modeling_tf_utils.py 74.5 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# 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.
"""TF general model utils."""
Julien Plu's avatar
Julien Plu committed
17

18
import functools
Julien Plu's avatar
Julien Plu committed
19
import inspect
thomwolf's avatar
thomwolf committed
20
import os
21
import re
Julien Plu's avatar
Julien Plu committed
22
import warnings
Sylvain Gugger's avatar
Sylvain Gugger committed
23
from typing import Dict, List, Optional, Union
thomwolf's avatar
thomwolf committed
24

Aymeric Augustin's avatar
Aymeric Augustin committed
25
import h5py
Julien Chaumond's avatar
Julien Chaumond committed
26
import numpy as np
thomwolf's avatar
thomwolf committed
27
import tensorflow as tf
Julien Plu's avatar
Julien Plu committed
28
from tensorflow.python.keras import backend as K
thomwolf's avatar
thomwolf committed
29
from tensorflow.python.keras.saving import hdf5_format
thomwolf's avatar
thomwolf committed
30
31

from .configuration_utils import PretrainedConfig
Julien Plu's avatar
Julien Plu committed
32
33
34
35
36
37
38
39
40
from .file_utils import (
    DUMMY_INPUTS,
    TF2_WEIGHTS_NAME,
    WEIGHTS_NAME,
    ModelOutput,
    cached_path,
    hf_bucket_url,
    is_remote_url,
)
41
from .generation_tf_utils import TFGenerationMixin
Julien Plu's avatar
Julien Plu committed
42
from .tokenization_utils_base import BatchEncoding
Lysandre Debut's avatar
Lysandre Debut committed
43
from .utils import logging
thomwolf's avatar
thomwolf committed
44

Aymeric Augustin's avatar
Aymeric Augustin committed
45

Lysandre Debut's avatar
Lysandre Debut committed
46
logger = logging.get_logger(__name__)
47
tf_logger = tf.get_logger()
thomwolf's avatar
thomwolf committed
48

Julien Plu's avatar
Julien Plu committed
49
50
51
52
TFModelInputType = Union[
    List[tf.Tensor], List[np.ndarray], Dict[str, tf.Tensor], Dict[str, np.ndarray], np.ndarray, tf.Tensor
]

53

54
class TFModelUtilsMixin:
Julien Chaumond's avatar
Julien Chaumond committed
55
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
56
    A few utilities for :obj:`tf.keras.Model`, to be used as a mixin.
Julien Chaumond's avatar
Julien Chaumond committed
57
58
59
60
    """

    def num_parameters(self, only_trainable: bool = False) -> int:
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
61
62
63
64
65
66
67
68
        Get the number of (optionally, trainable) parameters in the model.

        Args:
            only_trainable (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to return only the number of trainable parameters

        Returns:
            :obj:`int`: The number of parameters.
Julien Chaumond's avatar
Julien Chaumond committed
69
70
71
72
73
74
75
        """
        if only_trainable:
            return int(sum(np.prod(w.shape.as_list()) for w in self.trainable_variables))
        else:
            return self.count_params()


76
def keras_serializable(cls):
77
78
79
80
    """
    Decorate a Keras Layer class to support Keras serialization.

    This is done by:
Sylvain Gugger's avatar
Sylvain Gugger committed
81
82
83
84
85

    1. Adding a :obj:`transformers_config` dict to the Keras config dictionary in :obj:`get_config` (called by Keras at
       serialization time.
    2. Wrapping :obj:`__init__` to accept that :obj:`transformers_config` dict (passed by Keras at deserialization
       time) and convert it to a config object for the actual layer initializer.
Sylvain Gugger's avatar
Sylvain Gugger committed
86
87
    3. Registering the class as a custom object in Keras (if the Tensorflow version supports this), so that it does not
       need to be supplied in :obj:`custom_objects` in the call to :obj:`tf.keras.models.load_model`.
Sylvain Gugger's avatar
Sylvain Gugger committed
88
89
90
91
92
93
94
95

    Args:
        cls (a :obj:`tf.keras.layers.Layers subclass`):
            Typically a :obj:`TF.MainLayer` class in this project, in general must accept a :obj:`config` argument to
            its initializer.

    Returns:
        The same class object, with modifications for Keras deserialization.
96
    """
97
    initializer = cls.__init__
98

99
100
101
102
    config_class = getattr(cls, "config_class", None)
    if config_class is None:
        raise AttributeError("Must set `config_class` to use @keras_serializable")

103
    @functools.wraps(initializer)
104
    def wrapped_init(self, *args, **kwargs):
105
106
107
108
        config = args[0] if args and isinstance(args[0], PretrainedConfig) else kwargs.pop("config", None)

        if isinstance(config, dict):
            config = config_class.from_dict(config)
109
            initializer(self, config, *args, **kwargs)
110
111
112
113
114
        elif isinstance(config, PretrainedConfig):
            if len(args) > 0:
                initializer(self, *args, **kwargs)
            else:
                initializer(self, config, *args, **kwargs)
115
        else:
116
117
118
            raise ValueError("Must pass either `config` (PretrainedConfig) or `config` (dict)")

        self._config = config
Julien Plu's avatar
Julien Plu committed
119
        self._kwargs = kwargs
120

121
122
123
124
125
126
127
128
    cls.__init__ = wrapped_init

    if not hasattr(cls, "get_config"):
        raise TypeError("Only use @keras_serializable on tf.keras.layers.Layer subclasses")
    if hasattr(cls.get_config, "_is_default"):

        def get_config(self):
            cfg = super(cls, self).get_config()
129
            cfg["config"] = self._config.to_dict()
Julien Plu's avatar
Julien Plu committed
130
            cfg.update(self._kwargs)
131
132
133
134
            return cfg

        cls.get_config = get_config

135
    cls._keras_serializable = True
136
137
138
    if hasattr(tf.keras.utils, "register_keras_serializable"):
        cls = tf.keras.utils.register_keras_serializable()(cls)
    return cls
139
140


141
class TFCausalLanguageModelingLoss:
Sylvain Gugger's avatar
Sylvain Gugger committed
142
143
144
145
146
147
148
149
150
    """
    Loss function suitable for causal language modeling (CLM), that is, the task of guessing the next token.

    .. note::

        Any label of -100 will be ignored (along with the corresponding logits) in the loss computation.

    """

151
152
153
154
    def compute_loss(self, labels, logits):
        loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
        )
Muennighoff's avatar
Muennighoff committed
155
        # make sure only labels that are not equal to -100 affect the loss
156
        active_loss = tf.not_equal(tf.reshape(labels, (-1,)), -100)
157
158
159
160
161
        reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, shape_list(logits)[2])), active_loss)
        labels = tf.boolean_mask(tf.reshape(labels, (-1,)), active_loss)
        return loss_fn(labels, reduced_logits)


Julien Plu's avatar
Julien Plu committed
162
class TFQuestionAnsweringLoss:
Sylvain Gugger's avatar
Sylvain Gugger committed
163
    """
164
    Loss function suitable for question answering.
Sylvain Gugger's avatar
Sylvain Gugger committed
165
166
    """

Julien Plu's avatar
Julien Plu committed
167
168
169
170
171
172
173
174
175
176
177
    def compute_loss(self, labels, logits):
        loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
        )
        start_loss = loss_fn(labels["start_position"], logits[0])
        end_loss = loss_fn(labels["end_position"], logits[1])

        return (start_loss + end_loss) / 2.0


class TFTokenClassificationLoss:
Sylvain Gugger's avatar
Sylvain Gugger committed
178
179
180
181
182
183
184
185
186
    """
    Loss function suitable for token classification.

    .. note::

        Any label of -100 will be ignored (along with the corresponding logits) in the loss computation.

    """

Julien Plu's avatar
Julien Plu committed
187
188
189
190
    def compute_loss(self, labels, logits):
        loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
        )
191
192
        # make sure only labels that are not equal to -100
        # are taken into account as loss
193
        if tf.math.reduce_any(labels == -1):
Julien Plu's avatar
Julien Plu committed
194
195
196
197
            warnings.warn("Using `-1` to mask the loss for the token is deprecated. Please use `-100` instead.")
            active_loss = tf.reshape(labels, (-1,)) != -1
        else:
            active_loss = tf.reshape(labels, (-1,)) != -100
Julien Plu's avatar
Julien Plu committed
198
199
200
201
202
203
204
        reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, shape_list(logits)[2])), active_loss)
        labels = tf.boolean_mask(tf.reshape(labels, (-1,)), active_loss)

        return loss_fn(labels, reduced_logits)


class TFSequenceClassificationLoss:
Sylvain Gugger's avatar
Sylvain Gugger committed
205
206
207
208
    """
    Loss function suitable for sequence classification.
    """

Julien Plu's avatar
Julien Plu committed
209
    def compute_loss(self, labels, logits):
210
        if len(shape_list(logits)) == 1 or shape_list(logits)[1] == 1:
Julien Plu's avatar
Julien Plu committed
211
212
213
214
215
216
217
218
219
            loss_fn = tf.keras.losses.MeanSquaredError(reduction=tf.keras.losses.Reduction.NONE)
        else:
            loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
                from_logits=True, reduction=tf.keras.losses.Reduction.NONE
            )

        return loss_fn(labels, logits)


Sylvain Gugger's avatar
Sylvain Gugger committed
220
221
222
223
224
225
class TFMultipleChoiceLoss(TFSequenceClassificationLoss):
    """Loss function suitable for multiple choice tasks."""


class TFMaskedLanguageModelingLoss(TFCausalLanguageModelingLoss):
    """
Lysandre's avatar
Lysandre committed
226
    Loss function suitable for masked language modeling (MLM), that is, the task of guessing the masked tokens.
Sylvain Gugger's avatar
Sylvain Gugger committed
227

Lysandre's avatar
Lysandre committed
228
    .. note::
Sylvain Gugger's avatar
Sylvain Gugger committed
229

Lysandre's avatar
Lysandre committed
230
231
         Any label of -100 will be ignored (along with the corresponding logits) in the loss computation.
    """
Julien Plu's avatar
Julien Plu committed
232
233


234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
class TFNextSentencePredictionLoss:
    """
    Loss function suitable for next sentence prediction (NSP), that is, the task of guessing the next sentence.

