modeling_xlnet.py 70.7 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
# coding=utf-8
# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch XLNet model.
"""
from __future__ import (absolute_import, division, print_function,
                        unicode_literals)
from __future__ import absolute_import, division, print_function, unicode_literals

import copy
import json
import logging
import math
import os
import sys
from io import open

import torch
from torch import nn
thomwolf's avatar
thomwolf committed
32
from torch.nn import functional as F
33
from torch.nn import CrossEntropyLoss, MSELoss
thomwolf's avatar
thomwolf committed
34
35
36
37
38
39
40
41
42
43
44
45
46
47

from .file_utils import cached_path, WEIGHTS_NAME, CONFIG_NAME

logger = logging.getLogger(__name__)

PRETRAINED_MODEL_ARCHIVE_MAP = {
    'xlnet-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/xlnet-large-cased-pytorch_model.bin",
}
PRETRAINED_CONFIG_ARCHIVE_MAP = {
    'xlnet-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/xlnet-large-cased-config.json",
}
XLNET_CONFIG_NAME = 'xlnet_config.json'
TF_WEIGHTS_NAME = 'model.ckpt'

thomwolf's avatar
thomwolf committed
48

49
def build_tf_xlnet_to_pytorch_map(model, config, tf_weights=None, finetuning_task=None):
thomwolf's avatar
thomwolf committed
50
51
52
53
54
55
56
57
    """ A map of modules from TF to PyTorch.
        I use a map to keep the PyTorch model as
        identical to the original PyTorch model as possible.
    """

    tf_to_pt_map = {}

    if hasattr(model, 'transformer'):
58
59
60
        if hasattr(model, 'lm_loss'):
            # We will load also the output bias
            tf_to_pt_map['model/lm_loss/bias'] = model.lm_loss.bias
61
        if hasattr(model, 'sequence_summary') and 'model/sequnece_summary/summary/kernel' in tf_weights:
62
63
64
            # We will load also the sequence summary
            tf_to_pt_map['model/sequnece_summary/summary/kernel'] = model.sequence_summary.summary.weight
            tf_to_pt_map['model/sequnece_summary/summary/bias'] = model.sequence_summary.summary.bias
65
        if hasattr(model, 'logits_proj') and finetuning_task is not None and 'model/regression_{}/logit/kernel'.format(finetuning_task) in tf_weights:
66
67
68
            tf_to_pt_map['model/regression_{}/logit/kernel'.format(finetuning_task)] = model.logits_proj.weight
            tf_to_pt_map['model/regression_{}/logit/bias'.format(finetuning_task)] = model.logits_proj.bias

thomwolf's avatar
thomwolf committed
69
70
71
72
73
        # Now load the rest of the transformer
        model = model.transformer

    # Embeddings and output
    tf_to_pt_map.update({'model/transformer/word_embedding/lookup_table': model.word_embedding.weight,
74
                         'model/transformer/mask_emb/mask_emb': model.mask_emb})
thomwolf's avatar
thomwolf committed
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117

    # Transformer blocks
    for i, b in enumerate(model.layer):
        layer_str = "model/transformer/layer_%d/" % i
        tf_to_pt_map.update({
            layer_str + "rel_attn/LayerNorm/gamma": b.rel_attn.layer_norm.weight,
            layer_str + "rel_attn/LayerNorm/beta": b.rel_attn.layer_norm.bias,
            layer_str + "rel_attn/o/kernel": b.rel_attn.o,
            layer_str + "rel_attn/q/kernel": b.rel_attn.q,
            layer_str + "rel_attn/k/kernel": b.rel_attn.k,
            layer_str + "rel_attn/r/kernel": b.rel_attn.r,
            layer_str + "rel_attn/v/kernel": b.rel_attn.v,
            layer_str + "ff/LayerNorm/gamma": b.ff.layer_norm.weight,
            layer_str + "ff/LayerNorm/beta": b.ff.layer_norm.bias,
            layer_str + "ff/layer_1/kernel": b.ff.layer_1.weight,
            layer_str + "ff/layer_1/bias": b.ff.layer_1.bias,
            layer_str + "ff/layer_2/kernel": b.ff.layer_2.weight,
            layer_str + "ff/layer_2/bias": b.ff.layer_2.bias,
        })

