File size: 12,992 Bytes
d9115cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 | import abc
from dataclasses import dataclass, field, fields
from typing import Any, Dict, List, Optional, Union
import torch
import transformers
from torch.utils.data import IterableDataset
from src.data.loading.components.iterators import RawDataIterator
class BaseDatasetConfig:
"""
Class to provide base typing for dataset configurations. Should be inherited by all dataset configurations.
"""
def __init__(self):
pass
def get(self, attribute: str, default=None):
return getattr(self, attribute, default)
class BaseDataloaderConfig:
"""
Class to provide base typing for dataloader configurations. Should be inherited by all dataloader configurations.
"""
def __init__(self):
pass
def get(self, attribute: str, default=None):
return getattr(self, attribute, default)
@dataclass
class SequenceDatasetConfig(BaseDatasetConfig):
"""The generic dataset configuration class for datasets of sequence data.
Parameters:
----------
user_id_field: str
The user id field name.
data_iterator: RawDataIterator
The raw data iterator.
preprocessing_functions: list[callable]
The list of preprocessing functions. Should be in the order they must be applied.
num_placeholder_tokens_map: Optional[dict]
The number of placeholder tokens map.
iterate_per_row: bool
Whether to iterate per row or per batches.
keep_user_id: bool
Whether to keep the user id feature in the batches.
field_type_map: Optional[dict]
The field type map.
min_sequence_length: int
The minimum sequence length. Only works if iterating per row.
feature_map: Optional[dict]
maps the feature names to the desired feature names.
iterate_per_row: bool
Whether to iterate per row or per batches.
features_to_consider: list[str]
List of features to consider. If not specified, consider all features.
file_format: str
The file format of the dataset files. If not specified, the data iterator's `get_file_suffix`
method will be used to determine the file format.
For example, if the data iterator reads tfrecord files at the first level of the data_folder,
this can be set to "tfrecord.gz". If we want to retrieve all tfrecord files in subdirectories as well,
we can set it to "*/*tfrecord.gz".
"""
user_id_field: str
data_iterator: RawDataIterator
preprocessing_functions: list[callable] # type: ignore
iterate_per_row: bool = False
keep_user_id: bool = False
num_placeholder_tokens_map: Optional[dict] = field(default_factory=dict)
field_type_map: Optional[dict] = field(default_factory=dict)
min_sequence_length: int = 10
feature_map: Optional[dict] = None
features_to_consider: list[str] = field(default_factory=list)
file_format: str = None
@dataclass
class SequenceDataloaderConfig(BaseDataloaderConfig):
"""The generic dataloader configuration class for datasets of sequence data.
Each instance of this class is run on one device.
Parameters:
----------
dataset_class: IterableDataset
The dataset class.
data_folder: str
Path to the folder containingthe dataset files.
dataset_config: SequenceDatasetConfig
The dataset configuration.
labels: Dict[str, callable]
A dictionary mapping from feature names to
batch_size_per_device: list[callable]
The batch size per dataloader, also per device (GPU).
num_workers: int
The number of workers per dataloader, also per device (GPU).
assign_files_by_size: Optional[dict]
Whether to assign files to workers by file size to balance computation
across workers.
oov_token: Optional[int]
The token used to represent OOV items.
masking_token: int
The token used to represent masked items.
collate_fn: callable
Collate function used to construct batches.
sequence_length: int = 200
The length of sequences the dataloader should return. If raw sequences
are shorter, the dataloader will pad them to reach sequence_length.
padding_token: int = 0
The token used for padding sequences.
drop_last: bool = True
Whether to drop the last batch if it is smaller than
batch_size_per_device.
pin_memory: bool = True
Whether to allocate memory on CPU to ensure data is always available for
fast transfer to GPU.
should_shuffle_rows: bool = False
Whether to shuffle rows between epochs.
persistent_workers: bool = False
Whether to maintain worker processes across epochs.
assign_all_files_per_worker: bool = False
Whether to assign all files to each worker.
(NOTE: this should only be activated for training, not for evaluation,
as it will cause the workers to have overlapping files.)
"""
dataset_class: IterableDataset
data_folder: str
dataset_config: SequenceDatasetConfig
batch_size_per_device: int
num_workers: int
assign_files_by_size: bool
masking_token: int
collate_fn: callable # type: ignore
labels: Dict[str, callable] = field(default_factory=dict) # type: ignore
oov_token: Optional[int] = -1
sequence_length: int = 200
padding_token: int = 0
drop_last: bool = True
pin_memory: bool = True
should_shuffle_rows: bool = False
persistent_workers: bool = False
timeout: int = 0
assign_all_files_per_worker: bool = False
@dataclass
class LabelFunctionOutput:
"""Class to unify the output of label functions, making it easier to merge those
into SequentialModelInputData and SequentialModuleLabelData
Parameters:
-----------
sequence: torch.Tensor
The sequence tensor of shape (batch_size, sequence_length).
labels: torch.Tensor
The labels tensor of shape (num_labels,).
label_location: torch.Tensor
The label location tensor of shape (num_labels, 2).
This is used to indicate the position of the labels in `sequence`.
"""
sequence: torch.Tensor
labels: torch.Tensor
label_location: torch.Tensor = None
attention_mask: torch.Tensor = None
@dataclass
class SequentialModuleLabelData:
"""The label data class used to wrap the label data for training/testing.
Parameters
----------
labels: Dict[str, torch.Tensor]
Dictionary of label_name to label tensor.