    .. note::
         Any label of -100 will be ignored (along with the corresponding logits) in the loss computation.
    """

    def compute_loss(self, labels, logits):
        loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
        )
        # make sure only labels that are not equal to -100
        # are taken into account as loss
        next_sentence_active_loss = tf.not_equal(tf.reshape(labels, (-1,)), -100)
        next_sentence_reduced_logits = tf.boolean_mask(tf.reshape(logits, (-1, 2)), next_sentence_active_loss)
        next_sentence_label = tf.boolean_mask(tf.reshape(labels, (-1,)), next_sentence_active_loss)

        return loss_fn(next_sentence_label, next_sentence_reduced_logits)


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
def booleans_processing(config, **kwargs):
    """
    Process the input booleans of each model in order to be sure they are compliant with the execution mode (eager or
    graph)

    Args:
        config (:class:`~transformers.PretrainedConfig`):
            The config of the running model.
        **kwargs:
            The boolean parameters

    Returns:
        A dictionary with the proper values for each boolean
    """
    final_booleans = {}

    if tf.executing_eagerly():
        final_booleans["output_attentions"] = (
            kwargs["output_attentions"] if kwargs["output_attentions"] is not None else config.output_attentions
        )
        final_booleans["output_hidden_states"] = (
            kwargs["output_hidden_states"]
            if kwargs["output_hidden_states"] is not None
            else config.output_hidden_states
        )
Julien Plu's avatar
Julien Plu committed
280
281
282
        final_booleans["return_dict"] = (
            kwargs["return_dict"] if kwargs["return_dict"] is not None else config.return_dict
        )
283
284
285

        if "use_cache" in kwargs:
            final_booleans["use_cache"] = kwargs["use_cache"] if kwargs["use_cache"] is not None else config.use_cache
Julien Plu's avatar
Julien Plu committed
286

287
288
289
290
291
292
    else:
        if (
            kwargs["output_attentions"] is not None
            or kwargs["output_hidden_states"] is not None
            or ("use_cache" in kwargs and kwargs["use_cache"] is not None)
        ):
293
            tf_logger.warn(
294
295
296
297
298
299
300
                "The parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model."
                "They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`)."
            )

        final_booleans["output_attentions"] = config.output_attentions
        final_booleans["output_hidden_states"] = config.output_hidden_states

Julien Plu's avatar
Julien Plu committed
301
        if kwargs["return_dict"] is not None:
302
            tf_logger.warn("The parameter `return_dict` cannot be set in graph mode and will always be set to `True`.")
Julien Plu's avatar
Julien Plu committed
303
        final_booleans["return_dict"] = True
304
305
306
307
308
309
310
311
312

        if "use_cache" in kwargs:
            final_booleans["use_cache"] = config.use_cache

    return final_booleans


def input_processing(func, config, input_ids, **kwargs):
    """
Julien Plu's avatar
Julien Plu committed
313
314
315
    Process the input of each TensorFlow model including the booleans. In case of a list of symbolic inputs, each input
    has to be named accordingly to the parameters name, i.e. `input_ids = tf.keras.Input(shape=(128,), dtype='int32',
    name="input_ids")` otherwise the order of the tensors will not be guaranteed during the training.
316
317
318
319
320
321
322
323
324
325
326
327

    Args:
        func (:obj:`callable`):
            The callable function of the TensorFlow model.
        config (:class:`~transformers.PretrainedConfig`):
            The config of the running model.
        **kwargs:
            The inputs of the model.

    Returns:
        Two lists, one for the missing layers, and another one for the unexpected layers.
    """
Julien Plu's avatar
Julien Plu committed
328
329
    signature = dict(inspect.signature(func).parameters)
    signature.pop("kwargs", None)
Julien Plu's avatar
Julien Plu committed
330
    signature.pop("self", None)
Julien Plu's avatar
Julien Plu committed
331
332
    parameter_names = list(signature.keys())
    output = {}
333
    allowed_types = (tf.Tensor, bool, int, ModelOutput, tuple, list, dict, np.ndarray)
Julien Plu's avatar
Julien Plu committed
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354

    if "inputs" in kwargs["kwargs_call"]:
        warnings.warn(
            "The `inputs` argument is deprecated and will be removed in a future version, use `input_ids` instead.",
            FutureWarning,
        )

        output["input_ids"] = kwargs["kwargs_call"].pop("inputs")

    if "decoder_cached_states" in kwargs["kwargs_call"]:
        warnings.warn(
            "The `decoder_cached_states` argument is deprecated and will be removed in a future version, use `past_key_values` instead.",
            FutureWarning,
        )
        output["past_key_values"] = kwargs["kwargs_call"].pop("decoder_cached_states")

    if len(kwargs["kwargs_call"]) > 0:
        raise ValueError(
            f"The following keyword arguments are not supported by this model: {list(kwargs['kwargs_call'].keys())}."
        )

Julien Plu's avatar
Julien Plu committed
355
356
    kwargs.pop("kwargs_call")

Julien Plu's avatar
Julien Plu committed
357
358
359
360
    for k, v in kwargs.items():
        if isinstance(v, allowed_types) or v is None:
            output[k] = v
        else:
Julien Plu's avatar
Julien Plu committed
361
            raise ValueError(f"Data of type {type(v)} is not allowed only {allowed_types} is accepted for {k}.")
Julien Plu's avatar
Julien Plu committed
362
363
364
365
366

    if isinstance(input_ids, (tuple, list)):
        for i, input in enumerate(input_ids):
            # EagerTensors don't allow to use the .name property so we check for a real Tensor
            if type(input) == tf.Tensor:
Julien Plu's avatar
Julien Plu committed
367
368
                # Tensor names have always the pattern `name:id` then we check only the
                # `name` part
Julien Plu's avatar
Julien Plu committed
369
370
371
372
373
                tensor_name = input.name.split(":")[0]

                if tensor_name in parameter_names:
                    output[tensor_name] = input
                else:
Julien Plu's avatar
Julien Plu committed
374
                    output[parameter_names[i]] = input
Julien Plu's avatar
Julien Plu committed
375
376
377
378
            elif isinstance(input, allowed_types) or input is None:
                output[parameter_names[i]] = input
            else:
                raise ValueError(
Julien Plu's avatar
Julien Plu committed
379
                    f"Data of type {type(input)} is not allowed only {allowed_types} is accepted for {parameter_names[i]}."
Julien Plu's avatar
Julien Plu committed
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
                )
    elif isinstance(input_ids, (dict, BatchEncoding)):
        if "inputs" in input_ids:
            warnings.warn(
                "The `inputs` argument is deprecated and will be removed in a future version, use `input_ids` instead.",
                FutureWarning,
            )

            output["input_ids"] = input_ids.pop("inputs")

        if "decoder_cached_states" in input_ids:
            warnings.warn(
                "The `decoder_cached_states` argument is deprecated and will be removed in a future version, use `past_key_values` instead.",
                FutureWarning,
            )
            output["past_key_values"] = input_ids.pop("decoder_cached_states")

        for k, v in dict(input_ids).items():
398
            if isinstance(v, allowed_types) or v is None:
Julien Plu's avatar
Julien Plu committed
399
                output[k] = v
400
401
402
403
404
405
            elif k not in parameter_names and "args" not in parameter_names:
                logger.warn(
                    f"The parameter {k} does not belongs to the parameter list {parameter_names} and will be ignored."
                )
                continue
            else:
Julien Plu's avatar
Julien Plu committed
406
                raise ValueError(f"Data of type {type(v)} is not allowed only {allowed_types} is accepted for {k}.")
Julien Plu's avatar
Julien Plu committed
407
408
409
410
411
    else:
        if isinstance(input_ids, tf.Tensor) or input_ids is None:
            output[parameter_names[0]] = input_ids
        else:
            raise ValueError(
Julien Plu's avatar
Julien Plu committed
412
                f"Data of type {type(input_ids)} is not allowed only {allowed_types} is accepted for {parameter_names[0]}."
Julien Plu's avatar
Julien Plu committed
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
            )

    for name in parameter_names:
        if name not in list(output.keys()) and name != "args":
            output[name] = kwargs.pop(name, signature[name].default)

    # When creating a SavedModel TF calls the method with LayerCall.__call__(args, **kwargs)
    # So to respect the proper output we have to add this exception
    if "args" in output:
        if output["args"] is not None and type(output["args"]) == tf.Tensor:
            tensor_name = output["args"].name.split(":")[0]
            output[tensor_name] = output["args"]
        else:
            # `args` in this case is always the first parameter, then `input_ids`
            output["input_ids"] = output["args"]

        del output["args"]

    if "kwargs" in output:
        del output["kwargs"]

434
435
436
437
438
439
440
441
442
443
444
445
446
    boolean_dict = {
        k: v
        for k, v in output.items()
        if k in ["return_dict", "output_attentions", "output_hidden_states", "use_cache"]
    }

    output.update(
        booleans_processing(
            config=config,
            **boolean_dict,
        )
    )

Julien Plu's avatar
Julien Plu committed
447
448
449
    return output


Julien Plu's avatar
Julien Plu committed
450
def load_tf_weights(model, resolved_archive_file):
Julien Plu's avatar
Julien Plu committed
451
    """
Julien Plu's avatar
Julien Plu committed
452
    Detect missing and unexpected layers and load the TF weights accordingly to their names and shapes.
Julien Plu's avatar
Julien Plu committed
453
454
455
456
457
458
459
460
461
462
463
464
465

    Args:
        model (:obj:`tf.keras.models.Model`):
            The model to load the weights into.
        resolved_archive_file (:obj:`str`):
            The location of the H5 file.