    # Relative positioning biases
    if config.untie_r:
        r_r_list = []
        r_w_list = []
        r_s_list = []
        seg_embed_list = []
        for b in model.layer:
            r_r_list.append(b.rel_attn.r_r_bias)
            r_w_list.append(b.rel_attn.r_w_bias)
            r_s_list.append(b.rel_attn.r_s_bias)
            seg_embed_list.append(b.rel_attn.seg_embed)
    else:
        r_r_list = [model.r_r_bias]
        r_w_list = [model.r_w_bias]
        r_s_list = [model.r_s_bias]
        seg_embed_list = [model.seg_embed]
    tf_to_pt_map.update({
        'model/transformer/r_r_bias': r_r_list,
        'model/transformer/r_w_bias': r_w_list,
        'model/transformer/r_s_bias': r_s_list,
        'model/transformer/seg_embed': seg_embed_list})
    return tf_to_pt_map

118
def load_tf_weights_in_xlnet(model, config, tf_path, finetuning_task=None):
thomwolf's avatar
thomwolf committed
119
120
121
122
123
124
125
126
127
128
129
    """ Load tf checkpoints in a pytorch model
    """
    try:
        import numpy as np
        import tensorflow as tf
    except ImportError:
        print("Loading a TensorFlow models in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions.")
        raise
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
thomwolf's avatar
thomwolf committed
130
    tf_weights = {}
thomwolf's avatar
thomwolf committed
131
132
133
    for name, shape in init_vars:
        print("Loading TF weight {} with shape {}".format(name, shape))
        array = tf.train.load_variable(tf_path, name)
thomwolf's avatar
thomwolf committed
134
        tf_weights[name] = array
thomwolf's avatar
thomwolf committed
135

136
137
    input("Press Enter to continue...")

138
    # Build TF to PyTorch weights loading map
139
    tf_to_pt_map = build_tf_xlnet_to_pytorch_map(model, config, tf_weights, finetuning_task)
140

thomwolf's avatar
thomwolf committed
141
142
    for name, pointer in tf_to_pt_map.items():
        print("Importing {}".format(name))
143
144
145
        if name not in tf_weights:
            print("{} not in tf pre-trained weights, skipping".format(name))
            continue
thomwolf's avatar
thomwolf committed
146
        array = tf_weights[name]
thomwolf's avatar
thomwolf committed
147
148
        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
        # which are not required for using pretrained model
149
        if 'kernel' in name and ('ff' in name or 'summary' in name or 'logit' in name):
thomwolf's avatar
thomwolf committed
150
            print("Transposing")
thomwolf's avatar
thomwolf committed
151
            array = np.transpose(array)
thomwolf's avatar
thomwolf committed
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
        if isinstance(pointer, list):
            # Here we will split the TF weigths
            assert len(pointer) == array.shape[0]
            for i, p_i in enumerate(pointer):
                arr_i = array[i, ...]
                try:
                    assert p_i.shape == arr_i.shape
                except AssertionError as e:
                    e.args += (p_i.shape, arr_i.shape)
                    raise
                print("Initialize PyTorch weight {} for layer {}".format(name, i))
                p_i.data = torch.from_numpy(arr_i)
        else:
            try:
                assert pointer.shape == array.shape
            except AssertionError as e:
                e.args += (pointer.shape, array.shape)
                raise
            print("Initialize PyTorch weight {}".format(name))
            pointer.data = torch.from_numpy(array)
        tf_weights.pop(name, None)
        tf_weights.pop(name + '/Adam', None)
        tf_weights.pop(name + '/Adam_1', None)

    print("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
thomwolf's avatar
thomwolf committed
177
178
179
180
    return model


def gelu(x):
181
182
    """ Implementation of the gelu activation function.
        XLNet is using OpenAI GPT's gelu (not exactly the same as BERT)
thomwolf's avatar
thomwolf committed
183
184
        Also see https://arxiv.org/abs/1606.08415
    """
185
186
    cdf = 0.5 * (1.0 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
    return x * cdf
thomwolf's avatar
thomwolf committed
187
188
189
190
191
192
193
194
195
196
197


def swish(x):
    return x * torch.sigmoid(x)


ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish}

class XLNetBaseConfig(object):
    @classmethod
    def from_dict(cls, json_object):
thomwolf's avatar
thomwolf committed
198
199
        """Constructs a `XLNetBaseConfig` from a Python dictionary of parameters."""
        config = cls(vocab_size_or_config_json_file=-1)
thomwolf's avatar
thomwolf committed
200
201
202
203
204
205
        for key, value in json_object.items():
            config.__dict__[key] = value
        return config

    @classmethod
    def from_json_file(cls, json_file):
thomwolf's avatar
thomwolf committed
206
        """Constructs a `XLNetBaseConfig` from a json file of parameters."""
thomwolf's avatar
thomwolf committed
207
208
209
210
        with open(json_file, "r", encoding='utf-8') as reader:
            text = reader.read()
        return cls.from_dict(json.loads(text))

thomwolf's avatar
thomwolf committed
211
212
213
214
215
    def update(self, other):
        dict_b = other.to_dict()
        for key, value in dict_b.items():
            self.__dict__[key] = value

thomwolf's avatar
thomwolf committed
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
    def __repr__(self):
        return str(self.to_json_string())

    def to_dict(self):
        """Serializes this instance to a Python dictionary."""
        output = copy.deepcopy(self.__dict__)
        return output

    def to_json_string(self):
        """Serializes this instance to a JSON string."""
        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"

    def to_json_file(self, json_file_path):
        """ Save this instance to a json file."""
        with open(json_file_path, "w", encoding='utf-8') as writer:
            writer.write(self.to_json_string())


class XLNetConfig(XLNetBaseConfig):
    """Configuration class to store the configuration of a `XLNetModel`.
    """
    def __init__(self,
                 vocab_size_or_config_json_file,
thomwolf's avatar
thomwolf committed
239
240
241
242
                 d_model=1024,
                 n_layer=24,
                 n_head=16,
                 d_inner=4096,
thomwolf's avatar
thomwolf committed
243
244
                 ff_activation="gelu",
                 untie_r=True,
thomwolf's avatar
thomwolf committed
245
                 attn_type="bi",
thomwolf's avatar
thomwolf committed
246
247
248

                 max_position_embeddings=512,
                 initializer_range=0.02,
thomwolf's avatar
thomwolf committed
249
250
251
252
253
254
255
256
257
258
259
260
                 layer_norm_eps=1e-12,

                 dropout=0.1,
                 dropatt=0.1,
                 init="normal",
                 init_range=0.1,
                 init_std=0.02,
                 mem_len=None,
                 reuse_len=None,
                 bi_data=False,
                 clamp_len=-1,
                 same_length=False):
thomwolf's avatar
thomwolf committed
261
262
263
264
265
266
267
268
269
270
271
272
273
        """Constructs XLNetConfig.

        Args:
            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XLNetModel`.
            d_model: Size of the encoder layers and the pooler layer.
            n_layer: Number of hidden layers in the Transformer encoder.
            n_head: Number of attention heads for each attention layer in
                the Transformer encoder.
            d_inner: The size of the "intermediate" (i.e., feed-forward)
                layer in the Transformer encoder.
            ff_activation: The non-linear activation function (function or string) in the
                encoder and pooler. If string, "gelu", "relu" and "swish" are supported.
            untie_r: untie relative position biases
thomwolf's avatar
thomwolf committed
274
            attn_type: 'bi' for XLNet, 'uni' for Transformer-XL
thomwolf's avatar
thomwolf committed
275
276
277
278
279
280
281
282
283
284
285

