstance(module, AltCLIPAttention):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            out_proj_std = (module.embed_dim**-0.5) * factor
+            nn.init.normal_(module.q_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.k_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.v_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.out_proj.weight, std=out_proj_std)
+        elif isinstance(module, AltCLIPMLP):
+            factor = self.config.initializer_factor
+            in_proj_std = (
+                (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            )
+            fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
+            nn.init.normal_(module.fc1.weight, std=fc_std)
+            nn.init.normal_(module.fc2.weight, std=in_proj_std)
+        elif isinstance(module, AltCLIPModel):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+            module.text_projection._is_hf_initialized = True
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+            module.visual_projection._is_hf_initialized = True
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_factor)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_factor)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, AltCLIPEncoder):
+            module.gradient_checkpointing = value
+        if isinstance(module, AltRobertaEncoder):
+            module.gradient_checkpointing = value
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPVisionTransformer with CLIPVisionTransformer->AltCLIPVisionTransformer,CLIPVisionConfig->AltCLIPVisionConfig,CLIPVisionEmbeddings->AltCLIPVisionEmbeddings,CLIPEncoder->AltCLIPEncoder,CLIP_VISION_INPUTS_DOCSTRING->ALTCLIP_VISION_INPUTS_DOCSTRING
+class AltCLIPVisionTransformer(nn.Module):
+    def __init__(self, config: AltCLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = AltCLIPVisionEmbeddings(config)
+        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self.encoder = AltCLIPEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+    @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=AltCLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        hidden_states = self.embeddings(pixel_values)
+        hidden_states = self.pre_layrnorm(hidden_states)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class AltCLIPVisionModel(AltCLIPPreTrainedModel):
+    config_class = AltCLIPVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: AltCLIPVisionConfig):
+        super().__init__(config)
+        self.vision_model = AltCLIPVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=AltCLIPVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AltCLIPVisionModel
+
+        >>> model = AltCLIPVisionModel.from_pretrained("BAAI/AltCLIP")
+        >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        return self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class AltRobertaModel(AltCLIPPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    config_class = AltCLIPTextConfig
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->AltRoberta
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = AltRobertaEmbeddings(config)
+        self.encoder = AltRobertaEncoder(config)
+
+        self.pooler = AltRobertaPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class AltCLIPTextModel(AltCLIPPreTrainedModel):
+    config_class = AltCLIPTextConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.roberta = AltRobertaModel(config, add_pooling_layer=False)
+        self.transformation = nn.Linear(config.hidden_size, config.project_dim)
+        self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.roberta.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value: nn.Embedding) -> None:
+        self.roberta.embeddings.word_embeddings = value
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> nn.Embedding:
+        return super().resize_token_embeddings(new_num_tokens)
+
+    @add_start_docstrings_to_model_forward(ALTCLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndProjection, config_class=AltCLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ):
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoProcessor, AltCLIPTextModel
+
+        >>> model = AltCLIPTextModel.from_pretrained("BAAI/AltCLIP")
+        >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP")
+
+        >>> texts = ["it's a cat", "it's a dog"]
+
+        >>> inputs = processor(text=texts, padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled CLS states
+        ```"""
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.roberta(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        # last module outputs
+        sequence_output = outputs[0]
+
+        # project every module
+        sequence_output = self.pre_LN(sequence_output)
+
+        # pooler
+        projection_state = self.transformation(sequence_output)
+        pooler_output = projection_state[:, 0]
+
+        if not return_dict:
+            return (projection_state, pooler_output) + outputs[2:4]
+
+        return BaseModelOutputWithPoolingAndProjection(
+            last_hidden_state=projection_state,
+            pooler_output=pooler_output,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class AltCLIPModel(AltCLIPPreTrainedModel):
+    config_class = AltCLIPConfig
+
+    def __init__(self, config: AltCLIPConfig):
+        super().__init__(config)
+
+        if not isinstance(config.vision_config, AltCLIPVisionConfig):
+            raise ValueError(
+                "config.vision_config is expected to be of type AltCLIPVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+        if not isinstance(config.text_config, AltCLIPTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type AltCLIPTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.project_dim
+        self.vision_embed_dim = vision_config.hidden_size
+
+        self.text_model = AltCLIPTextModel(text_config)
+        self.vision_model = AltCLIPVisionTransformer(vision_config)
+
+        self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
+        self.logit_scale = nn.Parameter(torch.ones([]) * self.config.logit_scale_init_value)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(ALTCLIP_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        token_type_ids=None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`AltCLIPTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoProcessor, AltCLIPModel
+
+        >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP")
+        >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP")
+        >>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        pooled_output = text_outputs[1]
+        text_features = self.text_projection(pooled_output)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`AltCLIPVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AltCLIPModel
+
+        >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP")
+        >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP")
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> inputs = processor(images=image, return_tensors="pt")
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+        image_features = self.visual_projection(pooled_output)
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(ALTCLIP_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=AltCLIPOutput, config_class=AltCLIPConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        token_type_ids=None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, AltCLIPOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, AltCLIPModel
+
+        >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP")
+        >>> processor = AutoProcessor.from_pretrained("BAAI/AltCLIP")
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        image_embeds = self.visual_projection(image_embeds)
+
+        text_embeds = text_outputs[1]
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
+        logits_per_image = logits_per_text.T
+
+        loss = None
+        if return_loss:
+            loss = clip_loss(logits_per_text)
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return AltCLIPOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.create_position_ids_from_input_ids
+def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
+    return incremental_indices.long() + padding_idx

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/altclip/processing_altclip.py

@@ -0,0 +1,130 @@
+# coding=utf-8
+# Copyright 2022 WenXiang ZhongzhiCheng LedellWu LiuGuang BoWenZhang The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image/Text processor class for AltCLIP
+"""
+import warnings
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class AltCLIPProcessor(ProcessorMixin):
+    r"""
+    Constructs a AltCLIP processor which wraps a CLIP image processor and a XLM-Roberta tokenizer into a single
+    processor.
+
+    [`AltCLIPProcessor`] offers all the functionalities of [`CLIPImageProcessor`] and [`XLMRobertaTokenizerFast`]. See
+    the [`~AltCLIPProcessor.__call__`] and [`~AltCLIPProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`CLIPImageProcessor`]):
+            The image processor is a required input.
+        tokenizer ([`XLMRobertaTokenizerFast`]):
+            The tokenizer is a required input.
+    """
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "CLIPImageProcessor"
+    tokenizer_class = ("XLMRobertaTokenizer", "XLMRobertaTokenizerFast")
+
+    def __init__(self, image_processor=None, tokenizer=None, **kwargs):
+        if "feature_extractor" in kwargs:
+            warnings.warn(
+                "The `feature_extractor` argument is deprecated and will be removed in v5, use `image_processor`"
+                " instead.",
+                FutureWarning,
+            )
+            feature_extractor = kwargs.pop("feature_extractor")
+
+        image_processor = image_processor if image_processor is not None else feature_extractor
+        if image_processor is None:
+            raise ValueError("You need to specify an `image_processor`.")
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(self, text=None, images=None, return_tensors=None, **kwargs):
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to XLMRobertaTokenizerFast's [`~XLMRobertaTokenizerFast.__call__`] if `text` is not
+        `None` to encode the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(text, return_tensors=return_tensors, **kwargs)
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors, **kwargs)
+
+        if text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to XLMRobertaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`].
+        Please refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to XLMRobertaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/audio_spectrogram_transformer/__init__.py

@@ -0,0 +1,78 @@
+# Copyright 2021 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_speech_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_audio_spectrogram_transformer": [
+        "AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "ASTConfig",
+    ]
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_audio_spectrogram_transformer"] = [
+        "AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ASTForAudioClassification",
+        "ASTModel",
+        "ASTPreTrainedModel",
+    ]
+
+try:
+    if not is_speech_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["feature_extraction_audio_spectrogram_transformer"] = ["ASTFeatureExtractor"]
+
+if TYPE_CHECKING:
+    from .configuration_audio_spectrogram_transformer import (
+        AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        ASTConfig,
+    )
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_audio_spectrogram_transformer import (
+            AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ASTForAudioClassification,
+            ASTModel,
+            ASTPreTrainedModel,
+        )
+
+    try:
+        if not is_speech_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .feature_extraction_audio_spectrogram_transformer import ASTFeatureExtractor
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/audio_spectrogram_transformer/configuration_audio_spectrogram_transformer.py

@@ -0,0 +1,126 @@
+# coding=utf-8
+# Copyright 2022 Google AI and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Audio Spectogram Transformer (AST) model configuration"""
+
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "MIT/ast-finetuned-audioset-10-10-0.4593": (
+        "https://huggingface.co/MIT/ast-finetuned-audioset-10-10-0.4593/resolve/main/config.json"
+    ),
+}
+
+
+class ASTConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ASTModel`]. It is used to instantiate an AST
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the AST
+    [MIT/ast-finetuned-audioset-10-10-0.4593](https://huggingface.co/MIT/ast-finetuned-audioset-10-10-0.4593)
+    architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        patch_size (`int`, *optional*, defaults to `16`):
+            The size (resolution) of each patch.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        frequency_stride (`int`, *optional*, defaults to 10):
+            Frequency stride to use when patchifying the spectrograms.
+        time_stride (`int`, *optional*, defaults to 10):
+            Temporal stride to use when patchifying the spectrograms.
+        max_length (`int`, *optional*, defaults to 1024):
+            Temporal dimension of the spectrograms.
+        num_mel_bins (`int`, *optional*, defaults to 128):
+            Frequency dimension of the spectrograms (number of Mel-frequency bins).
+
+    Example:
+
+    ```python
+    >>> from transformers import ASTConfig, ASTModel
+
+    >>> # Initializing a AST MIT/ast-finetuned-audioset-10-10-0.4593 style configuration
+    >>> configuration = ASTConfig()
+
+    >>> # Initializing a model (with random weights) from the MIT/ast-finetuned-audioset-10-10-0.4593 style configuration
+    >>> model = ASTModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "audio-spectrogram-transformer"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        patch_size=16,
+        qkv_bias=True,
+        frequency_stride=10,
+        time_stride=10,
+        max_length=1024,
+        num_mel_bins=128,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.patch_size = patch_size
+        self.qkv_bias = qkv_bias
+        self.frequency_stride = frequency_stride
+        self.time_stride = time_stride
+        self.max_length = max_length
+        self.num_mel_bins = num_mel_bins

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/audio_spectrogram_transformer/convert_audio_spectrogram_transformer_original_to_pytorch.py

@@ -0,0 +1,279 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert Audio Spectrogram Transformer checkpoints from the original repository. URL: https://github.com/YuanGongND/ast"""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+import torchaudio
+from datasets import load_dataset
+from huggingface_hub import hf_hub_download
+
+from transformers import ASTConfig, ASTFeatureExtractor, ASTForAudioClassification
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_audio_spectrogram_transformer_config(model_name):
+    config = ASTConfig()
+
+    if "10-10" in model_name:
+        pass
+    elif "speech-commands" in model_name:
+        config.max_length = 128
+    elif "12-12" in model_name:
+        config.time_stride = 12
+        config.frequency_stride = 12
+    elif "14-14" in model_name:
+        config.time_stride = 14
+        config.frequency_stride = 14
+    elif "16-16" in model_name:
+        config.time_stride = 16
+        config.frequency_stride = 16
+    else:
+        raise ValueError("Model not supported")
+
+    repo_id = "huggingface/label-files"
+    if "speech-commands" in model_name:
+        config.num_labels = 35
+        filename = "speech-commands-v2-id2label.json"
+    else:
+        config.num_labels = 527
+        filename = "audioset-id2label.json"
+
+    id2label = json.load(open(hf_hub_download(repo_id, filename, repo_type="dataset"), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    config.id2label = id2label
+    config.label2id = {v: k for k, v in id2label.items()}
+
+    return config
+
+
+def rename_key(name):
+    if "module.v" in name:
+        name = name.replace("module.v", "audio_spectrogram_transformer")
+    if "cls_token" in name:
+        name = name.replace("cls_token", "embeddings.cls_token")
+    if "dist_token" in name:
+        name = name.replace("dist_token", "embeddings.distillation_token")
+    if "pos_embed" in name:
+        name = name.replace("pos_embed", "embeddings.position_embeddings")
+    if "patch_embed.proj" in name:
+        name = name.replace("patch_embed.proj", "embeddings.patch_embeddings.projection")
+    # transformer blocks
+    if "blocks" in name:
+        name = name.replace("blocks", "encoder.layer")
+    if "attn.proj" in name:
+        name = name.replace("attn.proj", "attention.output.dense")
+    if "attn" in name:
+        name = name.replace("attn", "attention.self")
+    if "norm1" in name:
+        name = name.replace("norm1", "layernorm_before")
+    if "norm2" in name:
+        name = name.replace("norm2", "layernorm_after")
+    if "mlp.fc1" in name:
+        name = name.replace("mlp.fc1", "intermediate.dense")
+    if "mlp.fc2" in name:
+        name = name.replace("mlp.fc2", "output.dense")
+    # final layernorm
+    if "audio_spectrogram_transformer.norm" in name:
+        name = name.replace("audio_spectrogram_transformer.norm", "audio_spectrogram_transformer.layernorm")
+    # classifier head
+    if "module.mlp_head.0" in name:
+        name = name.replace("module.mlp_head.0", "classifier.layernorm")
+    if "module.mlp_head.1" in name:
+        name = name.replace("module.mlp_head.1", "classifier.dense")
+
+    return name
+
+
+def convert_state_dict(orig_state_dict, config):
+    for key in orig_state_dict.copy().keys():
+        val = orig_state_dict.pop(key)
+
+        if "qkv" in key:
+            key_split = key.split(".")
+            layer_num = int(key_split[3])
+            dim = config.hidden_size
+            if "weight" in key:
+                orig_state_dict[
+                    f"audio_spectrogram_transformer.encoder.layer.{layer_num}.attention.attention.query.weight"
+                ] = val[:dim, :]
+                orig_state_dict[
+                    f"audio_spectrogram_transformer.encoder.layer.{layer_num}.attention.attention.key.weight"
+                ] = val[dim : dim * 2, :]
+                orig_state_dict[
+                    f"audio_spectrogram_transformer.encoder.layer.{layer_num}.attention.attention.value.weight"
+                ] = val[-dim:, :]
+            else:
+                orig_state_dict[
+                    f"audio_spectrogram_transformer.encoder.layer.{layer_num}.attention.attention.query.bias"
+                ] = val[:dim]
+                orig_state_dict[
+                    f"audio_spectrogram_transformer.encoder.layer.{layer_num}.attention.attention.key.bias"
+                ] = val[dim : dim * 2]
+                orig_state_dict[
+                    f"audio_spectrogram_transformer.encoder.layer.{layer_num}.attention.attention.value.bias"
+                ] = val[-dim:]
+        else:
+            orig_state_dict[rename_key(key)] = val
+
+    return orig_state_dict
+
+
+def remove_keys(state_dict):
+    ignore_keys = [
+        "module.v.head.weight",
+        "module.v.head.bias",
+        "module.v.head_dist.weight",
+        "module.v.head_dist.bias",
+    ]
+    for k in ignore_keys:
+        state_dict.pop(k, None)
+
+
+@torch.no_grad()
+def convert_audio_spectrogram_transformer_checkpoint(model_name, pytorch_dump_folder_path, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to our Audio Spectrogram Transformer structure.
+    """
+    config = get_audio_spectrogram_transformer_config(model_name)
+
+    model_name_to_url = {
+        "ast-finetuned-audioset-10-10-0.4593": (
+            "https://www.dropbox.com/s/ca0b1v2nlxzyeb4/audioset_10_10_0.4593.pth?dl=1"
+        ),
+        "ast-finetuned-audioset-10-10-0.450": (
+            "https://www.dropbox.com/s/1tv0hovue1bxupk/audioset_10_10_0.4495.pth?dl=1"
+        ),
+        "ast-finetuned-audioset-10-10-0.448": (
+            "https://www.dropbox.com/s/6u5sikl4b9wo4u5/audioset_10_10_0.4483.pth?dl=1"
+        ),
+        "ast-finetuned-audioset-10-10-0.448-v2": (
+            "https://www.dropbox.com/s/kt6i0v9fvfm1mbq/audioset_10_10_0.4475.pth?dl=1"
+        ),
+        "ast-finetuned-audioset-12-12-0.447": (
+            "https://www.dropbox.com/s/snfhx3tizr4nuc8/audioset_12_12_0.4467.pth?dl=1"
+        ),
+        "ast-finetuned-audioset-14-14-0.443": (
+            "https://www.dropbox.com/s/z18s6pemtnxm4k7/audioset_14_14_0.4431.pth?dl=1"
+        ),
+        "ast-finetuned-audioset-16-16-0.442": (
+            "https://www.dropbox.com/s/mdsa4t1xmcimia6/audioset_16_16_0.4422.pth?dl=1"
+        ),
+        "ast-finetuned-speech-commands-v2": (
+            "https://www.dropbox.com/s/q0tbqpwv44pquwy/speechcommands_10_10_0.9812.pth?dl=1"
+        ),
+    }
+
+    # load original state_dict
+    checkpoint_url = model_name_to_url[model_name]
+    state_dict = torch.hub.load_state_dict_from_url(checkpoint_url, map_location="cpu")
+    # remove some keys
+    remove_keys(state_dict)
+    # rename some keys
+    new_state_dict = convert_state_dict(state_dict, config)
+
+    # load 🤗 model
+    model = ASTForAudioClassification(config)
+    model.eval()
+
+    model.load_state_dict(new_state_dict)
+
+    # verify outputs on dummy input
+    # source: https://github.com/YuanGongND/ast/blob/79e873b8a54d0a3b330dd522584ff2b9926cd581/src/run.py#L62
+    mean = -4.2677393 if "speech-commands" not in model_name else -6.845978
+    std = 4.5689974 if "speech-commands" not in model_name else 5.5654526
+    max_length = 1024 if "speech-commands" not in model_name else 128
+    feature_extractor = ASTFeatureExtractor(mean=mean, std=std, max_length=max_length)
+
+    if "speech-commands" in model_name:
+        dataset = load_dataset("speech_commands", "v0.02", split="validation")
+        waveform = dataset[0]["audio"]["array"]
+    else:
+        filepath = hf_hub_download(
+            repo_id="nielsr/audio-spectogram-transformer-checkpoint",
+            filename="sample_audio.flac",
+            repo_type="dataset",
+        )
+
+        waveform, _ = torchaudio.load(filepath)
+        waveform = waveform.squeeze().numpy()
+
+    inputs = feature_extractor(waveform, sampling_rate=16000, return_tensors="pt")
+
+    # forward pass
+    outputs = model(**inputs)
+    logits = outputs.logits
+
+    if model_name == "ast-finetuned-audioset-10-10-0.4593":
+        expected_slice = torch.tensor([-0.8760, -7.0042, -8.6602])
+    elif model_name == "ast-finetuned-audioset-10-10-0.450":
+        expected_slice = torch.tensor([-1.1986, -7.0903, -8.2718])
+    elif model_name == "ast-finetuned-audioset-10-10-0.448":
+        expected_slice = torch.tensor([-2.6128, -8.0080, -9.4344])
+    elif model_name == "ast-finetuned-audioset-10-10-0.448-v2":
+        expected_slice = torch.tensor([-1.5080, -7.4534, -8.8917])
+    elif model_name == "ast-finetuned-audioset-12-12-0.447":
+        expected_slice = torch.tensor([-0.5050, -6.5833, -8.0843])
+    elif model_name == "ast-finetuned-audioset-14-14-0.443":
+        expected_slice = torch.tensor([-0.3826, -7.0336, -8.2413])
+    elif model_name == "ast-finetuned-audioset-16-16-0.442":
+        expected_slice = torch.tensor([-1.2113, -6.9101, -8.3470])
+    elif model_name == "ast-finetuned-speech-commands-v2":
+        expected_slice = torch.tensor([6.1589, -8.0566, -8.7984])
+    else:
+        raise ValueError("Unknown model name")
+    if not torch.allclose(logits[0, :3], expected_slice, atol=1e-4):
+        raise ValueError("Logits don't match")
+    print("Looks ok!")
+
+    if pytorch_dump_folder_path is not None:
+        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+        print(f"Saving model {model_name} to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        print(f"Saving feature extractor to {pytorch_dump_folder_path}")
+        feature_extractor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        print("Pushing model and feature extractor to the hub...")
+        model.push_to_hub(f"MIT/{model_name}")
+        feature_extractor.push_to_hub(f"MIT/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--model_name",
+        default="ast-finetuned-audioset-10-10-0.4593",
+        type=str,
+        help="Name of the Audio Spectrogram Transformer model you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
+    )
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+
+    args = parser.parse_args()
+    convert_audio_spectrogram_transformer_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.py

@@ -0,0 +1,197 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Feature extractor class for Audio Spectrogram Transformer.
+"""
+
+from typing import List, Optional, Union
+
+import numpy as np
+import torch
+import torchaudio.compliance.kaldi as ta_kaldi
+
+from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
+from ...feature_extraction_utils import BatchFeature
+from ...utils import TensorType, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class ASTFeatureExtractor(SequenceFeatureExtractor):
+    r"""
+    Constructs a Audio Spectrogram Transformer (AST) feature extractor.
+
+    This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
+    most of the main methods. Users should refer to this superclass for more information regarding those methods.
+
+    This class extracts mel-filter bank features from raw speech using TorchAudio, pads/truncates them to a fixed
+    length and normalizes them using a mean and standard deviation.
+
+    Args:
+        feature_size (`int`, *optional*, defaults to 1):
+            The feature dimension of the extracted features.
+        sampling_rate (`int`, *optional*, defaults to 16000):
+            The sampling rate at which the audio files should be digitalized expressed in hertz (Hz).
+        num_mel_bins (`int`, *optional*, defaults to 128):
+            Number of Mel-frequency bins.
+        max_length (`int`, *optional*, defaults to 1024):
+            Maximum length to which to pad/truncate the extracted features.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether or not to normalize the log-Mel features using `mean` and `std`.
+        mean (`float`, *optional*, defaults to -4.2677393):
+            The mean value used to normalize the log-Mel features. Uses the AudioSet mean by default.
+        std (`float`, *optional*, defaults to 4.5689974):
+            The standard deviation value used to normalize the log-Mel features. Uses the AudioSet standard deviation
+            by default.
+        return_attention_mask (`bool`, *optional*, defaults to `False`):
+            Whether or not [`~ASTFeatureExtractor.__call__`] should return `attention_mask`.
+    """
+
+    model_input_names = ["input_values", "attention_mask"]
+
+    def __init__(
+        self,
+        feature_size=1,
+        sampling_rate=16000,
+        num_mel_bins=128,
+        max_length=1024,
+        padding_value=0.0,
+        do_normalize=True,
+        mean=-4.2677393,
+        std=4.5689974,
+        return_attention_mask=False,
+        **kwargs,
+    ):
+        super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)
+        self.num_mel_bins = num_mel_bins
+        self.max_length = max_length
+        self.do_normalize = do_normalize
+        self.mean = mean
+        self.std = std
+        self.return_attention_mask = return_attention_mask
+
+    def _extract_fbank_features(
+        self,
+        waveform: np.ndarray,
+        max_length: int,
+    ) -> np.ndarray:
+        """
+        Get mel-filter bank features using TorchAudio. Note that TorchAudio requires 16-bit signed integers as inputs
+        and hence the waveform should not be normalized before feature extraction.
+        """
+        # waveform = waveform * (2**15)  # Kaldi compliance: 16-bit signed integers
+        waveform = torch.from_numpy(waveform).unsqueeze(0)
+        fbank = ta_kaldi.fbank(
+            waveform,
+            htk_compat=True,
+            sample_frequency=self.sampling_rate,
+            use_energy=False,
+            window_type="hanning",
+            num_mel_bins=self.num_mel_bins,
+            dither=0.0,
+            frame_shift=10,
+        )
+
+        n_frames = fbank.shape[0]
+        difference = max_length - n_frames
+
+        # pad or truncate, depending on difference
+        if difference > 0:
+            pad_module = torch.nn.ZeroPad2d((0, 0, 0, difference))
+            fbank = pad_module(fbank)
+        elif difference < 0:
+            fbank = fbank[0:max_length, :]
+
+        fbank = fbank.numpy()
+
+        return fbank
+
+    def normalize(self, input_values: np.ndarray) -> np.ndarray:
+        return (input_values - (self.mean)) / (self.std * 2)
+
+    def __call__(
+        self,
+        raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
+        sampling_rate: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Main method to featurize and prepare for the model one or several sequence(s).
+
+        Args:
+            raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
+                The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
+                values, a list of numpy arrays or a list of list of float values.
+            sampling_rate (`int`, *optional*):
+                The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
+                `sampling_rate` at the forward call to prevent silent errors.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors instead of list of python integers. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return Numpy `np.ndarray` objects.
+        """
+
+        if sampling_rate is not None:
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(
+                    f"The model corresponding to this feature extractor: {self} was trained using a sampling rate of"
+                    f" {self.sampling_rate}. Please make sure that the provided `raw_speech` input was sampled with"
+                    f" {self.sampling_rate} and not {sampling_rate}."
+                )
+        else:
+            logger.warning(
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
+                "Failing to do so can result in silent errors that might be hard to debug."
+            )
+
+        is_batched = bool(
+            isinstance(raw_speech, (list, tuple))
+            and (isinstance(raw_speech[0], np.ndarray) or isinstance(raw_speech[0], (tuple, list)))
+        )
+
+        if is_batched:
+            raw_speech = [np.asarray(speech, dtype=np.float32) for speech in raw_speech]
+        elif not is_batched and not isinstance(raw_speech, np.ndarray):
+            raw_speech = np.asarray(raw_speech, dtype=np.float32)
+        elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
+            raw_speech = raw_speech.astype(np.float32)
+
+        # always return batch
+        if not is_batched:
+            raw_speech = [raw_speech]
+
+        # extract fbank features and pad/truncate to max_length
+        features = [self._extract_fbank_features(waveform, max_length=self.max_length) for waveform in raw_speech]
+
+        # convert into BatchFeature
+        padded_inputs = BatchFeature({"input_values": features})
+
+        # make sure list is in array format
+        input_values = padded_inputs.get("input_values")
+        if isinstance(input_values[0], list):
+            padded_inputs["input_values"] = [np.asarray(feature, dtype=np.float32) for feature in input_values]
+
+        # normalization
+        if self.do_normalize:
+            padded_inputs["input_values"] = [self.normalize(feature) for feature in input_values]
+
+        if return_tensors is not None:
+            padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
+
+        return padded_inputs

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py

@@ -0,0 +1,627 @@
+# coding=utf-8
+# Copyright 2022 MIT and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch Audio Spectrogram Transformer (AST) model."""
+
+import math
+from typing import Dict, List, Optional, Set, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, SequenceClassifierOutput
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_audio_spectrogram_transformer import ASTConfig
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "ASTConfig"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "MIT/ast-finetuned-audioset-10-10-0.4593"
+_EXPECTED_OUTPUT_SHAPE = [1, 1214, 768]
+
+# Audio classification docstring
+_SEQ_CLASS_CHECKPOINT = "MIT/ast-finetuned-audioset-10-10-0.4593"
+_SEQ_CLASS_EXPECTED_OUTPUT = "'Speech'"
+_SEQ_CLASS_EXPECTED_LOSS = 0.17
+
+
+AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "MIT/ast-finetuned-audioset-10-10-0.4593",
+    # See all Audio Spectrogram Transformer models at https://huggingface.co/models?filter=ast
+]
+
+
+class ASTEmbeddings(nn.Module):
+    """
+    Construct the CLS token, position and patch embeddings.
+    """
+
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.distillation_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.patch_embeddings = ASTPatchEmbeddings(config)
+
+        frequency_out_dimension, time_out_dimension = self.get_shape(config)
+        num_patches = frequency_out_dimension * time_out_dimension
+        self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches + 2, config.hidden_size))
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def get_shape(self, config):
+        # see Karpathy's cs231n blog on how to calculate the output dimensions
+        # https://cs231n.github.io/convolutional-networks/#conv
+        frequency_out_dimension = (config.num_mel_bins - config.patch_size) // config.frequency_stride + 1
+        time_out_dimension = (config.max_length - config.patch_size) // config.time_stride + 1
+
+        return frequency_out_dimension, time_out_dimension
+
+    def forward(self, input_values: torch.Tensor) -> torch.Tensor:
+        batch_size = input_values.shape[0]
+        embeddings = self.patch_embeddings(input_values)
+
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
+        distillation_tokens = self.distillation_token.expand(batch_size, -1, -1)
+        embeddings = torch.cat((cls_tokens, distillation_tokens, embeddings), dim=1)
+        embeddings = embeddings + self.position_embeddings
+        embeddings = self.dropout(embeddings)
+
+        return embeddings
+
+
+class ASTPatchEmbeddings(nn.Module):
+    """
+    This class turns `input_values` into the initial `hidden_states` (patch embeddings) of shape `(batch_size,
+    seq_length, hidden_size)` to be consumed by a Transformer.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+
+        patch_size = config.patch_size
+        frequency_stride = config.frequency_stride
+        time_stride = config.time_stride
+
+        self.projection = nn.Conv2d(
+            1, config.hidden_size, kernel_size=(patch_size, patch_size), stride=(frequency_stride, time_stride)
+        )
+
+    def forward(self, input_values: torch.Tensor) -> torch.Tensor:
+        input_values = input_values.unsqueeze(1)
+        input_values = input_values.transpose(2, 3)
+        embeddings = self.projection(input_values).flatten(2).transpose(1, 2)
+        return embeddings
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfAttention with ViT->AST
+class ASTSelfAttention(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size {config.hidden_size,} is not a multiple of the number of attention "
+                f"heads {config.num_attention_heads}."
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self, hidden_states, head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->AST
+class ASTSelfOutput(nn.Module):
+    """
+    The residual connection is defined in ASTLayer instead of here (as is the case with other models), due to the
+    layernorm applied before each block.
+    """
+
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTAttention with ViT->AST
+class ASTAttention(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.attention = ASTSelfAttention(config)
+        self.output = ASTSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads: Set[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_outputs = self.attention(hidden_states, head_mask, output_attentions)
+
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTIntermediate with ViT->AST
+class ASTIntermediate(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTOutput with ViT->AST
+class ASTOutput(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        hidden_states = hidden_states + input_tensor
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTLayer with ViT->AST
+class ASTLayer(nn.Module):
+    """This corresponds to the Block class in the timm implementation."""
+
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = ASTAttention(config)
+        self.intermediate = ASTIntermediate(config)
+        self.output = ASTOutput(config)
+        self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_attention_outputs = self.attention(
+            self.layernorm_before(hidden_states),  # in AST, layernorm is applied before self-attention
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        # first residual connection
+        hidden_states = attention_output + hidden_states
+
+        # in AST, layernorm is also applied after self-attention
+        layer_output = self.layernorm_after(hidden_states)
+        layer_output = self.intermediate(layer_output)
+
+        # second residual connection is done here
+        layer_output = self.output(layer_output, hidden_states)
+
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTEncoder with ViT->AST
+class ASTEncoder(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([ASTLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[tuple, BaseModelOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    layer_head_mask,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ASTPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ASTConfig
+    base_model_prefix = "audio_spectrogram_transformer"
+    main_input_name = "input_values"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.deit.modeling_deit.DeiTPreTrainedModel._init_weights
+    def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> None:
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            # Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
+            # `trunc_normal_cpu` not implemented in `half` issues
+            module.weight.data = nn.init.trunc_normal_(
+                module.weight.data.to(torch.float32), mean=0.0, std=self.config.initializer_range
+            ).to(module.weight.dtype)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    # Copied from transformers.models.vit.modeling_vit.ViTPreTrainedModel._set_gradient_checkpointing with ViT->AST
+    def _set_gradient_checkpointing(self, module: ASTEncoder, value: bool = False) -> None:
+        if isinstance(module, ASTEncoder):
+            module.gradient_checkpointing = value
+
+
+AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
+    as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`ASTConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+AUDIO_SPECTROGRAM_TRANSFORMER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_values (`torch.FloatTensor` of shape `(batch_size, max_length, num_mel_bins)`):
+            Float values mel features extracted from the raw audio waveform. Raw audio waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `torch.FloatTensor`. See [`~ASTFeatureExtractor.__call__`]
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare AST Model transformer outputting raw hidden-states without any specific head on top.",
+    AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING,
+)
+class ASTModel(ASTPreTrainedModel):
+    def __init__(self, config: ASTConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = ASTEmbeddings(config)
+        self.encoder = ASTEncoder(config)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> ASTPatchEmbeddings:
+        return self.embeddings.patch_embeddings
+
+    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(AUDIO_SPECTROGRAM_TRANSFORMER_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+        modality="audio",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        input_values: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_values is None:
+            raise ValueError("You have to specify input_values")
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(input_values)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        sequence_output = self.layernorm(sequence_output)
+
+        pooled_output = (sequence_output[:, 0] + sequence_output[:, 1]) / 2
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+class ASTMLPHead(nn.Module):
+    def __init__(self, config: ASTConfig):
+        super().__init__()
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dense = nn.Linear(config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
+
+    def forward(self, hidden_state):
+        hidden_state = self.layernorm(hidden_state)
+        hidden_state = self.dense(hidden_state)
+        return hidden_state
+
+
+@add_start_docstrings(
+    """
+    Audio Spectrogram Transformer model with an audio classification head on top (a linear layer on top of the pooled
+    output) e.g. for datasets like AudioSet, Speech Commands v2.