    Returns:
        Two lists, one for the missing layers, and another one for the unexpected layers.
    """
    missing_layers = []
    unexpected_layers = []

Julien Plu's avatar
Julien Plu committed
466
    # Read the H5 file
Julien Plu's avatar
Julien Plu committed
467
    with h5py.File(resolved_archive_file, "r") as f:
Julien Plu's avatar
Julien Plu committed
468
469
        # Retrieve the name of each layer from the H5 file
        saved_h5_model_layers_name = set(hdf5_format.load_attributes_from_hdf5_group(f, "layer_names"))
Julien Plu's avatar
Julien Plu committed
470

Julien Plu's avatar
Julien Plu committed
471
472
        # Find the missing layers from the high level list of layers
        missing_layers = list(set([layer.name for layer in model.layers]) - saved_h5_model_layers_name)
Julien Plu's avatar
Julien Plu committed
473

Julien Plu's avatar
Julien Plu committed
474
475
476
477
        # Find the unexpected layers from the high level list of layers
        unexpected_layers = list(saved_h5_model_layers_name - set([layer.name for layer in model.layers]))
        saved_weight_names_set = set()
        symbolic_weights_names = set()
Julien Plu's avatar
Julien Plu committed
478
479
        weight_value_tuples = []

Julien Plu's avatar
Julien Plu committed
480
481
        # Compute missing and unexpected sub layers
        # Store the weights in list of tuples that looks like [(weight_object, value_of_weight),...]
Julien Plu's avatar
Julien Plu committed
482
        for layer in model.layers:
Julien Plu's avatar
Julien Plu committed
483
484
485
486
487
            # if layer_name from the H5 file belongs to the layers from the instantiated model
            if layer.name in saved_h5_model_layers_name:
                # Get the H5 layer object from its name
                h5_layer_object = f[layer.name]
                # Get all the weights as a list from the layer object
Julien Plu's avatar
Julien Plu committed
488
                symbolic_weights = layer.trainable_weights + layer.non_trainable_weights
Julien Plu's avatar
Julien Plu committed
489
                saved_weights = {}
Julien Plu's avatar
Julien Plu committed
490

Julien Plu's avatar
Julien Plu committed
491
492
493
494
                # Create a dict from the H5 saved model that looks like {"weight_name": weight_value}
                # And a set with only the names
                for weight_name in hdf5_format.load_attributes_from_hdf5_group(h5_layer_object, "weight_names"):
                    # TF names always start with the model name so we ignore it
Julien Plu's avatar
Julien Plu committed
495
                    name = "/".join(weight_name.split("/")[1:])
Julien Plu's avatar
Julien Plu committed
496
                    saved_weights[name] = np.asarray(h5_layer_object[weight_name])
Julien Plu's avatar
Julien Plu committed
497

Julien Plu's avatar
Julien Plu committed
498
499
500
501
                    # Add the updated name to the final list for computing missing/unexpected values
                    saved_weight_names_set.add(name)

                # Loop over each weights from the instantiated model and compare with the weights from the H5 file
Julien Plu's avatar
Julien Plu committed
502
                for symbolic_weight in symbolic_weights:
Julien Plu's avatar
Julien Plu committed
503
504
505
506
507
508
509
                    # TF names always start with the model name so we ignore it
                    symbolic_weight_name = "/".join(symbolic_weight.name.split("/")[1:])

                    # here we check if the current weight is among the weights from the H5 file
                    # If yes, get the weight_value of the corresponding weight from the H5 file
                    # If not, make the value to None
                    saved_weight_value = saved_weights.get(symbolic_weight_name, None)
Julien Plu's avatar
Julien Plu committed
510

Julien Plu's avatar
Julien Plu committed
511
512
                    # Add the updated name to the final list for computing missing/unexpected values
                    symbolic_weights_names.add(symbolic_weight_name)
Julien Plu's avatar
Julien Plu committed
513

Julien Plu's avatar
Julien Plu committed
514
515
516
                    # If the current weight is found
                    if saved_weight_value is not None:
                        # Check if the shape of the current weight and the one from the H5 file are different
Julien Plu's avatar
Julien Plu committed
517
                        if K.int_shape(symbolic_weight) != saved_weight_value.shape:
Julien Plu's avatar
Julien Plu committed
518
519
                            # If yes we reshape the weight from the H5 file accordingly to the current weight
                            # If the two shapes are not compatible we raise an issue
Julien Plu's avatar
Julien Plu committed
520
521
522
523
524
525
526
527
                            try:
                                array = np.reshape(saved_weight_value, K.int_shape(symbolic_weight))
                            except AssertionError as e:
                                e.args += (K.int_shape(symbolic_weight), saved_weight_value.shape)
                                raise e
                        else:
                            array = saved_weight_value

Julien Plu's avatar
Julien Plu committed
528
                        # We create the tuple that will be loaded and add it to the final list
Julien Plu's avatar
Julien Plu committed
529
530
                        weight_value_tuples.append((symbolic_weight, array))

Julien Plu's avatar
Julien Plu committed
531
    # Load all the weights
Julien Plu's avatar
Julien Plu committed
532
533
    K.batch_set_value(weight_value_tuples)

Julien Plu's avatar
Julien Plu committed
534
535
536
537
538
539
    # Compute the missing and unexpected layers
    missing_layers.extend(list(symbolic_weights_names - saved_weight_names_set))
    unexpected_layers.extend(list(saved_weight_names_set - symbolic_weights_names))

    return missing_layers, unexpected_layers

Julien Plu's avatar
Julien Plu committed
540

541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
def init_copy_embeddings(old_embeddings, new_num_tokens):
    r"""
    This function aims to reduce the embeddings in case new_num_tokens < old_num_tokens or to pad with -1 in case
    new_num_tokens > old_num_tokens. A mask is also computed in order to know which weight in the embeddings should be
    kept or not. Example:

        - if new_num_tokens=5 and old_num_tokens=4 and old_embeddings=[w1,w2,w3,w4]

            -  mask=[True,True,True,True,False] and current_weights=[w1,w2,w3,w4,-1]
        - if new_num_tokens=4 and old_num_tokens=5 and old_embeddings=[w1,w2,w3,w4,w5]

            - mask=[True,True,True,True] and current_weights=[w1,w2,w3,w4]
    """
    old_num_tokens, old_embedding_dim = shape_list(old_embeddings)
    size_diff = new_num_tokens - old_num_tokens

    # initialize new embeddings
    # Copy token embeddings from the previous ones
    if tf.math.greater(size_diff, 0):
        # if the new size is greater than the old one, we extend the current embeddings with a padding until getting new size
        # and we create a mask to properly identify the padded values and be replaced by the values of the newly created
        # embeddings
        current_weights = tf.pad(
            old_embeddings.value(), tf.convert_to_tensor([[0, size_diff], [0, 0]]), constant_values=-1
        )
        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
        mask = tf.fill(tf.convert_to_tensor([num_tokens_to_copy, 1]), True)
        mask = tf.pad(mask, tf.convert_to_tensor([[0, size_diff], [0, 0]]), constant_values=False)
    else:
        # if the new size if lower than the old one, we take the current embeddings until the new size
        current_weights = tf.slice(
            old_embeddings.value(),
            tf.convert_to_tensor([0, 0]),
            tf.convert_to_tensor([new_num_tokens, old_embedding_dim]),
        )
        mask = tf.fill(tf.convert_to_tensor([new_num_tokens, 1]), True)

    return mask, current_weights


581
class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin):
582
583
    r"""
    Base class for all TF models.
thomwolf's avatar
thomwolf committed
584

585
586
    :class:`~transformers.TFPreTrainedModel` takes care of storing the configuration of the models and handles methods
    for loading, downloading and saving models as well as a few methods common to all models to:
thomwolf's avatar
thomwolf committed
587

588
589
        * resize the input embeddings,
        * prune heads in the self-attention heads.
thomwolf's avatar
thomwolf committed
590

591
    Class attributes (overridden by derived classes):
Sylvain Gugger's avatar
Sylvain Gugger committed
592

593
594
595
596
        - **config_class** (:class:`~transformers.PretrainedConfig`) -- A subclass of
          :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
        - **base_model_prefix** (:obj:`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.
thomwolf's avatar
thomwolf committed
597
598
599
    """
    config_class = None
    base_model_prefix = ""
600
601
602
603
604
605
    # a list of re pattern of tensor names to ignore from the model when loading the model weights
    # (and avoid unnecessary warnings).
    _keys_to_ignore_on_load_missing = None
    # a list of re pattern of tensor names to ignore from the weights when loading the model weights
    # (and avoid unnecessary warnings).
    _keys_to_ignore_on_load_unexpected = None
thomwolf's avatar
thomwolf committed
606

607
    @property
608
609
    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
        """
Julien Plu's avatar
Julien Plu committed
610
611
612
613
        Dummy inputs to build the network.

        Returns:
            :obj:`Dict[str, tf.Tensor]`: The dummy inputs.
614
        """
Julien Plu's avatar
Julien Plu committed
615
616
617
        return {
            "input_ids": tf.constant(DUMMY_INPUTS),
        }
thomwolf's avatar
thomwolf committed
618
619

    def __init__(self, config, *inputs, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
620
        super().__init__(*inputs, **kwargs)
thomwolf's avatar
thomwolf committed
621
622
623
624
625
626
        if not isinstance(config, PretrainedConfig):
            raise ValueError(
                "Parameter config in `{}(config)` should be an instance of class `PretrainedConfig`. "
                "To create a model from a pretrained model use "
                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
                    self.__class__.__name__, self.__class__.__name__
627
628
                )
            )
629
        # Save config and origin of the pretrained weights if given in model
thomwolf's avatar
thomwolf committed
630
        self.config = config
631
        self.name_or_path = config.name_or_path
thomwolf's avatar
thomwolf committed
632

Julien Plu's avatar
Julien Plu committed
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
    @tf.function(
        input_signature=[
            {
                "input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
                "attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
                "token_type_ids": tf.TensorSpec((None, None), tf.int32, name="token_type_ids"),
            }
        ]
    )
    def serving(self, inputs):
        """
        Method used for serving the model.