            dropout: The dropout probabilitiy for all fully connected
                layers in the embeddings, encoder, and pooler.
            dropatt: The dropout ratio for the attention
                probabilities.
            max_position_embeddings: The maximum sequence length that this model might
                ever be used with. Typically set this to something large just in case
                (e.g., 512 or 1024 or 2048).
            initializer_range: The sttdev of the truncated_normal_initializer for
                initializing all weight matrices.
            layer_norm_eps: The epsilon used by LayerNorm.
thomwolf's avatar
thomwolf committed
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301

            dropout: float, dropout rate.
            dropatt: float, dropout rate on attention probabilities.
            init: str, the initialization scheme, either "normal" or "uniform".
            init_range: float, initialize the parameters with a uniform distribution
                in [-init_range, init_range]. Only effective when init="uniform".
            init_std: float, initialize the parameters with a normal distribution
                with mean 0 and stddev init_std. Only effective when init="normal".
            mem_len: int, the number of tokens to cache.
            reuse_len: int, the number of tokens in the currect batch to be cached
                and reused in the future.
            bi_data: bool, whether to use bidirectional input pipeline.
                Usually set to True during pretraining and False during finetuning.
            clamp_len: int, clamp all relative distances larger than clamp_len.
                -1 means no clamping.
            same_length: bool, whether to use the same attention length for each token.
thomwolf's avatar
thomwolf committed
302
303
304
305
306
307
308
309
        """
        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                        and isinstance(vocab_size_or_config_json_file, unicode)):
            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
                json_config = json.loads(reader.read())
            for key, value in json_config.items():
                self.__dict__[key] = value
        elif isinstance(vocab_size_or_config_json_file, int):
thomwolf's avatar
thomwolf committed
310
            self.n_token = vocab_size_or_config_json_file
thomwolf's avatar
thomwolf committed
311
312
313
            self.d_model = d_model
            self.n_layer = n_layer
            self.n_head = n_head
thomwolf's avatar
thomwolf committed
314
315
            assert d_model % n_head == 0
            self.d_head = d_model // n_head
thomwolf's avatar
thomwolf committed
316
317
318
            self.ff_activation = ff_activation
            self.d_inner = d_inner
            self.untie_r = untie_r
thomwolf's avatar
thomwolf committed
319
            self.attn_type = attn_type
thomwolf's avatar
thomwolf committed
320

thomwolf's avatar
thomwolf committed
321
322
323
            self.max_position_embeddings = max_position_embeddings
            self.initializer_range = initializer_range
            self.layer_norm_eps = layer_norm_eps
thomwolf's avatar
thomwolf committed
324
325
326
327
328
329
330
331
332
333
334

            self.init = init
            self.init_range = init_range
            self.init_std = init_std
            self.dropout = dropout
            self.dropatt = dropatt
            self.mem_len = mem_len
            self.reuse_len = reuse_len
            self.bi_data = bi_data
            self.clamp_len = clamp_len
            self.same_length = same_length
thomwolf's avatar
thomwolf committed
335
336
337
338
339
340
341
342
343
344
345
346
        else:
            raise ValueError("First argument must be either a vocabulary size (int)"
                             "or the path to a pretrained model config file (str)")