+    """,
+    AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING,
+)
+class ASTForAudioClassification(ASTPreTrainedModel):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.audio_spectrogram_transformer = ASTModel(config)
+
+        # Classifier head
+        self.classifier = ASTMLPHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(AUDIO_SPECTROGRAM_TRANSFORMER_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_SEQ_CLASS_CHECKPOINT,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        modality="audio",
+        expected_output=_SEQ_CLASS_EXPECTED_OUTPUT,
+        expected_loss=_SEQ_CLASS_EXPECTED_LOSS,
+    )
+    def forward(
+        self,
+        input_values: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the audio classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.audio_spectrogram_transformer(
+            input_values,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/__init__.py

@@ -0,0 +1,359 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "auto_factory": ["get_values"],
+    "configuration_auto": ["ALL_PRETRAINED_CONFIG_ARCHIVE_MAP", "CONFIG_MAPPING", "MODEL_NAMES_MAPPING", "AutoConfig"],
+    "feature_extraction_auto": ["FEATURE_EXTRACTOR_MAPPING", "AutoFeatureExtractor"],
+    "image_processing_auto": ["IMAGE_PROCESSOR_MAPPING", "AutoImageProcessor"],
+    "processing_auto": ["PROCESSOR_MAPPING", "AutoProcessor"],
+    "tokenization_auto": ["TOKENIZER_MAPPING", "AutoTokenizer"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_auto"] = [
+        "MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_AUDIO_XVECTOR_MAPPING",
+        "MODEL_FOR_BACKBONE_MAPPING",
+        "MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING",
+        "MODEL_FOR_CAUSAL_LM_MAPPING",
+        "MODEL_FOR_CTC_MAPPING",
+        "MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING",
+        "MODEL_FOR_DEPTH_ESTIMATION_MAPPING",
+        "MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_IMAGE_SEGMENTATION_MAPPING",
+        "MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING",
+        "MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING",
+        "MODEL_FOR_MASKED_LM_MAPPING",
+        "MODEL_FOR_MASK_GENERATION_MAPPING",
+        "MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
+        "MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
+        "MODEL_FOR_OBJECT_DETECTION_MAPPING",
+        "MODEL_FOR_PRETRAINING_MAPPING",
+        "MODEL_FOR_QUESTION_ANSWERING_MAPPING",
+        "MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING",
+        "MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
+        "MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING",
+        "MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING",
+        "MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING",
+        "MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_VISION_2_SEQ_MAPPING",
+        "MODEL_FOR_VISUAL_QUESTION_ANSWERING_MAPPING",
+        "MODEL_MAPPING",
+        "MODEL_WITH_LM_HEAD_MAPPING",
+        "MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING",
+        "MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING",
+        "AutoModel",
+        "AutoBackbone",
+        "AutoModelForAudioClassification",
+        "AutoModelForAudioFrameClassification",
+        "AutoModelForAudioXVector",
+        "AutoModelForCausalLM",
+        "AutoModelForCTC",
+        "AutoModelForDepthEstimation",
+        "AutoModelForImageClassification",
+        "AutoModelForImageSegmentation",
+        "AutoModelForInstanceSegmentation",
+        "AutoModelForMaskGeneration",
+        "AutoModelForMaskedImageModeling",
+        "AutoModelForMaskedLM",
+        "AutoModelForMultipleChoice",
+        "AutoModelForNextSentencePrediction",
+        "AutoModelForObjectDetection",
+        "AutoModelForPreTraining",
+        "AutoModelForQuestionAnswering",
+        "AutoModelForSemanticSegmentation",
+        "AutoModelForSeq2SeqLM",
+        "AutoModelForSequenceClassification",
+        "AutoModelForSpeechSeq2Seq",
+        "AutoModelForTableQuestionAnswering",
+        "AutoModelForTokenClassification",
+        "AutoModelForUniversalSegmentation",
+        "AutoModelForVideoClassification",
+        "AutoModelForVision2Seq",
+        "AutoModelForVisualQuestionAnswering",
+        "AutoModelForDocumentQuestionAnswering",
+        "AutoModelWithLMHead",
+        "AutoModelForZeroShotImageClassification",
+        "AutoModelForZeroShotObjectDetection",
+    ]
+
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_tf_auto"] = [
+        "TF_MODEL_FOR_CAUSAL_LM_MAPPING",
+        "TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING",
+        "TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING",
+        "TF_MODEL_FOR_MASKED_LM_MAPPING",
+        "TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
+        "TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
+        "TF_MODEL_FOR_PRETRAINING_MAPPING",
+        "TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING",
+        "TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING",
+        "TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING",
+        "TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
+        "TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING",
+        "TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING",
+        "TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING",
+        "TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING",
+        "TF_MODEL_FOR_VISION_2_SEQ_MAPPING",
+        "TF_MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING",
+        "TF_MODEL_MAPPING",
+        "TF_MODEL_WITH_LM_HEAD_MAPPING",
+        "TFAutoModel",
+        "TFAutoModelForCausalLM",
+        "TFAutoModelForImageClassification",
+        "TFAutoModelForMaskedLM",
+        "TFAutoModelForMultipleChoice",
+        "TFAutoModelForNextSentencePrediction",
+        "TFAutoModelForPreTraining",
+        "TFAutoModelForDocumentQuestionAnswering",
+        "TFAutoModelForQuestionAnswering",
+        "TFAutoModelForSemanticSegmentation",
+        "TFAutoModelForSeq2SeqLM",
+        "TFAutoModelForSequenceClassification",
+        "TFAutoModelForSpeechSeq2Seq",
+        "TFAutoModelForTableQuestionAnswering",
+        "TFAutoModelForTokenClassification",
+        "TFAutoModelForVision2Seq",
+        "TFAutoModelForZeroShotImageClassification",
+        "TFAutoModelWithLMHead",
+    ]
+
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_flax_auto"] = [
+        "FLAX_MODEL_FOR_CAUSAL_LM_MAPPING",
+        "FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING",
+        "FLAX_MODEL_FOR_MASKED_LM_MAPPING",
+        "FLAX_MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
+        "FLAX_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
+        "FLAX_MODEL_FOR_PRETRAINING_MAPPING",
+        "FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING",
+        "FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
+        "FLAX_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING",
+        "FLAX_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING",
+        "FLAX_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING",
+        "FLAX_MODEL_FOR_VISION_2_SEQ_MAPPING",
+        "FLAX_MODEL_MAPPING",
+        "FlaxAutoModel",
+        "FlaxAutoModelForCausalLM",
+        "FlaxAutoModelForImageClassification",
+        "FlaxAutoModelForMaskedLM",
+        "FlaxAutoModelForMultipleChoice",
+        "FlaxAutoModelForNextSentencePrediction",
+        "FlaxAutoModelForPreTraining",
+        "FlaxAutoModelForQuestionAnswering",
+        "FlaxAutoModelForSeq2SeqLM",
+        "FlaxAutoModelForSequenceClassification",
+        "FlaxAutoModelForSpeechSeq2Seq",
+        "FlaxAutoModelForTokenClassification",
+        "FlaxAutoModelForVision2Seq",
+    ]
+
+
+if TYPE_CHECKING:
+    from .auto_factory import get_values
+    from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP, CONFIG_MAPPING, MODEL_NAMES_MAPPING, AutoConfig
+    from .feature_extraction_auto import FEATURE_EXTRACTOR_MAPPING, AutoFeatureExtractor
+    from .image_processing_auto import IMAGE_PROCESSOR_MAPPING, AutoImageProcessor
+    from .processing_auto import PROCESSOR_MAPPING, AutoProcessor
+    from .tokenization_auto import TOKENIZER_MAPPING, AutoTokenizer
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_auto import (
+            MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING,
+            MODEL_FOR_AUDIO_XVECTOR_MAPPING,
+            MODEL_FOR_BACKBONE_MAPPING,
+            MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
+            MODEL_FOR_CAUSAL_LM_MAPPING,
+            MODEL_FOR_CTC_MAPPING,
+            MODEL_FOR_DEPTH_ESTIMATION_MAPPING,
+            MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING,
+            MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+            MODEL_FOR_IMAGE_SEGMENTATION_MAPPING,
+            MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING,
+            MODEL_FOR_MASK_GENERATION_MAPPING,
+            MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING,
+            MODEL_FOR_MASKED_LM_MAPPING,
+            MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
+            MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
+            MODEL_FOR_OBJECT_DETECTION_MAPPING,
+            MODEL_FOR_PRETRAINING_MAPPING,
+            MODEL_FOR_QUESTION_ANSWERING_MAPPING,
+            MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING,
+            MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
+            MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
+            MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
+            MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING,
+            MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
+            MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING,
+            MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING,
+            MODEL_FOR_VISION_2_SEQ_MAPPING,
+            MODEL_FOR_VISUAL_QUESTION_ANSWERING_MAPPING,
+            MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING,
+            MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING,
+            MODEL_MAPPING,
+            MODEL_WITH_LM_HEAD_MAPPING,
+            AutoBackbone,
+            AutoModel,
+            AutoModelForAudioClassification,
+            AutoModelForAudioFrameClassification,
+            AutoModelForAudioXVector,
+            AutoModelForCausalLM,
+            AutoModelForCTC,
+            AutoModelForDepthEstimation,
+            AutoModelForDocumentQuestionAnswering,
+            AutoModelForImageClassification,
+            AutoModelForImageSegmentation,
+            AutoModelForInstanceSegmentation,
+            AutoModelForMaskedImageModeling,
+            AutoModelForMaskedLM,
+            AutoModelForMaskGeneration,
+            AutoModelForMultipleChoice,
+            AutoModelForNextSentencePrediction,
+            AutoModelForObjectDetection,
+            AutoModelForPreTraining,
+            AutoModelForQuestionAnswering,
+            AutoModelForSemanticSegmentation,
+            AutoModelForSeq2SeqLM,
+            AutoModelForSequenceClassification,
+            AutoModelForSpeechSeq2Seq,
+            AutoModelForTableQuestionAnswering,
+            AutoModelForTokenClassification,
+            AutoModelForUniversalSegmentation,
+            AutoModelForVideoClassification,
+            AutoModelForVision2Seq,
+            AutoModelForVisualQuestionAnswering,
+            AutoModelForZeroShotImageClassification,
+            AutoModelForZeroShotObjectDetection,
+            AutoModelWithLMHead,
+        )
+
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_tf_auto import (
+            TF_MODEL_FOR_CAUSAL_LM_MAPPING,
+            TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING,
+            TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+            TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING,
+            TF_MODEL_FOR_MASKED_LM_MAPPING,
+            TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
+            TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
+            TF_MODEL_FOR_PRETRAINING_MAPPING,
+            TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING,
+            TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING,
+            TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
+            TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
+            TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
+            TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING,
+            TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
+            TF_MODEL_FOR_VISION_2_SEQ_MAPPING,
+            TF_MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING,
+            TF_MODEL_MAPPING,
+            TF_MODEL_WITH_LM_HEAD_MAPPING,
+            TFAutoModel,
+            TFAutoModelForCausalLM,
+            TFAutoModelForDocumentQuestionAnswering,
+            TFAutoModelForImageClassification,
+            TFAutoModelForMaskedLM,
+            TFAutoModelForMultipleChoice,
+            TFAutoModelForNextSentencePrediction,
+            TFAutoModelForPreTraining,
+            TFAutoModelForQuestionAnswering,
+            TFAutoModelForSemanticSegmentation,
+            TFAutoModelForSeq2SeqLM,
+            TFAutoModelForSequenceClassification,
+            TFAutoModelForSpeechSeq2Seq,
+            TFAutoModelForTableQuestionAnswering,
+            TFAutoModelForTokenClassification,
+            TFAutoModelForVision2Seq,
+            TFAutoModelForZeroShotImageClassification,
+            TFAutoModelWithLMHead,
+        )
+
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_flax_auto import (
+            FLAX_MODEL_FOR_CAUSAL_LM_MAPPING,
+            FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+            FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+            FLAX_MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
+            FLAX_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
+            FLAX_MODEL_FOR_PRETRAINING_MAPPING,
+            FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING,
+            FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
+            FLAX_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
+            FLAX_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
+            FLAX_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
+            FLAX_MODEL_FOR_VISION_2_SEQ_MAPPING,
+            FLAX_MODEL_MAPPING,
+            FlaxAutoModel,
+            FlaxAutoModelForCausalLM,
+            FlaxAutoModelForImageClassification,
+            FlaxAutoModelForMaskedLM,
+            FlaxAutoModelForMultipleChoice,
+            FlaxAutoModelForNextSentencePrediction,
+            FlaxAutoModelForPreTraining,
+            FlaxAutoModelForQuestionAnswering,
+            FlaxAutoModelForSeq2SeqLM,
+            FlaxAutoModelForSequenceClassification,
+            FlaxAutoModelForSpeechSeq2Seq,
+            FlaxAutoModelForTokenClassification,
+            FlaxAutoModelForVision2Seq,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/auto_factory.py

@@ -0,0 +1,670 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Factory function to build auto-model classes."""
+import copy
+import importlib
+from collections import OrderedDict
+
+from ...configuration_utils import PretrainedConfig
+from ...dynamic_module_utils import get_class_from_dynamic_module
+from ...utils import copy_func, logging
+from .configuration_auto import AutoConfig, model_type_to_module_name, replace_list_option_in_docstrings
+
+
+logger = logging.get_logger(__name__)
+
+
+CLASS_DOCSTRING = """
+    This is a generic model class that will be instantiated as one of the model classes of the library when created
+    with the [`~BaseAutoModelClass.from_pretrained`] class method or the [`~BaseAutoModelClass.from_config`] class
+    method.
+
+    This class cannot be instantiated directly using `__init__()` (throws an error).
+"""
+
+FROM_CONFIG_DOCSTRING = """
+        Instantiates one of the model classes of the library from a configuration.
+
+        Note:
+            Loading a model from its configuration file does **not** load the model weights. It only affects the
+            model's configuration. Use [`~BaseAutoModelClass.from_pretrained`] to load the model weights.
+
+        Args:
+            config ([`PretrainedConfig`]):
+                The model class to instantiate is selected based on the configuration class:
+
+                List options
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoConfig, BaseAutoModelClass
+
+        >>> # Download configuration from huggingface.co and cache.
+        >>> config = AutoConfig.from_pretrained("checkpoint_placeholder")
+        >>> model = BaseAutoModelClass.from_config(config)
+        ```
+"""
+
+FROM_PRETRAINED_TORCH_DOCSTRING = """
+        Instantiate one of the model classes of the library from a pretrained model.
+
+        The model class to instantiate is selected based on the `model_type` property of the config object (either
+        passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's missing, by
+        falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        The model is set in evaluation mode by default using `model.eval()` (so for instance, dropout modules are
+        deactivated). To train the model, you should first set it back in training mode with `model.train()`
+
+        Args:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                Can be either:
+
+                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                      user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing model weights saved using
+                      [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
+                    - A path or url to a *tensorflow index checkpoint file* (e.g, `./tf_model/model.ckpt.index`). In
+                      this case, `from_tf` should be set to `True` and a configuration object should be provided as
+                      `config` argument. This loading path is slower than converting the TensorFlow checkpoint in a
+                      PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+            model_args (additional positional arguments, *optional*):
+                Will be passed along to the underlying model `__init__()` method.
+            config ([`PretrainedConfig`], *optional*):
+                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
+                be automatically loaded when:
+
+                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
+                      model).
+                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
+                      save directory.
+                    - The model is loaded by supplying a local directory as `pretrained_model_name_or_path` and a
+                      configuration JSON file named *config.json* is found in the directory.
+            state_dict (*Dict[str, torch.Tensor]*, *optional*):
+                A state dictionary to use instead of a state dictionary loaded from saved weights file.
+
+                This option can be used if you want to create a model from a pretrained configuration but load your own
+                weights. In this case though, you should check if using [`~PreTrainedModel.save_pretrained`] and
+                [`~PreTrainedModel.from_pretrained`] is not a simpler option.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            from_tf (`bool`, *optional*, defaults to `False`):
+                Load the model weights from a TensorFlow checkpoint save file (see docstring of
+                `pretrained_model_name_or_path` argument).
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            local_files_only(`bool`, *optional*, defaults to `False`):
+                Whether or not to only look at local files (e.g., not try downloading the model).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (additional keyword arguments, *optional*):
+                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
+                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
+                automatically loaded:
+
+                    - If a configuration is provided with `config`, `**kwargs` will be directly passed to the
+                      underlying model's `__init__` method (we assume all relevant updates to the configuration have
+                      already been done)
+                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
+                      initialization function ([`~PretrainedConfig.from_pretrained`]). Each key of `kwargs` that
+                      corresponds to a configuration attribute will be used to override said attribute with the
+                      supplied `kwargs` value. Remaining keys that do not correspond to any configuration attribute
+                      will be passed to the underlying model's `__init__` function.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoConfig, BaseAutoModelClass
+
+        >>> # Download model and configuration from huggingface.co and cache.
+        >>> model = BaseAutoModelClass.from_pretrained("checkpoint_placeholder")
+
+        >>> # Update configuration during loading
+        >>> model = BaseAutoModelClass.from_pretrained("checkpoint_placeholder", output_attentions=True)
+        >>> model.config.output_attentions
+        True
+
+        >>> # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+        >>> config = AutoConfig.from_pretrained("./tf_model/shortcut_placeholder_tf_model_config.json")
+        >>> model = BaseAutoModelClass.from_pretrained(
+        ...     "./tf_model/shortcut_placeholder_tf_checkpoint.ckpt.index", from_tf=True, config=config
+        ... )
+        ```
+"""
+
+FROM_PRETRAINED_TF_DOCSTRING = """
+        Instantiate one of the model classes of the library from a pretrained model.
+
+        The model class to instantiate is selected based on the `model_type` property of the config object (either
+        passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's missing, by
+        falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        Args:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                Can be either:
+
+                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                      user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing model weights saved using
+                      [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
+                    - A path or url to a *PyTorch state_dict save file* (e.g, `./pt_model/pytorch_model.bin`). In this
+                      case, `from_pt` should be set to `True` and a configuration object should be provided as `config`
+                      argument. This loading path is slower than converting the PyTorch model in a TensorFlow model
+                      using the provided conversion scripts and loading the TensorFlow model afterwards.
+            model_args (additional positional arguments, *optional*):
+                Will be passed along to the underlying model `__init__()` method.
+            config ([`PretrainedConfig`], *optional*):
+                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
+                be automatically loaded when:
+
+                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
+                      model).
+                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
+                      save directory.
+                    - The model is loaded by supplying a local directory as `pretrained_model_name_or_path` and a
+                      configuration JSON file named *config.json* is found in the directory.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            from_pt (`bool`, *optional*, defaults to `False`):
+                Load the model weights from a PyTorch checkpoint save file (see docstring of
+                `pretrained_model_name_or_path` argument).
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            local_files_only(`bool`, *optional*, defaults to `False`):
+                Whether or not to only look at local files (e.g., not try downloading the model).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (additional keyword arguments, *optional*):
+                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
+                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
+                automatically loaded:
+
+                    - If a configuration is provided with `config`, `**kwargs` will be directly passed to the
+                      underlying model's `__init__` method (we assume all relevant updates to the configuration have
+                      already been done)
+                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
+                      initialization function ([`~PretrainedConfig.from_pretrained`]). Each key of `kwargs` that
+                      corresponds to a configuration attribute will be used to override said attribute with the
+                      supplied `kwargs` value. Remaining keys that do not correspond to any configuration attribute
+                      will be passed to the underlying model's `__init__` function.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoConfig, BaseAutoModelClass
+
+        >>> # Download model and configuration from huggingface.co and cache.
+        >>> model = BaseAutoModelClass.from_pretrained("checkpoint_placeholder")
+
+        >>> # Update configuration during loading
+        >>> model = BaseAutoModelClass.from_pretrained("checkpoint_placeholder", output_attentions=True)
+        >>> model.config.output_attentions
+        True
+
+        >>> # Loading from a PyTorch checkpoint file instead of a TensorFlow model (slower)
+        >>> config = AutoConfig.from_pretrained("./pt_model/shortcut_placeholder_pt_model_config.json")
+        >>> model = BaseAutoModelClass.from_pretrained(
+        ...     "./pt_model/shortcut_placeholder_pytorch_model.bin", from_pt=True, config=config
+        ... )
+        ```
+"""
+
+FROM_PRETRAINED_FLAX_DOCSTRING = """
+        Instantiate one of the model classes of the library from a pretrained model.
+
+        The model class to instantiate is selected based on the `model_type` property of the config object (either
+        passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's missing, by
+        falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        Args:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                Can be either:
+
+                    - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                      user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing model weights saved using
+                      [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`.
+                    - A path or url to a *PyTorch state_dict save file* (e.g, `./pt_model/pytorch_model.bin`). In this
+                      case, `from_pt` should be set to `True` and a configuration object should be provided as `config`
+                      argument. This loading path is slower than converting the PyTorch model in a TensorFlow model
+                      using the provided conversion scripts and loading the TensorFlow model afterwards.
+            model_args (additional positional arguments, *optional*):
+                Will be passed along to the underlying model `__init__()` method.
+            config ([`PretrainedConfig`], *optional*):
+                Configuration for the model to use instead of an automatically loaded configuration. Configuration can
+                be automatically loaded when:
+
+                    - The model is a model provided by the library (loaded with the *model id* string of a pretrained
+                      model).
+                    - The model was saved using [`~PreTrainedModel.save_pretrained`] and is reloaded by supplying the
+                      save directory.
+                    - The model is loaded by supplying a local directory as `pretrained_model_name_or_path` and a
+                      configuration JSON file named *config.json* is found in the directory.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            from_pt (`bool`, *optional*, defaults to `False`):
+                Load the model weights from a PyTorch checkpoint save file (see docstring of
+                `pretrained_model_name_or_path` argument).
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            output_loading_info(`bool`, *optional*, defaults to `False`):
+                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            local_files_only(`bool`, *optional*, defaults to `False`):
+                Whether or not to only look at local files (e.g., not try downloading the model).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (additional keyword arguments, *optional*):
+                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
+                `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
+                automatically loaded:
+
+                    - If a configuration is provided with `config`, `**kwargs` will be directly passed to the
+                      underlying model's `__init__` method (we assume all relevant updates to the configuration have
+                      already been done)
+                    - If a configuration is not provided, `kwargs` will be first passed to the configuration class
+                      initialization function ([`~PretrainedConfig.from_pretrained`]). Each key of `kwargs` that
+                      corresponds to a configuration attribute will be used to override said attribute with the
+                      supplied `kwargs` value. Remaining keys that do not correspond to any configuration attribute
+                      will be passed to the underlying model's `__init__` function.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoConfig, BaseAutoModelClass
+
+        >>> # Download model and configuration from huggingface.co and cache.
+        >>> model = BaseAutoModelClass.from_pretrained("checkpoint_placeholder")
+
+        >>> # Update configuration during loading
+        >>> model = BaseAutoModelClass.from_pretrained("checkpoint_placeholder", output_attentions=True)
+        >>> model.config.output_attentions
+        True
+
+        >>> # Loading from a PyTorch checkpoint file instead of a TensorFlow model (slower)
+        >>> config = AutoConfig.from_pretrained("./pt_model/shortcut_placeholder_pt_model_config.json")
+        >>> model = BaseAutoModelClass.from_pretrained(
+        ...     "./pt_model/shortcut_placeholder_pytorch_model.bin", from_pt=True, config=config
+        ... )
+        ```
+"""
+
+
+def _get_model_class(config, model_mapping):
+    supported_models = model_mapping[type(config)]
+    if not isinstance(supported_models, (list, tuple)):
+        return supported_models
+
+    name_to_model = {model.__name__: model for model in supported_models}
+    architectures = getattr(config, "architectures", [])
+    for arch in architectures:
+        if arch in name_to_model:
+            return name_to_model[arch]
+        elif f"TF{arch}" in name_to_model:
+            return name_to_model[f"TF{arch}"]
+        elif f"Flax{arch}" in name_to_model:
+            return name_to_model[f"Flax{arch}"]
+
+    # If not architecture is set in the config or match the supported models, the first element of the tuple is the
+    # defaults.
+    return supported_models[0]
+
+
+class _BaseAutoModelClass:
+    # Base class for auto models.
+    _model_mapping = None
+
+    def __init__(self, *args, **kwargs):
+        raise EnvironmentError(
+            f"{self.__class__.__name__} is designed to be instantiated "
+            f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
+            f"`{self.__class__.__name__}.from_config(config)` methods."
+        )
+
+    @classmethod
+    def from_config(cls, config, **kwargs):
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+        if hasattr(config, "auto_map") and cls.__name__ in config.auto_map:
+            if not trust_remote_code:
+                raise ValueError(
+                    "Loading this model requires you to execute the modeling file in that repo "
+                    "on your local machine. Make sure you have read the code there to avoid malicious use, then set "
+                    "the option `trust_remote_code=True` to remove this error."
+                )
+            if kwargs.get("revision", None) is None:
+                logger.warning(
+                    "Explicitly passing a `revision` is encouraged when loading a model with custom code to ensure "
+                    "no malicious code has been contributed in a newer revision."
+                )
+            class_ref = config.auto_map[cls.__name__]
+            if "--" in class_ref:
+                repo_id, class_ref = class_ref.split("--")
+            else:
+                repo_id = config.name_or_path
+            model_class = get_class_from_dynamic_module(class_ref, repo_id, **kwargs)
+            return model_class._from_config(config, **kwargs)
+        elif type(config) in cls._model_mapping.keys():
+            model_class = _get_model_class(config, cls._model_mapping)
+            return model_class._from_config(config, **kwargs)
+
+        raise ValueError(
+            f"Unrecognized configuration class {config.__class__} for this kind of AutoModel: {cls.__name__}.\n"
+            f"Model type should be one of {', '.join(c.__name__ for c in cls._model_mapping.keys())}."
+        )
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        config = kwargs.pop("config", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+        kwargs["_from_auto"] = True
+        hub_kwargs_names = [
+            "cache_dir",
+            "force_download",
+            "local_files_only",
+            "proxies",
+            "resume_download",
+            "revision",
+            "subfolder",
+            "use_auth_token",
+        ]
+        hub_kwargs = {name: kwargs.pop(name) for name in hub_kwargs_names if name in kwargs}
+        if not isinstance(config, PretrainedConfig):
+            kwargs_copy = copy.deepcopy(kwargs)
+            # ensure not to pollute the config object with torch_dtype="auto" - since it's
+            # meaningless in the context of the config object - torch.dtype values are acceptable
+            if kwargs_copy.get("torch_dtype", None) == "auto":
+                _ = kwargs_copy.pop("torch_dtype")
+
+            config, kwargs = AutoConfig.from_pretrained(
+                pretrained_model_name_or_path,
+                return_unused_kwargs=True,
+                trust_remote_code=trust_remote_code,
+                **hub_kwargs,
+                **kwargs_copy,
+            )
+        if hasattr(config, "auto_map") and cls.__name__ in config.auto_map:
+            if not trust_remote_code:
+                raise ValueError(
+                    f"Loading {pretrained_model_name_or_path} requires you to execute the modeling file in that repo "
+                    "on your local machine. Make sure you have read the code there to avoid malicious use, then set "
+                    "the option `trust_remote_code=True` to remove this error."
+                )
+            class_ref = config.auto_map[cls.__name__]
+            model_class = get_class_from_dynamic_module(
+                class_ref, pretrained_model_name_or_path, **hub_kwargs, **kwargs
+            )
+            return model_class.from_pretrained(
+                pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
+            )
+        elif type(config) in cls._model_mapping.keys():
+            model_class = _get_model_class(config, cls._model_mapping)
+            return model_class.from_pretrained(
+                pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
+            )
+        raise ValueError(
+            f"Unrecognized configuration class {config.__class__} for this kind of AutoModel: {cls.__name__}.\n"
+            f"Model type should be one of {', '.join(c.__name__ for c in cls._model_mapping.keys())}."
+        )
+
+    @classmethod
+    def register(cls, config_class, model_class):
+        """
+        Register a new model for this class.
+
+        Args:
+            config_class ([`PretrainedConfig`]):
+                The configuration corresponding to the model to register.
+            model_class ([`PreTrainedModel`]):
+                The model to register.
+        """
+        if hasattr(model_class, "config_class") and model_class.config_class != config_class:
+            raise ValueError(
+                "The model class you are passing has a `config_class` attribute that is not consistent with the "
+                f"config class you passed (model has {model_class.config_class} and you passed {config_class}. Fix "
+                "one of those so they match!"
+            )
+        cls._model_mapping.register(config_class, model_class)
+
+
+def insert_head_doc(docstring, head_doc=""):
+    if len(head_doc) > 0:
+        return docstring.replace(
+            "one of the model classes of the library ",
+            f"one of the model classes of the library (with a {head_doc} head) ",
+        )
+    return docstring.replace(
+        "one of the model classes of the library ", "one of the base model classes of the library "
+    )
+
+
+def auto_class_update(cls, checkpoint_for_example="bert-base-cased", head_doc=""):
+    # Create a new class with the right name from the base class
+    model_mapping = cls._model_mapping
+    name = cls.__name__
+    class_docstring = insert_head_doc(CLASS_DOCSTRING, head_doc=head_doc)
+    cls.__doc__ = class_docstring.replace("BaseAutoModelClass", name)
+
+    # Now we need to copy and re-register `from_config` and `from_pretrained` as class methods otherwise we can't
+    # have a specific docstrings for them.
+    from_config = copy_func(_BaseAutoModelClass.from_config)
+    from_config_docstring = insert_head_doc(FROM_CONFIG_DOCSTRING, head_doc=head_doc)
+    from_config_docstring = from_config_docstring.replace("BaseAutoModelClass", name)
+    from_config_docstring = from_config_docstring.replace("checkpoint_placeholder", checkpoint_for_example)
+    from_config.__doc__ = from_config_docstring
+    from_config = replace_list_option_in_docstrings(model_mapping._model_mapping, use_model_types=False)(from_config)
+    cls.from_config = classmethod(from_config)
+
+    if name.startswith("TF"):
+        from_pretrained_docstring = FROM_PRETRAINED_TF_DOCSTRING
+    elif name.startswith("Flax"):
+        from_pretrained_docstring = FROM_PRETRAINED_FLAX_DOCSTRING
+    else:
+        from_pretrained_docstring = FROM_PRETRAINED_TORCH_DOCSTRING
+    from_pretrained = copy_func(_BaseAutoModelClass.from_pretrained)
+    from_pretrained_docstring = insert_head_doc(from_pretrained_docstring, head_doc=head_doc)
+    from_pretrained_docstring = from_pretrained_docstring.replace("BaseAutoModelClass", name)
+    from_pretrained_docstring = from_pretrained_docstring.replace("checkpoint_placeholder", checkpoint_for_example)
+    shortcut = checkpoint_for_example.split("/")[-1].split("-")[0]
+    from_pretrained_docstring = from_pretrained_docstring.replace("shortcut_placeholder", shortcut)
+    from_pretrained.__doc__ = from_pretrained_docstring
+    from_pretrained = replace_list_option_in_docstrings(model_mapping._model_mapping)(from_pretrained)
+    cls.from_pretrained = classmethod(from_pretrained)
+    return cls
+
+
+def get_values(model_mapping):
+    result = []
+    for model in model_mapping.values():
+        if isinstance(model, (list, tuple)):
+            result += list(model)
+        else:
+            result.append(model)
+
+    return result
+
+
+def getattribute_from_module(module, attr):
+    if attr is None:
+        return None
+    if isinstance(attr, tuple):
+        return tuple(getattribute_from_module(module, a) for a in attr)
+    if hasattr(module, attr):
+        return getattr(module, attr)
+    # Some of the mappings have entries model_type -> object of another model type. In that case we try to grab the
+    # object at the top level.
+    transformers_module = importlib.import_module("transformers")
+
+    if module != transformers_module:
+        try:
+            return getattribute_from_module(transformers_module, attr)
+        except ValueError:
+            raise ValueError(f"Could not find {attr} neither in {module} nor in {transformers_module}!")
+    else:
+        raise ValueError(f"Could not find {attr} in {transformers_module}!")
+
+
+class _LazyAutoMapping(OrderedDict):
+    """
+    " A mapping config to object (model or tokenizer for instance) that will load keys and values when it is accessed.
+
+    Args:
+        - config_mapping: The map model type to config class
+        - model_mapping: The map model type to model (or tokenizer) class
+    """
+
+    def __init__(self, config_mapping, model_mapping):
+        self._config_mapping = config_mapping
+        self._reverse_config_mapping = {v: k for k, v in config_mapping.items()}
+        self._model_mapping = model_mapping
+        self._extra_content = {}
+        self._modules = {}
+
+    def __len__(self):
+        common_keys = set(self._config_mapping.keys()).intersection(self._model_mapping.keys())
+        return len(common_keys) + len(self._extra_content)
+
+    def __getitem__(self, key):
+        if key in self._extra_content:
+            return self._extra_content[key]
+        model_type = self._reverse_config_mapping[key.__name__]
+        if model_type in self._model_mapping:
+            model_name = self._model_mapping[model_type]
+            return self._load_attr_from_module(model_type, model_name)
+
+        # Maybe there was several model types associated with this config.