        Args:
            inputs (:obj:`Dict[str, tf.Tensor]`):
                The input of the saved model as a dictionnary of tensors.
        """
        output = self.call(inputs)

        return self.serving_output(output)

    def serving_output(output):
        """
        Prepare the output of the saved model. Each model must implement this function.

        Args:
            output (:obj:`~transformers.TFBaseModelOutput`):
                The output returned by the model.
        """
        raise NotImplementedError

664
    def get_input_embeddings(self) -> tf.keras.layers.Layer:
665
        """
666
        Returns the model's input embeddings layer.
667
668

        Returns:
669
            :obj:`tf.Variable`: The embeddings layer mapping vocabulary to hidden states.
670
        """
671
        main_layer = getattr(self, self.base_model_prefix, self)
Julien Plu's avatar
Julien Plu committed
672

673
674
        if main_layer is not self:
            return main_layer.get_input_embeddings()
675
676
677
        else:
            raise NotImplementedError

678
679
    def set_input_embeddings(self, value):
        """
680
        Set model's input embeddings
681
682

        Args:
683
684
            value (:obj:`tf.Variable`):
                The new weights mapping hidden states to vocabulary.
685
        """
686
        main_layer = getattr(self, self.base_model_prefix)
687

688
689
690
691
692
693
694
695
696
697
698
        if main_layer is None:
            raise NotImplementedError("The model does not implements the base_model_prefix attribute.")

        try:
            main_layer.set_input_embeddings(value)
        except AttributeError:
            logger.info("Building the model")
            self(self.dummy_inputs)
            main_layer.set_input_embeddings(value)

    def get_output_embeddings(self) -> Union[None, tf.keras.layers.Layer]:
699
        """
700
        Returns the model's output embeddings
701
702

        Returns:
703
            :obj:`tf.Variable`: The new weights mapping vocabulary to hidden states.
704
        """
705
706
707
708
709
        if self.get_lm_head() is not None:
            lm_head = self.get_lm_head()

            return lm_head.get_output_embeddings()

710
711
        return None  # Overwrite for models with output embeddings

712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
    def set_output_embeddings(self, value):
        """
        Set model's output embeddings

        Args:
            value (:obj:`tf.Variable`):
                The new weights mapping hidden states to vocabulary.
        """
        if self.get_lm_head() is not None:
            lm_head = self.get_lm_head()
            try:
                lm_head.set_output_embeddings(value)
            except AttributeError:
                logger.info("Building the model")
                self(self.dummy_inputs)
                lm_head.set_output_embeddings(value)

729
730
731
    def get_output_layer_with_bias(self) -> Union[None, tf.keras.layers.Layer]:
        """
        Get the layer that handles a bias attribute in case the model has an LM head with weights tied to the
732
        embeddings
733
734
735
736

        Return:
            :obj:`tf.keras.layers.Layer`: The layer that handles the bias, None if not an LM model.
        """
737
738
739
740
        warnings.warn(
            "The method get_output_layer_with_bias is deprecated. Please use `get_lm_head` instead.", FutureWarning
        )
        return self.get_lm_head()
741
742
743

    def get_prefix_bias_name(self) -> Union[None, str]:
        """
744
        Get the concatenated prefix name of the bias from the model name to the parent layer
745
746
747
748

        Return:
            :obj:`str`: The prefix name of the bias.
        """
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
        warnings.warn("The method get_prefix_bias_name is deprecated. Please use `get_bias` instead.", FutureWarning)
        return None

    def get_bias(self) -> Union[None, Dict[str, tf.Variable]]:
        """
        Dict of bias attached to an LM head. The key represents the name of the bias attribute.

        Return:
            :obj:`tf.Variable`: The weights representing the bias, None if not an LM model.
        """
        if self.get_lm_head() is not None:
            lm_head = self.get_lm_head()
            try:
                return lm_head.get_bias()
            except AttributeError:
                self(self.dummy_inputs)

                return lm_head.get_bias()
        return None

    def set_bias(self, value):
        """
        Set all the bias in the LM head.

        Args:
            value (:obj:`Dict[tf.Variable]`):
                All the new bias attached to an LM head.
        """
        if self.get_lm_head() is not None:
            lm_head = self.get_lm_head()
            try:
                lm_head.set_bias(value)
            except AttributeError:
                self(self.dummy_inputs)
                lm_head.set_bias(value)

    def get_lm_head(self) -> tf.keras.layers.Layer:
        """
        The LM Head layer. This method must be overwritten by all the models that have a lm head.

        Return:
            :obj:`tf.keras.layers.Layer`: The LM head layer if the model has one, None if not.
        """
792
793
        return None

794
795
796
    def resize_token_embeddings(self, new_num_tokens=None) -> tf.Variable:
        """
        Resizes input token embeddings matrix of the model if :obj:`new_num_tokens != config.vocab_size`.
797

798
        Takes care of tying weights embeddings afterwards if the model class has a :obj:`tie_weights()` method.
799

800
801
802
803
        Arguments:
            new_num_tokens (:obj:`int`, `optional`):
                The number of new tokens in the embedding matrix. Increasing the size will add newly initialized
                vectors at the end. Reducing the size will remove vectors from the end. If not provided or :obj:`None`,
804
                just returns a pointer to the input tokens :obj:`tf.Variable` module of the model without doing
805
806
807
808
                anything.

        Return:
            :obj:`tf.Variable`: Pointer to the input tokens Embeddings Module of the model.
809
        """
810
811
        if new_num_tokens is None or new_num_tokens == self.config.vocab_size:
            return self._get_word_embedding_weight(self.get_input_embeddings())
812

813
        model_embeds = self._resize_token_embeddings(new_num_tokens)
814
815
816

        # Update base model and current model config
        self.config.vocab_size = new_num_tokens
817
818
819

        return model_embeds

820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
    def _get_word_embedding_weight(model, embedding_layer):
        embeds = getattr(embedding_layer, "weight", None)
        if embeds is not None:
            return embeds

        embeds = getattr(embedding_layer, "decoder", None)
        if embeds is not None:
            return embeds

        # The reason why the attributes don't exist might be
        # because the model is not built, so retry getting
        # the argument after building the model
        model(model.dummy_inputs)

        embeds = getattr(embedding_layer, "weight", None)
        if embeds is not None:
            return embeds

        embeds = getattr(embedding_layer, "decoder", None)
        if embeds is not None:
            return embeds

        return None
843

844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
    def _resize_token_embeddings(self, new_num_tokens):
        old_embeddings = self._get_word_embedding_weight(self.get_input_embeddings())
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)

        # if word embeddings are not tied, make sure that lm head bias is resized as well
        if self.get_bias() is not None:
            old_lm_head_bias = self.get_bias()
            new_lm_head_bias = self._get_resized_lm_head_bias(old_lm_head_bias, new_num_tokens)

            self.set_bias(new_lm_head_bias)

        # if word embeddings are not tied, make sure that lm head decoder is resized as well
        if self.get_output_embeddings() is not None:
            old_lm_head_decoder = self._get_word_embedding_weight(self.get_output_embeddings())
            new_lm_head_decoder = self._get_resized_lm_head_decoder(old_lm_head_decoder, new_num_tokens)

            self.set_output_embeddings(new_lm_head_decoder)

        self.set_input_embeddings(new_embeddings)

        return self.get_input_embeddings()

    def _get_resized_lm_head_bias(self, old_lm_head_bias, new_num_tokens):
867
        """
868
869
        Build a resized bias from the old ones. Increasing the size will add newly initialized vectors at the end.
        Reducing the size will remove vectors from the end
thomwolf's avatar
thomwolf committed
870
871

        Args:
872
873
            old_lm_head_bias (:obj:`tf.Variable`):
                Old lm head bias to be resized.
874
            new_num_tokens (:obj:`int`, `optional`):
875
                New number of tokens in the linear matrix.
876
877

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
878
                vectors from the end. If not provided or :obj:`None`, just returns None
879
880

        Return:
881
            :obj:`tf.Variable`: Pointer to the resized bias.
thomwolf's avatar
thomwolf committed
882
        """
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
        new_lm_head_bias = {}

        for attr, weight in old_lm_head_bias.items():
            first_dim, old_num_tokens = (None, shape_list(weight)[0]) if tf.rank(weight) == 1 else shape_list(weight)
            size_diff = new_num_tokens - old_num_tokens
            final_shape = [new_num_tokens] if first_dim is None else [first_dim, new_num_tokens]

            # initialize new bias
            if tf.math.greater(size_diff, 0):
                padding_shape = [[0, size_diff]] if first_dim is None else [[0, 0], [0, size_diff]]
                current_bias = tf.pad(weight.value(), tf.convert_to_tensor(padding_shape), constant_values=-1)
                num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
                mask_shape = [num_tokens_to_copy] if first_dim is None else [1, num_tokens_to_copy]
                bias_mask = tf.fill(tf.convert_to_tensor(mask_shape), True)
                bias_mask = tf.pad(bias_mask, tf.convert_to_tensor(padding_shape), constant_values=False)
            else:
                slice_from = [0] if first_dim is None else [0, 0]
                current_bias = tf.slice(
                    weight.value(), tf.convert_to_tensor(slice_from), tf.convert_to_tensor(final_shape)
                )
                bias_mask = tf.fill(tf.convert_to_tensor(final_shape), True)
904

905
906
907
908
909
910
911
            new_bias = self.add_weight(
                shape=final_shape,
                initializer="zeros",
                trainable=True,
                name=weight.name.split(":")[0],
            )
            init_bias = tf.where(bias_mask, current_bias, new_bias.value())
912

913
914
            new_bias.assign(init_bias)
            new_lm_head_bias[attr] = new_bias
915

916
        return new_lm_head_bias
thomwolf's avatar
thomwolf committed
917

918
919
920
921
    def _get_resized_lm_head_decoder(self, old_lm_head_decoder, new_num_tokens):
        """
        Build a resized decoder from the old ones. Increasing the size will add newly initialized vectors at the end.
        Reducing the size will remove vectors from the end
thomwolf's avatar
thomwolf committed
922

