File size: 14,083 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 | from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import torch
from torch.nn.utils.rnn import pad_sequence
from src.data.loading.components.interfaces import (
LabelFunctionOutput,
SequentialModelInputData,
SequentialModuleLabelData,
)
from src.data.loading.utils import combine_list_of_tensor_dicts, pad_or_trim_sequence
from src.utils.tensor_utils import extract_locations
from src.data.loading.components.interfaces import ItemData
def identity_collate_fn(batch: Any) -> Any:
"""The default collate function that does nothing."""
return batch
def collate_with_sid_causal_duplicate(
# batch can be a list or a dict
# this function is used to create the generate contiguous sequences as data augmentation to improve the performance
batch: Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]],
sequence_field_name: str,
sid_hierarchy: int,
labels: Dict[str, callable], # type: ignore
sequence_length: int = 200,
masking_token: int = 1,
padding_token: int = 0,
oov_token: Optional[
int
] = None, # If oov_token is passed, we remove it from the sequence
max_batch_size: int = 128,
) -> Tuple[SequentialModelInputData, SequentialModuleLabelData]:
"""
this collate fn is used to create the generate contiguous sequences as data augmentation to improve the performance.
It does three things
1. augment the input sequences by creating all possible contiguous sequences
2. random sample max_batch_size sequences from the augmented sequences to prevent OOM
3. run regular collate_fn_train
Parameters
----------
batch : Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]]
The batch of data to be collated. Can be a list of dictionaries, in the case we were
loading the data per row, or a dictionary of tensors, in the case we were loading the data per batch.
sequence_field_name : str
The name of the field in the batch that contains the sequence to be augmented.
sid_hierarchy : int
The length of Semantic IDs
labels : List[Dict[str, callable]]
The list of functions to apply to generate the labels.
sequence_length : int
The length of the sequence to be padded or trimmed to. (not used in this function, passed to collate_fn_train)
masking_token : int
The token used for masking. (not used in this function, passed to collate_fn_train)
padding_token : int
The token used for padding. (not used in this function, passed to collate_fn_train)
oov_token : Optional[int]
If oov_token is passed, we remove it from the sequence. (not used in this function, passed to collate_fn_train)
max_batch_size : int
The maximum batch size to be used after the data augmentation.
"""
if isinstance(batch, list):
batch = combine_list_of_tensor_dicts(batch) # type: ignore
# calculating the total number of contiguous sub-sequences in the batch
total_num_seqs = torch.sum(
(
(
k := torch.tensor([s.shape[0] for s in batch[sequence_field_name]])
// sid_hierarchy
)
- 1
)
* k
// 2
)
if total_num_seqs > max_batch_size:
select_seqs = torch.randint(
low=0,
high=total_num_seqs,
size=(max_batch_size,),
)
else:
select_seqs = torch.arange(total_num_seqs)
new_batch = {field_name: [] for field_name in batch}
current_idx = 0
for row_index, sequence in enumerate(batch[sequence_field_name]):
end_indices = torch.arange(
2 * sid_hierarchy, sequence.shape[0] + 1, sid_hierarchy
)
for end_index in end_indices:
start_indices = torch.arange(
0, end_index - 2 * sid_hierarchy + 1, sid_hierarchy
) # we have a -2 here because we want to have at least two items in the sequence
for start_index in start_indices:
if current_idx in select_seqs:
new_batch[sequence_field_name].append(
sequence[start_index:end_index]
)
for field_name in new_batch:
if field_name != sequence_field_name:
new_batch[field_name].append(batch[field_name][row_index])
current_idx += 1
return collate_fn_train(
batch=new_batch,
labels=labels,
sequence_length=sequence_length,
masking_token=masking_token,
padding_token=padding_token,
oov_token=oov_token,
)
def collate_fn_inference_for_sequence(
# batch can be a list or a dict
batch: Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]],
id_field_name: str,
sequence_length: int = 200,
padding_token: int = 0,
oov_token: Optional[
int
] = None, # If oov_token is passed, we remove it from the sequence
**kwargs,
) -> SequentialModelInputData:
"""The collate function passed to inference dataloader for inference with sequential data.
It handles id_field_name for saving model outputs
Parameters
----------
batch : Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]]
The batch of data to be collated. Can be a list of dictionaries, in the case we were
loading the data per row, or a dictionary of tensors, in the case we were loading the data per batch.
sequence_length : int
The length of the sequence to be padded or trimmed to.
masking_token : int
The token used for masking.
padding_token : int
The token used for padding.
oov_token : Optional[int]
If oov_token is passed, we remove it from the sequence.
id_field_name : str
The name of the field that contains the id of the user/item. This is used to
map the predictions back to the original id.