+        model_types = [k for k, v in self._config_mapping.items() if v == key.__name__]
+        for mtype in model_types:
+            if mtype in self._model_mapping:
+                model_name = self._model_mapping[mtype]
+                return self._load_attr_from_module(mtype, model_name)
+        raise KeyError(key)
+
+    def _load_attr_from_module(self, model_type, attr):
+        module_name = model_type_to_module_name(model_type)
+        if module_name not in self._modules:
+            self._modules[module_name] = importlib.import_module(f".{module_name}", "transformers.models")
+        return getattribute_from_module(self._modules[module_name], attr)
+
+    def keys(self):
+        mapping_keys = [
+            self._load_attr_from_module(key, name)
+            for key, name in self._config_mapping.items()
+            if key in self._model_mapping.keys()
+        ]
+        return mapping_keys + list(self._extra_content.keys())
+
+    def get(self, key, default):
+        try:
+            return self.__getitem__(key)
+        except KeyError:
+            return default
+
+    def __bool__(self):
+        return bool(self.keys())
+
+    def values(self):
+        mapping_values = [
+            self._load_attr_from_module(key, name)
+            for key, name in self._model_mapping.items()
+            if key in self._config_mapping.keys()
+        ]
+        return mapping_values + list(self._extra_content.values())
+
+    def items(self):
+        mapping_items = [
+            (
+                self._load_attr_from_module(key, self._config_mapping[key]),
+                self._load_attr_from_module(key, self._model_mapping[key]),
+            )
+            for key in self._model_mapping.keys()
+            if key in self._config_mapping.keys()
+        ]
+        return mapping_items + list(self._extra_content.items())
+
+    def __iter__(self):
+        return iter(self.keys())
+
+    def __contains__(self, item):
+        if item in self._extra_content:
+            return True
+        if not hasattr(item, "__name__") or item.__name__ not in self._reverse_config_mapping:
+            return False
+        model_type = self._reverse_config_mapping[item.__name__]
+        return model_type in self._model_mapping
+
+    def register(self, key, value):
+        """
+        Register a new model in this mapping.
+        """
+        if hasattr(key, "__name__") and key.__name__ in self._reverse_config_mapping:
+            model_type = self._reverse_config_mapping[key.__name__]
+            if model_type in self._model_mapping.keys():
+                raise ValueError(f"'{key}' is already used by a Transformers model.")
+
+        self._extra_content[key] = value

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/configuration_auto.py

@@ -0,0 +1,970 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Config class."""
+import importlib
+import re
+import warnings
+from collections import OrderedDict
+from typing import List, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...dynamic_module_utils import get_class_from_dynamic_module
+from ...utils import CONFIG_NAME, logging
+
+
+logger = logging.get_logger(__name__)
+
+CONFIG_MAPPING_NAMES = OrderedDict(
+    [
+        # Add configs here
+        ("albert", "AlbertConfig"),
+        ("align", "AlignConfig"),
+        ("altclip", "AltCLIPConfig"),
+        ("audio-spectrogram-transformer", "ASTConfig"),
+        ("bart", "BartConfig"),
+        ("beit", "BeitConfig"),
+        ("bert", "BertConfig"),
+        ("bert-generation", "BertGenerationConfig"),
+        ("big_bird", "BigBirdConfig"),
+        ("bigbird_pegasus", "BigBirdPegasusConfig"),
+        ("biogpt", "BioGptConfig"),
+        ("bit", "BitConfig"),
+        ("blenderbot", "BlenderbotConfig"),
+        ("blenderbot-small", "BlenderbotSmallConfig"),
+        ("blip", "BlipConfig"),
+        ("blip-2", "Blip2Config"),
+        ("bloom", "BloomConfig"),
+        ("bridgetower", "BridgeTowerConfig"),
+        ("camembert", "CamembertConfig"),
+        ("canine", "CanineConfig"),
+        ("chinese_clip", "ChineseCLIPConfig"),
+        ("clap", "ClapConfig"),
+        ("clip", "CLIPConfig"),
+        ("clipseg", "CLIPSegConfig"),
+        ("codegen", "CodeGenConfig"),
+        ("conditional_detr", "ConditionalDetrConfig"),
+        ("convbert", "ConvBertConfig"),
+        ("convnext", "ConvNextConfig"),
+        ("convnextv2", "ConvNextV2Config"),
+        ("cpmant", "CpmAntConfig"),
+        ("ctrl", "CTRLConfig"),
+        ("cvt", "CvtConfig"),
+        ("data2vec-audio", "Data2VecAudioConfig"),
+        ("data2vec-text", "Data2VecTextConfig"),
+        ("data2vec-vision", "Data2VecVisionConfig"),
+        ("deberta", "DebertaConfig"),
+        ("deberta-v2", "DebertaV2Config"),
+        ("decision_transformer", "DecisionTransformerConfig"),
+        ("deformable_detr", "DeformableDetrConfig"),
+        ("deit", "DeiTConfig"),
+        ("deta", "DetaConfig"),
+        ("detr", "DetrConfig"),
+        ("dinat", "DinatConfig"),
+        ("distilbert", "DistilBertConfig"),
+        ("donut-swin", "DonutSwinConfig"),
+        ("dpr", "DPRConfig"),
+        ("dpt", "DPTConfig"),
+        ("efficientformer", "EfficientFormerConfig"),
+        ("efficientnet", "EfficientNetConfig"),
+        ("electra", "ElectraConfig"),
+        ("encoder-decoder", "EncoderDecoderConfig"),
+        ("ernie", "ErnieConfig"),
+        ("ernie_m", "ErnieMConfig"),
+        ("esm", "EsmConfig"),
+        ("flaubert", "FlaubertConfig"),
+        ("flava", "FlavaConfig"),
+        ("fnet", "FNetConfig"),
+        ("focalnet", "FocalNetConfig"),
+        ("fsmt", "FSMTConfig"),
+        ("funnel", "FunnelConfig"),
+        ("git", "GitConfig"),
+        ("glpn", "GLPNConfig"),
+        ("gpt-sw3", "GPT2Config"),
+        ("gpt2", "GPT2Config"),
+        ("gpt_bigcode", "GPTBigCodeConfig"),
+        ("gpt_neo", "GPTNeoConfig"),
+        ("gpt_neox", "GPTNeoXConfig"),
+        ("gpt_neox_japanese", "GPTNeoXJapaneseConfig"),
+        ("gptj", "GPTJConfig"),
+        ("gptsan-japanese", "GPTSanJapaneseConfig"),
+        ("graphormer", "GraphormerConfig"),
+        ("groupvit", "GroupViTConfig"),
+        ("hubert", "HubertConfig"),
+        ("ibert", "IBertConfig"),
+        ("imagegpt", "ImageGPTConfig"),
+        ("informer", "InformerConfig"),
+        ("jukebox", "JukeboxConfig"),
+        ("layoutlm", "LayoutLMConfig"),
+        ("layoutlmv2", "LayoutLMv2Config"),
+        ("layoutlmv3", "LayoutLMv3Config"),
+        ("led", "LEDConfig"),
+        ("levit", "LevitConfig"),
+        ("lilt", "LiltConfig"),
+        ("llama", "LlamaConfig"),
+        ("longformer", "LongformerConfig"),
+        ("longt5", "LongT5Config"),
+        ("luke", "LukeConfig"),
+        ("lxmert", "LxmertConfig"),
+        ("m2m_100", "M2M100Config"),
+        ("marian", "MarianConfig"),
+        ("markuplm", "MarkupLMConfig"),
+        ("mask2former", "Mask2FormerConfig"),
+        ("maskformer", "MaskFormerConfig"),
+        ("maskformer-swin", "MaskFormerSwinConfig"),
+        ("mbart", "MBartConfig"),
+        ("mctct", "MCTCTConfig"),
+        ("mega", "MegaConfig"),
+        ("megatron-bert", "MegatronBertConfig"),
+        ("mgp-str", "MgpstrConfig"),
+        ("mobilebert", "MobileBertConfig"),
+        ("mobilenet_v1", "MobileNetV1Config"),
+        ("mobilenet_v2", "MobileNetV2Config"),
+        ("mobilevit", "MobileViTConfig"),
+        ("mpnet", "MPNetConfig"),
+        ("mt5", "MT5Config"),
+        ("mvp", "MvpConfig"),
+        ("nat", "NatConfig"),
+        ("nezha", "NezhaConfig"),
+        ("nllb-moe", "NllbMoeConfig"),
+        ("nystromformer", "NystromformerConfig"),
+        ("oneformer", "OneFormerConfig"),
+        ("open-llama", "OpenLlamaConfig"),
+        ("openai-gpt", "OpenAIGPTConfig"),
+        ("opt", "OPTConfig"),
+        ("owlvit", "OwlViTConfig"),
+        ("pegasus", "PegasusConfig"),
+        ("pegasus_x", "PegasusXConfig"),
+        ("perceiver", "PerceiverConfig"),
+        ("pix2struct", "Pix2StructConfig"),
+        ("plbart", "PLBartConfig"),
+        ("poolformer", "PoolFormerConfig"),
+        ("prophetnet", "ProphetNetConfig"),
+        ("qdqbert", "QDQBertConfig"),
+        ("rag", "RagConfig"),
+        ("realm", "RealmConfig"),
+        ("reformer", "ReformerConfig"),
+        ("regnet", "RegNetConfig"),
+        ("rembert", "RemBertConfig"),
+        ("resnet", "ResNetConfig"),
+        ("retribert", "RetriBertConfig"),
+        ("roberta", "RobertaConfig"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormConfig"),
+        ("roc_bert", "RoCBertConfig"),
+        ("roformer", "RoFormerConfig"),
+        ("rwkv", "RwkvConfig"),
+        ("sam", "SamConfig"),
+        ("segformer", "SegformerConfig"),
+        ("sew", "SEWConfig"),
+        ("sew-d", "SEWDConfig"),
+        ("speech-encoder-decoder", "SpeechEncoderDecoderConfig"),
+        ("speech_to_text", "Speech2TextConfig"),
+        ("speech_to_text_2", "Speech2Text2Config"),
+        ("speecht5", "SpeechT5Config"),
+        ("splinter", "SplinterConfig"),
+        ("squeezebert", "SqueezeBertConfig"),
+        ("swin", "SwinConfig"),
+        ("swin2sr", "Swin2SRConfig"),
+        ("swinv2", "Swinv2Config"),
+        ("switch_transformers", "SwitchTransformersConfig"),
+        ("t5", "T5Config"),
+        ("table-transformer", "TableTransformerConfig"),
+        ("tapas", "TapasConfig"),
+        ("time_series_transformer", "TimeSeriesTransformerConfig"),
+        ("timesformer", "TimesformerConfig"),
+        ("trajectory_transformer", "TrajectoryTransformerConfig"),
+        ("transfo-xl", "TransfoXLConfig"),
+        ("trocr", "TrOCRConfig"),
+        ("tvlt", "TvltConfig"),
+        ("unispeech", "UniSpeechConfig"),
+        ("unispeech-sat", "UniSpeechSatConfig"),
+        ("upernet", "UperNetConfig"),
+        ("van", "VanConfig"),
+        ("videomae", "VideoMAEConfig"),
+        ("vilt", "ViltConfig"),
+        ("vision-encoder-decoder", "VisionEncoderDecoderConfig"),
+        ("vision-text-dual-encoder", "VisionTextDualEncoderConfig"),
+        ("visual_bert", "VisualBertConfig"),
+        ("vit", "ViTConfig"),
+        ("vit_hybrid", "ViTHybridConfig"),
+        ("vit_mae", "ViTMAEConfig"),
+        ("vit_msn", "ViTMSNConfig"),
+        ("wav2vec2", "Wav2Vec2Config"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerConfig"),
+        ("wavlm", "WavLMConfig"),
+        ("whisper", "WhisperConfig"),
+        ("xclip", "XCLIPConfig"),
+        ("xglm", "XGLMConfig"),
+        ("xlm", "XLMConfig"),
+        ("xlm-prophetnet", "XLMProphetNetConfig"),
+        ("xlm-roberta", "XLMRobertaConfig"),
+        ("xlm-roberta-xl", "XLMRobertaXLConfig"),
+        ("xlnet", "XLNetConfig"),
+        ("xmod", "XmodConfig"),
+        ("yolos", "YolosConfig"),
+        ("yoso", "YosoConfig"),
+    ]
+)
+
+CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
+    [
+        # Add archive maps here)
+        ("albert", "ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("align", "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("altclip", "ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("audio-spectrogram-transformer", "AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bart", "BART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("beit", "BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bert", "BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("big_bird", "BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bigbird_pegasus", "BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("biogpt", "BIOGPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bit", "BIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blenderbot", "BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blenderbot-small", "BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blip", "BLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blip-2", "BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bloom", "BLOOM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bridgetower", "BRIDGETOWER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("camembert", "CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("canine", "CANINE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("chinese_clip", "CHINESE_CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("clap", "CLAP_PRETRAINED_MODEL_ARCHIVE_LIST"),
+        ("clip", "CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("clipseg", "CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("codegen", "CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("conditional_detr", "CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("convbert", "CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("convnext", "CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("convnextv2", "CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("cpmant", "CPMANT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("ctrl", "CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("cvt", "CVT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("data2vec-audio", "DATA2VEC_AUDIO_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("data2vec-text", "DATA2VEC_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("data2vec-vision", "DATA2VEC_VISION_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deberta", "DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deberta-v2", "DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deformable_detr", "DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deit", "DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deta", "DETA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("detr", "DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("dinat", "DINAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("distilbert", "DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("donut-swin", "DONUT_SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("dpr", "DPR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("dpt", "DPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("efficientformer", "EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("efficientnet", "EFFICIENTNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("electra", "ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("ernie", "ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("ernie_m", "ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("esm", "ESM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("flaubert", "FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("flava", "FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("fnet", "FNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("focalnet", "FOCALNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("fsmt", "FSMT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("funnel", "FUNNEL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("git", "GIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("glpn", "GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt2", "GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt_bigcode", "GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt_neo", "GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt_neox", "GPT_NEOX_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gpt_neox_japanese", "GPT_NEOX_JAPANESE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gptj", "GPTJ_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("gptsan-japanese", "GPTSAN_JAPANESE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("graphormer", "GRAPHORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("groupvit", "GROUPVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("hubert", "HUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("ibert", "IBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("imagegpt", "IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("informer", "INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("jukebox", "JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("layoutlm", "LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("layoutlmv2", "LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("layoutlmv3", "LAYOUTLMV3_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("led", "LED_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("levit", "LEVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("lilt", "LILT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("llama", "LLAMA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("longformer", "LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("longt5", "LONGT5_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("luke", "LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("lxmert", "LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("m2m_100", "M2M_100_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("markuplm", "MARKUPLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mask2former", "MASK2FORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("maskformer", "MASKFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mbart", "MBART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mctct", "MCTCT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mega", "MEGA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("megatron-bert", "MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mgp-str", "MGP_STR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mobilenet_v1", "MOBILENET_V1_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mobilenet_v2", "MOBILENET_V2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mobilevit", "MOBILEVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mpnet", "MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("mvp", "MVP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("nat", "NAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("nezha", "NEZHA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("nllb-moe", "NLLB_MOE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("nystromformer", "NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("oneformer", "ONEFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("open-llama", "OPEN_LLAMA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("openai-gpt", "OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("opt", "OPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("owlvit", "OWLVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("pegasus", "PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("pegasus_x", "PEGASUS_X_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("perceiver", "PERCEIVER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("pix2struct", "PIX2STRUCT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("plbart", "PLBART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("poolformer", "POOLFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("prophetnet", "PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("qdqbert", "QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("realm", "REALM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("regnet", "REGNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("rembert", "REMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("resnet", "RESNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("retribert", "RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("roberta", "ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("roberta-prelayernorm", "ROBERTA_PRELAYERNORM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("roc_bert", "ROC_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("roformer", "ROFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("rwkv", "RWKV_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("sam", "SAM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("segformer", "SEGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("sew", "SEW_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("sew-d", "SEW_D_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("speech_to_text", "SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("speech_to_text_2", "SPEECH_TO_TEXT_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("speecht5", "SPEECHT5_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("splinter", "SPLINTER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("squeezebert", "SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("swin", "SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("swin2sr", "SWIN2SR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("swinv2", "SWINV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("switch_transformers", "SWITCH_TRANSFORMERS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("t5", "T5_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("table-transformer", "TABLE_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("tapas", "TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("time_series_transformer", "TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("timesformer", "TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("transfo-xl", "TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("tvlt", "TVLT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("unispeech", "UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("unispeech-sat", "UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("van", "VAN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("videomae", "VIDEOMAE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vilt", "VILT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("visual_bert", "VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vit", "VIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vit_hybrid", "VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vit_mae", "VIT_MAE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vit_msn", "VIT_MSN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("wav2vec2", "WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("wav2vec2-conformer", "WAV2VEC2_CONFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("whisper", "WHISPER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xclip", "XCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xglm", "XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlm", "XLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlm-prophetnet", "XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlm-roberta", "XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xlnet", "XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xmod", "XMOD_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("yolos", "YOLOS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("yoso", "YOSO_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+    ]
+)
+
+MODEL_NAMES_MAPPING = OrderedDict(
+    [
+        # Add full (and cased) model names here
+        ("albert", "ALBERT"),
+        ("align", "ALIGN"),
+        ("altclip", "AltCLIP"),
+        ("audio-spectrogram-transformer", "Audio Spectrogram Transformer"),
+        ("bart", "BART"),
+        ("barthez", "BARThez"),
+        ("bartpho", "BARTpho"),
+        ("beit", "BEiT"),
+        ("bert", "BERT"),
+        ("bert-generation", "Bert Generation"),
+        ("bert-japanese", "BertJapanese"),
+        ("bertweet", "BERTweet"),
+        ("big_bird", "BigBird"),
+        ("bigbird_pegasus", "BigBird-Pegasus"),
+        ("biogpt", "BioGpt"),
+        ("bit", "BiT"),
+        ("blenderbot", "Blenderbot"),
+        ("blenderbot-small", "BlenderbotSmall"),
+        ("blip", "BLIP"),
+        ("blip-2", "BLIP-2"),
+        ("bloom", "BLOOM"),
+        ("bort", "BORT"),
+        ("bridgetower", "BridgeTower"),
+        ("byt5", "ByT5"),
+        ("camembert", "CamemBERT"),
+        ("canine", "CANINE"),
+        ("chinese_clip", "Chinese-CLIP"),
+        ("clap", "CLAP"),
+        ("clip", "CLIP"),
+        ("clipseg", "CLIPSeg"),
+        ("codegen", "CodeGen"),
+        ("conditional_detr", "Conditional DETR"),
+        ("convbert", "ConvBERT"),
+        ("convnext", "ConvNeXT"),
+        ("convnextv2", "ConvNeXTV2"),
+        ("cpm", "CPM"),
+        ("cpmant", "CPM-Ant"),
+        ("ctrl", "CTRL"),
+        ("cvt", "CvT"),
+        ("data2vec-audio", "Data2VecAudio"),
+        ("data2vec-text", "Data2VecText"),
+        ("data2vec-vision", "Data2VecVision"),
+        ("deberta", "DeBERTa"),
+        ("deberta-v2", "DeBERTa-v2"),
+        ("decision_transformer", "Decision Transformer"),
+        ("deformable_detr", "Deformable DETR"),
+        ("deit", "DeiT"),
+        ("deplot", "DePlot"),
+        ("deta", "DETA"),
+        ("detr", "DETR"),
+        ("dialogpt", "DialoGPT"),
+        ("dinat", "DiNAT"),
+        ("distilbert", "DistilBERT"),
+        ("dit", "DiT"),
+        ("donut-swin", "DonutSwin"),
+        ("dpr", "DPR"),
+        ("dpt", "DPT"),
+        ("efficientformer", "EfficientFormer"),
+        ("efficientnet", "EfficientNet"),
+        ("electra", "ELECTRA"),
+        ("encoder-decoder", "Encoder decoder"),
+        ("ernie", "ERNIE"),
+        ("ernie_m", "ErnieM"),
+        ("esm", "ESM"),
+        ("flan-t5", "FLAN-T5"),
+        ("flan-ul2", "FLAN-UL2"),
+        ("flaubert", "FlauBERT"),
+        ("flava", "FLAVA"),
+        ("fnet", "FNet"),
+        ("focalnet", "FocalNet"),
+        ("fsmt", "FairSeq Machine-Translation"),
+        ("funnel", "Funnel Transformer"),
+        ("git", "GIT"),
+        ("glpn", "GLPN"),
+        ("gpt-sw3", "GPT-Sw3"),
+        ("gpt2", "OpenAI GPT-2"),
+        ("gpt_bigcode", "GPTBigCode"),
+        ("gpt_neo", "GPT Neo"),
+        ("gpt_neox", "GPT NeoX"),
+        ("gpt_neox_japanese", "GPT NeoX Japanese"),
+        ("gptj", "GPT-J"),
+        ("gptsan-japanese", "GPTSAN-japanese"),
+        ("graphormer", "Graphormer"),
+        ("groupvit", "GroupViT"),
+        ("herbert", "HerBERT"),
+        ("hubert", "Hubert"),
+        ("ibert", "I-BERT"),
+        ("imagegpt", "ImageGPT"),
+        ("informer", "Informer"),
+        ("jukebox", "Jukebox"),
+        ("layoutlm", "LayoutLM"),
+        ("layoutlmv2", "LayoutLMv2"),
+        ("layoutlmv3", "LayoutLMv3"),
+        ("layoutxlm", "LayoutXLM"),
+        ("led", "LED"),
+        ("levit", "LeViT"),
+        ("lilt", "LiLT"),
+        ("llama", "LLaMA"),
+        ("longformer", "Longformer"),
+        ("longt5", "LongT5"),
+        ("luke", "LUKE"),
+        ("lxmert", "LXMERT"),
+        ("m2m_100", "M2M100"),
+        ("marian", "Marian"),
+        ("markuplm", "MarkupLM"),
+        ("mask2former", "Mask2Former"),
+        ("maskformer", "MaskFormer"),
+        ("maskformer-swin", "MaskFormerSwin"),
+        ("matcha", "MatCha"),
+        ("mbart", "mBART"),
+        ("mbart50", "mBART-50"),
+        ("mctct", "M-CTC-T"),
+        ("mega", "MEGA"),
+        ("megatron-bert", "Megatron-BERT"),
+        ("megatron_gpt2", "Megatron-GPT2"),
+        ("mgp-str", "MGP-STR"),
+        ("mluke", "mLUKE"),
+        ("mobilebert", "MobileBERT"),
+        ("mobilenet_v1", "MobileNetV1"),
+        ("mobilenet_v2", "MobileNetV2"),
+        ("mobilevit", "MobileViT"),
+        ("mpnet", "MPNet"),
+        ("mt5", "MT5"),
+        ("mvp", "MVP"),
+        ("nat", "NAT"),
+        ("nezha", "Nezha"),
+        ("nllb", "NLLB"),
+        ("nllb-moe", "NLLB-MOE"),
+        ("nystromformer", "Nyströmformer"),
+        ("oneformer", "OneFormer"),
+        ("open-llama", "OpenLlama"),
+        ("openai-gpt", "OpenAI GPT"),
+        ("opt", "OPT"),
+        ("owlvit", "OWL-ViT"),
+        ("pegasus", "Pegasus"),
+        ("pegasus_x", "PEGASUS-X"),
+        ("perceiver", "Perceiver"),
+        ("phobert", "PhoBERT"),
+        ("pix2struct", "Pix2Struct"),
+        ("plbart", "PLBart"),
+        ("poolformer", "PoolFormer"),
+        ("prophetnet", "ProphetNet"),
+        ("qdqbert", "QDQBert"),
+        ("rag", "RAG"),
+        ("realm", "REALM"),
+        ("reformer", "Reformer"),
+        ("regnet", "RegNet"),
+        ("rembert", "RemBERT"),
+        ("resnet", "ResNet"),
+        ("retribert", "RetriBERT"),
+        ("roberta", "RoBERTa"),
+        ("roberta-prelayernorm", "RoBERTa-PreLayerNorm"),
+        ("roc_bert", "RoCBert"),
+        ("roformer", "RoFormer"),
+        ("rwkv", "RWKV"),
+        ("sam", "SAM"),
+        ("segformer", "SegFormer"),
+        ("sew", "SEW"),
+        ("sew-d", "SEW-D"),
+        ("speech-encoder-decoder", "Speech Encoder decoder"),
+        ("speech_to_text", "Speech2Text"),
+        ("speech_to_text_2", "Speech2Text2"),
+        ("speecht5", "SpeechT5"),
+        ("splinter", "Splinter"),
+        ("squeezebert", "SqueezeBERT"),
+        ("swin", "Swin Transformer"),
+        ("swin2sr", "Swin2SR"),
+        ("swinv2", "Swin Transformer V2"),
+        ("switch_transformers", "SwitchTransformers"),
+        ("t5", "T5"),
+        ("t5v1.1", "T5v1.1"),
+        ("table-transformer", "Table Transformer"),
+        ("tapas", "TAPAS"),
+        ("tapex", "TAPEX"),
+        ("time_series_transformer", "Time Series Transformer"),
+        ("timesformer", "TimeSformer"),
+        ("trajectory_transformer", "Trajectory Transformer"),
+        ("transfo-xl", "Transformer-XL"),
+        ("trocr", "TrOCR"),
+        ("tvlt", "TVLT"),
+        ("ul2", "UL2"),
+        ("unispeech", "UniSpeech"),
+        ("unispeech-sat", "UniSpeechSat"),
+        ("upernet", "UPerNet"),
+        ("van", "VAN"),
+        ("videomae", "VideoMAE"),
+        ("vilt", "ViLT"),
+        ("vision-encoder-decoder", "Vision Encoder decoder"),
+        ("vision-text-dual-encoder", "VisionTextDualEncoder"),
+        ("visual_bert", "VisualBERT"),
+        ("vit", "ViT"),
+        ("vit_hybrid", "ViT Hybrid"),
+        ("vit_mae", "ViTMAE"),
+        ("vit_msn", "ViTMSN"),
+        ("wav2vec2", "Wav2Vec2"),
+        ("wav2vec2-conformer", "Wav2Vec2-Conformer"),
+        ("wav2vec2_phoneme", "Wav2Vec2Phoneme"),
+        ("wavlm", "WavLM"),
+        ("whisper", "Whisper"),
+        ("xclip", "X-CLIP"),
+        ("xglm", "XGLM"),
+        ("xlm", "XLM"),
+        ("xlm-prophetnet", "XLM-ProphetNet"),
+        ("xlm-roberta", "XLM-RoBERTa"),
+        ("xlm-roberta-xl", "XLM-RoBERTa-XL"),
+        ("xlm-v", "XLM-V"),
+        ("xlnet", "XLNet"),
+        ("xls_r", "XLS-R"),
+        ("xlsr_wav2vec2", "XLSR-Wav2Vec2"),
+        ("xmod", "X-MOD"),
+        ("yolos", "YOLOS"),
+        ("yoso", "YOSO"),
+    ]
+)
+
+SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict(
+    [
+        ("openai-gpt", "openai"),
+        ("data2vec-audio", "data2vec"),
+        ("data2vec-text", "data2vec"),
+        ("data2vec-vision", "data2vec"),
+        ("donut-swin", "donut"),
+        ("maskformer-swin", "maskformer"),
+        ("xclip", "x_clip"),
+    ]
+)
+
+
+def model_type_to_module_name(key):
+    """Converts a config key to the corresponding module."""
+    # Special treatment
+    if key in SPECIAL_MODEL_TYPE_TO_MODULE_NAME:
+        return SPECIAL_MODEL_TYPE_TO_MODULE_NAME[key]
+
+    return key.replace("-", "_")
+
+
+def config_class_to_model_type(config):
+    """Converts a config class name to the corresponding model type"""
+    for key, cls in CONFIG_MAPPING_NAMES.items():
+        if cls == config:
+            return key
+    return None
+
+
+class _LazyConfigMapping(OrderedDict):
+    """
+    A dictionary that lazily load its values when they are requested.
+    """
+
+    def __init__(self, mapping):
+        self._mapping = mapping
+        self._extra_content = {}
+        self._modules = {}
+
+    def __getitem__(self, key):
+        if key in self._extra_content:
+            return self._extra_content[key]
+        if key not in self._mapping:
+            raise KeyError(key)
+        value = self._mapping[key]
+        module_name = model_type_to_module_name(key)
+        if module_name not in self._modules:
+            self._modules[module_name] = importlib.import_module(f".{module_name}", "transformers.models")
+        if hasattr(self._modules[module_name], value):
+            return getattr(self._modules[module_name], value)
+
+        # Some of the mappings have entries model_type -> config of another model type. In that case we try to grab the
+        # object at the top level.
+        transformers_module = importlib.import_module("transformers")
+        return getattr(transformers_module, value)
+
+    def keys(self):
+        return list(self._mapping.keys()) + list(self._extra_content.keys())
+
+    def values(self):
+        return [self[k] for k in self._mapping.keys()] + list(self._extra_content.values())
+
+    def items(self):
+        return [(k, self[k]) for k in self._mapping.keys()] + list(self._extra_content.items())
+
+    def __iter__(self):
+        return iter(list(self._mapping.keys()) + list(self._extra_content.keys()))
+
+    def __contains__(self, item):
+        return item in self._mapping or item in self._extra_content
+
+    def register(self, key, value):
+        """
+        Register a new configuration in this mapping.
+        """
+        if key in self._mapping.keys():
+            raise ValueError(f"'{key}' is already used by a Transformers config, pick another name.")
+        self._extra_content[key] = value
+
+
+CONFIG_MAPPING = _LazyConfigMapping(CONFIG_MAPPING_NAMES)
+
+
+class _LazyLoadAllMappings(OrderedDict):
+    """
+    A mapping that will load all pairs of key values at the first access (either by indexing, requestions keys, values,
+    etc.)
+
+    Args:
+        mapping: The mapping to load.
+    """
+
+    def __init__(self, mapping):
+        self._mapping = mapping
+        self._initialized = False
+        self._data = {}
+
+    def _initialize(self):
+        if self._initialized:
+            return
+        warnings.warn(
+            "ALL_PRETRAINED_CONFIG_ARCHIVE_MAP is deprecated and will be removed in v5 of Transformers. "
+            "It does not contain all available model checkpoints, far from it. Checkout hf.co/models for that.",
+            FutureWarning,
+        )
+
+        for model_type, map_name in self._mapping.items():
+            module_name = model_type_to_module_name(model_type)
+            module = importlib.import_module(f".{module_name}", "transformers.models")
+            mapping = getattr(module, map_name)
+            self._data.update(mapping)
+
+        self._initialized = True
+
+    def __getitem__(self, key):
+        self._initialize()
+        return self._data[key]
+
+    def keys(self):
+        self._initialize()
+        return self._data.keys()
+
+    def values(self):
+        self._initialize()
+        return self._data.values()
+
+    def items(self):
+        self._initialize()
+        return self._data.keys()
+
+    def __iter__(self):
+        self._initialize()
+        return iter(self._data)
+
+    def __contains__(self, item):
+        self._initialize()
+        return item in self._data
+
+
+ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = _LazyLoadAllMappings(CONFIG_ARCHIVE_MAP_MAPPING_NAMES)
+
+
+def _get_class_name(model_class: Union[str, List[str]]):
+    if isinstance(model_class, (list, tuple)):
+        return " or ".join([f"[`{c}`]" for c in model_class if c is not None])
+    return f"[`{model_class}`]"
+
+
+def _list_model_options(indent, config_to_class=None, use_model_types=True):
+    if config_to_class is None and not use_model_types:
+        raise ValueError("Using `use_model_types=False` requires a `config_to_class` dictionary.")
+    if use_model_types:
+        if config_to_class is None:
+            model_type_to_name = {model_type: f"[`{config}`]" for model_type, config in CONFIG_MAPPING_NAMES.items()}
+        else:
+            model_type_to_name = {
+                model_type: _get_class_name(model_class)
+                for model_type, model_class in config_to_class.items()
+                if model_type in MODEL_NAMES_MAPPING
+            }
+        lines = [
+            f"{indent}- **{model_type}** -- {model_type_to_name[model_type]} ({MODEL_NAMES_MAPPING[model_type]} model)"
+            for model_type in sorted(model_type_to_name.keys())
+        ]
+    else:
+        config_to_name = {
+            CONFIG_MAPPING_NAMES[config]: _get_class_name(clas)
+            for config, clas in config_to_class.items()
+            if config in CONFIG_MAPPING_NAMES
+        }
+        config_to_model_name = {
+            config: MODEL_NAMES_MAPPING[model_type] for model_type, config in CONFIG_MAPPING_NAMES.items()
+        }
+        lines = [
+            f"{indent}- [`{config_name}`] configuration class:"
+            f" {config_to_name[config_name]} ({config_to_model_name[config_name]} model)"
+            for config_name in sorted(config_to_name.keys())
+        ]
+    return "\n".join(lines)
+
+
+def replace_list_option_in_docstrings(config_to_class=None, use_model_types=True):
+    def docstring_decorator(fn):
+        docstrings = fn.__doc__
+        lines = docstrings.split("\n")
+        i = 0
+        while i < len(lines) and re.search(r"^(\s*)List options\s*$", lines[i]) is None:
+            i += 1
+        if i < len(lines):
+            indent = re.search(r"^(\s*)List options\s*$", lines[i]).groups()[0]
+            if use_model_types:
+                indent = f"{indent}    "
+            lines[i] = _list_model_options(indent, config_to_class=config_to_class, use_model_types=use_model_types)
+            docstrings = "\n".join(lines)
+        else:
+            raise ValueError(
+                f"The function {fn} should have an empty 'List options' in its docstring as placeholder, current"
+                f" docstring is:\n{docstrings}"
+            )
+        fn.__doc__ = docstrings
+        return fn
+
+    return docstring_decorator
+
+
+class AutoConfig:
+    r"""
+    This is a generic configuration class that will be instantiated as one of the configuration classes of the library
+    when created with the [`~AutoConfig.from_pretrained`] class method.
+
+    This class cannot be instantiated directly using `__init__()` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "AutoConfig is designed to be instantiated "
+            "using the `AutoConfig.from_pretrained(pretrained_model_name_or_path)` method."