923
924
925
926
927
        Args:
            old_lm_head_decoder (:obj:`tf.Variable`):
                Old lm head decoder to be resized.
            new_num_tokens (:obj:`int`, `optional`):
                New number of tokens in the linear matrix.
928

929
930
                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
                vectors from the end. If not provided or :obj:`None`, just returns None
931

932
933
934
935
936
937
938
939
        Return:
            :obj:`tf.Variable`: Pointer to the resized decoder or None if the output embeddings are differents of the
            input ones.
        """
        new_lm_head_decoder = old_lm_head_decoder
        is_input_output_equals = tf.reduce_any(
            self._get_word_embedding_weight(self.get_input_embeddings()) == old_lm_head_decoder
        )
940

941
942
943
944
945
        if old_lm_head_decoder is not None and not is_input_output_equals:
            old_embedding_dim = shape_list(old_lm_head_decoder)[1]
            decoder_mask, current_decoder = init_copy_embeddings(old_lm_head_decoder, new_num_tokens)
            new_lm_head_decoder = self.add_weight(
                shape=(new_num_tokens, old_embedding_dim),
946
947
                initializer="zeros",
                trainable=True,
948
                name=old_lm_head_decoder.name.split(":")[0],
949
            )
950
951
952
            init_decoder = tf.where(decoder_mask, current_decoder, new_lm_head_decoder.value())

            new_lm_head_decoder.assign(init_decoder)
953

954
        return new_lm_head_decoder
955

956
957
958
959
    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None) -> tf.Variable:
        """
        Build a resized Embedding weights from a provided token Embedding weights. Increasing the size will add newly
        initialized vectors at the end. Reducing the size will remove vectors from the end
960

961
962
963
964
965
        Args:
            old_embeddings (:obj:`tf.Variable`):
                Old embeddings to be resized.
            new_num_tokens (:obj:`int`, `optional`):
                New number of tokens in the embedding matrix.
966

967
968
969
                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
                vectors from the end. If not provided or :obj:`None`, just returns a pointer to the input tokens
                :obj:`tf.Variable`` module of the model without doing anything.
970

971
972
973
974
975
976
977
978
979
980
981
982
983
984
        Return:
            :obj:`tf.Variable`: Pointer to the resized Embedding Module or the old Embedding Module if
            :obj:`new_num_tokens` is :obj:`None`
        """
        old_embedding_dim = shape_list(old_embeddings)[1]
        init_range = getattr(self.config, "initializer_range", 0.02)
        embeddings_mask, current_embeddings = init_copy_embeddings(old_embeddings, new_num_tokens)
        new_embeddings = self.add_weight(
            name=old_embeddings.name.split(":")[0],
            shape=[new_num_tokens, old_embedding_dim],
            initializer=get_initializer(init_range),
            dtype=tf.float32,
        )
        init_embeddings = tf.where(embeddings_mask, current_embeddings, new_embeddings.value())
985

986
        new_embeddings.assign(init_embeddings)
987

988
        return new_embeddings
thomwolf's avatar
thomwolf committed
989
990

    def prune_heads(self, heads_to_prune):
991
992
        """
        Prunes heads of the base model.
thomwolf's avatar
thomwolf committed
993

994
995
        Arguments:
            heads_to_prune (:obj:`Dict[int, List[int]]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
996
997
998
                Dictionary with keys being selected layer indices (:obj:`int`) and associated values being the list of
                heads to prune in said layer (list of :obj:`int`). For instance {1: [0, 2], 2: [2, 3]} will prune heads
                0 and 2 on layer 1 and heads 2 and 3 on layer 2.
thomwolf's avatar
thomwolf committed
999
1000
1001
        """
        raise NotImplementedError

Julien Plu's avatar
Julien Plu committed
1002
    def save_pretrained(self, save_directory, saved_model=False, version=1):
1003
1004
        """
        Save a model and its configuration file to a directory, so that it can be re-loaded using the
Sylvain Gugger's avatar
Sylvain Gugger committed
1005
        :func:`~transformers.TFPreTrainedModel.from_pretrained` class method.
1006
1007
1008
1009

        Arguments:
            save_directory (:obj:`str`):
                Directory to which to save. Will be created if it doesn't exist.
Julien Plu's avatar
Julien Plu committed
1010
1011
1012
1013
1014
1015
            saved_model (:obj:`bool`, `optional`, defaults to :obj:`False`):
                If the model has to be saved in saved model format as well or not.
            version (:obj:`int`, `optional`, defaults to 1):
                The version of the saved model. A saved model needs to be versioned in order to be properly loaded by
                TensorFlow Serving as detailed in the official documentation
                https://www.tensorflow.org/tfx/serving/serving_basic
thomwolf's avatar
thomwolf committed
1016
        """
1017
1018
1019
1020
        if os.path.isfile(save_directory):
            logger.error("Provided path ({}) should be a directory, not a file".format(save_directory))
            return
        os.makedirs(save_directory, exist_ok=True)
thomwolf's avatar
thomwolf committed
1021

Julien Plu's avatar
Julien Plu committed
1022
1023
1024
1025
1026
        if saved_model:
            saved_model_dir = os.path.join(save_directory, "saved_model", str(version))
            self.save(saved_model_dir, include_optimizer=False, signatures=self.serving)
            logger.info(f"Saved model created in {saved_model_dir}")

thomwolf's avatar
thomwolf committed
1027
1028
1029
1030
1031
1032
        # Save configuration file
        self.config.save_pretrained(save_directory)

        # If we save using the predefined names, we can load using `from_pretrained`
        output_model_file = os.path.join(save_directory, TF2_WEIGHTS_NAME)
        self.save_weights(output_model_file)
thomwolf's avatar
thomwolf committed
1033
        logger.info("Model weights saved in {}".format(output_model_file))
thomwolf's avatar
thomwolf committed
1034
1035
1036

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
1037
1038
        r"""
        Instantiate a pretrained TF 2.0 model from a pre-trained model configuration.
thomwolf's avatar
thomwolf committed
1039

1040
1041
1042
        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.
thomwolf's avatar
thomwolf committed
1043

1044
1045
        The warning `Weights from XXX not used in YYY` means that the layer XXX is not used by YYY, therefore those
        weights are discarded.
thomwolf's avatar
thomwolf committed
1046
1047

        Parameters:
1048
1049
1050
            pretrained_model_name_or_path (:obj:`str`, `optional`):
                Can be either:

1051
1052
1053
                    - 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``.
1054
1055
                    - A path to a `directory` containing model weights saved using
                      :func:`~transformersTF.PreTrainedModel.save_pretrained`, e.g., ``./my_model_directory/``.
Sylvain Gugger's avatar
Sylvain Gugger committed
1056
                    - A path or url to a `PyTorch state_dict save file` (e.g, ``./pt_model/pytorch_model.bin``). In
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
                      this case, ``from_pt`` should be set to :obj:`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.
                    - :obj:`None` if you are both providing the configuration and state dictionary (resp. with keyword
                      arguments ``config`` and ``state_dict``).
            model_args (sequence of positional arguments, `optional`):
                All remaning positional arguments will be passed to the underlying model's ``__init__`` method.
            config (:obj:`Union[PretrainedConfig, str]`, `optional`):
                Can be either:

                    - an instance of a class derived from :class:`~transformers.PretrainedConfig`,
                    - a string valid as input to :func:`~transformers.PretrainedConfig.from_pretrained`.

                Configuration for the model to use instead of an automatically loaded configuation. Configuration can
                be automatically loaded when:

1074
1075
                    - The model is a model provided by the library (loaded with the `model id` string of a pretrained
                      model).
1076
                    - The model was saved using :func:`~transformers.TFPreTrainedModel.save_pretrained` and is reloaded
1077
1078
                      by supplying the save directory.
                    - The model is loaded by supplying a local directory as ``pretrained_model_name_or_path`` and a
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
                      configuration JSON file named `config.json` is found in the directory.
            from_pt: (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Load the model weights from a PyTorch state_dict save file (see docstring of
                ``pretrained_model_name_or_path`` argument).
            cache_dir (:obj:`str`, `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 (:obj:`bool`, `optional`, defaults to :obj:`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 (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
                file exists.
            proxies: (:obj:`Dict[str, str], `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1093
1094
                A dictionary of proxy servers to use by protocol or endpoint, e.g., :obj:`{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
1095
            output_loading_info(:obj:`bool`, `optional`, defaults to :obj:`False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1096
                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
1097
1098
            local_files_only(:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to only look at local files (e.g., not try doanloading the model).
1099
1100
1101
            use_auth_token (:obj:`str` or `bool`, `optional`):
                The token to use as HTTP bearer authorization for remote files. If :obj:`True`, will use the token
                generated when running :obj:`transformers-cli login` (stored in :obj:`~/.huggingface`).
Julien Chaumond's avatar
Julien Chaumond committed
1102
1103
1104
1105
            revision(:obj:`str`, `optional`, defaults to :obj:`"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.
1106
            mirror(:obj:`str`, `optional`, defaults to :obj:`None`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1107
1108
1109
                Mirror source to accelerate downloads in China. If you are from China and have an accessibility
                problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
                Please refer to the mirror site for more information.
1110
1111
            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.,
1112
                :obj:`output_attentions=True`). Behaves differently depending on whether a ``config`` is provided or
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
                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 (:func:`~transformers.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.
thomwolf's avatar
thomwolf committed
1123

1124
1125
1126
1127
        .. note::

            Passing :obj:`use_auth_token=True` is required when you want to use a private model.

thomwolf's avatar
thomwolf committed
1128
1129
        Examples::

1130
            >>> from transformers import BertConfig, TFBertModel
1131
            >>> # Download model and configuration from huggingface.co and cache.
1132
1133
1134
1135
1136
1137
1138
1139
1140
            >>> model = TFBertModel.from_pretrained('bert-base-uncased')
            >>> # Model was saved using `save_pretrained('./test/saved_model/')` (for example purposes, not runnable).
            >>> model = TFBertModel.from_pretrained('./test/saved_model/')
            >>> # Update configuration during loading.
            >>> model = TFBertModel.from_pretrained('bert-base-uncased', output_attentions=True)
            >>> assert model.config.output_attentions == True
            >>> # Loading from a Pytorch model file instead of a TensorFlow checkpoint (slower, for example purposes, not runnable).
            >>> config = BertConfig.from_json_file('./pt_model/my_pt_model_config.json')
            >>> model = TFBertModel.from_pretrained('./pt_model/my_pytorch_model.bin', from_pt=True, config=config)
thomwolf's avatar
thomwolf committed
1141
1142