"""
if isinstance(batch, list):
batch = combine_list_of_tensor_dicts(batch) # type: ignore
model_input_data = SequentialModelInputData()
for field_name, field_sequence in batch.items(): # type: ignore
if field_name in id_field_name:
# We use the id field as the user_id_list so predictions can be mapped back to the original id.
model_input_data.user_id_list = field_sequence
# TODO (lneves): Allow for non-sequential data to be passed as a feature.
current_sequence = field_sequence # type: ignore
if oov_token:
# removing the oov token # TODO (Clark): in the future we can add special OOV handling
current_sequence = [
sequence[sequence != oov_token] for sequence in field_sequence
]
# 1. in-batch padding s.t. all sequences have the same length and in the format of pt tensor
current_sequence = pad_sequence(
current_sequence, batch_first=True, padding_value=padding_token
)
# 2. padding or trimming the sequence to the desired length for training
current_sequence = pad_or_trim_sequence(
padded_sequence=current_sequence,
sequence_length=sequence_length,
padding_token=padding_token,
)
model_input_data.transformed_sequences[field_name] = current_sequence
if field_name not in id_field_name and model_input_data.mask is None:
# if a field is not id, then it means its the real sequence we want calculate attention mask for it
model_input_data.mask = (current_sequence != padding_token).long()
return model_input_data # type: ignore
def collate_fn_train(
# batch can be a list or a dict
batch: Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]],
labels: Dict[str, callable], # type: ignore
sequence_length: int = 200,
masking_token: int = 1,
padding_token: int = 0,
oov_token: Optional[
int
] = None, # If oov_token is passed, we remove it from the sequence
data_augmentation_functions: Optional[
List[Dict[str, callable]]
] = None, # type: ignore
) -> Tuple[SequentialModelInputData, SequentialModuleLabelData]:
"""The collate function passed to dataloader. It can do training masking and padding for the input sequence.
Parameters
----------
batch : Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]]
The batch of data to be collated. Can be a list of dictionaries, in the case we were
loading the data per row, or a dictionary of tensors, in the case we were loading the data per batch.
labels : List[Dict[str, callable]]
The list of functions to apply to generate the labels.
sequence_length : int
The length of the sequence to be padded or trimmed to.
masking_token : int
The token used for masking.
padding_token : int
The token used for padding.
oov_token : Optional[int]
If oov_token is passed, we remove it from the sequence.
data_augmentation_functions : Optional[List[Dict[str, callable]]]
The list of functions to apply to augment the data.
"""
if isinstance(batch, list):
batch = combine_list_of_tensor_dicts(batch) # type: ignore
if data_augmentation_functions:
for data_augmentation_function in data_augmentation_functions:
batch = data_augmentation_function(batch)
model_input_data = SequentialModelInputData()
model_label_data = SequentialModuleLabelData()
for field_name, field_sequence in batch.items(): # type: ignore
# TODO (lneves): Allow for non-sequential data to be passed as a feature.
current_sequence = field_sequence # type: ignore
if oov_token:
# removing the oov token # TODO (Clark): in the future we can add special OOV handling
current_sequence = [
sequence[sequence != oov_token] for sequence in field_sequence
]
# 1. in-batch padding s.t. all sequences have the same length and in the format of pt tensor
current_sequence = pad_sequence(
current_sequence, batch_first=True, padding_value=padding_token
)
# 2. padding or trimming the sequence to the desired length for training
current_sequence = pad_or_trim_sequence(
padded_sequence=current_sequence,
sequence_length=sequence_length,
padding_token=padding_token,
)
# creating labels if the field is in the labels list
if field_name in labels:
label_function = labels[field_name].transform
label_function_output: LabelFunctionOutput = label_function.transform_label(
sequence=current_sequence,
padding_token=padding_token,
masking_token=masking_token,
)
model_label_data.labels[field_name] = label_function_output.labels
model_label_data.label_location[
field_name
] = label_function_output.label_location
model_label_data.attention_mask[
field_name
] = label_function_output.attention_mask
model_input_data.transformed_sequences[
field_name
] = label_function_output.sequence
else:
model_input_data.transformed_sequences[field_name] = current_sequence
# Currently supports a single masking per sequence
# TODO (lneves): Evaluate if this works or if we should have one mask per feature.
if model_input_data.mask is None:
model_input_data.mask = (current_sequence != padding_token).long()
return model_input_data, model_label_data # type: ignore
def collate_fn_items(
# batch can be a list or a dict
batch: Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]],
item_id_field: str,
feature_to_input_name: Dict[str, str], # type: ignore
) -> ItemData:
"""The collate function passed to the item dataloader.
Parameters
----------
batch : Union[List[Dict[str, torch.Tensor]], Dict[str, torch.Tensor]]
The batch of data to be collated. Can be a list of dictionaries, in the case we
loaded the data per row, or a dictionary of tensors, in the case loaded the data
per batch.
item_id_field : str
The name of the field in the batch that contains the item IDs.
feature_to_input_name : Dict[str, str]
The mapping from raw feature name to input feature name in ItemData.
Returns:
--------
model_input_data : ItemData
An ItemData object, which stores a batch of item features via a list of item IDs
in the field `item_ids` and a dictionary mapping feature names to value tensors
stacked along the batch dimension.
"""
if isinstance(batch, list):
batch = combine_list_of_tensor_dicts(batch) # type: ignore
# does not change shape of text tokens
model_input_data = ItemData()
for field_name, field_value in batch.items(): # type: ignore
if field_name == item_id_field:
model_input_data.item_ids = list(field_value)
else:
# In this case, field_value is a list of tensors, each representing the
# features of a single item. We stack these tensors along the batch
# dimension to create a single tensor for the batch of items.
field_value = torch.stack(field_value, dim=0)
model_input_data.transformed_features[
feature_to_input_name[field_name]
] = field_value
return model_input_data |