+        )
+
+    @classmethod
+    def for_model(cls, model_type: str, *args, **kwargs):
+        if model_type in CONFIG_MAPPING:
+            config_class = CONFIG_MAPPING[model_type]
+            return config_class(*args, **kwargs)
+        raise ValueError(
+            f"Unrecognized model identifier: {model_type}. Should contain one of {', '.join(CONFIG_MAPPING.keys())}"
+        )
+
+    @classmethod
+    @replace_list_option_in_docstrings()
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r"""
+        Instantiate one of the configuration classes of the library from a pretrained model configuration.
+
+        The configuration class to instantiate is selected based on the `model_type` property of the config object that
+        is loaded, or when it's missing, by falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        Args:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                Can be either:
+
+                    - A string, the *model id* of a pretrained model configuration hosted inside a model repo on
+                      huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or
+                      namespaced under a user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing a configuration file saved using the
+                      [`~PretrainedConfig.save_pretrained`] method, or the [`~PreTrainedModel.save_pretrained`] method,
+                      e.g., `./my_model_directory/`.
+                    - A path or url to a saved configuration JSON *file*, e.g.,
+                      `./my_model_directory/configuration.json`.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download the model weights and configuration files and override the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+                If `False`, then this function returns just the final configuration object.
+
+                If `True`, then this functions returns a `Tuple(config, unused_kwargs)` where *unused_kwargs* is a
+                dictionary consisting of the key/value pairs whose keys are not configuration attributes: i.e., the
+                part of `kwargs` which has not been used to update `config` and is otherwise ignored.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs(additional keyword arguments, *optional*):
+                The values in kwargs of any keys which are configuration attributes will be used to override the loaded
+                values. Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled
+                by the `return_unused_kwargs` keyword parameter.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoConfig
+
+        >>> # Download configuration from huggingface.co and cache.
+        >>> config = AutoConfig.from_pretrained("bert-base-uncased")
+
+        >>> # Download configuration from huggingface.co (user-uploaded) and cache.
+        >>> config = AutoConfig.from_pretrained("dbmdz/bert-base-german-cased")
+
+        >>> # If configuration file is in a directory (e.g., was saved using *save_pretrained('./test/saved_model/')*).
+        >>> config = AutoConfig.from_pretrained("./test/bert_saved_model/")
+
+        >>> # Load a specific configuration file.
+        >>> config = AutoConfig.from_pretrained("./test/bert_saved_model/my_configuration.json")
+
+        >>> # Change some config attributes when loading a pretrained config.
+        >>> config = AutoConfig.from_pretrained("bert-base-uncased", output_attentions=True, foo=False)
+        >>> config.output_attentions
+        True
+
+        >>> config, unused_kwargs = AutoConfig.from_pretrained(
+        ...     "bert-base-uncased", output_attentions=True, foo=False, return_unused_kwargs=True
+        ... )
+        >>> config.output_attentions
+        True
+
+        >>> unused_kwargs
+        {'foo': False}
+        ```"""
+        kwargs["_from_auto"] = True
+        kwargs["name_or_path"] = pretrained_model_name_or_path
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+        config_dict, unused_kwargs = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs)
+        if "auto_map" in config_dict and "AutoConfig" in config_dict["auto_map"]:
+            if not trust_remote_code:
+                raise ValueError(
+                    f"Loading {pretrained_model_name_or_path} requires you to execute the configuration file in that"
+                    " repo on your local machine. Make sure you have read the code there to avoid malicious use, then"
+                    " set the option `trust_remote_code=True` to remove this error."
+                )
+            class_ref = config_dict["auto_map"]["AutoConfig"]
+            config_class = get_class_from_dynamic_module(class_ref, pretrained_model_name_or_path, **kwargs)
+            return config_class.from_pretrained(pretrained_model_name_or_path, **kwargs)
+        elif "model_type" in config_dict:
+            config_class = CONFIG_MAPPING[config_dict["model_type"]]
+            return config_class.from_dict(config_dict, **unused_kwargs)
+        else:
+            # Fallback: use pattern matching on the string.
+            # We go from longer names to shorter names to catch roberta before bert (for instance)
+            for pattern in sorted(CONFIG_MAPPING.keys(), key=len, reverse=True):
+                if pattern in str(pretrained_model_name_or_path):
+                    return CONFIG_MAPPING[pattern].from_dict(config_dict, **unused_kwargs)
+
+        raise ValueError(
+            f"Unrecognized model in {pretrained_model_name_or_path}. "
+            f"Should have a `model_type` key in its {CONFIG_NAME}, or contain one of the following strings "
+            f"in its name: {', '.join(CONFIG_MAPPING.keys())}"
+        )
+
+    @staticmethod
+    def register(model_type, config):
+        """
+        Register a new configuration for this class.
+
+        Args:
+            model_type (`str`): The model type like "bert" or "gpt".
+            config ([`PretrainedConfig`]): The config to register.
+        """
+        if issubclass(config, PretrainedConfig) and config.model_type != model_type:
+            raise ValueError(
+                "The config you are passing has a `model_type` attribute that is not consistent with the model type "
+                f"you passed (config has {config.model_type} and you passed {model_type}. Fix one of those so they "
+                "match!"
+            )
+        CONFIG_MAPPING.register(model_type, config)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/feature_extraction_auto.py

@@ -0,0 +1,364 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" AutoFeatureExtractor class."""
+import importlib
+import json
+import os
+from collections import OrderedDict
+from typing import Dict, Optional, Union
+
+# Build the list of all feature extractors
+from ...configuration_utils import PretrainedConfig
+from ...dynamic_module_utils import get_class_from_dynamic_module
+from ...feature_extraction_utils import FeatureExtractionMixin
+from ...utils import CONFIG_NAME, FEATURE_EXTRACTOR_NAME, get_file_from_repo, logging
+from .auto_factory import _LazyAutoMapping
+from .configuration_auto import (
+    CONFIG_MAPPING_NAMES,
+    AutoConfig,
+    model_type_to_module_name,
+    replace_list_option_in_docstrings,
+)
+
+
+logger = logging.get_logger(__name__)
+
+FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
+    [
+        ("audio-spectrogram-transformer", "ASTFeatureExtractor"),
+        ("beit", "BeitFeatureExtractor"),
+        ("chinese_clip", "ChineseCLIPFeatureExtractor"),
+        ("clap", "ClapFeatureExtractor"),
+        ("clip", "CLIPFeatureExtractor"),
+        ("clipseg", "ViTFeatureExtractor"),
+        ("conditional_detr", "ConditionalDetrFeatureExtractor"),
+        ("convnext", "ConvNextFeatureExtractor"),
+        ("cvt", "ConvNextFeatureExtractor"),
+        ("data2vec-audio", "Wav2Vec2FeatureExtractor"),
+        ("data2vec-vision", "BeitFeatureExtractor"),
+        ("deformable_detr", "DeformableDetrFeatureExtractor"),
+        ("deit", "DeiTFeatureExtractor"),
+        ("detr", "DetrFeatureExtractor"),
+        ("dinat", "ViTFeatureExtractor"),
+        ("donut-swin", "DonutFeatureExtractor"),
+        ("dpt", "DPTFeatureExtractor"),
+        ("flava", "FlavaFeatureExtractor"),
+        ("glpn", "GLPNFeatureExtractor"),
+        ("groupvit", "CLIPFeatureExtractor"),
+        ("hubert", "Wav2Vec2FeatureExtractor"),
+        ("imagegpt", "ImageGPTFeatureExtractor"),
+        ("layoutlmv2", "LayoutLMv2FeatureExtractor"),
+        ("layoutlmv3", "LayoutLMv3FeatureExtractor"),
+        ("levit", "LevitFeatureExtractor"),
+        ("maskformer", "MaskFormerFeatureExtractor"),
+        ("mctct", "MCTCTFeatureExtractor"),
+        ("mobilenet_v1", "MobileNetV1FeatureExtractor"),
+        ("mobilenet_v2", "MobileNetV2FeatureExtractor"),
+        ("mobilevit", "MobileViTFeatureExtractor"),
+        ("nat", "ViTFeatureExtractor"),
+        ("owlvit", "OwlViTFeatureExtractor"),
+        ("perceiver", "PerceiverFeatureExtractor"),
+        ("poolformer", "PoolFormerFeatureExtractor"),
+        ("regnet", "ConvNextFeatureExtractor"),
+        ("resnet", "ConvNextFeatureExtractor"),
+        ("segformer", "SegformerFeatureExtractor"),
+        ("sew", "Wav2Vec2FeatureExtractor"),
+        ("sew-d", "Wav2Vec2FeatureExtractor"),
+        ("speech_to_text", "Speech2TextFeatureExtractor"),
+        ("speecht5", "SpeechT5FeatureExtractor"),
+        ("swin", "ViTFeatureExtractor"),
+        ("swinv2", "ViTFeatureExtractor"),
+        ("table-transformer", "DetrFeatureExtractor"),
+        ("timesformer", "VideoMAEFeatureExtractor"),
+        ("tvlt", "TvltFeatureExtractor"),
+        ("unispeech", "Wav2Vec2FeatureExtractor"),
+        ("unispeech-sat", "Wav2Vec2FeatureExtractor"),
+        ("van", "ConvNextFeatureExtractor"),
+        ("videomae", "VideoMAEFeatureExtractor"),
+        ("vilt", "ViltFeatureExtractor"),
+        ("vit", "ViTFeatureExtractor"),
+        ("vit_mae", "ViTFeatureExtractor"),
+        ("vit_msn", "ViTFeatureExtractor"),
+        ("wav2vec2", "Wav2Vec2FeatureExtractor"),
+        ("wav2vec2-conformer", "Wav2Vec2FeatureExtractor"),
+        ("wavlm", "Wav2Vec2FeatureExtractor"),
+        ("whisper", "WhisperFeatureExtractor"),
+        ("xclip", "CLIPFeatureExtractor"),
+        ("yolos", "YolosFeatureExtractor"),
+    ]
+)
+
+FEATURE_EXTRACTOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FEATURE_EXTRACTOR_MAPPING_NAMES)
+
+
+def feature_extractor_class_from_name(class_name: str):
+    for module_name, extractors in FEATURE_EXTRACTOR_MAPPING_NAMES.items():
+        if class_name in extractors:
+            module_name = model_type_to_module_name(module_name)
+
+            module = importlib.import_module(f".{module_name}", "transformers.models")
+            try:
+                return getattr(module, class_name)
+            except AttributeError:
+                continue
+
+    for _, extractor in FEATURE_EXTRACTOR_MAPPING._extra_content.items():
+        if getattr(extractor, "__name__", None) == class_name:
+            return extractor
+
+    # We did not fine the class, but maybe it's because a dep is missing. In that case, the class will be in the main
+    # init and we return the proper dummy to get an appropriate error message.
+    main_module = importlib.import_module("transformers")
+    if hasattr(main_module, class_name):
+        return getattr(main_module, class_name)
+
+    return None
+
+
+def get_feature_extractor_config(
+    pretrained_model_name_or_path: Union[str, os.PathLike],
+    cache_dir: Optional[Union[str, os.PathLike]] = None,
+    force_download: bool = False,
+    resume_download: bool = False,
+    proxies: Optional[Dict[str, str]] = None,
+    use_auth_token: Optional[Union[bool, str]] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
+    **kwargs,
+):
+    """
+    Loads the tokenizer configuration from a pretrained model tokenizer configuration.
+
+    Args:
+        pretrained_model_name_or_path (`str` or `os.PathLike`):
+            This can be either:
+
+            - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
+              huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
+              under a user or organization name, like `dbmdz/bert-base-german-cased`.
+            - a path to a *directory* containing a configuration file saved using the
+              [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
+
+        cache_dir (`str` or `os.PathLike`, *optional*):
+            Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+            cache should not be used.
+        force_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to force to (re-)download the configuration files and override the cached versions if they
+            exist.
+        resume_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
+        proxies (`Dict[str, str]`, *optional*):
+            A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+            'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+        use_auth_token (`str` or *bool*, *optional*):
+            The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+            when running `huggingface-cli login` (stored in `~/.huggingface`).
+        revision (`str`, *optional*, defaults to `"main"`):
+            The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+            git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+            identifier allowed by git.
+        local_files_only (`bool`, *optional*, defaults to `False`):
+            If `True`, will only try to load the tokenizer configuration from local files.
+
+    <Tip>
+
+    Passing `use_auth_token=True` is required when you want to use a private model.
+
+    </Tip>
+
+    Returns:
+        `Dict`: The configuration of the tokenizer.
+
+    Examples:
+
+    ```python
+    # Download configuration from huggingface.co and cache.
+    tokenizer_config = get_tokenizer_config("bert-base-uncased")
+    # This model does not have a tokenizer config so the result will be an empty dict.
+    tokenizer_config = get_tokenizer_config("xlm-roberta-base")
+
+    # Save a pretrained tokenizer locally and you can reload its config
+    from transformers import AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+    tokenizer.save_pretrained("tokenizer-test")
+    tokenizer_config = get_tokenizer_config("tokenizer-test")
+    ```"""
+    resolved_config_file = get_file_from_repo(
+        pretrained_model_name_or_path,
+        FEATURE_EXTRACTOR_NAME,
+        cache_dir=cache_dir,
+        force_download=force_download,
+        resume_download=resume_download,
+        proxies=proxies,
+        use_auth_token=use_auth_token,
+        revision=revision,
+        local_files_only=local_files_only,
+    )
+    if resolved_config_file is None:
+        logger.info(
+            "Could not locate the feature extractor configuration file, will try to use the model config instead."
+        )
+        return {}
+
+    with open(resolved_config_file, encoding="utf-8") as reader:
+        return json.load(reader)
+
+
+class AutoFeatureExtractor:
+    r"""
+    This is a generic feature extractor class that will be instantiated as one of the feature extractor classes of the
+    library when created with the [`AutoFeatureExtractor.from_pretrained`] class method.
+
+    This class cannot be instantiated directly using `__init__()` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "AutoFeatureExtractor is designed to be instantiated "
+            "using the `AutoFeatureExtractor.from_pretrained(pretrained_model_name_or_path)` method."
+        )
+
+    @classmethod
+    @replace_list_option_in_docstrings(FEATURE_EXTRACTOR_MAPPING_NAMES)
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r"""
+        Instantiate one of the feature extractor classes of the library from a pretrained model vocabulary.
+
+        The feature extractor class to instantiate is selected based on the `model_type` property of the config object
+        (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's
+        missing, by falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        Params:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                This can be either:
+
+                - a string, the *model id* of a pretrained feature_extractor hosted inside a model repo on
+                  huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or
+                  namespaced under a user or organization name, like `dbmdz/bert-base-german-cased`.
+                - a path to a *directory* containing a feature extractor file saved using the
+                  [`~feature_extraction_utils.FeatureExtractionMixin.save_pretrained`] method, e.g.,
+                  `./my_model_directory/`.
+                - a path or url to a saved feature extractor JSON *file*, e.g.,
+                  `./my_model_directory/preprocessor_config.json`.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model feature extractor should be cached if the
+                standard cache should not be used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force to (re-)download the feature extractor files and override the cached versions
+                if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received file. Attempts to resume the download if such a file
+                exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+                when running `huggingface-cli login` (stored in `~/.huggingface`).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+                If `False`, then this function returns just the final feature extractor object. If `True`, then this
+                functions returns a `Tuple(feature_extractor, unused_kwargs)` where *unused_kwargs* is a dictionary
+                consisting of the key/value pairs whose keys are not feature extractor attributes: i.e., the part of
+                `kwargs` which has not been used to update `feature_extractor` and is otherwise ignored.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (`Dict[str, Any]`, *optional*):
+                The values in kwargs of any keys which are feature extractor attributes will be used to override the
+                loaded values. Behavior concerning key/value pairs whose keys are *not* feature extractor attributes is
+                controlled by the `return_unused_kwargs` keyword parameter.
+
+        <Tip>
+
+        Passing `use_auth_token=True` is required when you want to use a private model.
+
+        </Tip>
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoFeatureExtractor
+
+        >>> # Download feature extractor from huggingface.co and cache.
+        >>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base-960h")
+
+        >>> # If feature extractor files are in a directory (e.g. feature extractor was saved using *save_pretrained('./test/saved_model/')*)
+        >>> # feature_extractor = AutoFeatureExtractor.from_pretrained("./test/saved_model/")
+        ```"""
+        config = kwargs.pop("config", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+        kwargs["_from_auto"] = True
+
+        config_dict, _ = FeatureExtractionMixin.get_feature_extractor_dict(pretrained_model_name_or_path, **kwargs)
+        feature_extractor_class = config_dict.get("feature_extractor_type", None)
+        feature_extractor_auto_map = None
+        if "AutoFeatureExtractor" in config_dict.get("auto_map", {}):
+            feature_extractor_auto_map = config_dict["auto_map"]["AutoFeatureExtractor"]
+
+        # If we don't find the feature extractor class in the feature extractor config, let's try the model config.
+        if feature_extractor_class is None and feature_extractor_auto_map is None:
+            if not isinstance(config, PretrainedConfig):
+                config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+            # It could be in `config.feature_extractor_type``
+            feature_extractor_class = getattr(config, "feature_extractor_type", None)
+            if hasattr(config, "auto_map") and "AutoFeatureExtractor" in config.auto_map:
+                feature_extractor_auto_map = config.auto_map["AutoFeatureExtractor"]
+
+        if feature_extractor_class is not None:
+            # If we have custom code for a feature extractor, we get the proper class.
+            if feature_extractor_auto_map is not None:
+                if not trust_remote_code:
+                    raise ValueError(
+                        f"Loading {pretrained_model_name_or_path} requires you to execute the feature extractor file "
+                        "in that repo on your local machine. Make sure you have read the code there to avoid "
+                        "malicious use, then set the option `trust_remote_code=True` to remove this error."
+                    )
+                feature_extractor_class = get_class_from_dynamic_module(
+                    feature_extractor_auto_map, pretrained_model_name_or_path, **kwargs
+                )
+            else:
+                feature_extractor_class = feature_extractor_class_from_name(feature_extractor_class)
+
+            return feature_extractor_class.from_dict(config_dict, **kwargs)
+        # Last try: we use the FEATURE_EXTRACTOR_MAPPING.
+        elif type(config) in FEATURE_EXTRACTOR_MAPPING:
+            feature_extractor_class = FEATURE_EXTRACTOR_MAPPING[type(config)]
+            return feature_extractor_class.from_dict(config_dict, **kwargs)
+
+        raise ValueError(
+            f"Unrecognized feature extractor in {pretrained_model_name_or_path}. Should have a "
+            f"`feature_extractor_type` key in its {FEATURE_EXTRACTOR_NAME} of {CONFIG_NAME}, or one of the following "
+            f"`model_type` keys in its {CONFIG_NAME}: {', '.join(c for c in FEATURE_EXTRACTOR_MAPPING_NAMES.keys())}"
+        )
+
+    @staticmethod
+    def register(config_class, feature_extractor_class):
+        """
+        Register a new feature extractor for this class.
+
+        Args:
+            config_class ([`PretrainedConfig`]):
+                The configuration corresponding to the model to register.
+            feature_extractor_class ([`FeatureExtractorMixin`]): The feature extractor to register.
+        """
+        FEATURE_EXTRACTOR_MAPPING.register(config_class, feature_extractor_class)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/image_processing_auto.py

@@ -0,0 +1,388 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" AutoImageProcessor class."""
+import importlib
+import json
+import os
+from collections import OrderedDict
+from typing import Dict, Optional, Union
+
+# Build the list of all image processors
+from ...configuration_utils import PretrainedConfig
+from ...dynamic_module_utils import get_class_from_dynamic_module
+from ...image_processing_utils import ImageProcessingMixin
+from ...utils import CONFIG_NAME, IMAGE_PROCESSOR_NAME, get_file_from_repo, logging
+from .auto_factory import _LazyAutoMapping
+from .configuration_auto import (
+    CONFIG_MAPPING_NAMES,
+    AutoConfig,
+    model_type_to_module_name,
+    replace_list_option_in_docstrings,
+)
+
+
+logger = logging.get_logger(__name__)
+
+IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
+    [
+        ("align", "EfficientNetImageProcessor"),
+        ("beit", "BeitImageProcessor"),
+        ("bit", "BitImageProcessor"),
+        ("blip", "BlipImageProcessor"),
+        ("blip-2", "BlipImageProcessor"),
+        ("bridgetower", "BridgeTowerImageProcessor"),
+        ("chinese_clip", "ChineseCLIPImageProcessor"),
+        ("clip", "CLIPImageProcessor"),
+        ("clipseg", "ViTImageProcessor"),
+        ("conditional_detr", "ConditionalDetrImageProcessor"),
+        ("convnext", "ConvNextImageProcessor"),
+        ("convnextv2", "ConvNextImageProcessor"),
+        ("cvt", "ConvNextImageProcessor"),
+        ("data2vec-vision", "BeitImageProcessor"),
+        ("deformable_detr", "DeformableDetrImageProcessor"),
+        ("deit", "DeiTImageProcessor"),
+        ("deta", "DetaImageProcessor"),
+        ("detr", "DetrImageProcessor"),
+        ("dinat", "ViTImageProcessor"),
+        ("donut-swin", "DonutImageProcessor"),
+        ("dpt", "DPTImageProcessor"),
+        ("efficientformer", "EfficientFormerImageProcessor"),
+        ("efficientnet", "EfficientNetImageProcessor"),
+        ("flava", "FlavaImageProcessor"),
+        ("focalnet", "BitImageProcessor"),
+        ("git", "CLIPImageProcessor"),
+        ("glpn", "GLPNImageProcessor"),
+        ("groupvit", "CLIPImageProcessor"),
+        ("imagegpt", "ImageGPTImageProcessor"),
+        ("layoutlmv2", "LayoutLMv2ImageProcessor"),
+        ("layoutlmv3", "LayoutLMv3ImageProcessor"),
+        ("levit", "LevitImageProcessor"),
+        ("mask2former", "Mask2FormerImageProcessor"),
+        ("maskformer", "MaskFormerImageProcessor"),
+        ("mgp-str", "ViTImageProcessor"),
+        ("mobilenet_v1", "MobileNetV1ImageProcessor"),
+        ("mobilenet_v2", "MobileNetV2ImageProcessor"),
+        ("mobilenet_v2", "MobileNetV2ImageProcessor"),
+        ("mobilevit", "MobileViTImageProcessor"),
+        ("mobilevit", "MobileViTImageProcessor"),
+        ("nat", "ViTImageProcessor"),
+        ("oneformer", "OneFormerImageProcessor"),
+        ("owlvit", "OwlViTImageProcessor"),
+        ("perceiver", "PerceiverImageProcessor"),
+        ("pix2struct", "Pix2StructImageProcessor"),
+        ("poolformer", "PoolFormerImageProcessor"),
+        ("regnet", "ConvNextImageProcessor"),
+        ("resnet", "ConvNextImageProcessor"),
+        ("sam", "SamImageProcessor"),
+        ("segformer", "SegformerImageProcessor"),
+        ("swin", "ViTImageProcessor"),
+        ("swin2sr", "Swin2SRImageProcessor"),
+        ("swinv2", "ViTImageProcessor"),
+        ("table-transformer", "DetrImageProcessor"),
+        ("timesformer", "VideoMAEImageProcessor"),
+        ("tvlt", "TvltImageProcessor"),
+        ("upernet", "SegformerImageProcessor"),
+        ("van", "ConvNextImageProcessor"),
+        ("videomae", "VideoMAEImageProcessor"),
+        ("vilt", "ViltImageProcessor"),
+        ("vit", "ViTImageProcessor"),
+        ("vit_hybrid", "ViTHybridImageProcessor"),
+        ("vit_mae", "ViTImageProcessor"),
+        ("vit_msn", "ViTImageProcessor"),
+        ("xclip", "CLIPImageProcessor"),
+        ("yolos", "YolosImageProcessor"),
+    ]
+)
+
+IMAGE_PROCESSOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, IMAGE_PROCESSOR_MAPPING_NAMES)
+
+
+def image_processor_class_from_name(class_name: str):
+    for module_name, extractors in IMAGE_PROCESSOR_MAPPING_NAMES.items():
+        if class_name in extractors:
+            module_name = model_type_to_module_name(module_name)
+
+            module = importlib.import_module(f".{module_name}", "transformers.models")
+            try:
+                return getattr(module, class_name)
+            except AttributeError:
+                continue
+
+    for _, extractor in IMAGE_PROCESSOR_MAPPING._extra_content.items():
+        if getattr(extractor, "__name__", None) == class_name:
+            return extractor
+
+    # We did not fine the class, but maybe it's because a dep is missing. In that case, the class will be in the main
+    # init and we return the proper dummy to get an appropriate error message.
+    main_module = importlib.import_module("transformers")
+    if hasattr(main_module, class_name):
+        return getattr(main_module, class_name)
+
+    return None
+
+
+def get_image_processor_config(
+    pretrained_model_name_or_path: Union[str, os.PathLike],
+    cache_dir: Optional[Union[str, os.PathLike]] = None,
+    force_download: bool = False,
+    resume_download: bool = False,
+    proxies: Optional[Dict[str, str]] = None,
+    use_auth_token: Optional[Union[bool, str]] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
+    **kwargs,
+):
+    """
+    Loads the image processor configuration from a pretrained model image processor configuration.
+
+    Args:
+        pretrained_model_name_or_path (`str` or `os.PathLike`):
+            This can be either:
+
+            - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
+              huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
+              under a user or organization name, like `dbmdz/bert-base-german-cased`.
+            - a path to a *directory* containing a configuration file saved using the
+              [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
+
+        cache_dir (`str` or `os.PathLike`, *optional*):
+            Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+            cache should not be used.
+        force_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to force to (re-)download the configuration files and override the cached versions if they
+            exist.
+        resume_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
+        proxies (`Dict[str, str]`, *optional*):
+            A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+            'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+        use_auth_token (`str` or *bool*, *optional*):
+            The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+            when running `huggingface-cli login` (stored in `~/.huggingface`).
+        revision (`str`, *optional*, defaults to `"main"`):
+            The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+            git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+            identifier allowed by git.
+        local_files_only (`bool`, *optional*, defaults to `False`):
+            If `True`, will only try to load the image processor configuration from local files.
+
+    <Tip>
+
+    Passing `use_auth_token=True` is required when you want to use a private model.
+
+    </Tip>
+
+    Returns:
+        `Dict`: The configuration of the image processor.
+
+    Examples:
+
+    ```python
+    # Download configuration from huggingface.co and cache.
+    image_processor_config = get_image_processor_config("bert-base-uncased")
+    # This model does not have a image processor config so the result will be an empty dict.
+    image_processor_config = get_image_processor_config("xlm-roberta-base")
+
+    # Save a pretrained image processor locally and you can reload its config
+    from transformers import AutoTokenizer
+
+    image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
+    image_processor.save_pretrained("image-processor-test")
+    image_processor_config = get_image_processor_config("image-processor-test")
+    ```"""
+    resolved_config_file = get_file_from_repo(
+        pretrained_model_name_or_path,
+        IMAGE_PROCESSOR_NAME,
+        cache_dir=cache_dir,
+        force_download=force_download,
+        resume_download=resume_download,
+        proxies=proxies,
+        use_auth_token=use_auth_token,
+        revision=revision,
+        local_files_only=local_files_only,
+    )
+    if resolved_config_file is None:
+        logger.info(
+            "Could not locate the image processor configuration file, will try to use the model config instead."
+        )
+        return {}
+
+    with open(resolved_config_file, encoding="utf-8") as reader:
+        return json.load(reader)
+
+
+class AutoImageProcessor:
+    r"""
+    This is a generic image processor class that will be instantiated as one of the image processor classes of the
+    library when created with the [`AutoImageProcessor.from_pretrained`] class method.
+
+    This class cannot be instantiated directly using `__init__()` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "AutoImageProcessor is designed to be instantiated "
+            "using the `AutoImageProcessor.from_pretrained(pretrained_model_name_or_path)` method."
+        )
+
+    @classmethod
+    @replace_list_option_in_docstrings(IMAGE_PROCESSOR_MAPPING_NAMES)
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r"""
+        Instantiate one of the image processor classes of the library from a pretrained model vocabulary.
+
+        The image processor class to instantiate is selected based on the `model_type` property of the config object
+        (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's
+        missing, by falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        Params:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                This can be either:
+
+                - a string, the *model id* of a pretrained image_processor hosted inside a model repo on
+                  huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or
+                  namespaced under a user or organization name, like `dbmdz/bert-base-german-cased`.
+                - a path to a *directory* containing a image processor file saved using the
+                  [`~image_processing_utils.ImageProcessingMixin.save_pretrained`] method, e.g.,
+                  `./my_model_directory/`.
+                - a path or url to a saved image processor JSON *file*, e.g.,
+                  `./my_model_directory/preprocessor_config.json`.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model image processor should be cached if the
+                standard cache should not be used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force to (re-)download the image processor files and override the cached versions if
+                they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received file. Attempts to resume the download if such a file
+                exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+                when running `huggingface-cli login` (stored in `~/.huggingface`).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+                If `False`, then this function returns just the final image processor object. If `True`, then this
+                functions returns a `Tuple(image_processor, unused_kwargs)` where *unused_kwargs* is a dictionary
+                consisting of the key/value pairs whose keys are not image processor attributes: i.e., the part of
+                `kwargs` which has not been used to update `image_processor` and is otherwise ignored.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (`Dict[str, Any]`, *optional*):
+                The values in kwargs of any keys which are image processor attributes will be used to override the
+                loaded values. Behavior concerning key/value pairs whose keys are *not* image processor attributes is
+                controlled by the `return_unused_kwargs` keyword parameter.
+
+        <Tip>
+
+        Passing `use_auth_token=True` is required when you want to use a private model.
+
+        </Tip>
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor
+
+        >>> # Download image processor from huggingface.co and cache.
+        >>> image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
+
+        >>> # If image processor files are in a directory (e.g. image processor was saved using *save_pretrained('./test/saved_model/')*)
+        >>> # image_processor = AutoImageProcessor.from_pretrained("./test/saved_model/")
+        ```"""
+        config = kwargs.pop("config", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+        kwargs["_from_auto"] = True
+
+        config_dict, _ = ImageProcessingMixin.get_image_processor_dict(pretrained_model_name_or_path, **kwargs)
+        image_processor_class = config_dict.get("image_processor_type", None)
+        image_processor_auto_map = None
+        if "AutoImageProcessor" in config_dict.get("auto_map", {}):
+            image_processor_auto_map = config_dict["auto_map"]["AutoImageProcessor"]
+
+        # If we still don't have the image processor class, check if we're loading from a previous feature extractor config
+        # and if so, infer the image processor class from there.
+        if image_processor_class is None and image_processor_auto_map is None:
+            feature_extractor_class = config_dict.pop("feature_extractor_type", None)
+            if feature_extractor_class is not None:
+                logger.warning(
+                    "Could not find image processor class in the image processor config or the model config. Loading"
+                    " based on pattern matching with the model's feature extractor configuration."
+                )
+                image_processor_class = feature_extractor_class.replace("FeatureExtractor", "ImageProcessor")
+            if "AutoFeatureExtractor" in config_dict.get("auto_map", {}):
+                feature_extractor_auto_map = config_dict["auto_map"]["AutoFeatureExtractor"]
+                image_processor_auto_map = feature_extractor_auto_map.replace("FeatureExtractor", "ImageProcessor")
+                logger.warning(
+                    "Could not find image processor auto map in the image processor config or the model config."
+                    " Loading based on pattern matching with the model's feature extractor configuration."
+                )
+
+        # If we don't find the image processor class in the image processor config, let's try the model config.
+        if image_processor_class is None and image_processor_auto_map is None:
+            if not isinstance(config, PretrainedConfig):
+                config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+            # It could be in `config.image_processor_type``
+            image_processor_class = getattr(config, "image_processor_type", None)
+            if hasattr(config, "auto_map") and "AutoImageProcessor" in config.auto_map:
+                image_processor_auto_map = config.auto_map["AutoImageProcessor"]
+
+        if image_processor_class is not None:
+            # If we have custom code for a image processor, we get the proper class.
+            if image_processor_auto_map is not None:
+                if not trust_remote_code:
+                    raise ValueError(
+                        f"Loading {pretrained_model_name_or_path} requires you to execute the image processor file "
+                        "in that repo on your local machine. Make sure you have read the code there to avoid "
+                        "malicious use, then set the option `trust_remote_code=True` to remove this error."
+                    )
+                image_processor_class = get_class_from_dynamic_module(
+                    image_processor_auto_map, pretrained_model_name_or_path, **kwargs
+                )
+            else:
+                image_processor_class = image_processor_class_from_name(image_processor_class)
+
+            return image_processor_class.from_dict(config_dict, **kwargs)
+        # Last try: we use the IMAGE_PROCESSOR_MAPPING.
+        elif type(config) in IMAGE_PROCESSOR_MAPPING:
+            image_processor_class = IMAGE_PROCESSOR_MAPPING[type(config)]
+            return image_processor_class.from_dict(config_dict, **kwargs)
+
+        raise ValueError(
+            f"Unrecognized image processor in {pretrained_model_name_or_path}. Should have a "
+            f"`image_processor_type` key in its {IMAGE_PROCESSOR_NAME} of {CONFIG_NAME}, or one of the following "
+            f"`model_type` keys in its {CONFIG_NAME}: {', '.join(c for c in IMAGE_PROCESSOR_MAPPING_NAMES.keys())}"
+        )
+
+    @staticmethod
+    def register(config_class, image_processor_class):
+        """
+        Register a new image processor for this class.
+
+        Args:
+            config_class ([`PretrainedConfig`]):
+                The configuration corresponding to the model to register.
+            image_processor_class ([`ImageProcessingMixin`]): The image processor to register.