        """
1143
1144
1145
1146
1147
1148
1149
        config = kwargs.pop("config", None)
        cache_dir = kwargs.pop("cache_dir", None)
        from_pt = kwargs.pop("from_pt", False)
        force_download = kwargs.pop("force_download", False)
        resume_download = kwargs.pop("resume_download", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
1150
        local_files_only = kwargs.pop("local_files_only", False)
1151
        use_auth_token = kwargs.pop("use_auth_token", None)
Julien Chaumond's avatar
Julien Chaumond committed
1152
        revision = kwargs.pop("revision", None)
1153
        mirror = kwargs.pop("mirror", None)
thomwolf's avatar
thomwolf committed
1154

1155
1156
1157
        # 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
thomwolf's avatar
thomwolf committed
1158
            config, model_kwargs = cls.config_class.from_pretrained(
1159
1160
1161
1162
                config_path,
                *model_args,
                cache_dir=cache_dir,
                return_unused_kwargs=True,
thomwolf's avatar
thomwolf committed
1163
                force_download=force_download,
1164
                resume_download=resume_download,
1165
1166
                proxies=proxies,
                local_files_only=local_files_only,
1167
                use_auth_token=use_auth_token,
Julien Chaumond's avatar
Julien Chaumond committed
1168
                revision=revision,
1169
                **kwargs,
thomwolf's avatar
thomwolf committed
1170
1171
1172
1173
1174
            )
        else:
            model_kwargs = kwargs

        # Load model
thomwolf's avatar
thomwolf committed
1175
        if pretrained_model_name_or_path is not None:
1176
            if os.path.isdir(pretrained_model_name_or_path):
1177
1178
1179
1180
                if from_pt and os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
                    # Load from a PyTorch checkpoint in priority if from_pt
                    archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
thomwolf's avatar
thomwolf committed
1181
1182
1183
                    # Load from a TF 2.0 checkpoint
                    archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
                else:
1184
1185
1186
1187
1188
                    raise EnvironmentError(
                        "Error no file named {} found in directory {} or `from_pt` set to False".format(
                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME], pretrained_model_name_or_path
                        )
                    )
Julien Chaumond's avatar
Julien Chaumond committed
1189
            elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
thomwolf's avatar
thomwolf committed
1190
                archive_file = pretrained_model_name_or_path
1191
1192
            elif os.path.isfile(pretrained_model_name_or_path + ".index"):
                archive_file = pretrained_model_name_or_path + ".index"
thomwolf's avatar
thomwolf committed
1193
            else:
thomwolf's avatar
thomwolf committed
1194
                archive_file = hf_bucket_url(
Julien Chaumond's avatar
Julien Chaumond committed
1195
1196
                    pretrained_model_name_or_path,
                    filename=(WEIGHTS_NAME if from_pt else TF2_WEIGHTS_NAME),
Julien Chaumond's avatar
Julien Chaumond committed
1197
                    revision=revision,
1198
                    mirror=mirror,
thomwolf's avatar
thomwolf committed
1199
                )
thomwolf's avatar
thomwolf committed
1200
1201

            try:
1202
                # Load from URL or cache if already cached
1203
1204
1205
1206
1207
                resolved_archive_file = cached_path(
                    archive_file,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
1208
1209
                    resume_download=resume_download,
                    local_files_only=local_files_only,
1210
                    use_auth_token=use_auth_token,
1211
                )
Julien Chaumond's avatar
Julien Chaumond committed
1212
1213
            except EnvironmentError as err:
                logger.error(err)
1214
1215
1216
1217
1218
1219
                msg = (
                    f"Can't load weights for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
                    f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
                    f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a file named one of {TF2_WEIGHTS_NAME}, {WEIGHTS_NAME}.\n\n"
                )
                raise EnvironmentError(msg)
thomwolf's avatar
thomwolf committed
1220
1221
            if resolved_archive_file == archive_file:
                logger.info("loading weights file {}".format(archive_file))
thomwolf's avatar
thomwolf committed
1222
            else:
1223
                logger.info("loading weights file {} from cache at {}".format(archive_file, resolved_archive_file))
thomwolf's avatar
thomwolf committed
1224
        else:
thomwolf's avatar
thomwolf committed
1225
            resolved_archive_file = None
thomwolf's avatar
thomwolf committed
1226

1227
1228
        config.name_or_path = pretrained_model_name_or_path

thomwolf's avatar
thomwolf committed
1229
1230
1231
1232
        # Instantiate model.
        model = cls(config, *model_args, **model_kwargs)

        if from_pt:
Julien Plu's avatar
Julien Plu committed
1233
1234
            from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model

thomwolf's avatar
thomwolf committed
1235
            # Load from a PyTorch checkpoint
thomwolf's avatar
thomwolf committed
1236
            return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file, allow_missing_keys=True)
thomwolf's avatar
thomwolf committed
1237

Julien Plu's avatar
Julien Plu committed
1238
        model(model.dummy_inputs)  # build the network with dummy inputs
thomwolf's avatar
thomwolf committed
1239

thomwolf's avatar
thomwolf committed
1240
        assert os.path.isfile(resolved_archive_file), "Error retrieving file {}".format(resolved_archive_file)
thomwolf's avatar
thomwolf committed
1241
1242
        # 'by_name' allow us to do transfer learning by skipping/adding layers
        # see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1339-L1357
1243
        try:
Julien Plu's avatar
Julien Plu committed
1244
            missing_keys, unexpected_keys = load_tf_weights(model, resolved_archive_file)
1245
        except OSError:
1246
1247
1248
1249
            raise OSError(
                "Unable to load weights from h5 file. "
                "If you tried to load a TF 2.0 model from a PyTorch checkpoint, please set from_pt=True. "
            )
thomwolf's avatar
thomwolf committed
1250

Julien Plu's avatar
Julien Plu committed
1251
        model(model.dummy_inputs)  # Make sure restore ops are run
thomwolf's avatar
thomwolf committed
1252

1253
1254
        if cls._keys_to_ignore_on_load_missing is not None:
            for pat in cls._keys_to_ignore_on_load_missing:
1255
1256
                missing_keys = [k for k in missing_keys if re.search(pat, k) is None]

1257
1258
        if cls._keys_to_ignore_on_load_unexpected is not None:
            for pat in cls._keys_to_ignore_on_load_unexpected:
Julien Plu's avatar
Julien Plu committed
1259
1260
                unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]

1261
1262
        if len(unexpected_keys) > 0:
            logger.warning(
Julien Plu's avatar
Julien Plu committed
1263
                f"Some layers from the model checkpoint at {pretrained_model_name_or_path} were not used when "
1264
1265
                f"initializing {model.__class__.__name__}: {unexpected_keys}\n"
                f"- This IS expected if you are initializing {model.__class__.__name__} from the checkpoint of a model trained on another task "
1266
                f"or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n"
1267
1268
1269
1270
                f"- This IS NOT expected if you are initializing {model.__class__.__name__} from the checkpoint of a model that you expect "
                f"to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
            )
        else:
Julien Plu's avatar
Julien Plu committed
1271
1272
            logger.warning(f"All model checkpoint layers were used when initializing {model.__class__.__name__}.\n")

thomwolf's avatar
thomwolf committed
1273
        if len(missing_keys) > 0:
1274
            logger.warning(
Julien Plu's avatar
Julien Plu committed
1275
                f"Some layers of {model.__class__.__name__} were not initialized from the model checkpoint at {pretrained_model_name_or_path} "
1276
1277
                f"and are newly initialized: {missing_keys}\n"
                f"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference."
1278
            )
1279
1280
        else:
            logger.warning(
Julien Plu's avatar
Julien Plu committed
1281
                f"All the layers of {model.__class__.__name__} were initialized from the model checkpoint at {pretrained_model_name_or_path}.\n"
1282
                f"If your task is similar to the task the model of the checkpoint was trained on, "
1283
                f"you can already use {model.__class__.__name__} for predictions without further training."
1284
            )
Julien Plu's avatar
Julien Plu committed
1285

thomwolf's avatar
thomwolf committed
1286
        if output_loading_info:
Julien Plu's avatar
Julien Plu committed
1287
1288
            loading_info = {"missing_keys": missing_keys, "unexpected_keys": unexpected_keys}

thomwolf's avatar
thomwolf committed
1289
1290
            return model, loading_info

thomwolf's avatar
thomwolf committed
1291
        return model
thomwolf's avatar
WIP  
thomwolf committed
1292

1293

thomwolf's avatar
WIP  
thomwolf committed
1294
class TFConv1D(tf.keras.layers.Layer):
Sylvain Gugger's avatar
Sylvain Gugger committed
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
    """
    1D-convolutional layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2).

    Basically works like a linear layer but the weights are transposed.