class XLNetRunConfig(XLNetBaseConfig):
    """XLNetRunConfig contains hyperparameters that could be different
    between pretraining and finetuning.
    These hyperparameters can also be changed from run to run.
    We store them separately from XLNetConfig for flexibility.
    """
    def __init__(self, 
thomwolf's avatar
thomwolf committed
347
348
                 dropout=0.1,
                 dropatt=0.1,
thomwolf's avatar
thomwolf committed
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
                 init="normal",
                 init_range=0.1,
                 init_std=0.02,
                 mem_len=None,
                 reuse_len=None,
                 bi_data=False,
                 clamp_len=-1,
                 same_length=False):
        """
        Args:
        dropout: float, dropout rate.
        dropatt: float, dropout rate on attention probabilities.
        init: str, the initialization scheme, either "normal" or "uniform".
        init_range: float, initialize the parameters with a uniform distribution
            in [-init_range, init_range]. Only effective when init="uniform".
        init_std: float, initialize the parameters with a normal distribution
            with mean 0 and stddev init_std. Only effective when init="normal".
        mem_len: int, the number of tokens to cache.
        reuse_len: int, the number of tokens in the currect batch to be cached
            and reused in the future.
        bi_data: bool, whether to use bidirectional input pipeline.
            Usually set to True during pretraining and False during finetuning.
        clamp_len: int, clamp all relative distances larger than clamp_len.
            -1 means no clamping.
        same_length: bool, whether to use the same attention length for each token.
        """

        self.init = init
        self.init_range = init_range
        self.init_std = init_std
        self.dropout = dropout
        self.dropatt = dropatt
        self.mem_len = mem_len
        self.reuse_len = reuse_len
        self.bi_data = bi_data
        self.clamp_len = clamp_len
        self.same_length = same_length

try:
    from apex.normalization.fused_layer_norm import FusedLayerNorm as XLNetLayerNorm
except ImportError:
    logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
    class XLNetLayerNorm(nn.Module):
thomwolf's avatar
thomwolf committed
392
        def __init__(self, d_model, eps=1e-12):
thomwolf's avatar
thomwolf committed
393
394
395
            """Construct a layernorm module in the TF style (epsilon inside the square root).
            """
            super(XLNetLayerNorm, self).__init__()
thomwolf's avatar
thomwolf committed
396
397
            self.weight = nn.Parameter(torch.ones(d_model))
            self.bias = nn.Parameter(torch.zeros(d_model))
thomwolf's avatar
thomwolf committed
398
399
400
401
402
403
404
405
            self.variance_epsilon = eps

        def forward(self, x):
            u = x.mean(-1, keepdim=True)
            s = (x - u).pow(2).mean(-1, keepdim=True)
            x = (x - u) / torch.sqrt(s + self.variance_epsilon)
            return self.weight * x + self.bias

thomwolf's avatar
thomwolf committed
406
407
408
409
class XLNetRelativeAttention(nn.Module):
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(XLNetRelativeAttention, self).__init__()
        self.output_attentions = output_attentions
thomwolf's avatar
thomwolf committed
410
        if config.d_model % config.n_head != 0:
thomwolf's avatar
thomwolf committed
411
412
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention "
thomwolf's avatar
thomwolf committed
413
                "heads (%d)" % (config.d_model, config.n_head))
thomwolf's avatar
thomwolf committed
414
415
416
417
        self.output_attentions = output_attentions
        self.keep_multihead_output = keep_multihead_output
        self.multihead_output = None

thomwolf's avatar
thomwolf committed
418
        self.n_head = config.n_head
thomwolf's avatar
thomwolf committed
419
420
421
422
423
424
425
426
427
428
429
430
431
        self.d_head = config.d_head
        self.d_model = config.d_model
        self.scale = 1 / (config.d_head ** 0.5)

        self.q = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
        self.k = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
        self.v = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
        self.o = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
        self.r = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))

        self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
        self.r_s_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
        self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
thomwolf's avatar
thomwolf committed
432
        self.seg_embed = nn.Parameter(torch.Tensor(2, self.n_head, self.d_head))
thomwolf's avatar
thomwolf committed
433

thomwolf's avatar
thomwolf committed
434
        self.layer_norm = XLNetLayerNorm(config.d_model, eps=config.layer_norm_eps)
thomwolf's avatar
thomwolf committed
435
436
437
438
439
        self.dropout = nn.Dropout(config.dropout)