+        """
+        IMAGE_PROCESSOR_MAPPING.register(config_class, image_processor_class)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/modeling_auto.py

@@ -0,0 +1,1358 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Model class."""
+
+import warnings
+from collections import OrderedDict
+
+from ...utils import logging
+from .auto_factory import _BaseAutoModelClass, _LazyAutoMapping, auto_class_update
+from .configuration_auto import CONFIG_MAPPING_NAMES
+
+
+logger = logging.get_logger(__name__)
+
+
+MODEL_MAPPING_NAMES = OrderedDict(
+    [
+        # Base model mapping
+        ("albert", "AlbertModel"),
+        ("align", "AlignModel"),
+        ("altclip", "AltCLIPModel"),
+        ("audio-spectrogram-transformer", "ASTModel"),
+        ("bart", "BartModel"),
+        ("beit", "BeitModel"),
+        ("bert", "BertModel"),
+        ("bert-generation", "BertGenerationEncoder"),
+        ("big_bird", "BigBirdModel"),
+        ("bigbird_pegasus", "BigBirdPegasusModel"),
+        ("biogpt", "BioGptModel"),
+        ("bit", "BitModel"),
+        ("blenderbot", "BlenderbotModel"),
+        ("blenderbot-small", "BlenderbotSmallModel"),
+        ("blip", "BlipModel"),
+        ("blip-2", "Blip2Model"),
+        ("bloom", "BloomModel"),
+        ("bridgetower", "BridgeTowerModel"),
+        ("camembert", "CamembertModel"),
+        ("canine", "CanineModel"),
+        ("chinese_clip", "ChineseCLIPModel"),
+        ("clap", "ClapModel"),
+        ("clip", "CLIPModel"),
+        ("clipseg", "CLIPSegModel"),
+        ("codegen", "CodeGenModel"),
+        ("conditional_detr", "ConditionalDetrModel"),
+        ("convbert", "ConvBertModel"),
+        ("convnext", "ConvNextModel"),
+        ("convnextv2", "ConvNextV2Model"),
+        ("cpmant", "CpmAntModel"),
+        ("ctrl", "CTRLModel"),
+        ("cvt", "CvtModel"),
+        ("data2vec-audio", "Data2VecAudioModel"),
+        ("data2vec-text", "Data2VecTextModel"),
+        ("data2vec-vision", "Data2VecVisionModel"),
+        ("deberta", "DebertaModel"),
+        ("deberta-v2", "DebertaV2Model"),
+        ("decision_transformer", "DecisionTransformerModel"),
+        ("deformable_detr", "DeformableDetrModel"),
+        ("deit", "DeiTModel"),
+        ("deta", "DetaModel"),
+        ("detr", "DetrModel"),
+        ("dinat", "DinatModel"),
+        ("distilbert", "DistilBertModel"),
+        ("donut-swin", "DonutSwinModel"),
+        ("dpr", "DPRQuestionEncoder"),
+        ("dpt", "DPTModel"),
+        ("efficientformer", "EfficientFormerModel"),
+        ("efficientnet", "EfficientNetModel"),
+        ("electra", "ElectraModel"),
+        ("ernie", "ErnieModel"),
+        ("ernie_m", "ErnieMModel"),
+        ("esm", "EsmModel"),
+        ("flaubert", "FlaubertModel"),
+        ("flava", "FlavaModel"),
+        ("fnet", "FNetModel"),
+        ("focalnet", "FocalNetModel"),
+        ("fsmt", "FSMTModel"),
+        ("funnel", ("FunnelModel", "FunnelBaseModel")),
+        ("git", "GitModel"),
+        ("glpn", "GLPNModel"),
+        ("gpt-sw3", "GPT2Model"),
+        ("gpt2", "GPT2Model"),
+        ("gpt_bigcode", "GPTBigCodeModel"),
+        ("gpt_neo", "GPTNeoModel"),
+        ("gpt_neox", "GPTNeoXModel"),
+        ("gpt_neox_japanese", "GPTNeoXJapaneseModel"),
+        ("gptj", "GPTJModel"),
+        ("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
+        ("graphormer", "GraphormerModel"),
+        ("groupvit", "GroupViTModel"),
+        ("hubert", "HubertModel"),
+        ("ibert", "IBertModel"),
+        ("imagegpt", "ImageGPTModel"),
+        ("informer", "InformerModel"),
+        ("jukebox", "JukeboxModel"),
+        ("layoutlm", "LayoutLMModel"),
+        ("layoutlmv2", "LayoutLMv2Model"),
+        ("layoutlmv3", "LayoutLMv3Model"),
+        ("led", "LEDModel"),
+        ("levit", "LevitModel"),
+        ("lilt", "LiltModel"),
+        ("llama", "LlamaModel"),
+        ("longformer", "LongformerModel"),
+        ("longt5", "LongT5Model"),
+        ("luke", "LukeModel"),
+        ("lxmert", "LxmertModel"),
+        ("m2m_100", "M2M100Model"),
+        ("marian", "MarianModel"),
+        ("markuplm", "MarkupLMModel"),
+        ("mask2former", "Mask2FormerModel"),
+        ("maskformer", "MaskFormerModel"),
+        ("maskformer-swin", "MaskFormerSwinModel"),
+        ("mbart", "MBartModel"),
+        ("mctct", "MCTCTModel"),
+        ("mega", "MegaModel"),
+        ("megatron-bert", "MegatronBertModel"),
+        ("mgp-str", "MgpstrForSceneTextRecognition"),
+        ("mobilebert", "MobileBertModel"),
+        ("mobilenet_v1", "MobileNetV1Model"),
+        ("mobilenet_v2", "MobileNetV2Model"),
+        ("mobilevit", "MobileViTModel"),
+        ("mpnet", "MPNetModel"),
+        ("mt5", "MT5Model"),
+        ("mvp", "MvpModel"),
+        ("nat", "NatModel"),
+        ("nezha", "NezhaModel"),
+        ("nllb-moe", "NllbMoeModel"),
+        ("nystromformer", "NystromformerModel"),
+        ("oneformer", "OneFormerModel"),
+        ("open-llama", "OpenLlamaModel"),
+        ("openai-gpt", "OpenAIGPTModel"),
+        ("opt", "OPTModel"),
+        ("owlvit", "OwlViTModel"),
+        ("pegasus", "PegasusModel"),
+        ("pegasus_x", "PegasusXModel"),
+        ("perceiver", "PerceiverModel"),
+        ("plbart", "PLBartModel"),
+        ("poolformer", "PoolFormerModel"),
+        ("prophetnet", "ProphetNetModel"),
+        ("qdqbert", "QDQBertModel"),
+        ("reformer", "ReformerModel"),
+        ("regnet", "RegNetModel"),
+        ("rembert", "RemBertModel"),
+        ("resnet", "ResNetModel"),
+        ("retribert", "RetriBertModel"),
+        ("roberta", "RobertaModel"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormModel"),
+        ("roc_bert", "RoCBertModel"),
+        ("roformer", "RoFormerModel"),
+        ("rwkv", "RwkvModel"),
+        ("sam", "SamModel"),
+        ("segformer", "SegformerModel"),
+        ("sew", "SEWModel"),
+        ("sew-d", "SEWDModel"),
+        ("speech_to_text", "Speech2TextModel"),
+        ("speecht5", "SpeechT5Model"),
+        ("splinter", "SplinterModel"),
+        ("squeezebert", "SqueezeBertModel"),
+        ("swin", "SwinModel"),
+        ("swin2sr", "Swin2SRModel"),
+        ("swinv2", "Swinv2Model"),
+        ("switch_transformers", "SwitchTransformersModel"),
+        ("t5", "T5Model"),
+        ("table-transformer", "TableTransformerModel"),
+        ("tapas", "TapasModel"),
+        ("time_series_transformer", "TimeSeriesTransformerModel"),
+        ("timesformer", "TimesformerModel"),
+        ("trajectory_transformer", "TrajectoryTransformerModel"),
+        ("transfo-xl", "TransfoXLModel"),
+        ("tvlt", "TvltModel"),
+        ("unispeech", "UniSpeechModel"),
+        ("unispeech-sat", "UniSpeechSatModel"),
+        ("van", "VanModel"),
+        ("videomae", "VideoMAEModel"),
+        ("vilt", "ViltModel"),
+        ("vision-text-dual-encoder", "VisionTextDualEncoderModel"),
+        ("visual_bert", "VisualBertModel"),
+        ("vit", "ViTModel"),
+        ("vit_hybrid", "ViTHybridModel"),
+        ("vit_mae", "ViTMAEModel"),
+        ("vit_msn", "ViTMSNModel"),
+        ("wav2vec2", "Wav2Vec2Model"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerModel"),
+        ("wavlm", "WavLMModel"),
+        ("whisper", "WhisperModel"),
+        ("xclip", "XCLIPModel"),
+        ("xglm", "XGLMModel"),
+        ("xlm", "XLMModel"),
+        ("xlm-prophetnet", "XLMProphetNetModel"),
+        ("xlm-roberta", "XLMRobertaModel"),
+        ("xlm-roberta-xl", "XLMRobertaXLModel"),
+        ("xlnet", "XLNetModel"),
+        ("xmod", "XmodModel"),
+        ("yolos", "YolosModel"),
+        ("yoso", "YosoModel"),
+    ]
+)
+
+MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for pre-training mapping
+        ("albert", "AlbertForPreTraining"),
+        ("bart", "BartForConditionalGeneration"),
+        ("bert", "BertForPreTraining"),
+        ("big_bird", "BigBirdForPreTraining"),
+        ("bloom", "BloomForCausalLM"),
+        ("camembert", "CamembertForMaskedLM"),
+        ("ctrl", "CTRLLMHeadModel"),
+        ("data2vec-text", "Data2VecTextForMaskedLM"),
+        ("deberta", "DebertaForMaskedLM"),
+        ("deberta-v2", "DebertaV2ForMaskedLM"),
+        ("distilbert", "DistilBertForMaskedLM"),
+        ("electra", "ElectraForPreTraining"),
+        ("ernie", "ErnieForPreTraining"),
+        ("flaubert", "FlaubertWithLMHeadModel"),
+        ("flava", "FlavaForPreTraining"),
+        ("fnet", "FNetForPreTraining"),
+        ("fsmt", "FSMTForConditionalGeneration"),
+        ("funnel", "FunnelForPreTraining"),
+        ("gpt-sw3", "GPT2LMHeadModel"),
+        ("gpt2", "GPT2LMHeadModel"),
+        ("gpt_bigcode", "GPTBigCodeForCausalLM"),
+        ("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
+        ("ibert", "IBertForMaskedLM"),
+        ("layoutlm", "LayoutLMForMaskedLM"),
+        ("longformer", "LongformerForMaskedLM"),
+        ("luke", "LukeForMaskedLM"),
+        ("lxmert", "LxmertForPreTraining"),
+        ("mega", "MegaForMaskedLM"),
+        ("megatron-bert", "MegatronBertForPreTraining"),
+        ("mobilebert", "MobileBertForPreTraining"),
+        ("mpnet", "MPNetForMaskedLM"),
+        ("mvp", "MvpForConditionalGeneration"),
+        ("nezha", "NezhaForPreTraining"),
+        ("nllb-moe", "NllbMoeForConditionalGeneration"),
+        ("openai-gpt", "OpenAIGPTLMHeadModel"),
+        ("retribert", "RetriBertModel"),
+        ("roberta", "RobertaForMaskedLM"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForMaskedLM"),
+        ("roc_bert", "RoCBertForPreTraining"),
+        ("rwkv", "RwkvForCausalLM"),
+        ("splinter", "SplinterForPreTraining"),
+        ("squeezebert", "SqueezeBertForMaskedLM"),
+        ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
+        ("t5", "T5ForConditionalGeneration"),
+        ("tapas", "TapasForMaskedLM"),
+        ("transfo-xl", "TransfoXLLMHeadModel"),
+        ("tvlt", "TvltForPreTraining"),
+        ("unispeech", "UniSpeechForPreTraining"),
+        ("unispeech-sat", "UniSpeechSatForPreTraining"),
+        ("videomae", "VideoMAEForPreTraining"),
+        ("visual_bert", "VisualBertForPreTraining"),
+        ("vit_mae", "ViTMAEForPreTraining"),
+        ("wav2vec2", "Wav2Vec2ForPreTraining"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerForPreTraining"),
+        ("xlm", "XLMWithLMHeadModel"),
+        ("xlm-roberta", "XLMRobertaForMaskedLM"),
+        ("xlm-roberta-xl", "XLMRobertaXLForMaskedLM"),
+        ("xlnet", "XLNetLMHeadModel"),
+        ("xmod", "XmodForMaskedLM"),
+    ]
+)
+
+MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
+    [
+        # Model with LM heads mapping
+        ("albert", "AlbertForMaskedLM"),
+        ("bart", "BartForConditionalGeneration"),
+        ("bert", "BertForMaskedLM"),
+        ("big_bird", "BigBirdForMaskedLM"),
+        ("bigbird_pegasus", "BigBirdPegasusForConditionalGeneration"),
+        ("blenderbot-small", "BlenderbotSmallForConditionalGeneration"),
+        ("bloom", "BloomForCausalLM"),
+        ("camembert", "CamembertForMaskedLM"),
+        ("codegen", "CodeGenForCausalLM"),
+        ("convbert", "ConvBertForMaskedLM"),
+        ("cpmant", "CpmAntForCausalLM"),
+        ("ctrl", "CTRLLMHeadModel"),
+        ("data2vec-text", "Data2VecTextForMaskedLM"),
+        ("deberta", "DebertaForMaskedLM"),
+        ("deberta-v2", "DebertaV2ForMaskedLM"),
+        ("distilbert", "DistilBertForMaskedLM"),
+        ("electra", "ElectraForMaskedLM"),
+        ("encoder-decoder", "EncoderDecoderModel"),
+        ("ernie", "ErnieForMaskedLM"),
+        ("esm", "EsmForMaskedLM"),
+        ("flaubert", "FlaubertWithLMHeadModel"),
+        ("fnet", "FNetForMaskedLM"),
+        ("fsmt", "FSMTForConditionalGeneration"),
+        ("funnel", "FunnelForMaskedLM"),
+        ("git", "GitForCausalLM"),
+        ("gpt-sw3", "GPT2LMHeadModel"),
+        ("gpt2", "GPT2LMHeadModel"),
+        ("gpt_bigcode", "GPTBigCodeForCausalLM"),
+        ("gpt_neo", "GPTNeoForCausalLM"),
+        ("gpt_neox", "GPTNeoXForCausalLM"),
+        ("gpt_neox_japanese", "GPTNeoXJapaneseForCausalLM"),
+        ("gptj", "GPTJForCausalLM"),
+        ("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
+        ("ibert", "IBertForMaskedLM"),
+        ("layoutlm", "LayoutLMForMaskedLM"),
+        ("led", "LEDForConditionalGeneration"),
+        ("longformer", "LongformerForMaskedLM"),
+        ("longt5", "LongT5ForConditionalGeneration"),
+        ("luke", "LukeForMaskedLM"),
+        ("m2m_100", "M2M100ForConditionalGeneration"),
+        ("marian", "MarianMTModel"),
+        ("mega", "MegaForMaskedLM"),
+        ("megatron-bert", "MegatronBertForCausalLM"),
+        ("mobilebert", "MobileBertForMaskedLM"),
+        ("mpnet", "MPNetForMaskedLM"),
+        ("mvp", "MvpForConditionalGeneration"),
+        ("nezha", "NezhaForMaskedLM"),
+        ("nllb-moe", "NllbMoeForConditionalGeneration"),
+        ("nystromformer", "NystromformerForMaskedLM"),
+        ("openai-gpt", "OpenAIGPTLMHeadModel"),
+        ("pegasus_x", "PegasusXForConditionalGeneration"),
+        ("plbart", "PLBartForConditionalGeneration"),
+        ("qdqbert", "QDQBertForMaskedLM"),
+        ("reformer", "ReformerModelWithLMHead"),
+        ("rembert", "RemBertForMaskedLM"),
+        ("roberta", "RobertaForMaskedLM"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForMaskedLM"),
+        ("roc_bert", "RoCBertForMaskedLM"),
+        ("roformer", "RoFormerForMaskedLM"),
+        ("rwkv", "RwkvForCausalLM"),
+        ("speech_to_text", "Speech2TextForConditionalGeneration"),
+        ("squeezebert", "SqueezeBertForMaskedLM"),
+        ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
+        ("t5", "T5ForConditionalGeneration"),
+        ("tapas", "TapasForMaskedLM"),
+        ("transfo-xl", "TransfoXLLMHeadModel"),
+        ("wav2vec2", "Wav2Vec2ForMaskedLM"),
+        ("whisper", "WhisperForConditionalGeneration"),
+        ("xlm", "XLMWithLMHeadModel"),
+        ("xlm-roberta", "XLMRobertaForMaskedLM"),
+        ("xlm-roberta-xl", "XLMRobertaXLForMaskedLM"),
+        ("xlnet", "XLNetLMHeadModel"),
+        ("xmod", "XmodForMaskedLM"),
+        ("yoso", "YosoForMaskedLM"),
+    ]
+)
+
+MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Causal LM mapping
+        ("bart", "BartForCausalLM"),
+        ("bert", "BertLMHeadModel"),
+        ("bert-generation", "BertGenerationDecoder"),
+        ("big_bird", "BigBirdForCausalLM"),
+        ("bigbird_pegasus", "BigBirdPegasusForCausalLM"),
+        ("biogpt", "BioGptForCausalLM"),
+        ("blenderbot", "BlenderbotForCausalLM"),
+        ("blenderbot-small", "BlenderbotSmallForCausalLM"),
+        ("bloom", "BloomForCausalLM"),
+        ("camembert", "CamembertForCausalLM"),
+        ("codegen", "CodeGenForCausalLM"),
+        ("cpmant", "CpmAntForCausalLM"),
+        ("ctrl", "CTRLLMHeadModel"),
+        ("data2vec-text", "Data2VecTextForCausalLM"),
+        ("electra", "ElectraForCausalLM"),
+        ("ernie", "ErnieForCausalLM"),
+        ("git", "GitForCausalLM"),
+        ("gpt-sw3", "GPT2LMHeadModel"),
+        ("gpt2", "GPT2LMHeadModel"),
+        ("gpt_bigcode", "GPTBigCodeForCausalLM"),
+        ("gpt_neo", "GPTNeoForCausalLM"),
+        ("gpt_neox", "GPTNeoXForCausalLM"),
+        ("gpt_neox_japanese", "GPTNeoXJapaneseForCausalLM"),
+        ("gptj", "GPTJForCausalLM"),
+        ("llama", "LlamaForCausalLM"),
+        ("marian", "MarianForCausalLM"),
+        ("mbart", "MBartForCausalLM"),
+        ("mega", "MegaForCausalLM"),
+        ("megatron-bert", "MegatronBertForCausalLM"),
+        ("mvp", "MvpForCausalLM"),
+        ("open-llama", "OpenLlamaForCausalLM"),
+        ("openai-gpt", "OpenAIGPTLMHeadModel"),
+        ("opt", "OPTForCausalLM"),
+        ("pegasus", "PegasusForCausalLM"),
+        ("plbart", "PLBartForCausalLM"),
+        ("prophetnet", "ProphetNetForCausalLM"),
+        ("qdqbert", "QDQBertLMHeadModel"),
+        ("reformer", "ReformerModelWithLMHead"),
+        ("rembert", "RemBertForCausalLM"),
+        ("roberta", "RobertaForCausalLM"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForCausalLM"),
+        ("roc_bert", "RoCBertForCausalLM"),
+        ("roformer", "RoFormerForCausalLM"),
+        ("rwkv", "RwkvForCausalLM"),
+        ("speech_to_text_2", "Speech2Text2ForCausalLM"),
+        ("transfo-xl", "TransfoXLLMHeadModel"),
+        ("trocr", "TrOCRForCausalLM"),
+        ("xglm", "XGLMForCausalLM"),
+        ("xlm", "XLMWithLMHeadModel"),
+        ("xlm-prophetnet", "XLMProphetNetForCausalLM"),
+        ("xlm-roberta", "XLMRobertaForCausalLM"),
+        ("xlm-roberta-xl", "XLMRobertaXLForCausalLM"),
+        ("xlnet", "XLNetLMHeadModel"),
+        ("xmod", "XmodForCausalLM"),
+    ]
+)
+
+MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES = OrderedDict(
+    [
+        ("deit", "DeiTForMaskedImageModeling"),
+        ("focalnet", "FocalNetForMaskedImageModeling"),
+        ("swin", "SwinForMaskedImageModeling"),
+        ("swinv2", "Swinv2ForMaskedImageModeling"),
+        ("vit", "ViTForMaskedImageModeling"),
+    ]
+)
+
+
+MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES = OrderedDict(
+    # Model for Causal Image Modeling mapping
+    [
+        ("imagegpt", "ImageGPTForCausalImageModeling"),
+    ]
+)
+
+MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Image Classification mapping
+        ("beit", "BeitForImageClassification"),
+        ("bit", "BitForImageClassification"),
+        ("convnext", "ConvNextForImageClassification"),
+        ("convnextv2", "ConvNextV2ForImageClassification"),
+        ("cvt", "CvtForImageClassification"),
+        ("data2vec-vision", "Data2VecVisionForImageClassification"),
+        ("deit", ("DeiTForImageClassification", "DeiTForImageClassificationWithTeacher")),
+        ("dinat", "DinatForImageClassification"),
+        (
+            "efficientformer",
+            (
+                "EfficientFormerForImageClassification",
+                "EfficientFormerForImageClassificationWithTeacher",
+            ),
+        ),
+        ("efficientnet", "EfficientNetForImageClassification"),
+        ("focalnet", "FocalNetForImageClassification"),
+        ("imagegpt", "ImageGPTForImageClassification"),
+        ("levit", ("LevitForImageClassification", "LevitForImageClassificationWithTeacher")),
+        ("mobilenet_v1", "MobileNetV1ForImageClassification"),
+        ("mobilenet_v2", "MobileNetV2ForImageClassification"),
+        ("mobilevit", "MobileViTForImageClassification"),
+        ("nat", "NatForImageClassification"),
+        (
+            "perceiver",
+            (
+                "PerceiverForImageClassificationLearned",
+                "PerceiverForImageClassificationFourier",
+                "PerceiverForImageClassificationConvProcessing",
+            ),
+        ),
+        ("poolformer", "PoolFormerForImageClassification"),
+        ("regnet", "RegNetForImageClassification"),
+        ("resnet", "ResNetForImageClassification"),
+        ("segformer", "SegformerForImageClassification"),
+        ("swin", "SwinForImageClassification"),
+        ("swinv2", "Swinv2ForImageClassification"),
+        ("van", "VanForImageClassification"),
+        ("vit", "ViTForImageClassification"),
+        ("vit_hybrid", "ViTHybridForImageClassification"),
+        ("vit_msn", "ViTMSNForImageClassification"),
+    ]
+)
+
+MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Do not add new models here, this class will be deprecated in the future.
+        # Model for Image Segmentation mapping
+        ("detr", "DetrForSegmentation"),
+    ]
+)
+
+MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Semantic Segmentation mapping
+        ("beit", "BeitForSemanticSegmentation"),
+        ("data2vec-vision", "Data2VecVisionForSemanticSegmentation"),
+        ("dpt", "DPTForSemanticSegmentation"),
+        ("mobilenet_v2", "MobileNetV2ForSemanticSegmentation"),
+        ("mobilevit", "MobileViTForSemanticSegmentation"),
+        ("segformer", "SegformerForSemanticSegmentation"),
+        ("upernet", "UperNetForSemanticSegmentation"),
+    ]
+)
+
+MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Instance Segmentation mapping
+        # MaskFormerForInstanceSegmentation can be removed from this mapping in v5
+        ("maskformer", "MaskFormerForInstanceSegmentation"),
+    ]
+)
+
+MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Universal Segmentation mapping
+        ("detr", "DetrForSegmentation"),
+        ("mask2former", "Mask2FormerForUniversalSegmentation"),
+        ("maskformer", "MaskFormerForInstanceSegmentation"),
+        ("oneformer", "OneFormerForUniversalSegmentation"),
+    ]
+)
+
+MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        ("timesformer", "TimesformerForVideoClassification"),
+        ("videomae", "VideoMAEForVideoClassification"),
+    ]
+)
+
+MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
+    [
+        ("blip", "BlipForConditionalGeneration"),
+        ("blip-2", "Blip2ForConditionalGeneration"),
+        ("vision-encoder-decoder", "VisionEncoderDecoderModel"),
+    ]
+)
+
+MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Masked LM mapping
+        ("albert", "AlbertForMaskedLM"),
+        ("bart", "BartForConditionalGeneration"),
+        ("bert", "BertForMaskedLM"),
+        ("big_bird", "BigBirdForMaskedLM"),
+        ("camembert", "CamembertForMaskedLM"),
+        ("convbert", "ConvBertForMaskedLM"),
+        ("data2vec-text", "Data2VecTextForMaskedLM"),
+        ("deberta", "DebertaForMaskedLM"),
+        ("deberta-v2", "DebertaV2ForMaskedLM"),
+        ("distilbert", "DistilBertForMaskedLM"),
+        ("electra", "ElectraForMaskedLM"),
+        ("ernie", "ErnieForMaskedLM"),
+        ("esm", "EsmForMaskedLM"),
+        ("flaubert", "FlaubertWithLMHeadModel"),
+        ("fnet", "FNetForMaskedLM"),
+        ("funnel", "FunnelForMaskedLM"),
+        ("ibert", "IBertForMaskedLM"),
+        ("layoutlm", "LayoutLMForMaskedLM"),
+        ("longformer", "LongformerForMaskedLM"),
+        ("luke", "LukeForMaskedLM"),
+        ("mbart", "MBartForConditionalGeneration"),
+        ("mega", "MegaForMaskedLM"),
+        ("megatron-bert", "MegatronBertForMaskedLM"),
+        ("mobilebert", "MobileBertForMaskedLM"),
+        ("mpnet", "MPNetForMaskedLM"),
+        ("mvp", "MvpForConditionalGeneration"),
+        ("nezha", "NezhaForMaskedLM"),
+        ("nystromformer", "NystromformerForMaskedLM"),
+        ("perceiver", "PerceiverForMaskedLM"),
+        ("qdqbert", "QDQBertForMaskedLM"),
+        ("reformer", "ReformerForMaskedLM"),
+        ("rembert", "RemBertForMaskedLM"),
+        ("roberta", "RobertaForMaskedLM"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForMaskedLM"),
+        ("roc_bert", "RoCBertForMaskedLM"),
+        ("roformer", "RoFormerForMaskedLM"),
+        ("squeezebert", "SqueezeBertForMaskedLM"),
+        ("tapas", "TapasForMaskedLM"),
+        ("wav2vec2", "Wav2Vec2ForMaskedLM"),
+        ("xlm", "XLMWithLMHeadModel"),
+        ("xlm-roberta", "XLMRobertaForMaskedLM"),
+        ("xlm-roberta-xl", "XLMRobertaXLForMaskedLM"),
+        ("xmod", "XmodForMaskedLM"),
+        ("yoso", "YosoForMaskedLM"),
+    ]
+)
+
+MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Object Detection mapping
+        ("conditional_detr", "ConditionalDetrForObjectDetection"),
+        ("deformable_detr", "DeformableDetrForObjectDetection"),
+        ("deta", "DetaForObjectDetection"),
+        ("detr", "DetrForObjectDetection"),
+        ("table-transformer", "TableTransformerForObjectDetection"),
+        ("yolos", "YolosForObjectDetection"),
+    ]
+)
+
+MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Zero Shot Object Detection mapping
+        ("owlvit", "OwlViTForObjectDetection")
+    ]
+)
+
+MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for depth estimation mapping
+        ("dpt", "DPTForDepthEstimation"),
+        ("glpn", "GLPNForDepthEstimation"),
+    ]
+)
+MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Seq2Seq Causal LM mapping
+        ("bart", "BartForConditionalGeneration"),
+        ("bigbird_pegasus", "BigBirdPegasusForConditionalGeneration"),
+        ("blenderbot", "BlenderbotForConditionalGeneration"),
+        ("blenderbot-small", "BlenderbotSmallForConditionalGeneration"),
+        ("encoder-decoder", "EncoderDecoderModel"),
+        ("fsmt", "FSMTForConditionalGeneration"),
+        ("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
+        ("led", "LEDForConditionalGeneration"),
+        ("longt5", "LongT5ForConditionalGeneration"),
+        ("m2m_100", "M2M100ForConditionalGeneration"),
+        ("marian", "MarianMTModel"),
+        ("mbart", "MBartForConditionalGeneration"),
+        ("mt5", "MT5ForConditionalGeneration"),
+        ("mvp", "MvpForConditionalGeneration"),
+        ("nllb-moe", "NllbMoeForConditionalGeneration"),
+        ("pegasus", "PegasusForConditionalGeneration"),
+        ("pegasus_x", "PegasusXForConditionalGeneration"),
+        ("plbart", "PLBartForConditionalGeneration"),
+        ("prophetnet", "ProphetNetForConditionalGeneration"),
+        ("switch_transformers", "SwitchTransformersForConditionalGeneration"),
+        ("t5", "T5ForConditionalGeneration"),
+        ("xlm-prophetnet", "XLMProphetNetForConditionalGeneration"),
+    ]
+)
+
+MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict(
+    [
+        ("speech-encoder-decoder", "SpeechEncoderDecoderModel"),
+        ("speech_to_text", "Speech2TextForConditionalGeneration"),
+        ("speecht5", "SpeechT5ForSpeechToText"),
+        ("whisper", "WhisperForConditionalGeneration"),
+    ]
+)
+
+MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Sequence Classification mapping
+        ("albert", "AlbertForSequenceClassification"),
+        ("bart", "BartForSequenceClassification"),
+        ("bert", "BertForSequenceClassification"),
+        ("big_bird", "BigBirdForSequenceClassification"),
+        ("bigbird_pegasus", "BigBirdPegasusForSequenceClassification"),
+        ("biogpt", "BioGptForSequenceClassification"),
+        ("bloom", "BloomForSequenceClassification"),
+        ("camembert", "CamembertForSequenceClassification"),
+        ("canine", "CanineForSequenceClassification"),
+        ("convbert", "ConvBertForSequenceClassification"),
+        ("ctrl", "CTRLForSequenceClassification"),
+        ("data2vec-text", "Data2VecTextForSequenceClassification"),
+        ("deberta", "DebertaForSequenceClassification"),
+        ("deberta-v2", "DebertaV2ForSequenceClassification"),
+        ("distilbert", "DistilBertForSequenceClassification"),
+        ("electra", "ElectraForSequenceClassification"),
+        ("ernie", "ErnieForSequenceClassification"),
+        ("ernie_m", "ErnieMForSequenceClassification"),
+        ("esm", "EsmForSequenceClassification"),
+        ("flaubert", "FlaubertForSequenceClassification"),
+        ("fnet", "FNetForSequenceClassification"),
+        ("funnel", "FunnelForSequenceClassification"),
+        ("gpt-sw3", "GPT2ForSequenceClassification"),
+        ("gpt2", "GPT2ForSequenceClassification"),
+        ("gpt_bigcode", "GPTBigCodeForSequenceClassification"),
+        ("gpt_neo", "GPTNeoForSequenceClassification"),
+        ("gpt_neox", "GPTNeoXForSequenceClassification"),
+        ("gptj", "GPTJForSequenceClassification"),
+        ("ibert", "IBertForSequenceClassification"),
+        ("layoutlm", "LayoutLMForSequenceClassification"),
+        ("layoutlmv2", "LayoutLMv2ForSequenceClassification"),
+        ("layoutlmv3", "LayoutLMv3ForSequenceClassification"),
+        ("led", "LEDForSequenceClassification"),
+        ("lilt", "LiltForSequenceClassification"),
+        ("llama", "LlamaForSequenceClassification"),
+        ("longformer", "LongformerForSequenceClassification"),
+        ("luke", "LukeForSequenceClassification"),
+        ("markuplm", "MarkupLMForSequenceClassification"),
+        ("mbart", "MBartForSequenceClassification"),
+        ("mega", "MegaForSequenceClassification"),
+        ("megatron-bert", "MegatronBertForSequenceClassification"),
+        ("mobilebert", "MobileBertForSequenceClassification"),
+        ("mpnet", "MPNetForSequenceClassification"),
+        ("mvp", "MvpForSequenceClassification"),
+        ("nezha", "NezhaForSequenceClassification"),
+        ("nystromformer", "NystromformerForSequenceClassification"),
+        ("open-llama", "OpenLlamaForSequenceClassification"),
+        ("openai-gpt", "OpenAIGPTForSequenceClassification"),
+        ("opt", "OPTForSequenceClassification"),
+        ("perceiver", "PerceiverForSequenceClassification"),
+        ("plbart", "PLBartForSequenceClassification"),
+        ("qdqbert", "QDQBertForSequenceClassification"),
+        ("reformer", "ReformerForSequenceClassification"),
+        ("rembert", "RemBertForSequenceClassification"),
+        ("roberta", "RobertaForSequenceClassification"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForSequenceClassification"),
+        ("roc_bert", "RoCBertForSequenceClassification"),
+        ("roformer", "RoFormerForSequenceClassification"),
+        ("squeezebert", "SqueezeBertForSequenceClassification"),
+        ("tapas", "TapasForSequenceClassification"),
+        ("transfo-xl", "TransfoXLForSequenceClassification"),
+        ("xlm", "XLMForSequenceClassification"),
+        ("xlm-roberta", "XLMRobertaForSequenceClassification"),
+        ("xlm-roberta-xl", "XLMRobertaXLForSequenceClassification"),
+        ("xlnet", "XLNetForSequenceClassification"),
+        ("xmod", "XmodForSequenceClassification"),
+        ("yoso", "YosoForSequenceClassification"),
+    ]
+)
+
+MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Question Answering mapping
+        ("albert", "AlbertForQuestionAnswering"),
+        ("bart", "BartForQuestionAnswering"),
+        ("bert", "BertForQuestionAnswering"),
+        ("big_bird", "BigBirdForQuestionAnswering"),
+        ("bigbird_pegasus", "BigBirdPegasusForQuestionAnswering"),
+        ("bloom", "BloomForQuestionAnswering"),
+        ("camembert", "CamembertForQuestionAnswering"),
+        ("canine", "CanineForQuestionAnswering"),
+        ("convbert", "ConvBertForQuestionAnswering"),
+        ("data2vec-text", "Data2VecTextForQuestionAnswering"),
+        ("deberta", "DebertaForQuestionAnswering"),
+        ("deberta-v2", "DebertaV2ForQuestionAnswering"),
+        ("distilbert", "DistilBertForQuestionAnswering"),
+        ("electra", "ElectraForQuestionAnswering"),
+        ("ernie", "ErnieForQuestionAnswering"),
+        ("ernie_m", "ErnieMForQuestionAnswering"),
+        ("flaubert", "FlaubertForQuestionAnsweringSimple"),
+        ("fnet", "FNetForQuestionAnswering"),
+        ("funnel", "FunnelForQuestionAnswering"),
+        ("gpt2", "GPT2ForQuestionAnswering"),
+        ("gpt_neo", "GPTNeoForQuestionAnswering"),
+        ("gpt_neox", "GPTNeoXForQuestionAnswering"),
+        ("gptj", "GPTJForQuestionAnswering"),
+        ("ibert", "IBertForQuestionAnswering"),
+        ("layoutlmv2", "LayoutLMv2ForQuestionAnswering"),
+        ("layoutlmv3", "LayoutLMv3ForQuestionAnswering"),
+        ("led", "LEDForQuestionAnswering"),
+        ("lilt", "LiltForQuestionAnswering"),
+        ("longformer", "LongformerForQuestionAnswering"),
+        ("luke", "LukeForQuestionAnswering"),
+        ("lxmert", "LxmertForQuestionAnswering"),
+        ("markuplm", "MarkupLMForQuestionAnswering"),
+        ("mbart", "MBartForQuestionAnswering"),
+        ("mega", "MegaForQuestionAnswering"),
+        ("megatron-bert", "MegatronBertForQuestionAnswering"),
+        ("mobilebert", "MobileBertForQuestionAnswering"),
+        ("mpnet", "MPNetForQuestionAnswering"),
+        ("mvp", "MvpForQuestionAnswering"),
+        ("nezha", "NezhaForQuestionAnswering"),
+        ("nystromformer", "NystromformerForQuestionAnswering"),
+        ("opt", "OPTForQuestionAnswering"),
+        ("qdqbert", "QDQBertForQuestionAnswering"),
+        ("reformer", "ReformerForQuestionAnswering"),
+        ("rembert", "RemBertForQuestionAnswering"),
+        ("roberta", "RobertaForQuestionAnswering"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForQuestionAnswering"),
+        ("roc_bert", "RoCBertForQuestionAnswering"),
+        ("roformer", "RoFormerForQuestionAnswering"),
+        ("splinter", "SplinterForQuestionAnswering"),
+        ("squeezebert", "SqueezeBertForQuestionAnswering"),
+        ("xlm", "XLMForQuestionAnsweringSimple"),
+        ("xlm-roberta", "XLMRobertaForQuestionAnswering"),