    Args:
        nf (:obj:`int`):
            The number of output features.
        nx (:obj:`int`):
            The number of input features.
        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
            The standard deviation to use to initialize the weights.
        kwargs:
            Additional keyword arguments passed along to the :obj:`__init__` of :obj:`tf.keras.layers.Layer`.
    """

thomwolf's avatar
thomwolf committed
1311
    def __init__(self, nf, nx, initializer_range=0.02, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1312
        super().__init__(**kwargs)
thomwolf's avatar
WIP  
thomwolf committed
1313
        self.nf = nf
thomwolf's avatar
thomwolf committed
1314
        self.nx = nx
thomwolf's avatar
thomwolf committed
1315
        self.initializer_range = initializer_range
thomwolf's avatar
thomwolf committed
1316
1317
1318

    def build(self, input_shape):
        self.weight = self.add_weight(
1319
1320
1321
            "weight", shape=[self.nx, self.nf], initializer=get_initializer(self.initializer_range)
        )
        self.bias = self.add_weight("bias", shape=[1, self.nf], initializer=tf.zeros_initializer())
thomwolf's avatar
thomwolf committed
1322

thomwolf's avatar
WIP  
thomwolf committed
1323
    def call(self, x):
thomwolf's avatar
thomwolf committed
1324
        bz, sl = shape_list(x)[:2]
thomwolf's avatar
thomwolf committed
1325

thomwolf's avatar
thomwolf committed
1326
        x = tf.reshape(x, [-1, self.nx])
thomwolf's avatar
thomwolf committed
1327
        x = tf.matmul(x, self.weight) + self.bias
thomwolf's avatar
thomwolf committed
1328
1329

        x = tf.reshape(x, [bz, sl, self.nf])
thomwolf's avatar
thomwolf committed
1330

thomwolf's avatar
WIP  
thomwolf committed
1331
        return x
thomwolf's avatar
thomwolf committed
1332
1333


1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
class WordEmbeddings(tf.keras.layers.Layer):
    def __init__(self, vocab_size: int, hidden_size: int, initializer_range: float, **kwargs):
        super().__init__(**kwargs)

        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.initializer_range = initializer_range

    def build(self, input_shape):
        self.word_embeddings = self.add_weight(
            name="weight",
            shape=[self.vocab_size, self.hidden_size],
            initializer=get_initializer(initializer_range=self.initializer_range),
        )

        super().build(input_shape=input_shape)

    def get_config(self):
        config = {
            "vocab_size": self.vocab_size,
            "hidden_size": self.hidden_size,
            "initializer_range": self.initializer_range,
        }
        base_config = super().get_config()

        return dict(list(base_config.items()) + list(config.items()))

    def call(self, input_ids):
        flat_input_ids = tf.reshape(tensor=input_ids, shape=[-1])
        embeddings = tf.gather(params=self.word_embeddings, indices=flat_input_ids)
        embeddings = tf.reshape(
            tensor=embeddings, shape=tf.concat(values=[shape_list(tensor=input_ids), [self.hidden_size]], axis=0)
        )

        embeddings.set_shape(shape=input_ids.shape.as_list() + [self.hidden_size])

        return embeddings


class TokenTypeEmbeddings(tf.keras.layers.Layer):
    def __init__(self, type_vocab_size: int, hidden_size: int, initializer_range: float, **kwargs):
        super().__init__(**kwargs)

        self.type_vocab_size = type_vocab_size
        self.hidden_size = hidden_size
        self.initializer_range = initializer_range

    def build(self, input_shape):
        self.token_type_embeddings = self.add_weight(
            name="embeddings",
            shape=[self.type_vocab_size, self.hidden_size],
            initializer=get_initializer(initializer_range=self.initializer_range),
        )

        super().build(input_shape=input_shape)

    def get_config(self):
        config = {
            "type_vocab_size": self.type_vocab_size,
            "hidden_size": self.hidden_size,
            "initializer_range": self.initializer_range,
        }
        base_config = super().get_config()

        return dict(list(base_config.items()) + list(config.items()))

    def call(self, token_type_ids):
        flat_token_type_ids = tf.reshape(tensor=token_type_ids, shape=[-1])
        one_hot_data = tf.one_hot(indices=flat_token_type_ids, depth=self.type_vocab_size, dtype=self._compute_dtype)
        embeddings = tf.matmul(a=one_hot_data, b=self.token_type_embeddings)
        embeddings = tf.reshape(
            tensor=embeddings, shape=tf.concat(values=[shape_list(tensor=token_type_ids), [self.hidden_size]], axis=0)
        )

        embeddings.set_shape(shape=token_type_ids.shape.as_list() + [self.hidden_size])

        return embeddings


class PositionEmbeddings(tf.keras.layers.Layer):
    def __init__(self, max_position_embeddings: int, hidden_size: int, initializer_range: float, **kwargs):
        super().__init__(**kwargs)

        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.initializer_range = initializer_range

    def build(self, input_shape):
        self.position_embeddings = self.add_weight(
            name="embeddings",
            shape=[self.max_position_embeddings, self.hidden_size],
            initializer=get_initializer(initializer_range=self.initializer_range),
        )

        super().build(input_shape)

    def get_config(self):
        config = {
            "max_position_embeddings": self.max_position_embeddings,
            "hidden_size": self.hidden_size,
            "initializer_range": self.initializer_range,
        }
        base_config = super().get_config()

        return dict(list(base_config.items()) + list(config.items()))

    def call(self, position_ids):
        input_shape = shape_list(tensor=position_ids)
        position_embeddings = self.position_embeddings[: input_shape[1], :]

        return tf.broadcast_to(input=position_embeddings, shape=input_shape)


thomwolf's avatar
thomwolf committed
1447
class TFSharedEmbeddings(tf.keras.layers.Layer):
Stas Bekman's avatar
Stas Bekman committed
1448
    r"""
Sylvain Gugger's avatar
Sylvain Gugger committed
1449
    Construct shared token embeddings.
1450

Sylvain Gugger's avatar
Sylvain Gugger committed
1451
1452
    The weights of the embedding layer is usually shared with the weights of the linear decoder when doing language
    modeling.
Sylvain Gugger's avatar
Sylvain Gugger committed
1453
1454
1455

    Args:
        vocab_size (:obj:`int`):
1456
            The size of the vocabulary, e.g., the number of unique tokens.
Sylvain Gugger's avatar
Sylvain Gugger committed
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
        hidden_size (:obj:`int`):
            The size of the embedding vectors.
        initializer_range (:obj:`float`, `optional`):
            The standard deviation to use when initializing the weights. If no value is provided, it will default to
            :math:`1/\sqrt{hidden\_size}`.
        kwargs:
            Additional keyword arguments passed along to the :obj:`__init__` of :obj:`tf.keras.layers.Layer`.
    """

    def __init__(self, vocab_size: int, hidden_size: int, initializer_range: Optional[float] = None, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1467
        super().__init__(**kwargs)
thomwolf's avatar
thomwolf committed
1468
1469
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
1470
        self.initializer_range = hidden_size ** -0.5 if initializer_range is None else initializer_range
thomwolf's avatar
thomwolf committed
1471
1472

    def build(self, input_shape):
Sylvain Gugger's avatar
Sylvain Gugger committed
1473
1474
1475
        """
        Build shared token embedding layer Shared weights logic adapted from
        https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
thomwolf's avatar
thomwolf committed
1476
1477
        """
        self.weight = self.add_weight(
1478
1479
            "weight", shape=[self.vocab_size, self.hidden_size], initializer=get_initializer(self.initializer_range)
        )
Julien Chaumond's avatar
Julien Chaumond committed
1480
        super().build(input_shape)
thomwolf's avatar
thomwolf committed
1481

Julien Plu's avatar
Julien Plu committed
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
    def get_config(self):
        config = {
            "vocab_size": self.vocab_size,
            "hidden_size": self.hidden_size,
            "initializer_range": self.initializer_range,
        }
        base_config = super().get_config()

        return dict(list(base_config.items()) + list(config.items()))

Sylvain Gugger's avatar
Sylvain Gugger committed
1492
1493
1494
1495
    def call(self, inputs: tf.Tensor, mode: str = "embedding") -> tf.Tensor:
        """
        Get token embeddings of inputs or decode final hidden state.

thomwolf's avatar
thomwolf committed
1496
        Args:
Sylvain Gugger's avatar
Sylvain Gugger committed
1497
1498
1499
1500
1501
1502
1503
1504
            inputs (:obj:`tf.Tensor`):
                In embedding mode, should be an int64 tensor with shape :obj:`[batch_size, length]`.

                In linear mode, should be a float tensor with shape :obj:`[batch_size, length, hidden_size]`.
            mode (:obj:`str`, defaults to :obj:`"embedding"`):
               A valid value is either :obj:`"embedding"` or :obj:`"linear"`, the first one indicates that the layer
               should be used as an embedding layer, the second one that the layer should be used as a linear decoder.

thomwolf's avatar
thomwolf committed
1505
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1506
            :obj:`tf.Tensor`: In embedding mode, the output is a float32 embedding tensor, with shape
Sylvain Gugger's avatar
Sylvain Gugger committed
1507
1508
            :obj:`[batch_size, length, embedding_size]`.

1509
            In linear mode, the output is a float32 with shape :obj:`[batch_size, length, vocab_size]`.
Sylvain Gugger's avatar
Sylvain Gugger committed
1510

thomwolf's avatar
thomwolf committed
1511
        Raises:
Sylvain Gugger's avatar
Sylvain Gugger committed
1512
            ValueError: if :obj:`mode` is not valid.
1513

Sylvain Gugger's avatar
Sylvain Gugger committed
1514
1515
        Shared weights logic is adapted from `here
        <https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24>`__.
thomwolf's avatar
thomwolf committed
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
        """
        if mode == "embedding":
            return self._embedding(inputs)
        elif mode == "linear":
            return self._linear(inputs)
        else:
            raise ValueError("mode {} is not valid.".format(mode))

    def _embedding(self, input_ids):
        """Applies embedding based on inputs tensor."""
        return tf.gather(self.weight, input_ids)

    def _linear(self, inputs):
        """
Julien Plu's avatar
Julien Plu committed
1530
        Computes logits by running inputs through a linear layer.
thomwolf's avatar
thomwolf committed
1531

Julien Plu's avatar
Julien Plu committed
1532
1533
1534
1535
1536
1537
1538
        Args:
            inputs: A float32 tensor with shape [..., hidden_size]

        Returns:
            float32 tensor with shape [..., vocab_size].
        """
        first_dims = shape_list(inputs)[:-1]
thomwolf's avatar
thomwolf committed
1539
1540
1541
1542
1543
1544
        x = tf.reshape(inputs, [-1, self.hidden_size])
        logits = tf.matmul(x, self.weight, transpose_b=True)

        return tf.reshape(logits, first_dims + [self.vocab_size])


thomwolf's avatar
thomwolf committed
1545
class TFSequenceSummary(tf.keras.layers.Layer):
Julien Plu's avatar
Julien Plu committed
1546
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
1547
1548
1549
1550
    Compute a single vector summary of a sequence hidden states.