Label tensor is the shape of mask size # long tensor
label_location: Dict[str, torch.Tensor]
Dictionary of label_name to label location tensor.
Label location tensor is the shape of mask_size, 2 as it contains coordinates # long tensor
"""
labels: Dict[str, torch.Tensor] = field(default_factory=dict)
label_location: Dict[str, torch.Tensor] = field(default_factory=dict)
attention_mask: Dict[str, torch.Tensor] = field(default_factory=dict)
@dataclass
class SequentialModelInputData:
"""The model input data class used to wrap the model input data for training/testing.
Parameters
----------
user_id_list: Union[torch.Tensor, List[str], None]
Tensor or list of user_ids.
transformed_sequences: Dict[str, torch.Tensor]
Dictionary of sequence_name to sequence tensor.
Sequence tensor is (batch_size_per_device x sequence length)
mask: torch.Tensor
The mask for the sequence data.
(batch_size_per_device x sequence length)
"""
user_id_list: Union[torch.Tensor, List[str], None] = None
transformed_sequences: Dict[str, torch.Tensor] = field(default_factory=dict)
mask: torch.Tensor = (
None # Single mask if needed as all sequences are padded the same way.
)
@dataclass
class SemanticIDDatasetConfig(SequenceDatasetConfig):
"""The dataset configuration class used to store the dataset configuration for pipelines
that use semantic ids.
Note that this class inherits from SequenceDatasetConfig, thus inherits all of
its parameters.
Parameters:
-----------
semantic_id_map: Optional[Dict[str, torch.Tensor]]
The semantic id map from field name to a 2-D tensor.
keep_user_id: bool
Whether to keep the user id in the dataset. If set to True, the user id
will be included in the dataset and can be used for inference or evaluation.
"""
semantic_id_map: Optional[Dict[str, torch.Tensor]] = None
keep_user_id: bool = False
@dataclass
class TokenizerConfig:
"""The configuration class used to store the tokenizer configuration.
Parameters:
----------
tokenizer: transformers.PreTrainedTokenizer
The tokenizer.
max_length: int
The maximum length of the tokenized sequences.
padding: str
The padding strategy.
truncation: bool
Whether to truncate the sequences.
special_tokens: Optional[Dict[str, str]]
The special tokens.
add_special_tokens: bool
Whether to add special tokens.
postprocess_eos_token: Optional[bool]
Whether to postprocess the eos token.
"""
tokenizer: transformers.PreTrainedTokenizer
max_length: int
padding: str
truncation: bool
special_tokens: Optional[Dict[str, str]] = field(default_factory=dict)
add_special_tokens: bool = True
postprocess_eos_token: Optional[bool] = False
@dataclass
class ItemDatasetConfig(BaseDatasetConfig):
"""The configuration class used to store the item dataset configuration.
Parameters
----------
item_id_field: str
The item id field.
preprocessing_functions: list[callable]
The preprocessing functions to be applied to the data.
data_iterator: RawDataIterator
The data iterator.
iterate_per_row: bool
Whether to iterate per row or per batch.
keep_item_id: bool
Whether to keep the item id in the data.
features_to_consider: Optional[List[str]]
The features to consider.
feature_map: Optional[Dict[str, str]]
The map from raw feature name to the key to be used to store the feature in
ItemTextData.transformed_features.
field_type_map: Optional[Dict]
The map from field name to the type of the field.
embedding_map: Optional[Dict[str, torch.Tensor]]
The map from field name to the embedding tensor, where the embedding tensor
is N x d where N is the number of unique IDs in the field and d is the embedding
dimension.
num_placeholder_tokens_map: Optional[Dict[str, int]]
The map of the sparse ID field to the number of placeholder IDs in the field.
"""
item_id_field: str
preprocessing_functions: list[callable] # type: ignore
data_iterator: RawDataIterator
iterate_per_row: bool = True
keep_item_id: bool = True
features_to_consider: Optional[List[str]] = None
feature_map: Optional[Dict[str, str]] = None
field_type_map: Optional[Dict] = None
embedding_map: Optional[Dict[str, torch.Tensor]] = None
num_placeholder_tokens_map: Optional[Dict[str, int]] = None
@dataclass
class ItemDataloaderConfig(BaseDataloaderConfig):
dataset_class: IterableDataset
data_folder: str
dataset_config: ItemDatasetConfig
batch_size_per_device: int
num_workers: int
assign_files_by_size: bool
collate_fn: callable # type: ignore
drop_last: bool = True
pin_memory: bool = True
should_shuffle_rows: bool = False
persistent_workers: bool = False
timeout: int = 0
oov_token: Optional[int] = 0
limit_files: Optional[int] = None
assign_all_files_per_worker: bool = False
@dataclass
class ItemData:
"""The data class used to wrap a batch of item features.
Parameters
----------
item_ids: Union[torch.Tensor, List[str], None]
The item ids.
transformed_features: Dict[str, torch.Tensor]
The transformed features.
"""
item_ids: Union[torch.Tensor, List[str], None] = None
transformed_features: Dict[str, torch.Tensor] = field(default_factory=dict)
@dataclass
class ItemTextData(ItemData):
"""The data class used to wrap a batch of items with text features for training/testing.
It is a child class of ItemData, with additional text tokens and text masks.
Parameters
----------
text_tokens: Optional[torch.Tensor]
The text tokens.
text_masks: Optional[torch.Tensor]
The text masks.
"""
text_tokens: Optional[torch.Tensor] = None
text_masks: Optional[torch.Tensor] = None
|