+        ("xlm-roberta-xl", "XLMRobertaXLForQuestionAnswering"),
+        ("xlnet", "XLNetForQuestionAnsweringSimple"),
+        ("xmod", "XmodForQuestionAnswering"),
+        ("yoso", "YosoForQuestionAnswering"),
+    ]
+)
+
+MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Table Question Answering mapping
+        ("tapas", "TapasForQuestionAnswering"),
+    ]
+)
+
+MODEL_FOR_VISUAL_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        ("vilt", "ViltForQuestionAnswering"),
+    ]
+)
+
+MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        ("layoutlm", "LayoutLMForQuestionAnswering"),
+        ("layoutlmv2", "LayoutLMv2ForQuestionAnswering"),
+        ("layoutlmv3", "LayoutLMv3ForQuestionAnswering"),
+    ]
+)
+
+MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Token Classification mapping
+        ("albert", "AlbertForTokenClassification"),
+        ("bert", "BertForTokenClassification"),
+        ("big_bird", "BigBirdForTokenClassification"),
+        ("biogpt", "BioGptForTokenClassification"),
+        ("bloom", "BloomForTokenClassification"),
+        ("camembert", "CamembertForTokenClassification"),
+        ("canine", "CanineForTokenClassification"),
+        ("convbert", "ConvBertForTokenClassification"),
+        ("data2vec-text", "Data2VecTextForTokenClassification"),
+        ("deberta", "DebertaForTokenClassification"),
+        ("deberta-v2", "DebertaV2ForTokenClassification"),
+        ("distilbert", "DistilBertForTokenClassification"),
+        ("electra", "ElectraForTokenClassification"),
+        ("ernie", "ErnieForTokenClassification"),
+        ("ernie_m", "ErnieMForTokenClassification"),
+        ("esm", "EsmForTokenClassification"),
+        ("flaubert", "FlaubertForTokenClassification"),
+        ("fnet", "FNetForTokenClassification"),
+        ("funnel", "FunnelForTokenClassification"),
+        ("gpt-sw3", "GPT2ForTokenClassification"),
+        ("gpt2", "GPT2ForTokenClassification"),
+        ("gpt_bigcode", "GPTBigCodeForTokenClassification"),
+        ("gpt_neo", "GPTNeoForTokenClassification"),
+        ("gpt_neox", "GPTNeoXForTokenClassification"),
+        ("ibert", "IBertForTokenClassification"),
+        ("layoutlm", "LayoutLMForTokenClassification"),
+        ("layoutlmv2", "LayoutLMv2ForTokenClassification"),
+        ("layoutlmv3", "LayoutLMv3ForTokenClassification"),
+        ("lilt", "LiltForTokenClassification"),
+        ("longformer", "LongformerForTokenClassification"),
+        ("luke", "LukeForTokenClassification"),
+        ("markuplm", "MarkupLMForTokenClassification"),
+        ("mega", "MegaForTokenClassification"),
+        ("megatron-bert", "MegatronBertForTokenClassification"),
+        ("mobilebert", "MobileBertForTokenClassification"),
+        ("mpnet", "MPNetForTokenClassification"),
+        ("nezha", "NezhaForTokenClassification"),
+        ("nystromformer", "NystromformerForTokenClassification"),
+        ("qdqbert", "QDQBertForTokenClassification"),
+        ("rembert", "RemBertForTokenClassification"),
+        ("roberta", "RobertaForTokenClassification"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForTokenClassification"),
+        ("roc_bert", "RoCBertForTokenClassification"),
+        ("roformer", "RoFormerForTokenClassification"),
+        ("squeezebert", "SqueezeBertForTokenClassification"),
+        ("xlm", "XLMForTokenClassification"),
+        ("xlm-roberta", "XLMRobertaForTokenClassification"),
+        ("xlm-roberta-xl", "XLMRobertaXLForTokenClassification"),
+        ("xlnet", "XLNetForTokenClassification"),
+        ("xmod", "XmodForTokenClassification"),
+        ("yoso", "YosoForTokenClassification"),
+    ]
+)
+
+MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Multiple Choice mapping
+        ("albert", "AlbertForMultipleChoice"),
+        ("bert", "BertForMultipleChoice"),
+        ("big_bird", "BigBirdForMultipleChoice"),
+        ("camembert", "CamembertForMultipleChoice"),
+        ("canine", "CanineForMultipleChoice"),
+        ("convbert", "ConvBertForMultipleChoice"),
+        ("data2vec-text", "Data2VecTextForMultipleChoice"),
+        ("deberta-v2", "DebertaV2ForMultipleChoice"),
+        ("distilbert", "DistilBertForMultipleChoice"),
+        ("electra", "ElectraForMultipleChoice"),
+        ("ernie", "ErnieForMultipleChoice"),
+        ("ernie_m", "ErnieMForMultipleChoice"),
+        ("flaubert", "FlaubertForMultipleChoice"),
+        ("fnet", "FNetForMultipleChoice"),
+        ("funnel", "FunnelForMultipleChoice"),
+        ("ibert", "IBertForMultipleChoice"),
+        ("longformer", "LongformerForMultipleChoice"),
+        ("luke", "LukeForMultipleChoice"),
+        ("mega", "MegaForMultipleChoice"),
+        ("megatron-bert", "MegatronBertForMultipleChoice"),
+        ("mobilebert", "MobileBertForMultipleChoice"),
+        ("mpnet", "MPNetForMultipleChoice"),
+        ("nezha", "NezhaForMultipleChoice"),
+        ("nystromformer", "NystromformerForMultipleChoice"),
+        ("qdqbert", "QDQBertForMultipleChoice"),
+        ("rembert", "RemBertForMultipleChoice"),
+        ("roberta", "RobertaForMultipleChoice"),
+        ("roberta-prelayernorm", "RobertaPreLayerNormForMultipleChoice"),
+        ("roc_bert", "RoCBertForMultipleChoice"),
+        ("roformer", "RoFormerForMultipleChoice"),
+        ("squeezebert", "SqueezeBertForMultipleChoice"),
+        ("xlm", "XLMForMultipleChoice"),
+        ("xlm-roberta", "XLMRobertaForMultipleChoice"),
+        ("xlm-roberta-xl", "XLMRobertaXLForMultipleChoice"),
+        ("xlnet", "XLNetForMultipleChoice"),
+        ("xmod", "XmodForMultipleChoice"),
+        ("yoso", "YosoForMultipleChoice"),
+    ]
+)
+
+MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES = OrderedDict(
+    [
+        ("bert", "BertForNextSentencePrediction"),
+        ("ernie", "ErnieForNextSentencePrediction"),
+        ("fnet", "FNetForNextSentencePrediction"),
+        ("megatron-bert", "MegatronBertForNextSentencePrediction"),
+        ("mobilebert", "MobileBertForNextSentencePrediction"),
+        ("nezha", "NezhaForNextSentencePrediction"),
+        ("qdqbert", "QDQBertForNextSentencePrediction"),
+    ]
+)
+
+MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Audio Classification mapping
+        ("audio-spectrogram-transformer", "ASTForAudioClassification"),
+        ("data2vec-audio", "Data2VecAudioForSequenceClassification"),
+        ("hubert", "HubertForSequenceClassification"),
+        ("sew", "SEWForSequenceClassification"),
+        ("sew-d", "SEWDForSequenceClassification"),
+        ("unispeech", "UniSpeechForSequenceClassification"),
+        ("unispeech-sat", "UniSpeechSatForSequenceClassification"),
+        ("wav2vec2", "Wav2Vec2ForSequenceClassification"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerForSequenceClassification"),
+        ("wavlm", "WavLMForSequenceClassification"),
+        ("whisper", "WhisperForAudioClassification"),
+    ]
+)
+
+MODEL_FOR_CTC_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Connectionist temporal classification (CTC) mapping
+        ("data2vec-audio", "Data2VecAudioForCTC"),
+        ("hubert", "HubertForCTC"),
+        ("mctct", "MCTCTForCTC"),
+        ("sew", "SEWForCTC"),
+        ("sew-d", "SEWDForCTC"),
+        ("unispeech", "UniSpeechForCTC"),
+        ("unispeech-sat", "UniSpeechSatForCTC"),
+        ("wav2vec2", "Wav2Vec2ForCTC"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerForCTC"),
+        ("wavlm", "WavLMForCTC"),
+    ]
+)
+
+MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Audio Classification mapping
+        ("data2vec-audio", "Data2VecAudioForAudioFrameClassification"),
+        ("unispeech-sat", "UniSpeechSatForAudioFrameClassification"),
+        ("wav2vec2", "Wav2Vec2ForAudioFrameClassification"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerForAudioFrameClassification"),
+        ("wavlm", "WavLMForAudioFrameClassification"),
+    ]
+)
+
+MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Audio Classification mapping
+        ("data2vec-audio", "Data2VecAudioForXVector"),
+        ("unispeech-sat", "UniSpeechSatForXVector"),
+        ("wav2vec2", "Wav2Vec2ForXVector"),
+        ("wav2vec2-conformer", "Wav2Vec2ConformerForXVector"),
+        ("wavlm", "WavLMForXVector"),
+    ]
+)
+
+MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Zero Shot Image Classification mapping
+        ("align", "AlignModel"),
+        ("altclip", "AltCLIPModel"),
+        ("blip", "BlipModel"),
+        ("chinese_clip", "ChineseCLIPModel"),
+        ("clip", "CLIPModel"),
+        ("clipseg", "CLIPSegModel"),
+    ]
+)
+
+MODEL_FOR_BACKBONE_MAPPING_NAMES = OrderedDict(
+    [
+        # Backbone mapping
+        ("bit", "BitBackbone"),
+        ("convnext", "ConvNextBackbone"),
+        ("convnextv2", "ConvNextV2Backbone"),
+        ("dinat", "DinatBackbone"),
+        ("focalnet", "FocalNetBackbone"),
+        ("maskformer-swin", "MaskFormerSwinBackbone"),
+        ("nat", "NatBackbone"),
+        ("resnet", "ResNetBackbone"),
+        ("swin", "SwinBackbone"),
+    ]
+)
+
+MODEL_FOR_MASK_GENERATION_MAPPING_NAMES = OrderedDict(
+    [
+        ("sam", "SamModel"),
+    ]
+)
+
+MODEL_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_MAPPING_NAMES)
+MODEL_FOR_PRETRAINING_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_PRETRAINING_MAPPING_NAMES)
+MODEL_WITH_LM_HEAD_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_WITH_LM_HEAD_MAPPING_NAMES)
+MODEL_FOR_CAUSAL_LM_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_CAUSAL_LM_MAPPING_NAMES)
+MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES
+)
+MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_IMAGE_SEGMENTATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES
+)
+MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES
+)
+MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING_NAMES
+)
+MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES
+)
+MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_VISION_2_SEQ_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES)
+MODEL_FOR_VISUAL_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_VISUAL_QUESTION_ANSWERING_MAPPING_NAMES
+)
+MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES
+)
+MODEL_FOR_MASKED_LM_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_MASKED_LM_MAPPING_NAMES)
+MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES
+)
+MODEL_FOR_OBJECT_DETECTION_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES)
+MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES
+)
+MODEL_FOR_DEPTH_ESTIMATION_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES)
+MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES
+)
+MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES
+)
+MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES
+)
+MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_MULTIPLE_CHOICE_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES)
+MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES
+)
+MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_CTC_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_CTC_MAPPING_NAMES)
+MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES)
+MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES
+)
+MODEL_FOR_AUDIO_XVECTOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES)
+
+MODEL_FOR_BACKBONE_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_BACKBONE_MAPPING_NAMES)
+
+MODEL_FOR_MASK_GENERATION_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, MODEL_FOR_MASK_GENERATION_MAPPING_NAMES)
+
+
+class AutoModelForMaskGeneration(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_MASK_GENERATION_MAPPING
+
+
+class AutoModel(_BaseAutoModelClass):
+    _model_mapping = MODEL_MAPPING
+
+
+AutoModel = auto_class_update(AutoModel)
+
+
+class AutoModelForPreTraining(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_PRETRAINING_MAPPING
+
+
+AutoModelForPreTraining = auto_class_update(AutoModelForPreTraining, head_doc="pretraining")
+
+
+# Private on purpose, the public class will add the deprecation warnings.
+class _AutoModelWithLMHead(_BaseAutoModelClass):
+    _model_mapping = MODEL_WITH_LM_HEAD_MAPPING
+
+
+_AutoModelWithLMHead = auto_class_update(_AutoModelWithLMHead, head_doc="language modeling")
+
+
+class AutoModelForCausalLM(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_CAUSAL_LM_MAPPING
+
+
+AutoModelForCausalLM = auto_class_update(AutoModelForCausalLM, head_doc="causal language modeling")
+
+
+class AutoModelForMaskedLM(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_MASKED_LM_MAPPING
+
+
+AutoModelForMaskedLM = auto_class_update(AutoModelForMaskedLM, head_doc="masked language modeling")
+
+
+class AutoModelForSeq2SeqLM(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
+
+
+AutoModelForSeq2SeqLM = auto_class_update(
+    AutoModelForSeq2SeqLM, head_doc="sequence-to-sequence language modeling", checkpoint_for_example="t5-base"
+)
+
+
+class AutoModelForSequenceClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING
+
+
+AutoModelForSequenceClassification = auto_class_update(
+    AutoModelForSequenceClassification, head_doc="sequence classification"
+)
+
+
+class AutoModelForQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_QUESTION_ANSWERING_MAPPING
+
+
+AutoModelForQuestionAnswering = auto_class_update(AutoModelForQuestionAnswering, head_doc="question answering")
+
+
+class AutoModelForTableQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING
+
+
+AutoModelForTableQuestionAnswering = auto_class_update(
+    AutoModelForTableQuestionAnswering,
+    head_doc="table question answering",
+    checkpoint_for_example="google/tapas-base-finetuned-wtq",
+)
+
+
+class AutoModelForVisualQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_VISUAL_QUESTION_ANSWERING_MAPPING
+
+
+AutoModelForVisualQuestionAnswering = auto_class_update(
+    AutoModelForVisualQuestionAnswering,
+    head_doc="visual question answering",
+    checkpoint_for_example="dandelin/vilt-b32-finetuned-vqa",
+)
+
+
+class AutoModelForDocumentQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING
+
+
+AutoModelForDocumentQuestionAnswering = auto_class_update(
+    AutoModelForDocumentQuestionAnswering,
+    head_doc="document question answering",
+    checkpoint_for_example='impira/layoutlm-document-qa", revision="52e01b3',
+)
+
+
+class AutoModelForTokenClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING
+
+
+AutoModelForTokenClassification = auto_class_update(AutoModelForTokenClassification, head_doc="token classification")
+
+
+class AutoModelForMultipleChoice(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_MULTIPLE_CHOICE_MAPPING
+
+
+AutoModelForMultipleChoice = auto_class_update(AutoModelForMultipleChoice, head_doc="multiple choice")
+
+
+class AutoModelForNextSentencePrediction(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING
+
+
+AutoModelForNextSentencePrediction = auto_class_update(
+    AutoModelForNextSentencePrediction, head_doc="next sentence prediction"
+)
+
+
+class AutoModelForImageClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+
+
+AutoModelForImageClassification = auto_class_update(AutoModelForImageClassification, head_doc="image classification")
+
+
+class AutoModelForZeroShotImageClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING
+
+
+AutoModelForZeroShotImageClassification = auto_class_update(
+    AutoModelForZeroShotImageClassification, head_doc="zero-shot image classification"
+)
+
+
+class AutoModelForImageSegmentation(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_IMAGE_SEGMENTATION_MAPPING
+
+
+AutoModelForImageSegmentation = auto_class_update(AutoModelForImageSegmentation, head_doc="image segmentation")
+
+
+class AutoModelForSemanticSegmentation(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING
+
+
+AutoModelForSemanticSegmentation = auto_class_update(
+    AutoModelForSemanticSegmentation, head_doc="semantic segmentation"
+)
+
+
+class AutoModelForUniversalSegmentation(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING
+
+
+AutoModelForUniversalSegmentation = auto_class_update(
+    AutoModelForUniversalSegmentation, head_doc="universal image segmentation"
+)
+
+
+class AutoModelForInstanceSegmentation(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING
+
+
+AutoModelForInstanceSegmentation = auto_class_update(
+    AutoModelForInstanceSegmentation, head_doc="instance segmentation"
+)
+
+
+class AutoModelForObjectDetection(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_OBJECT_DETECTION_MAPPING
+
+
+AutoModelForObjectDetection = auto_class_update(AutoModelForObjectDetection, head_doc="object detection")
+
+
+class AutoModelForZeroShotObjectDetection(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING
+
+
+AutoModelForZeroShotObjectDetection = auto_class_update(
+    AutoModelForZeroShotObjectDetection, head_doc="zero-shot object detection"
+)
+
+
+class AutoModelForDepthEstimation(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_DEPTH_ESTIMATION_MAPPING
+
+
+AutoModelForDepthEstimation = auto_class_update(AutoModelForDepthEstimation, head_doc="depth estimation")
+
+
+class AutoModelForVideoClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING
+
+
+AutoModelForVideoClassification = auto_class_update(AutoModelForVideoClassification, head_doc="video classification")
+
+
+class AutoModelForVision2Seq(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_VISION_2_SEQ_MAPPING
+
+
+AutoModelForVision2Seq = auto_class_update(AutoModelForVision2Seq, head_doc="vision-to-text modeling")
+
+
+class AutoModelForAudioClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING
+
+
+AutoModelForAudioClassification = auto_class_update(AutoModelForAudioClassification, head_doc="audio classification")
+
+
+class AutoModelForCTC(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_CTC_MAPPING
+
+
+AutoModelForCTC = auto_class_update(AutoModelForCTC, head_doc="connectionist temporal classification")
+
+
+class AutoModelForSpeechSeq2Seq(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING
+
+
+AutoModelForSpeechSeq2Seq = auto_class_update(
+    AutoModelForSpeechSeq2Seq, head_doc="sequence-to-sequence speech-to-text modeling"
+)
+
+
+class AutoModelForAudioFrameClassification(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING
+
+
+AutoModelForAudioFrameClassification = auto_class_update(
+    AutoModelForAudioFrameClassification, head_doc="audio frame (token) classification"
+)
+
+
+class AutoModelForAudioXVector(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_AUDIO_XVECTOR_MAPPING
+
+
+class AutoBackbone(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_BACKBONE_MAPPING
+
+
+AutoModelForAudioXVector = auto_class_update(AutoModelForAudioXVector, head_doc="audio retrieval via x-vector")
+
+
+class AutoModelForMaskedImageModeling(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING
+
+
+AutoModelForMaskedImageModeling = auto_class_update(AutoModelForMaskedImageModeling, head_doc="masked image modeling")
+
+
+class AutoModelWithLMHead(_AutoModelWithLMHead):
+    @classmethod
+    def from_config(cls, config):
+        warnings.warn(
+            "The class `AutoModelWithLMHead` is deprecated and will be removed in a future version. Please use "
+            "`AutoModelForCausalLM` for causal language models, `AutoModelForMaskedLM` for masked language models and "
+            "`AutoModelForSeq2SeqLM` for encoder-decoder models.",
+            FutureWarning,
+        )
+        return super().from_config(config)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        warnings.warn(
+            "The class `AutoModelWithLMHead` is deprecated and will be removed in a future version. Please use "
+            "`AutoModelForCausalLM` for causal language models, `AutoModelForMaskedLM` for masked language models and "
+            "`AutoModelForSeq2SeqLM` for encoder-decoder models.",
+            FutureWarning,
+        )
+        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/modeling_flax_auto.py

@@ -0,0 +1,369 @@
+# coding=utf-8
+# Copyright 2018 The Google Flax Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Model class."""
+
+
+from collections import OrderedDict
+
+from ...utils import logging
+from .auto_factory import _BaseAutoModelClass, _LazyAutoMapping, auto_class_update
+from .configuration_auto import CONFIG_MAPPING_NAMES
+
+
+logger = logging.get_logger(__name__)
+
+
+FLAX_MODEL_MAPPING_NAMES = OrderedDict(
+    [
+        # Base model mapping
+        ("albert", "FlaxAlbertModel"),
+        ("bart", "FlaxBartModel"),
+        ("beit", "FlaxBeitModel"),
+        ("bert", "FlaxBertModel"),
+        ("big_bird", "FlaxBigBirdModel"),
+        ("blenderbot", "FlaxBlenderbotModel"),
+        ("blenderbot-small", "FlaxBlenderbotSmallModel"),
+        ("clip", "FlaxCLIPModel"),
+        ("distilbert", "FlaxDistilBertModel"),
+        ("electra", "FlaxElectraModel"),
+        ("gpt-sw3", "FlaxGPT2Model"),
+        ("gpt2", "FlaxGPT2Model"),
+        ("gpt_neo", "FlaxGPTNeoModel"),
+        ("gptj", "FlaxGPTJModel"),
+        ("longt5", "FlaxLongT5Model"),
+        ("marian", "FlaxMarianModel"),
+        ("mbart", "FlaxMBartModel"),
+        ("mt5", "FlaxMT5Model"),
+        ("opt", "FlaxOPTModel"),
+        ("pegasus", "FlaxPegasusModel"),
+        ("regnet", "FlaxRegNetModel"),
+        ("resnet", "FlaxResNetModel"),
+        ("roberta", "FlaxRobertaModel"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormModel"),
+        ("roformer", "FlaxRoFormerModel"),
+        ("t5", "FlaxT5Model"),
+        ("vision-text-dual-encoder", "FlaxVisionTextDualEncoderModel"),
+        ("vit", "FlaxViTModel"),
+        ("wav2vec2", "FlaxWav2Vec2Model"),
+        ("whisper", "FlaxWhisperModel"),
+        ("xglm", "FlaxXGLMModel"),
+        ("xlm-roberta", "FlaxXLMRobertaModel"),
+    ]
+)
+
+FLAX_MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for pre-training mapping
+        ("albert", "FlaxAlbertForPreTraining"),
+        ("bart", "FlaxBartForConditionalGeneration"),
+        ("bert", "FlaxBertForPreTraining"),
+        ("big_bird", "FlaxBigBirdForPreTraining"),
+        ("electra", "FlaxElectraForPreTraining"),
+        ("longt5", "FlaxLongT5ForConditionalGeneration"),
+        ("mbart", "FlaxMBartForConditionalGeneration"),
+        ("mt5", "FlaxMT5ForConditionalGeneration"),
+        ("roberta", "FlaxRobertaForMaskedLM"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForMaskedLM"),
+        ("roformer", "FlaxRoFormerForMaskedLM"),
+        ("t5", "FlaxT5ForConditionalGeneration"),
+        ("wav2vec2", "FlaxWav2Vec2ForPreTraining"),
+        ("whisper", "FlaxWhisperForConditionalGeneration"),
+        ("xlm-roberta", "FlaxXLMRobertaForMaskedLM"),
+    ]
+)
+
+FLAX_MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Masked LM mapping
+        ("albert", "FlaxAlbertForMaskedLM"),
+        ("bart", "FlaxBartForConditionalGeneration"),
+        ("bert", "FlaxBertForMaskedLM"),
+        ("big_bird", "FlaxBigBirdForMaskedLM"),
+        ("distilbert", "FlaxDistilBertForMaskedLM"),
+        ("electra", "FlaxElectraForMaskedLM"),
+        ("mbart", "FlaxMBartForConditionalGeneration"),
+        ("roberta", "FlaxRobertaForMaskedLM"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForMaskedLM"),
+        ("roformer", "FlaxRoFormerForMaskedLM"),
+        ("xlm-roberta", "FlaxXLMRobertaForMaskedLM"),
+    ]
+)
+
+FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Seq2Seq Causal LM mapping
+        ("bart", "FlaxBartForConditionalGeneration"),
+        ("blenderbot", "FlaxBlenderbotForConditionalGeneration"),
+        ("blenderbot-small", "FlaxBlenderbotSmallForConditionalGeneration"),
+        ("encoder-decoder", "FlaxEncoderDecoderModel"),
+        ("longt5", "FlaxLongT5ForConditionalGeneration"),
+        ("marian", "FlaxMarianMTModel"),
+        ("mbart", "FlaxMBartForConditionalGeneration"),
+        ("mt5", "FlaxMT5ForConditionalGeneration"),
+        ("pegasus", "FlaxPegasusForConditionalGeneration"),
+        ("t5", "FlaxT5ForConditionalGeneration"),
+    ]
+)
+
+FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Image-classsification
+        ("beit", "FlaxBeitForImageClassification"),
+        ("regnet", "FlaxRegNetForImageClassification"),
+        ("resnet", "FlaxResNetForImageClassification"),
+        ("vit", "FlaxViTForImageClassification"),
+    ]
+)
+
+FLAX_MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
+    [
+        ("vision-encoder-decoder", "FlaxVisionEncoderDecoderModel"),
+    ]
+)
+
+FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Causal LM mapping
+        ("bart", "FlaxBartForCausalLM"),
+        ("bert", "FlaxBertForCausalLM"),
+        ("big_bird", "FlaxBigBirdForCausalLM"),
+        ("electra", "FlaxElectraForCausalLM"),
+        ("gpt-sw3", "FlaxGPT2LMHeadModel"),
+        ("gpt2", "FlaxGPT2LMHeadModel"),
+        ("gpt_neo", "FlaxGPTNeoForCausalLM"),
+        ("gptj", "FlaxGPTJForCausalLM"),
+        ("opt", "FlaxOPTForCausalLM"),
+        ("roberta", "FlaxRobertaForCausalLM"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForCausalLM"),
+        ("xglm", "FlaxXGLMForCausalLM"),
+        ("xlm-roberta", "FlaxXLMRobertaForCausalLM"),
+    ]
+)
+
+FLAX_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Sequence Classification mapping
+        ("albert", "FlaxAlbertForSequenceClassification"),
+        ("bart", "FlaxBartForSequenceClassification"),
+        ("bert", "FlaxBertForSequenceClassification"),
+        ("big_bird", "FlaxBigBirdForSequenceClassification"),
+        ("distilbert", "FlaxDistilBertForSequenceClassification"),
+        ("electra", "FlaxElectraForSequenceClassification"),
+        ("mbart", "FlaxMBartForSequenceClassification"),
+        ("roberta", "FlaxRobertaForSequenceClassification"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForSequenceClassification"),
+        ("roformer", "FlaxRoFormerForSequenceClassification"),
+        ("xlm-roberta", "FlaxXLMRobertaForSequenceClassification"),
+    ]
+)
+
+FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Question Answering mapping
+        ("albert", "FlaxAlbertForQuestionAnswering"),
+        ("bart", "FlaxBartForQuestionAnswering"),
+        ("bert", "FlaxBertForQuestionAnswering"),
+        ("big_bird", "FlaxBigBirdForQuestionAnswering"),
+        ("distilbert", "FlaxDistilBertForQuestionAnswering"),
+        ("electra", "FlaxElectraForQuestionAnswering"),
+        ("mbart", "FlaxMBartForQuestionAnswering"),
+        ("roberta", "FlaxRobertaForQuestionAnswering"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForQuestionAnswering"),
+        ("roformer", "FlaxRoFormerForQuestionAnswering"),
+        ("xlm-roberta", "FlaxXLMRobertaForQuestionAnswering"),
+    ]
+)
+
+FLAX_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Token Classification mapping
+        ("albert", "FlaxAlbertForTokenClassification"),
+        ("bert", "FlaxBertForTokenClassification"),
+        ("big_bird", "FlaxBigBirdForTokenClassification"),
+        ("distilbert", "FlaxDistilBertForTokenClassification"),
+        ("electra", "FlaxElectraForTokenClassification"),
+        ("roberta", "FlaxRobertaForTokenClassification"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForTokenClassification"),
+        ("roformer", "FlaxRoFormerForTokenClassification"),
+        ("xlm-roberta", "FlaxXLMRobertaForTokenClassification"),
+    ]
+)
+
+FLAX_MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Multiple Choice mapping
+        ("albert", "FlaxAlbertForMultipleChoice"),
+        ("bert", "FlaxBertForMultipleChoice"),
+        ("big_bird", "FlaxBigBirdForMultipleChoice"),
+        ("distilbert", "FlaxDistilBertForMultipleChoice"),
+        ("electra", "FlaxElectraForMultipleChoice"),
+        ("roberta", "FlaxRobertaForMultipleChoice"),
+        ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForMultipleChoice"),
+        ("roformer", "FlaxRoFormerForMultipleChoice"),
+        ("xlm-roberta", "FlaxXLMRobertaForMultipleChoice"),
+    ]
+)
+
+FLAX_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES = OrderedDict(
+    [
+        ("bert", "FlaxBertForNextSentencePrediction"),
+    ]
+)
+
+FLAX_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict(
+    [
+        ("speech-encoder-decoder", "FlaxSpeechEncoderDecoderModel"),
+        ("whisper", "FlaxWhisperForConditionalGeneration"),
+    ]
+)
+
+FLAX_MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        ("whisper", "FlaxWhisperForAudioClassification"),
+    ]
+)
+
+FLAX_MODEL_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FLAX_MODEL_MAPPING_NAMES)
+FLAX_MODEL_FOR_PRETRAINING_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_PRETRAINING_MAPPING_NAMES)
+FLAX_MODEL_FOR_MASKED_LM_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_MASKED_LM_MAPPING_NAMES)
+FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_VISION_2_SEQ_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES)
+FLAX_MODEL_FOR_CAUSAL_LM_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES)
+FLAX_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_MULTIPLE_CHOICE_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES
+)
+FLAX_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, FLAX_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES
+)
+
+
+class FlaxAutoModel(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_MAPPING
+
+
+FlaxAutoModel = auto_class_update(FlaxAutoModel)
+
+
+class FlaxAutoModelForPreTraining(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_PRETRAINING_MAPPING
+
+
+FlaxAutoModelForPreTraining = auto_class_update(FlaxAutoModelForPreTraining, head_doc="pretraining")
+
+
+class FlaxAutoModelForCausalLM(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_CAUSAL_LM_MAPPING
+
+
+FlaxAutoModelForCausalLM = auto_class_update(FlaxAutoModelForCausalLM, head_doc="causal language modeling")
+
+
+class FlaxAutoModelForMaskedLM(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_MASKED_LM_MAPPING
+
+
+FlaxAutoModelForMaskedLM = auto_class_update(FlaxAutoModelForMaskedLM, head_doc="masked language modeling")
+
+
+class FlaxAutoModelForSeq2SeqLM(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
+
+
+FlaxAutoModelForSeq2SeqLM = auto_class_update(
+    FlaxAutoModelForSeq2SeqLM, head_doc="sequence-to-sequence language modeling", checkpoint_for_example="t5-base"
+)
+
+
+class FlaxAutoModelForSequenceClassification(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING
+
+
+FlaxAutoModelForSequenceClassification = auto_class_update(
+    FlaxAutoModelForSequenceClassification, head_doc="sequence classification"
+)
+
+
+class FlaxAutoModelForQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING
+
+
+FlaxAutoModelForQuestionAnswering = auto_class_update(FlaxAutoModelForQuestionAnswering, head_doc="question answering")
+
+
+class FlaxAutoModelForTokenClassification(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING
+
+
+FlaxAutoModelForTokenClassification = auto_class_update(
+    FlaxAutoModelForTokenClassification, head_doc="token classification"
+)
+
+
+class FlaxAutoModelForMultipleChoice(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_MULTIPLE_CHOICE_MAPPING
+
+
+FlaxAutoModelForMultipleChoice = auto_class_update(FlaxAutoModelForMultipleChoice, head_doc="multiple choice")
+
+
+class FlaxAutoModelForNextSentencePrediction(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING
+
+
+FlaxAutoModelForNextSentencePrediction = auto_class_update(
+    FlaxAutoModelForNextSentencePrediction, head_doc="next sentence prediction"
+)
+
+
+class FlaxAutoModelForImageClassification(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+
+
+FlaxAutoModelForImageClassification = auto_class_update(
+    FlaxAutoModelForImageClassification, head_doc="image classification"
+)
+
+
+class FlaxAutoModelForVision2Seq(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_VISION_2_SEQ_MAPPING
+
+
+FlaxAutoModelForVision2Seq = auto_class_update(FlaxAutoModelForVision2Seq, head_doc="vision-to-text modeling")
+
+
+class FlaxAutoModelForSpeechSeq2Seq(_BaseAutoModelClass):
+    _model_mapping = FLAX_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING
+
+
+FlaxAutoModelForSpeechSeq2Seq = auto_class_update(
+    FlaxAutoModelForSpeechSeq2Seq, head_doc="sequence-to-sequence speech-to-text modeling"
+)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/modeling_tf_auto.py

@@ -0,0 +1,668 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Model class."""