    Args:
        config (:class:`~transformers.PretrainedConfig`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1551
1552
            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
            config class of your model for the default values it uses):
Sylvain Gugger's avatar
Sylvain Gugger committed
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564

            - **summary_type** (:obj:`str`) -- The method to use to make this summary. Accepted values are:

                - :obj:`"last"` -- Take the last token hidden state (like XLNet)
                - :obj:`"first"` -- Take the first token hidden state (like Bert)
                - :obj:`"mean"` -- Take the mean of all tokens hidden states
                - :obj:`"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
                - :obj:`"attn"` -- Not implemented now, use multi-head attention

            - **summary_use_proj** (:obj:`bool`) -- Add a projection after the vector extraction.
            - **summary_proj_to_labels** (:obj:`bool`) -- If :obj:`True`, the projection outputs to
              :obj:`config.num_labels` classes (otherwise to :obj:`config.hidden_size`).
Sylvain Gugger's avatar
Sylvain Gugger committed
1565
            - **summary_activation** (:obj:`Optional[str]`) -- Set to :obj:`"tanh"` to add a tanh activation to the
Sylvain Gugger's avatar
Sylvain Gugger committed
1566
1567
1568
1569
1570
1571
1572
1573
1574
              output, another string or :obj:`None` will add no activation.
            - **summary_first_dropout** (:obj:`float`) -- Optional dropout probability before the projection and
              activation.
            - **summary_last_dropout** (:obj:`float`)-- Optional dropout probability after the projection and
              activation.

        initializer_range (:obj:`float`, defaults to 0.02): The standard deviation to use to initialize the weights.
        kwargs:
            Additional keyword arguments passed along to the :obj:`__init__` of :obj:`tf.keras.layers.Layer`.
thomwolf's avatar
thomwolf committed
1575
    """
1576

Sylvain Gugger's avatar
Sylvain Gugger committed
1577
    def __init__(self, config: PretrainedConfig, initializer_range: float = 0.02, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
1578
        super().__init__(**kwargs)
thomwolf's avatar
thomwolf committed
1579

1580
1581
        self.summary_type = config.summary_type if hasattr(config, "summary_use_proj") else "last"
        if self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
1582
1583
1584
1585
1586
            # We should use a standard multi-head attention module with absolute positional embedding for that.
            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
            raise NotImplementedError

1587
        self.has_summary = hasattr(config, "summary_use_proj") and config.summary_use_proj
1588
        if self.has_summary:
1589
            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
thomwolf's avatar
thomwolf committed
1590
1591
1592
                num_classes = config.num_labels
            else:
                num_classes = config.hidden_size
1593
1594
1595
            self.summary = tf.keras.layers.Dense(
                num_classes, kernel_initializer=get_initializer(initializer_range), name="summary"
            )
thomwolf's avatar
thomwolf committed
1596

1597
        self.has_activation = hasattr(config, "summary_activation") and config.summary_activation == "tanh"
1598
        if self.has_activation:
1599
            self.activation = tf.keras.activations.tanh
thomwolf's avatar
thomwolf committed
1600

1601
        self.has_first_dropout = hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0
1602
        if self.has_first_dropout:
thomwolf's avatar
thomwolf committed
1603
1604
            self.first_dropout = tf.keras.layers.Dropout(config.summary_first_dropout)

1605
        self.has_last_dropout = hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0
1606
        if self.has_last_dropout:
thomwolf's avatar
thomwolf committed
1607
1608
            self.last_dropout = tf.keras.layers.Dropout(config.summary_last_dropout)

Julien Plu's avatar
Julien Plu committed
1609
    def call(self, inputs, cls_index=None, training=False):
thomwolf's avatar
thomwolf committed
1610
1611
1612
1613
1614
1615
1616
        if not isinstance(inputs, (dict, tuple, list)):
            hidden_states = inputs
        elif isinstance(inputs, (tuple, list)):
            hidden_states = inputs[0]
            cls_index = inputs[1] if len(inputs) > 1 else None
            assert len(inputs) <= 2, "Too many inputs."
        else:
1617
            hidden_states = inputs.get("hidden_states")
1618
            cls_index = inputs.get("cls_index", None)
thomwolf's avatar
thomwolf committed
1619

1620
        if self.summary_type == "last":
thomwolf's avatar
thomwolf committed
1621
            output = hidden_states[:, -1]
1622
        elif self.summary_type == "first":
thomwolf's avatar
thomwolf committed
1623
            output = hidden_states[:, 0]
1624
        elif self.summary_type == "mean":
Lysandre's avatar
Lysandre committed
1625
            output = tf.reduce_mean(hidden_states, axis=1)
1626
        elif self.summary_type == "cls_index":
1627
            hidden_shape = shape_list(hidden_states)  # e.g. [batch, num choices, seq length, hidden dims]
thomwolf's avatar
thomwolf committed
1628
            if cls_index is None:
1629
1630
1631
                cls_index = tf.fill(
                    hidden_shape[:-2], hidden_shape[-2] - 1
                )  # A tensor full of shape [batch] or [batch, num choices] full of sequence length
1632
1633
1634
1635
            cls_shape = shape_list(cls_index)
            if len(cls_shape) <= len(hidden_shape) - 2:
                cls_index = cls_index[..., tf.newaxis]
            # else:
1636
1637
            # cls_index = cls_index[..., tf.newaxis]
            # cls_index = cls_index.expand((-1,) * (cls_index.dim()-1) + (hidden_states.size(-1),))
thomwolf's avatar
thomwolf committed
1638
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
1639
            output = tf.gather(hidden_states, cls_index, batch_dims=len(hidden_shape) - 2)
1640
1641
1642
1643
            output = tf.squeeze(
                output, axis=len(hidden_shape) - 2
            )  # shape of output: (batch, num choices, hidden_size)
        elif self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
1644
1645
            raise NotImplementedError

1646
1647
        if self.has_first_dropout:
            output = self.first_dropout(output, training=training)
thomwolf's avatar
thomwolf committed
1648

1649
        if self.has_summary:
1650
            output = self.summary(output)
thomwolf's avatar
thomwolf committed
1651

1652
        if self.has_activation:
thomwolf's avatar
thomwolf committed
1653
1654
            output = self.activation(output)

1655
1656
        if self.has_last_dropout:
            output = self.last_dropout(output, training=training)
thomwolf's avatar
thomwolf committed
1657
1658
1659

        return output

1660

Julien Plu's avatar
Julien Plu committed
1661
def shape_list(tensor: tf.Tensor) -> List[int]:
Sylvain Gugger's avatar
Sylvain Gugger committed
1662
1663
1664
1665
    """
    Deal with dynamic shape in tensorflow cleanly.

    Args:
Julien Plu's avatar
Julien Plu committed
1666
        tensor (:obj:`tf.Tensor`): The tensor we want the shape of.
Sylvain Gugger's avatar
Sylvain Gugger committed
1667
1668
1669
1670

    Returns:
        :obj:`List[int]`: The shape of the tensor as a list.
    """
Julien Plu's avatar
Julien Plu committed
1671
    dynamic = tf.shape(tensor)
Julien Plu's avatar
Julien Plu committed
1672
1673

    if tensor.shape == tf.TensorShape(None):
1674
        return dynamic
Julien Plu's avatar
Julien Plu committed
1675
1676
1677

    static = tensor.shape.as_list()

thomwolf's avatar
thomwolf committed
1678
    return [dynamic[i] if s is None else s for i, s in enumerate(static)]
thomwolf's avatar
thomwolf committed
1679

1680

Sylvain Gugger's avatar
Sylvain Gugger committed
1681
1682
1683
1684
def get_initializer(initializer_range: float = 0.02) -> tf.initializers.TruncatedNormal:
    """
    Creates a :obj:`tf.initializers.TruncatedNormal` with the given range.

Julien Chaumond's avatar
Julien Chaumond committed
1685
    Args:
Sylvain Gugger's avatar
Sylvain Gugger committed
1686
1687
        initializer_range (`float`, defaults to 0.02): Standard deviation of the initializer range.

Julien Chaumond's avatar
Julien Chaumond committed
1688
    Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1689
        :obj:`tf.initializers.TruncatedNormal`: The truncated normal initializer.
Julien Chaumond's avatar
Julien Chaumond committed
1690
1691
    """
    return tf.keras.initializers.TruncatedNormal(stddev=initializer_range)
1692
1693


Sam Shleifer's avatar
Sam Shleifer committed
1694
1695
class TFWrappedEmbeddings:
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
1696
1697
1698
    this class wraps a the TFSharedEmbeddingTokens layer into a python 'no-keras-layer' class to avoid problem with
    weight restoring. Also it makes sure that the layer is called from the correct scope to avoid problem with
    saving/storing the correct weights
Sam Shleifer's avatar
Sam Shleifer committed
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
    """

    def __init__(self, layer, abs_scope_name=None):
        self._layer = layer
        self._abs_scope_name = abs_scope_name

    def call(self, inputs, mode="embedding"):
        if self._abs_scope_name is None:
            return self._layer.call(inputs, mode)

        # if an abs scope name is given to the embedding variable, call variable from absolute scope
        with tf.compat.v1.variable_scope(self._abs_scope_name, auxiliary_name_scope=False) as abs_scope_name:
            with tf.name_scope(abs_scope_name.original_name_scope):
                return self._layer.call(inputs, mode)

    def __call__(self, inputs, mode="embedding"):
        if self._abs_scope_name is None:
            return self._layer(inputs, mode)

        # if an abs scope name is given to the embedding variable, call variable from absolute scope
        with tf.compat.v1.variable_scope(self._abs_scope_name, auxiliary_name_scope=False) as abs_scope_name:
            with tf.name_scope(abs_scope_name.original_name_scope):
                return self._layer(inputs, mode)