+
+
+import warnings
+from collections import OrderedDict
+
+from ...utils import logging
+from .auto_factory import _BaseAutoModelClass, _LazyAutoMapping, auto_class_update
+from .configuration_auto import CONFIG_MAPPING_NAMES
+
+
+logger = logging.get_logger(__name__)
+
+
+TF_MODEL_MAPPING_NAMES = OrderedDict(
+    [
+        # Base model mapping
+        ("albert", "TFAlbertModel"),
+        ("bart", "TFBartModel"),
+        ("bert", "TFBertModel"),
+        ("blenderbot", "TFBlenderbotModel"),
+        ("blenderbot-small", "TFBlenderbotSmallModel"),
+        ("blip", "TFBlipModel"),
+        ("camembert", "TFCamembertModel"),
+        ("clip", "TFCLIPModel"),
+        ("convbert", "TFConvBertModel"),
+        ("convnext", "TFConvNextModel"),
+        ("ctrl", "TFCTRLModel"),
+        ("cvt", "TFCvtModel"),
+        ("data2vec-vision", "TFData2VecVisionModel"),
+        ("deberta", "TFDebertaModel"),
+        ("deberta-v2", "TFDebertaV2Model"),
+        ("deit", "TFDeiTModel"),
+        ("distilbert", "TFDistilBertModel"),
+        ("dpr", "TFDPRQuestionEncoder"),
+        ("electra", "TFElectraModel"),
+        ("esm", "TFEsmModel"),
+        ("flaubert", "TFFlaubertModel"),
+        ("funnel", ("TFFunnelModel", "TFFunnelBaseModel")),
+        ("gpt-sw3", "TFGPT2Model"),
+        ("gpt2", "TFGPT2Model"),
+        ("gptj", "TFGPTJModel"),
+        ("groupvit", "TFGroupViTModel"),
+        ("hubert", "TFHubertModel"),
+        ("layoutlm", "TFLayoutLMModel"),
+        ("layoutlmv3", "TFLayoutLMv3Model"),
+        ("led", "TFLEDModel"),
+        ("longformer", "TFLongformerModel"),
+        ("lxmert", "TFLxmertModel"),
+        ("marian", "TFMarianModel"),
+        ("mbart", "TFMBartModel"),
+        ("mobilebert", "TFMobileBertModel"),
+        ("mobilevit", "TFMobileViTModel"),
+        ("mpnet", "TFMPNetModel"),
+        ("mt5", "TFMT5Model"),
+        ("openai-gpt", "TFOpenAIGPTModel"),
+        ("opt", "TFOPTModel"),
+        ("pegasus", "TFPegasusModel"),
+        ("regnet", "TFRegNetModel"),
+        ("rembert", "TFRemBertModel"),
+        ("resnet", "TFResNetModel"),
+        ("roberta", "TFRobertaModel"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormModel"),
+        ("roformer", "TFRoFormerModel"),
+        ("segformer", "TFSegformerModel"),
+        ("speech_to_text", "TFSpeech2TextModel"),
+        ("swin", "TFSwinModel"),
+        ("t5", "TFT5Model"),
+        ("tapas", "TFTapasModel"),
+        ("transfo-xl", "TFTransfoXLModel"),
+        ("vision-text-dual-encoder", "TFVisionTextDualEncoderModel"),
+        ("vit", "TFViTModel"),
+        ("vit_mae", "TFViTMAEModel"),
+        ("wav2vec2", "TFWav2Vec2Model"),
+        ("whisper", "TFWhisperModel"),
+        ("xglm", "TFXGLMModel"),
+        ("xlm", "TFXLMModel"),
+        ("xlm-roberta", "TFXLMRobertaModel"),
+        ("xlnet", "TFXLNetModel"),
+    ]
+)
+
+TF_MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for pre-training mapping
+        ("albert", "TFAlbertForPreTraining"),
+        ("bart", "TFBartForConditionalGeneration"),
+        ("bert", "TFBertForPreTraining"),
+        ("camembert", "TFCamembertForMaskedLM"),
+        ("ctrl", "TFCTRLLMHeadModel"),
+        ("distilbert", "TFDistilBertForMaskedLM"),
+        ("electra", "TFElectraForPreTraining"),
+        ("flaubert", "TFFlaubertWithLMHeadModel"),
+        ("funnel", "TFFunnelForPreTraining"),
+        ("gpt-sw3", "TFGPT2LMHeadModel"),
+        ("gpt2", "TFGPT2LMHeadModel"),
+        ("layoutlm", "TFLayoutLMForMaskedLM"),
+        ("lxmert", "TFLxmertForPreTraining"),
+        ("mobilebert", "TFMobileBertForPreTraining"),
+        ("mpnet", "TFMPNetForMaskedLM"),
+        ("openai-gpt", "TFOpenAIGPTLMHeadModel"),
+        ("roberta", "TFRobertaForMaskedLM"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForMaskedLM"),
+        ("t5", "TFT5ForConditionalGeneration"),
+        ("tapas", "TFTapasForMaskedLM"),
+        ("transfo-xl", "TFTransfoXLLMHeadModel"),
+        ("vit_mae", "TFViTMAEForPreTraining"),
+        ("xlm", "TFXLMWithLMHeadModel"),
+        ("xlm-roberta", "TFXLMRobertaForMaskedLM"),
+        ("xlnet", "TFXLNetLMHeadModel"),
+    ]
+)
+
+TF_MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
+    [
+        # Model with LM heads mapping
+        ("albert", "TFAlbertForMaskedLM"),
+        ("bart", "TFBartForConditionalGeneration"),
+        ("bert", "TFBertForMaskedLM"),
+        ("camembert", "TFCamembertForMaskedLM"),
+        ("convbert", "TFConvBertForMaskedLM"),
+        ("ctrl", "TFCTRLLMHeadModel"),
+        ("distilbert", "TFDistilBertForMaskedLM"),
+        ("electra", "TFElectraForMaskedLM"),
+        ("esm", "TFEsmForMaskedLM"),
+        ("flaubert", "TFFlaubertWithLMHeadModel"),
+        ("funnel", "TFFunnelForMaskedLM"),
+        ("gpt-sw3", "TFGPT2LMHeadModel"),
+        ("gpt2", "TFGPT2LMHeadModel"),
+        ("gptj", "TFGPTJForCausalLM"),
+        ("layoutlm", "TFLayoutLMForMaskedLM"),
+        ("led", "TFLEDForConditionalGeneration"),
+        ("longformer", "TFLongformerForMaskedLM"),
+        ("marian", "TFMarianMTModel"),
+        ("mobilebert", "TFMobileBertForMaskedLM"),
+        ("mpnet", "TFMPNetForMaskedLM"),
+        ("openai-gpt", "TFOpenAIGPTLMHeadModel"),
+        ("rembert", "TFRemBertForMaskedLM"),
+        ("roberta", "TFRobertaForMaskedLM"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForMaskedLM"),
+        ("roformer", "TFRoFormerForMaskedLM"),
+        ("speech_to_text", "TFSpeech2TextForConditionalGeneration"),
+        ("t5", "TFT5ForConditionalGeneration"),
+        ("tapas", "TFTapasForMaskedLM"),
+        ("transfo-xl", "TFTransfoXLLMHeadModel"),
+        ("whisper", "TFWhisperForConditionalGeneration"),
+        ("xlm", "TFXLMWithLMHeadModel"),
+        ("xlm-roberta", "TFXLMRobertaForMaskedLM"),
+        ("xlnet", "TFXLNetLMHeadModel"),
+    ]
+)
+
+TF_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Causal LM mapping
+        ("bert", "TFBertLMHeadModel"),
+        ("camembert", "TFCamembertForCausalLM"),
+        ("ctrl", "TFCTRLLMHeadModel"),
+        ("gpt-sw3", "TFGPT2LMHeadModel"),
+        ("gpt2", "TFGPT2LMHeadModel"),
+        ("gptj", "TFGPTJForCausalLM"),
+        ("openai-gpt", "TFOpenAIGPTLMHeadModel"),
+        ("opt", "TFOPTForCausalLM"),
+        ("rembert", "TFRemBertForCausalLM"),
+        ("roberta", "TFRobertaForCausalLM"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForCausalLM"),
+        ("roformer", "TFRoFormerForCausalLM"),
+        ("transfo-xl", "TFTransfoXLLMHeadModel"),
+        ("xglm", "TFXGLMForCausalLM"),
+        ("xlm", "TFXLMWithLMHeadModel"),
+        ("xlm-roberta", "TFXLMRobertaForCausalLM"),
+        ("xlnet", "TFXLNetLMHeadModel"),
+    ]
+)
+
+TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES = OrderedDict(
+    [
+        ("deit", "TFDeiTForMaskedImageModeling"),
+        ("swin", "TFSwinForMaskedImageModeling"),
+    ]
+)
+
+TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Image-classsification
+        ("convnext", "TFConvNextForImageClassification"),
+        ("cvt", "TFCvtForImageClassification"),
+        ("data2vec-vision", "TFData2VecVisionForImageClassification"),
+        ("deit", ("TFDeiTForImageClassification", "TFDeiTForImageClassificationWithTeacher")),
+        ("mobilevit", "TFMobileViTForImageClassification"),
+        ("regnet", "TFRegNetForImageClassification"),
+        ("resnet", "TFResNetForImageClassification"),
+        ("segformer", "TFSegformerForImageClassification"),
+        ("swin", "TFSwinForImageClassification"),
+        ("vit", "TFViTForImageClassification"),
+    ]
+)
+
+
+TF_MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Zero Shot Image Classification mapping
+        ("blip", "TFBlipModel"),
+        ("clip", "TFCLIPModel"),
+    ]
+)
+
+
+TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Semantic Segmentation mapping
+        ("data2vec-vision", "TFData2VecVisionForSemanticSegmentation"),
+        ("mobilevit", "TFMobileViTForSemanticSegmentation"),
+        ("segformer", "TFSegformerForSemanticSegmentation"),
+    ]
+)
+
+TF_MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
+    [
+        ("blip", "TFBlipForConditionalGeneration"),
+        ("vision-encoder-decoder", "TFVisionEncoderDecoderModel"),
+    ]
+)
+
+TF_MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Masked LM mapping
+        ("albert", "TFAlbertForMaskedLM"),
+        ("bert", "TFBertForMaskedLM"),
+        ("camembert", "TFCamembertForMaskedLM"),
+        ("convbert", "TFConvBertForMaskedLM"),
+        ("deberta", "TFDebertaForMaskedLM"),
+        ("deberta-v2", "TFDebertaV2ForMaskedLM"),
+        ("distilbert", "TFDistilBertForMaskedLM"),
+        ("electra", "TFElectraForMaskedLM"),
+        ("esm", "TFEsmForMaskedLM"),
+        ("flaubert", "TFFlaubertWithLMHeadModel"),
+        ("funnel", "TFFunnelForMaskedLM"),
+        ("layoutlm", "TFLayoutLMForMaskedLM"),
+        ("longformer", "TFLongformerForMaskedLM"),
+        ("mobilebert", "TFMobileBertForMaskedLM"),
+        ("mpnet", "TFMPNetForMaskedLM"),
+        ("rembert", "TFRemBertForMaskedLM"),
+        ("roberta", "TFRobertaForMaskedLM"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForMaskedLM"),
+        ("roformer", "TFRoFormerForMaskedLM"),
+        ("tapas", "TFTapasForMaskedLM"),
+        ("xlm", "TFXLMWithLMHeadModel"),
+        ("xlm-roberta", "TFXLMRobertaForMaskedLM"),
+    ]
+)
+
+TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Seq2Seq Causal LM mapping
+        ("bart", "TFBartForConditionalGeneration"),
+        ("blenderbot", "TFBlenderbotForConditionalGeneration"),
+        ("blenderbot-small", "TFBlenderbotSmallForConditionalGeneration"),
+        ("encoder-decoder", "TFEncoderDecoderModel"),
+        ("led", "TFLEDForConditionalGeneration"),
+        ("marian", "TFMarianMTModel"),
+        ("mbart", "TFMBartForConditionalGeneration"),
+        ("mt5", "TFMT5ForConditionalGeneration"),
+        ("pegasus", "TFPegasusForConditionalGeneration"),
+        ("t5", "TFT5ForConditionalGeneration"),
+    ]
+)
+
+TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict(
+    [
+        ("speech_to_text", "TFSpeech2TextForConditionalGeneration"),
+        ("whisper", "TFWhisperForConditionalGeneration"),
+    ]
+)
+
+TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Sequence Classification mapping
+        ("albert", "TFAlbertForSequenceClassification"),
+        ("bart", "TFBartForSequenceClassification"),
+        ("bert", "TFBertForSequenceClassification"),
+        ("camembert", "TFCamembertForSequenceClassification"),
+        ("convbert", "TFConvBertForSequenceClassification"),
+        ("ctrl", "TFCTRLForSequenceClassification"),
+        ("deberta", "TFDebertaForSequenceClassification"),
+        ("deberta-v2", "TFDebertaV2ForSequenceClassification"),
+        ("distilbert", "TFDistilBertForSequenceClassification"),
+        ("electra", "TFElectraForSequenceClassification"),
+        ("esm", "TFEsmForSequenceClassification"),
+        ("flaubert", "TFFlaubertForSequenceClassification"),
+        ("funnel", "TFFunnelForSequenceClassification"),
+        ("gpt-sw3", "TFGPT2ForSequenceClassification"),
+        ("gpt2", "TFGPT2ForSequenceClassification"),
+        ("gptj", "TFGPTJForSequenceClassification"),
+        ("layoutlm", "TFLayoutLMForSequenceClassification"),
+        ("layoutlmv3", "TFLayoutLMv3ForSequenceClassification"),
+        ("longformer", "TFLongformerForSequenceClassification"),
+        ("mobilebert", "TFMobileBertForSequenceClassification"),
+        ("mpnet", "TFMPNetForSequenceClassification"),
+        ("openai-gpt", "TFOpenAIGPTForSequenceClassification"),
+        ("rembert", "TFRemBertForSequenceClassification"),
+        ("roberta", "TFRobertaForSequenceClassification"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForSequenceClassification"),
+        ("roformer", "TFRoFormerForSequenceClassification"),
+        ("tapas", "TFTapasForSequenceClassification"),
+        ("transfo-xl", "TFTransfoXLForSequenceClassification"),
+        ("xlm", "TFXLMForSequenceClassification"),
+        ("xlm-roberta", "TFXLMRobertaForSequenceClassification"),
+        ("xlnet", "TFXLNetForSequenceClassification"),
+    ]
+)
+
+TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Question Answering mapping
+        ("albert", "TFAlbertForQuestionAnswering"),
+        ("bert", "TFBertForQuestionAnswering"),
+        ("camembert", "TFCamembertForQuestionAnswering"),
+        ("convbert", "TFConvBertForQuestionAnswering"),
+        ("deberta", "TFDebertaForQuestionAnswering"),
+        ("deberta-v2", "TFDebertaV2ForQuestionAnswering"),
+        ("distilbert", "TFDistilBertForQuestionAnswering"),
+        ("electra", "TFElectraForQuestionAnswering"),
+        ("flaubert", "TFFlaubertForQuestionAnsweringSimple"),
+        ("funnel", "TFFunnelForQuestionAnswering"),
+        ("gptj", "TFGPTJForQuestionAnswering"),
+        ("layoutlmv3", "TFLayoutLMv3ForQuestionAnswering"),
+        ("longformer", "TFLongformerForQuestionAnswering"),
+        ("mobilebert", "TFMobileBertForQuestionAnswering"),
+        ("mpnet", "TFMPNetForQuestionAnswering"),
+        ("rembert", "TFRemBertForQuestionAnswering"),
+        ("roberta", "TFRobertaForQuestionAnswering"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForQuestionAnswering"),
+        ("roformer", "TFRoFormerForQuestionAnswering"),
+        ("xlm", "TFXLMForQuestionAnsweringSimple"),
+        ("xlm-roberta", "TFXLMRobertaForQuestionAnswering"),
+        ("xlnet", "TFXLNetForQuestionAnsweringSimple"),
+    ]
+)
+TF_MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = OrderedDict([("wav2vec2", "TFWav2Vec2ForSequenceClassification")])
+
+TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        ("layoutlm", "TFLayoutLMForQuestionAnswering"),
+    ]
+)
+
+
+TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Table Question Answering mapping
+        ("tapas", "TFTapasForQuestionAnswering"),
+    ]
+)
+
+TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Token Classification mapping
+        ("albert", "TFAlbertForTokenClassification"),
+        ("bert", "TFBertForTokenClassification"),
+        ("camembert", "TFCamembertForTokenClassification"),
+        ("convbert", "TFConvBertForTokenClassification"),
+        ("deberta", "TFDebertaForTokenClassification"),
+        ("deberta-v2", "TFDebertaV2ForTokenClassification"),
+        ("distilbert", "TFDistilBertForTokenClassification"),
+        ("electra", "TFElectraForTokenClassification"),
+        ("esm", "TFEsmForTokenClassification"),
+        ("flaubert", "TFFlaubertForTokenClassification"),
+        ("funnel", "TFFunnelForTokenClassification"),
+        ("layoutlm", "TFLayoutLMForTokenClassification"),
+        ("layoutlmv3", "TFLayoutLMv3ForTokenClassification"),
+        ("longformer", "TFLongformerForTokenClassification"),
+        ("mobilebert", "TFMobileBertForTokenClassification"),
+        ("mpnet", "TFMPNetForTokenClassification"),
+        ("rembert", "TFRemBertForTokenClassification"),
+        ("roberta", "TFRobertaForTokenClassification"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForTokenClassification"),
+        ("roformer", "TFRoFormerForTokenClassification"),
+        ("xlm", "TFXLMForTokenClassification"),
+        ("xlm-roberta", "TFXLMRobertaForTokenClassification"),
+        ("xlnet", "TFXLNetForTokenClassification"),
+    ]
+)
+
+TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
+    [
+        # Model for Multiple Choice mapping
+        ("albert", "TFAlbertForMultipleChoice"),
+        ("bert", "TFBertForMultipleChoice"),
+        ("camembert", "TFCamembertForMultipleChoice"),
+        ("convbert", "TFConvBertForMultipleChoice"),
+        ("distilbert", "TFDistilBertForMultipleChoice"),
+        ("electra", "TFElectraForMultipleChoice"),
+        ("flaubert", "TFFlaubertForMultipleChoice"),
+        ("funnel", "TFFunnelForMultipleChoice"),
+        ("longformer", "TFLongformerForMultipleChoice"),
+        ("mobilebert", "TFMobileBertForMultipleChoice"),
+        ("mpnet", "TFMPNetForMultipleChoice"),
+        ("rembert", "TFRemBertForMultipleChoice"),
+        ("roberta", "TFRobertaForMultipleChoice"),
+        ("roberta-prelayernorm", "TFRobertaPreLayerNormForMultipleChoice"),
+        ("roformer", "TFRoFormerForMultipleChoice"),
+        ("xlm", "TFXLMForMultipleChoice"),
+        ("xlm-roberta", "TFXLMRobertaForMultipleChoice"),
+        ("xlnet", "TFXLNetForMultipleChoice"),
+    ]
+)
+
+TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES = OrderedDict(
+    [
+        ("bert", "TFBertForNextSentencePrediction"),
+        ("mobilebert", "TFMobileBertForNextSentencePrediction"),
+    ]
+)
+
+TF_MODEL_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_MAPPING_NAMES)
+TF_MODEL_FOR_PRETRAINING_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_FOR_PRETRAINING_MAPPING_NAMES)
+TF_MODEL_WITH_LM_HEAD_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_WITH_LM_HEAD_MAPPING_NAMES)
+TF_MODEL_FOR_CAUSAL_LM_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES)
+TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES
+)
+TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES
+)
+TF_MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES
+)
+TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES
+)
+TF_MODEL_FOR_VISION_2_SEQ_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES)
+TF_MODEL_FOR_MASKED_LM_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_FOR_MASKED_LM_MAPPING_NAMES)
+TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES
+)
+TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES
+)
+TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES
+)
+TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES
+)
+TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES
+)
+TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES
+)
+TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES
+)
+TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES
+)
+TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES
+)
+TF_MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING = _LazyAutoMapping(
+    CONFIG_MAPPING_NAMES, TF_MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES
+)
+
+
+class TFAutoModel(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_MAPPING
+
+
+TFAutoModel = auto_class_update(TFAutoModel)
+
+
+class TFAutoModelForAudioClassification(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING
+
+
+TFAutoModelForAudioClassification = auto_class_update(
+    TFAutoModelForAudioClassification, head_doc="audio classification"
+)
+
+
+class TFAutoModelForPreTraining(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_PRETRAINING_MAPPING
+
+
+TFAutoModelForPreTraining = auto_class_update(TFAutoModelForPreTraining, head_doc="pretraining")
+
+
+# Private on purpose, the public class will add the deprecation warnings.
+class _TFAutoModelWithLMHead(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_WITH_LM_HEAD_MAPPING
+
+
+_TFAutoModelWithLMHead = auto_class_update(_TFAutoModelWithLMHead, head_doc="language modeling")
+
+
+class TFAutoModelForCausalLM(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_CAUSAL_LM_MAPPING
+
+
+TFAutoModelForCausalLM = auto_class_update(TFAutoModelForCausalLM, head_doc="causal language modeling")
+
+
+class TFAutoModelForMaskedImageModeling(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING
+
+
+TFAutoModelForMaskedImageModeling = auto_class_update(
+    TFAutoModelForMaskedImageModeling, head_doc="masked image modeling"
+)
+
+
+class TFAutoModelForImageClassification(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+
+
+TFAutoModelForImageClassification = auto_class_update(
+    TFAutoModelForImageClassification, head_doc="image classification"
+)
+
+
+class TFAutoModelForZeroShotImageClassification(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING
+
+
+TFAutoModelForZeroShotImageClassification = auto_class_update(
+    TFAutoModelForZeroShotImageClassification, head_doc="zero-shot image classification"
+)
+
+
+class TFAutoModelForSemanticSegmentation(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING
+
+
+TF_AutoModelForSemanticSegmentation = auto_class_update(
+    TFAutoModelForSemanticSegmentation, head_doc="semantic segmentation"
+)
+
+
+class TFAutoModelForVision2Seq(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_VISION_2_SEQ_MAPPING
+
+
+TFAutoModelForVision2Seq = auto_class_update(TFAutoModelForVision2Seq, head_doc="vision-to-text modeling")
+
+
+class TFAutoModelForMaskedLM(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_MASKED_LM_MAPPING
+
+
+TFAutoModelForMaskedLM = auto_class_update(TFAutoModelForMaskedLM, head_doc="masked language modeling")
+
+
+class TFAutoModelForSeq2SeqLM(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
+
+
+TFAutoModelForSeq2SeqLM = auto_class_update(
+    TFAutoModelForSeq2SeqLM, head_doc="sequence-to-sequence language modeling", checkpoint_for_example="t5-base"
+)
+
+
+class TFAutoModelForSequenceClassification(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING
+
+
+TFAutoModelForSequenceClassification = auto_class_update(
+    TFAutoModelForSequenceClassification, head_doc="sequence classification"
+)
+
+
+class TFAutoModelForQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING
+
+
+TFAutoModelForQuestionAnswering = auto_class_update(TFAutoModelForQuestionAnswering, head_doc="question answering")
+
+
+class TFAutoModelForDocumentQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING
+
+
+TFAutoModelForDocumentQuestionAnswering = auto_class_update(
+    TFAutoModelForDocumentQuestionAnswering,
+    head_doc="document question answering",
+    checkpoint_for_example='impira/layoutlm-document-qa", revision="52e01b3',
+)
+
+
+class TFAutoModelForTableQuestionAnswering(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING
+
+
+TFAutoModelForTableQuestionAnswering = auto_class_update(
+    TFAutoModelForTableQuestionAnswering,
+    head_doc="table question answering",
+    checkpoint_for_example="google/tapas-base-finetuned-wtq",
+)
+
+
+class TFAutoModelForTokenClassification(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING
+
+
+TFAutoModelForTokenClassification = auto_class_update(
+    TFAutoModelForTokenClassification, head_doc="token classification"
+)
+
+
+class TFAutoModelForMultipleChoice(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING
+
+
+TFAutoModelForMultipleChoice = auto_class_update(TFAutoModelForMultipleChoice, head_doc="multiple choice")
+
+
+class TFAutoModelForNextSentencePrediction(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING
+
+
+TFAutoModelForNextSentencePrediction = auto_class_update(
+    TFAutoModelForNextSentencePrediction, head_doc="next sentence prediction"
+)
+
+
+class TFAutoModelForSpeechSeq2Seq(_BaseAutoModelClass):
+    _model_mapping = TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING
+
+
+TFAutoModelForSpeechSeq2Seq = auto_class_update(
+    TFAutoModelForSpeechSeq2Seq, head_doc="sequence-to-sequence speech-to-text modeling"
+)
+
+
+class TFAutoModelWithLMHead(_TFAutoModelWithLMHead):
+    @classmethod
+    def from_config(cls, config):
+        warnings.warn(
+            "The class `TFAutoModelWithLMHead` is deprecated and will be removed in a future version. Please use"
+            " `TFAutoModelForCausalLM` for causal language models, `TFAutoModelForMaskedLM` for masked language models"
+            " and `TFAutoModelForSeq2SeqLM` for encoder-decoder models.",
+            FutureWarning,
+        )
+        return super().from_config(config)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        warnings.warn(
+            "The class `TFAutoModelWithLMHead` is deprecated and will be removed in a future version. Please use"
+            " `TFAutoModelForCausalLM` for causal language models, `TFAutoModelForMaskedLM` for masked language models"
+            " and `TFAutoModelForSeq2SeqLM` for encoder-decoder models.",
+            FutureWarning,
+        )
+        return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/processing_auto.py

@@ -0,0 +1,310 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" AutoProcessor class."""
+import importlib
+import inspect
+import json
+from collections import OrderedDict
+
+# Build the list of all feature extractors
+from ...configuration_utils import PretrainedConfig
+from ...dynamic_module_utils import get_class_from_dynamic_module
+from ...feature_extraction_utils import FeatureExtractionMixin
+from ...image_processing_utils import ImageProcessingMixin
+from ...tokenization_utils import TOKENIZER_CONFIG_FILE
+from ...utils import FEATURE_EXTRACTOR_NAME, get_file_from_repo, logging
+from .auto_factory import _LazyAutoMapping
+from .configuration_auto import (
+    CONFIG_MAPPING_NAMES,
+    AutoConfig,
+    model_type_to_module_name,
+    replace_list_option_in_docstrings,
+)
+from .feature_extraction_auto import AutoFeatureExtractor
+from .image_processing_auto import AutoImageProcessor
+from .tokenization_auto import AutoTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+PROCESSOR_MAPPING_NAMES = OrderedDict(
+    [
+        ("align", "AlignProcessor"),
+        ("altclip", "AltCLIPProcessor"),
+        ("blip", "BlipProcessor"),
+        ("blip-2", "Blip2Processor"),
+        ("bridgetower", "BridgeTowerProcessor"),
+        ("chinese_clip", "ChineseCLIPProcessor"),
+        ("clap", "ClapProcessor"),
+        ("clip", "CLIPProcessor"),
+        ("clipseg", "CLIPSegProcessor"),
+        ("flava", "FlavaProcessor"),
+        ("git", "GitProcessor"),
+        ("groupvit", "CLIPProcessor"),
+        ("hubert", "Wav2Vec2Processor"),
+        ("layoutlmv2", "LayoutLMv2Processor"),
+        ("layoutlmv3", "LayoutLMv3Processor"),
+        ("markuplm", "MarkupLMProcessor"),
+        ("mgp-str", "MgpstrProcessor"),
+        ("oneformer", "OneFormerProcessor"),
+        ("owlvit", "OwlViTProcessor"),
+        ("pix2struct", "Pix2StructProcessor"),
+        ("sam", "SamProcessor"),
+        ("sew", "Wav2Vec2Processor"),
+        ("sew-d", "Wav2Vec2Processor"),
+        ("speech_to_text", "Speech2TextProcessor"),
+        ("speech_to_text_2", "Speech2Text2Processor"),
+        ("speecht5", "SpeechT5Processor"),
+        ("trocr", "TrOCRProcessor"),
+        ("tvlt", "TvltProcessor"),
+        ("unispeech", "Wav2Vec2Processor"),
+        ("unispeech-sat", "Wav2Vec2Processor"),
+        ("vilt", "ViltProcessor"),
+        ("vision-text-dual-encoder", "VisionTextDualEncoderProcessor"),
+        ("wav2vec2", "Wav2Vec2Processor"),
+        ("wav2vec2-conformer", "Wav2Vec2Processor"),
+        ("wavlm", "Wav2Vec2Processor"),
+        ("whisper", "WhisperProcessor"),
+        ("xclip", "XCLIPProcessor"),
+    ]
+)
+
+PROCESSOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, PROCESSOR_MAPPING_NAMES)
+
+
+def processor_class_from_name(class_name: str):
+    for module_name, processors in PROCESSOR_MAPPING_NAMES.items():
+        if class_name in processors:
+            module_name = model_type_to_module_name(module_name)
+
+            module = importlib.import_module(f".{module_name}", "transformers.models")
+            try:
+                return getattr(module, class_name)
+            except AttributeError:
+                continue
+
+    for processor in PROCESSOR_MAPPING._extra_content.values():
+        if getattr(processor, "__name__", None) == class_name:
+            return processor
+
+    # We did not fine the class, but maybe it's because a dep is missing. In that case, the class will be in the main
+    # init and we return the proper dummy to get an appropriate error message.
+    main_module = importlib.import_module("transformers")
+    if hasattr(main_module, class_name):
+        return getattr(main_module, class_name)
+
+    return None
+
+
+class AutoProcessor:
+    r"""
+    This is a generic processor class that will be instantiated as one of the processor classes of the library when
+    created with the [`AutoProcessor.from_pretrained`] class method.
+
+    This class cannot be instantiated directly using `__init__()` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "AutoProcessor is designed to be instantiated "
+            "using the `AutoProcessor.from_pretrained(pretrained_model_name_or_path)` method."
+        )
+
+    @classmethod
+    @replace_list_option_in_docstrings(PROCESSOR_MAPPING_NAMES)
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r"""
+        Instantiate one of the processor classes of the library from a pretrained model vocabulary.
+
+        The processor class to instantiate is selected based on the `model_type` property of the config object (either
+        passed as an argument or loaded from `pretrained_model_name_or_path` if possible):
+
+        List options
+
+        Params:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                This can be either:
+
+                - a string, the *model id* of a pretrained feature_extractor hosted inside a model repo on
+                  huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or
+                  namespaced under a user or organization name, like `dbmdz/bert-base-german-cased`.
+                - a path to a *directory* containing a processor files saved using the `save_pretrained()` method,
+                  e.g., `./my_model_directory/`.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model feature extractor should be cached if the
+                standard cache should not be used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force to (re-)download the feature extractor files and override the cached versions
+                if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received file. Attempts to resume the download if such a file
+                exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+                when running `huggingface-cli login` (stored in `~/.huggingface`).
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+                If `False`, then this function returns just the final feature extractor object. If `True`, then this
+                functions returns a `Tuple(feature_extractor, unused_kwargs)` where *unused_kwargs* is a dictionary
+                consisting of the key/value pairs whose keys are not feature extractor attributes: i.e., the part of
+                `kwargs` which has not been used to update `feature_extractor` and is otherwise ignored.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (`Dict[str, Any]`, *optional*):
+                The values in kwargs of any keys which are feature extractor attributes will be used to override the
+                loaded values. Behavior concerning key/value pairs whose keys are *not* feature extractor attributes is
+                controlled by the `return_unused_kwargs` keyword parameter.
+
+        <Tip>
+
+        Passing `use_auth_token=True` is required when you want to use a private model.
+
+        </Tip>
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoProcessor
+
+        >>> # Download processor from huggingface.co and cache.
+        >>> processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
+
+        >>> # If processor files are in a directory (e.g. processor was saved using *save_pretrained('./test/saved_model/')*)
+        >>> # processor = AutoProcessor.from_pretrained("./test/saved_model/")
+        ```"""
+        config = kwargs.pop("config", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+        kwargs["_from_auto"] = True
+
+        processor_class = None
+        processor_auto_map = None
+
+        # First, let's see if we have a preprocessor config.
+        # Filter the kwargs for `get_file_from_repo`.
+        get_file_from_repo_kwargs = {
+            key: kwargs[key] for key in inspect.signature(get_file_from_repo).parameters.keys() if key in kwargs
+        }
+        # Let's start by checking whether the processor class is saved in an image processor
+        preprocessor_config_file = get_file_from_repo(
+            pretrained_model_name_or_path, FEATURE_EXTRACTOR_NAME, **get_file_from_repo_kwargs
+        )
+        if preprocessor_config_file is not None:
+            config_dict, _ = ImageProcessingMixin.get_image_processor_dict(pretrained_model_name_or_path, **kwargs)
+            processor_class = config_dict.get("processor_class", None)
+            if "AutoProcessor" in config_dict.get("auto_map", {}):
+                processor_auto_map = config_dict["auto_map"]["AutoProcessor"]
+
+        # If not found, let's check whether the processor class is saved in a feature extractor config
+        if preprocessor_config_file is not None and processor_class is None:
+            config_dict, _ = FeatureExtractionMixin.get_feature_extractor_dict(pretrained_model_name_or_path, **kwargs)
+            processor_class = config_dict.get("processor_class", None)
+            if "AutoProcessor" in config_dict.get("auto_map", {}):
+                processor_auto_map = config_dict["auto_map"]["AutoProcessor"]
+
+        if processor_class is None:
+            # Next, let's check whether the processor class is saved in a tokenizer
+            tokenizer_config_file = get_file_from_repo(
+                pretrained_model_name_or_path, TOKENIZER_CONFIG_FILE, **get_file_from_repo_kwargs
+            )
+            if tokenizer_config_file is not None:
+                with open(tokenizer_config_file, encoding="utf-8") as reader:
+                    config_dict = json.load(reader)
+
+                processor_class = config_dict.get("processor_class", None)
+                if "AutoProcessor" in config_dict.get("auto_map", {}):
+                    processor_auto_map = config_dict["auto_map"]["AutoProcessor"]
+
+        if processor_class is None:
+            # Otherwise, load config, if it can be loaded.
+            if not isinstance(config, PretrainedConfig):
+                config = AutoConfig.from_pretrained(
+                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                )
+
+            # And check if the config contains the processor class.
+            processor_class = getattr(config, "processor_class", None)
+            if hasattr(config, "auto_map") and "AutoProcessor" in config.auto_map:
+                processor_auto_map = config.auto_map["AutoProcessor"]
+
+        if processor_class is not None:
+            # If we have custom code for a feature extractor, we get the proper class.
+            if processor_auto_map is not None:
+                if not trust_remote_code:
+                    raise ValueError(
+                        f"Loading {pretrained_model_name_or_path} requires you to execute the feature extractor file "
+                        "in that repo on your local machine. Make sure you have read the code there to avoid "
+                        "malicious use, then set the option `trust_remote_code=True` to remove this error."
+                    )
+
+                processor_class = get_class_from_dynamic_module(
+                    processor_auto_map, pretrained_model_name_or_path, **kwargs
+                )
+            else:
+                processor_class = processor_class_from_name(processor_class)
+
+            return processor_class.from_pretrained(
+                pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+            )
+
+        # Last try: we use the PROCESSOR_MAPPING.
+        if type(config) in PROCESSOR_MAPPING:
+            return PROCESSOR_MAPPING[type(config)].from_pretrained(pretrained_model_name_or_path, **kwargs)
+
+        # At this stage, there doesn't seem to be a `Processor` class available for this model, so let's try a
+        # tokenizer.
+        try:
+            return AutoTokenizer.from_pretrained(
+                pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+            )
+        except Exception:
+            try:
+                return AutoImageProcessor.from_pretrained(
+                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                )
+            except Exception:
+                pass
+
+            try:
+                return AutoFeatureExtractor.from_pretrained(
+                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                )
+            except Exception:
+                pass
+
+        raise ValueError(
+            f"Unrecognized processing class in {pretrained_model_name_or_path}. Can't instantiate a processor, a "
+            "tokenizer, an image processor or a feature extractor for this model. Make sure the repository contains"
+            "the files of at least one of those processing classes."
+        )
+
+    @staticmethod
+    def register(config_class, processor_class):
+        """
+        Register a new processor for this class.
+
+        Args:
+            config_class ([`PretrainedConfig`]):
+                The configuration corresponding to the model to register.
+            processor_class ([`FeatureExtractorMixin`]): The processor to register.
+        """
+        PROCESSOR_MAPPING.register(config_class, processor_class)

SupplementaryMaterial_Code/OPERA/transformers-4.29.2/src/transformers/models/auto/tokenization_auto.py

@@ -0,0 +1,767 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Auto Tokenizer class."""
+
+import importlib
+import json
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...dynamic_module_utils import get_class_from_dynamic_module
+from ...tokenization_utils import PreTrainedTokenizer
+from ...tokenization_utils_base import TOKENIZER_CONFIG_FILE
+from ...utils import cached_file, extract_commit_hash, is_sentencepiece_available, is_tokenizers_available, logging
+from ..encoder_decoder import EncoderDecoderConfig
+from .auto_factory import _LazyAutoMapping
+from .configuration_auto import (
+    CONFIG_MAPPING_NAMES,
+    AutoConfig,
+    config_class_to_model_type,
+    model_type_to_module_name,
+    replace_list_option_in_docstrings,
+)
+
+
+if is_tokenizers_available():
+    from ...tokenization_utils_fast import PreTrainedTokenizerFast
+else:
+    PreTrainedTokenizerFast = None
+
+
+logger = logging.get_logger(__name__)
+
+if TYPE_CHECKING:
+    # This significantly improves completion suggestion performance when
+    # the transformers package is used with Microsoft's Pylance language server.
+    TOKENIZER_MAPPING_NAMES: OrderedDict[str, Tuple[Optional[str], Optional[str]]] = OrderedDict()
+else:
+    TOKENIZER_MAPPING_NAMES = OrderedDict(
+        [
+            (
+                "albert",
+                (
+                    "AlbertTokenizer" if is_sentencepiece_available() else None,
+                    "AlbertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("align", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("bart", ("BartTokenizer", "BartTokenizerFast")),
+            (
+                "barthez",
+                (
+                    "BarthezTokenizer" if is_sentencepiece_available() else None,
+                    "BarthezTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("bartpho", ("BartphoTokenizer", None)),
+            ("bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("bert-generation", ("BertGenerationTokenizer" if is_sentencepiece_available() else None, None)),
+            ("bert-japanese", ("BertJapaneseTokenizer", None)),
+            ("bertweet", ("BertweetTokenizer", None)),
+            (
+                "big_bird",
+                (
+                    "BigBirdTokenizer" if is_sentencepiece_available() else None,
+                    "BigBirdTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("bigbird_pegasus", ("PegasusTokenizer", "PegasusTokenizerFast" if is_tokenizers_available() else None)),
+            ("biogpt", ("BioGptTokenizer", None)),
+            ("blenderbot", ("BlenderbotTokenizer", "BlenderbotTokenizerFast")),
+            ("blenderbot-small", ("BlenderbotSmallTokenizer", None)),
+            ("blip", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("blip-2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("bloom", (None, "BloomTokenizerFast" if is_tokenizers_available() else None)),
+            ("bridgetower", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("byt5", ("ByT5Tokenizer", None)),
+            (
+                "camembert",
+                (
+                    "CamembertTokenizer" if is_sentencepiece_available() else None,
+                    "CamembertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("canine", ("CanineTokenizer", None)),
+            ("chinese_clip", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "clap",
+                (
+                    "RobertaTokenizer",
+                    "RobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "clip",
+                (
+                    "CLIPTokenizer",
+                    "CLIPTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "clipseg",
+                (
+                    "CLIPTokenizer",
+                    "CLIPTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("codegen", ("CodeGenTokenizer", "CodeGenTokenizerFast" if is_tokenizers_available() else None)),
+            ("convbert", ("ConvBertTokenizer", "ConvBertTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "cpm",
+                (
+                    "CpmTokenizer" if is_sentencepiece_available() else None,
+                    "CpmTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("cpmant", ("CpmAntTokenizer", None)),
+            ("ctrl", ("CTRLTokenizer", None)),
+            ("data2vec-text", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("deberta", ("DebertaTokenizer", "DebertaTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "deberta-v2",
+                (
+                    "DebertaV2Tokenizer" if is_sentencepiece_available() else None,
+                    "DebertaV2TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("distilbert", ("DistilBertTokenizer", "DistilBertTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "dpr",
+                (
+                    "DPRQuestionEncoderTokenizer",
+                    "DPRQuestionEncoderTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("electra", ("ElectraTokenizer", "ElectraTokenizerFast" if is_tokenizers_available() else None)),
+            ("ernie", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("ernie_m", ("ErnieMTokenizer" if is_sentencepiece_available() else None, None)),
+            ("esm", ("EsmTokenizer", None)),
+            ("flaubert", ("FlaubertTokenizer", None)),
+            ("fnet", ("FNetTokenizer", "FNetTokenizerFast" if is_tokenizers_available() else None)),
+            ("fsmt", ("FSMTTokenizer", None)),
+            ("funnel", ("FunnelTokenizer", "FunnelTokenizerFast" if is_tokenizers_available() else None)),
+            ("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)),
+            ("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("gpt_bigcode", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("gpt_neo", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("gpt_neox", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
+            ("gpt_neox_japanese", ("GPTNeoXJapaneseTokenizer", None)),
+            ("gptj", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("gptsan-japanese", ("GPTSanJapaneseTokenizer", None)),
+            ("groupvit", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
+            ("herbert", ("HerbertTokenizer", "HerbertTokenizerFast" if is_tokenizers_available() else None)),
+            ("hubert", ("Wav2Vec2CTCTokenizer", None)),
+            ("ibert", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("jukebox", ("JukeboxTokenizer", None)),
+            ("layoutlm", ("LayoutLMTokenizer", "LayoutLMTokenizerFast" if is_tokenizers_available() else None)),
+            ("layoutlmv2", ("LayoutLMv2Tokenizer", "LayoutLMv2TokenizerFast" if is_tokenizers_available() else None)),
+            ("layoutlmv3", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),
+            ("layoutxlm", ("LayoutXLMTokenizer", "LayoutXLMTokenizerFast" if is_tokenizers_available() else None)),
+            ("led", ("LEDTokenizer", "LEDTokenizerFast" if is_tokenizers_available() else None)),
+            ("lilt", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "llama",
+                (
+                    "LlamaTokenizer" if is_sentencepiece_available() else None,
+                    "LlamaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("longformer", ("LongformerTokenizer", "LongformerTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "longt5",
+                (
+                    "T5Tokenizer" if is_sentencepiece_available() else None,
+                    "T5TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("luke", ("LukeTokenizer", None)),
+            ("lxmert", ("LxmertTokenizer", "LxmertTokenizerFast" if is_tokenizers_available() else None)),
+            ("m2m_100", ("M2M100Tokenizer" if is_sentencepiece_available() else None, None)),
+            ("marian", ("MarianTokenizer" if is_sentencepiece_available() else None, None)),
+            (
+                "mbart",
+                (
+                    "MBartTokenizer" if is_sentencepiece_available() else None,
+                    "MBartTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "mbart50",
+                (
+                    "MBart50Tokenizer" if is_sentencepiece_available() else None,
+                    "MBart50TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("mega", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            ("megatron-bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("mgp-str", ("MgpstrTokenizer", None)),
+            ("mluke", ("MLukeTokenizer" if is_sentencepiece_available() else None, None)),
+            ("mobilebert", ("MobileBertTokenizer", "MobileBertTokenizerFast" if is_tokenizers_available() else None)),
+            ("mpnet", ("MPNetTokenizer", "MPNetTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "mt5",
+                (
+                    "MT5Tokenizer" if is_sentencepiece_available() else None,
+                    "MT5TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("mvp", ("MvpTokenizer", "MvpTokenizerFast" if is_tokenizers_available() else None)),
+            ("nezha", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "nllb",
+                (
+                    "NllbTokenizer" if is_sentencepiece_available() else None,
+                    "NllbTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "nllb-moe",
+                (
+                    "NllbTokenizer" if is_sentencepiece_available() else None,
+                    "NllbTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "nystromformer",
+                (
+                    "AlbertTokenizer" if is_sentencepiece_available() else None,
+                    "AlbertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("oneformer", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
+            ("openai-gpt", ("OpenAIGPTTokenizer", "OpenAIGPTTokenizerFast" if is_tokenizers_available() else None)),
+            ("opt", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
+            ("owlvit", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "pegasus",
+                (
+                    "PegasusTokenizer" if is_sentencepiece_available() else None,
+                    "PegasusTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "pegasus_x",
+                (
+                    "PegasusTokenizer" if is_sentencepiece_available() else None,
+                    "PegasusTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "perceiver",
+                (
+                    "PerceiverTokenizer",
+                    None,
+                ),
+            ),
+            ("phobert", ("PhobertTokenizer", None)),
+            ("pix2struct", ("T5Tokenizer", "T5TokenizerFast" if is_tokenizers_available() else None)),
+            ("plbart", ("PLBartTokenizer" if is_sentencepiece_available() else None, None)),
+            ("prophetnet", ("ProphetNetTokenizer", None)),
+            ("qdqbert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("rag", ("RagTokenizer", None)),
+            ("realm", ("RealmTokenizer", "RealmTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "reformer",
+                (
+                    "ReformerTokenizer" if is_sentencepiece_available() else None,
+                    "ReformerTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "rembert",
+                (
+                    "RemBertTokenizer" if is_sentencepiece_available() else None,
+                    "RemBertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("retribert", ("RetriBertTokenizer", "RetriBertTokenizerFast" if is_tokenizers_available() else None)),
+            ("roberta", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "roberta-prelayernorm",
+                ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None),
+            ),
+            ("roc_bert", ("RoCBertTokenizer", None)),
+            ("roformer", ("RoFormerTokenizer", "RoFormerTokenizerFast" if is_tokenizers_available() else None)),
+            ("rwkv", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
+            ("speech_to_text", ("Speech2TextTokenizer" if is_sentencepiece_available() else None, None)),
+            ("speech_to_text_2", ("Speech2Text2Tokenizer", None)),
+            ("speecht5", ("SpeechT5Tokenizer" if is_sentencepiece_available() else None, None)),
+            ("splinter", ("SplinterTokenizer", "SplinterTokenizerFast")),
+            (
+                "squeezebert",
+                ("SqueezeBertTokenizer", "SqueezeBertTokenizerFast" if is_tokenizers_available() else None),
+            ),
+            (
+                "switch_transformers",
+                (
+                    "T5Tokenizer" if is_sentencepiece_available() else None,
+                    "T5TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "t5",
+                (
+                    "T5Tokenizer" if is_sentencepiece_available() else None,
+                    "T5TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("tapas", ("TapasTokenizer", None)),
+            ("tapex", ("TapexTokenizer", None)),
+            ("transfo-xl", ("TransfoXLTokenizer", None)),
+            ("vilt", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("visual_bert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("wav2vec2", ("Wav2Vec2CTCTokenizer", None)),
+            ("wav2vec2-conformer", ("Wav2Vec2CTCTokenizer", None)),
+            ("wav2vec2_phoneme", ("Wav2Vec2PhonemeCTCTokenizer", None)),
+            ("whisper", ("WhisperTokenizer", "WhisperTokenizerFast" if is_tokenizers_available() else None)),
+            ("xclip", ("CLIPTokenizer", "CLIPTokenizerFast" if is_tokenizers_available() else None)),
+            (
+                "xglm",
+                (
+                    "XGLMTokenizer" if is_sentencepiece_available() else None,
+                    "XGLMTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            ("xlm", ("XLMTokenizer", None)),
+            ("xlm-prophetnet", ("XLMProphetNetTokenizer" if is_sentencepiece_available() else None, None)),
+            (
+                "xlm-roberta",
+                (
+                    "XLMRobertaTokenizer" if is_sentencepiece_available() else None,
+                    "XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "xlm-roberta-xl",
+                (
+                    "XLMRobertaTokenizer" if is_sentencepiece_available() else None,
+                    "XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "xlnet",
+                (
+                    "XLNetTokenizer" if is_sentencepiece_available() else None,
+                    "XLNetTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "xmod",
+                (
+                    "XLMRobertaTokenizer" if is_sentencepiece_available() else None,
+                    "XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+            (
+                "yoso",
+                (
+                    "AlbertTokenizer" if is_sentencepiece_available() else None,
+                    "AlbertTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
+        ]
+    )
+
+TOKENIZER_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TOKENIZER_MAPPING_NAMES)
+
+CONFIG_TO_TYPE = {v: k for k, v in CONFIG_MAPPING_NAMES.items()}
+
+
+def tokenizer_class_from_name(class_name: str):
+    if class_name == "PreTrainedTokenizerFast":
+        return PreTrainedTokenizerFast
+
+    for module_name, tokenizers in TOKENIZER_MAPPING_NAMES.items():
+        if class_name in tokenizers:
+            module_name = model_type_to_module_name(module_name)
+
+            module = importlib.import_module(f".{module_name}", "transformers.models")
+            try:
+                return getattr(module, class_name)
+            except AttributeError:
+                continue
+
+    for config, tokenizers in TOKENIZER_MAPPING._extra_content.items():
+        for tokenizer in tokenizers:
+            if getattr(tokenizer, "__name__", None) == class_name:
+                return tokenizer
+
+    # We did not fine the class, but maybe it's because a dep is missing. In that case, the class will be in the main
+    # init and we return the proper dummy to get an appropriate error message.
+    main_module = importlib.import_module("transformers")
+    if hasattr(main_module, class_name):
+        return getattr(main_module, class_name)
+
+    return None
+
+
+def get_tokenizer_config(
+    pretrained_model_name_or_path: Union[str, os.PathLike],
+    cache_dir: Optional[Union[str, os.PathLike]] = None,
+    force_download: bool = False,
+    resume_download: bool = False,
+    proxies: Optional[Dict[str, str]] = None,
+    use_auth_token: Optional[Union[bool, str]] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
+    subfolder: str = "",
+    **kwargs,
+):
+    """
+    Loads the tokenizer configuration from a pretrained model tokenizer configuration.
+
+    Args:
+        pretrained_model_name_or_path (`str` or `os.PathLike`):
+            This can be either:
+
+            - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
+              huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
+              under a user or organization name, like `dbmdz/bert-base-german-cased`.
+            - a path to a *directory* containing a configuration file saved using the
+              [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
+
+        cache_dir (`str` or `os.PathLike`, *optional*):
+            Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+            cache should not be used.
+        force_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to force to (re-)download the configuration files and override the cached versions if they
+            exist.
+        resume_download (`bool`, *optional*, defaults to `False`):
+            Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
+        proxies (`Dict[str, str]`, *optional*):
+            A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+            'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+        use_auth_token (`str` or *bool*, *optional*):
+            The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+            when running `huggingface-cli login` (stored in `~/.huggingface`).
+        revision (`str`, *optional*, defaults to `"main"`):
+            The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+            git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+            identifier allowed by git.
+        local_files_only (`bool`, *optional*, defaults to `False`):
+            If `True`, will only try to load the tokenizer configuration from local files.
+        subfolder (`str`, *optional*, defaults to `""`):
+            In case the tokenizer config is located inside a subfolder of the model repo on huggingface.co, you can
+            specify the folder name here.
+
+    <Tip>
+
+    Passing `use_auth_token=True` is required when you want to use a private model.
+
+    </Tip>
+
+    Returns:
+        `Dict`: The configuration of the tokenizer.
+
+    Examples:
+
+    ```python
+    # Download configuration from huggingface.co and cache.
+    tokenizer_config = get_tokenizer_config("bert-base-uncased")
+    # This model does not have a tokenizer config so the result will be an empty dict.
+    tokenizer_config = get_tokenizer_config("xlm-roberta-base")
+
+    # Save a pretrained tokenizer locally and you can reload its config
+    from transformers import AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+    tokenizer.save_pretrained("tokenizer-test")
+    tokenizer_config = get_tokenizer_config("tokenizer-test")
+    ```"""
+    commit_hash = kwargs.get("_commit_hash", None)
+    resolved_config_file = cached_file(
+        pretrained_model_name_or_path,
+        TOKENIZER_CONFIG_FILE,
+        cache_dir=cache_dir,
+        force_download=force_download,
+        resume_download=resume_download,
+        proxies=proxies,
+        use_auth_token=use_auth_token,
+        revision=revision,
+        local_files_only=local_files_only,
+        subfolder=subfolder,
+        _raise_exceptions_for_missing_entries=False,
+        _raise_exceptions_for_connection_errors=False,
+        _commit_hash=commit_hash,
+    )
+    if resolved_config_file is None:
+        logger.info("Could not locate the tokenizer configuration file, will try to use the model config instead.")
+        return {}
+    commit_hash = extract_commit_hash(resolved_config_file, commit_hash)
+
+    with open(resolved_config_file, encoding="utf-8") as reader:
+        result = json.load(reader)
+    result["_commit_hash"] = commit_hash
+    return result
+
+
+class AutoTokenizer:
+    r"""
+    This is a generic tokenizer class that will be instantiated as one of the tokenizer classes of the library when
+    created with the [`AutoTokenizer.from_pretrained`] class method.
+
+    This class cannot be instantiated directly using `__init__()` (throws an error).
+    """
+
+    def __init__(self):
+        raise EnvironmentError(
+            "AutoTokenizer is designed to be instantiated "
+            "using the `AutoTokenizer.from_pretrained(pretrained_model_name_or_path)` method."
+        )
+
+    @classmethod
+    @replace_list_option_in_docstrings(TOKENIZER_MAPPING_NAMES)
+    def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
+        r"""
+        Instantiate one of the tokenizer classes of the library from a pretrained model vocabulary.
+
+        The tokenizer class to instantiate is selected based on the `model_type` property of the config object (either
+        passed as an argument or loaded from `pretrained_model_name_or_path` if possible), or when it's missing, by
+        falling back to using pattern matching on `pretrained_model_name_or_path`:
+
+        List options
+
+        Params:
+            pretrained_model_name_or_path (`str` or `os.PathLike`):
+                Can be either:
+
+                    - A string, the *model id* of a predefined tokenizer hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
+                      user or organization name, like `dbmdz/bert-base-german-cased`.
+                    - A path to a *directory* containing vocabulary files required by the tokenizer, for instance saved
+                      using the [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
+                    - A path or url to a single saved vocabulary file if and only if the tokenizer only requires a
+                      single vocabulary file (like Bert or XLNet), e.g.: `./my_model_directory/vocab.txt`. (Not
+                      applicable to all derived classes)
+            inputs (additional positional arguments, *optional*):
+                Will be passed along to the Tokenizer `__init__()` method.
+            config ([`PretrainedConfig`], *optional*)
+                The configuration object used to dertermine the tokenizer class to instantiate.
+            cache_dir (`str` or `os.PathLike`, *optional*):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download the model weights and configuration files and override the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+                identifier allowed by git.
+            subfolder (`str`, *optional*):
+                In case the relevant files are located inside a subfolder of the model repo on huggingface.co (e.g. for
+                facebook/rag-token-base), specify it here.
+            use_fast (`bool`, *optional*, defaults to `True`):
+                Use a [fast Rust-based tokenizer](https://huggingface.co/docs/tokenizers/index) if it is supported for
+                a given model. If a fast tokenizer is not available for a given model, a normal Python-based tokenizer
+                is returned instead.
+            tokenizer_type (`str`, *optional*):
+                Tokenizer type to be loaded.
+            trust_remote_code (`bool`, *optional*, defaults to `False`):
+                Whether or not to allow for custom models defined on the Hub in their own modeling files. This option
+                should only be set to `True` for repositories you trust and in which you have read the code, as it will
+                execute code present on the Hub on your local machine.
+            kwargs (additional keyword arguments, *optional*):
+                Will be passed to the Tokenizer `__init__()` method. Can be used to set special tokens like
+                `bos_token`, `eos_token`, `unk_token`, `sep_token`, `pad_token`, `cls_token`, `mask_token`,
+                `additional_special_tokens`. See parameters in the `__init__()` for more details.
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer
+
+        >>> # Download vocabulary from huggingface.co and cache.
+        >>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+
+        >>> # Download vocabulary from huggingface.co (user-uploaded) and cache.
+        >>> tokenizer = AutoTokenizer.from_pretrained("dbmdz/bert-base-german-cased")
+
+        >>> # If vocabulary files are in a directory (e.g. tokenizer was saved using *save_pretrained('./test/saved_model/')*)
+        >>> # tokenizer = AutoTokenizer.from_pretrained("./test/bert_saved_model/")
+
+        >>> # Download vocabulary from huggingface.co and define model-specific arguments
+        >>> tokenizer = AutoTokenizer.from_pretrained("roberta-base", add_prefix_space=True)
+        ```"""
+        config = kwargs.pop("config", None)
+        kwargs["_from_auto"] = True
+
+        use_fast = kwargs.pop("use_fast", True)
+        tokenizer_type = kwargs.pop("tokenizer_type", None)
+        trust_remote_code = kwargs.pop("trust_remote_code", False)
+
+        # First, let's see whether the tokenizer_type is passed so that we can leverage it
+        if tokenizer_type is not None:
+            tokenizer_class = None
+            tokenizer_class_tuple = TOKENIZER_MAPPING_NAMES.get(tokenizer_type, None)
+
+            if tokenizer_class_tuple is None:
+                raise ValueError(
+                    f"Passed `tokenizer_type` {tokenizer_type} does not exist. `tokenizer_type` should be one of "
+                    f"{', '.join(c for c in TOKENIZER_MAPPING_NAMES.keys())}."
+                )
+
+            tokenizer_class_name, tokenizer_fast_class_name = tokenizer_class_tuple
+
+            if use_fast:
+                if tokenizer_fast_class_name is not None:
+                    tokenizer_class = tokenizer_class_from_name(tokenizer_fast_class_name)
+                else:
+                    logger.warning(
+                        "`use_fast` is set to `True` but the tokenizer class does not have a fast version. "
+                        " Falling back to the slow version."
+                    )
+            if tokenizer_class is None:
+                tokenizer_class = tokenizer_class_from_name(tokenizer_class_name)
+
+            if tokenizer_class is None:
+                raise ValueError(f"Tokenizer class {tokenizer_class_name} is not currently imported.")
+
+            return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+
+        # Next, let's try to use the tokenizer_config file to get the tokenizer class.
+        tokenizer_config = get_tokenizer_config(pretrained_model_name_or_path, **kwargs)
+        if "_commit_hash" in tokenizer_config:
+            kwargs["_commit_hash"] = tokenizer_config["_commit_hash"]
+        config_tokenizer_class = tokenizer_config.get("tokenizer_class")
+        tokenizer_auto_map = None
+        if "auto_map" in tokenizer_config:
+            if isinstance(tokenizer_config["auto_map"], (tuple, list)):
+                # Legacy format for dynamic tokenizers
+                tokenizer_auto_map = tokenizer_config["auto_map"]
+            else:
+                tokenizer_auto_map = tokenizer_config["auto_map"].get("AutoTokenizer", None)
+
+        # If that did not work, let's try to use the config.
+        if config_tokenizer_class is None:
+            if not isinstance(config, PretrainedConfig):
+                config = AutoConfig.from_pretrained(
+                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                )
+            config_tokenizer_class = config.tokenizer_class
+            if hasattr(config, "auto_map") and "AutoTokenizer" in config.auto_map:
+                tokenizer_auto_map = config.auto_map["AutoTokenizer"]
+
+        # If we have the tokenizer class from the tokenizer config or the model config we're good!
+        if config_tokenizer_class is not None:
+            tokenizer_class = None
+            if tokenizer_auto_map is not None:
+                if not trust_remote_code:
+                    raise ValueError(
+                        f"Loading {pretrained_model_name_or_path} requires you to execute the tokenizer file in that"
+                        " repo on your local machine. Make sure you have read the code there to avoid malicious use,"
+                        " then set the option `trust_remote_code=True` to remove this error."
+                    )
+
+                if use_fast and tokenizer_auto_map[1] is not None:
+                    class_ref = tokenizer_auto_map[1]
+                else:
+                    class_ref = tokenizer_auto_map[0]
+                tokenizer_class = get_class_from_dynamic_module(class_ref, pretrained_model_name_or_path, **kwargs)
+
+            elif use_fast and not config_tokenizer_class.endswith("Fast"):
+                tokenizer_class_candidate = f"{config_tokenizer_class}Fast"
+                tokenizer_class = tokenizer_class_from_name(tokenizer_class_candidate)
+            if tokenizer_class is None:
+                tokenizer_class_candidate = config_tokenizer_class
+                tokenizer_class = tokenizer_class_from_name(tokenizer_class_candidate)
+
+            if tokenizer_class is None:
+                raise ValueError(
+                    f"Tokenizer class {tokenizer_class_candidate} does not exist or is not currently imported."
+                )
+            return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+
+        # Otherwise we have to be creative.
+        # if model is an encoder decoder, the encoder tokenizer class is used by default
+        if isinstance(config, EncoderDecoderConfig):
+            if type(config.decoder) is not type(config.encoder):  # noqa: E721
+                logger.warning(
+                    f"The encoder model config class: {config.encoder.__class__} is different from the decoder model "
+                    f"config class: {config.decoder.__class__}. It is not recommended to use the "
+                    "`AutoTokenizer.from_pretrained()` method in this case. Please use the encoder and decoder "
+                    "specific tokenizer classes."
+                )
+            config = config.encoder
+
+        model_type = config_class_to_model_type(type(config).__name__)
+        if model_type is not None:
+            tokenizer_class_py, tokenizer_class_fast = TOKENIZER_MAPPING[type(config)]
+            if tokenizer_class_fast and (use_fast or tokenizer_class_py is None):
+                return tokenizer_class_fast.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+            else:
+                if tokenizer_class_py is not None:
+                    return tokenizer_class_py.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+                else:
+                    raise ValueError(
+                        "This tokenizer cannot be instantiated. Please make sure you have `sentencepiece` installed "
+                        "in order to use this tokenizer."
+                    )
+
+        raise ValueError(
+            f"Unrecognized configuration class {config.__class__} to build an AutoTokenizer.\n"
+            f"Model type should be one of {', '.join(c.__name__ for c in TOKENIZER_MAPPING.keys())}."
+        )
+
+    def register(config_class, slow_tokenizer_class=None, fast_tokenizer_class=None):
+        """
+        Register a new tokenizer in this mapping.
+
+
+        Args:
+            config_class ([`PretrainedConfig`]):
+                The configuration corresponding to the model to register.
+            slow_tokenizer_class ([`PretrainedTokenizer`], *optional*):
+                The slow tokenizer to register.
+            slow_tokenizer_class ([`PretrainedTokenizerFast`], *optional*):
+                The fast tokenizer to register.
+        """
+        if slow_tokenizer_class is None and fast_tokenizer_class is None:
+            raise ValueError("You need to pass either a `slow_tokenizer_class` or a `fast_tokenizer_class")
+        if slow_tokenizer_class is not None and issubclass(slow_tokenizer_class, PreTrainedTokenizerFast):
+            raise ValueError("You passed a fast tokenizer in the `slow_tokenizer_class`.")
+        if fast_tokenizer_class is not None and issubclass(fast_tokenizer_class, PreTrainedTokenizer):
+            raise ValueError("You passed a slow tokenizer in the `fast_tokenizer_class`.")
+
+        if (
+            slow_tokenizer_class is not None
+            and fast_tokenizer_class is not None
+            and issubclass(fast_tokenizer_class, PreTrainedTokenizerFast)
+            and fast_tokenizer_class.slow_tokenizer_class != slow_tokenizer_class
+        ):
+            raise ValueError(
+                "The fast tokenizer class you are passing has a `slow_tokenizer_class` attribute that is not "
+                "consistent with the slow tokenizer class you passed (fast tokenizer has "
+                f"{fast_tokenizer_class.slow_tokenizer_class} and you passed {slow_tokenizer_class}. Fix one of those "
+                "so they match!"
+            )
+
+        # Avoid resetting a set slow/fast tokenizer if we are passing just the other ones.
+        if config_class in TOKENIZER_MAPPING._extra_content:
+            existing_slow, existing_fast = TOKENIZER_MAPPING[config_class]
+            if slow_tokenizer_class is None:
+                slow_tokenizer_class = existing_slow
+            if fast_tokenizer_class is None:
+                fast_tokenizer_class = existing_fast
+
+        TOKENIZER_MAPPING.register(config_class, (slow_tokenizer_class, fast_tokenizer_class))