    def prune_heads(self, heads):
        raise NotImplementedError

thomwolf's avatar
thomwolf committed
440
441
442
443
444
445
446
447
448
449
450
451
    @staticmethod
    def rel_shift(x, klen=-1):
        """perform relative shift to form the relative attention score."""
        x_size = x.shape

        x = x.reshape(x_size[1], x_size[0], x_size[2], x_size[3])
        x = x[1:, ...]
        x = x.reshape(x_size[0], x_size[1] - 1, x_size[2], x_size[3])
        x = x[:, 0:klen, :, :]

        return x

thomwolf's avatar
thomwolf committed
452
453
454
455
456
457
458
459
    def rel_attn_core(self, q_head, k_head_h, v_head_h, k_head_r, seg_mat=None, attn_mask=None):
        """Core relative positional attention operations."""

        # content based attention score
        ac = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_w_bias, k_head_h)

        # position based attention score
        bd = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_r_bias, k_head_r)
thomwolf's avatar
thomwolf committed
460
        bd = self.rel_shift(bd, klen=ac.shape[1])
thomwolf's avatar
thomwolf committed
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491

        # segment based attention score
        if seg_mat is None:
            ef = 0
        else:
            ef = torch.einsum('ibnd,snd->ibns', q_head + self.r_s_bias, self.seg_embed)
            ef = torch.einsum('ijbs,ibns->ijbn', seg_mat, ef)

        # merge attention scores and perform masking
        attn_score = (ac + bd + ef) * self.scale
        if attn_mask is not None:
            # attn_score = attn_score * (1 - attn_mask) - 1e30 * attn_mask
            attn_score = attn_score - 1e30 * attn_mask

        # attention probability
        attn_prob = F.softmax(attn_score, dim=1)
        attn_prob = self.dropout(attn_prob)

        # attention output
        attn_vec = torch.einsum('ijbn,jbnd->ibnd', attn_prob, v_head_h)

        return attn_vec

    def post_attention(self, h, attn_vec, residual=True):
        """Post-attention processing."""
        # post-attention projection (back to `d_model`)
        attn_out = torch.einsum('ibnd,hnd->ibh', attn_vec, self.o)

        attn_out = self.dropout(attn_out)
        if residual:
            attn_out = attn_out + h
thomwolf's avatar
thomwolf committed
492
        output = self.layer_norm(attn_out)
thomwolf's avatar
thomwolf committed
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563

        return output

    def forward(self, h, g,
                      attn_mask_h, attn_mask_g,
                      r, seg_mat,
                      mems=None, target_mapping=None, head_mask=None):
        if g is not None:
            ###### Two-stream attention with relative positional encoding.
            # content based attention score
            if mems is not None and mems.dim() > 1:
                cat = torch.cat([mems, h], dim=0)
            else:
                cat = h

            # content-based key head
            k_head_h = torch.einsum('ibh,hnd->ibnd', cat, self.k)

            # content-based value head
            v_head_h = torch.einsum('ibh,hnd->ibnd', cat, self.v)

            # position-based key head
            k_head_r = torch.einsum('ibh,hnd->ibnd', r, self.r)

            ##### h-stream
            # content-stream query head
            q_head_h = torch.einsum('ibh,hnd->ibnd', h, self.q)

            # core attention ops
            attn_vec_h = self.rel_attn_core(
                q_head_h, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_h)

            # post processing
            output_h = self.post_attention(h, attn_vec_h)

            ##### g-stream
            # query-stream query head
            q_head_g = torch.einsum('ibh,hnd->ibnd', g, self.q)

            # core attention ops
            if target_mapping is not None:
                q_head_g = torch.einsum('mbnd,mlb->lbnd', q_head_g, target_mapping)
                attn_vec_g = self.rel_attn_core(
                    q_head_g, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_g)
                attn_vec_g = torch.einsum('lbnd,mlb->mbnd', attn_vec_g, target_mapping)
            else:
                attn_vec_g = self.rel_attn_core(
                    q_head_g, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_g)

            # post processing
            output_g = self.post_attention(g, attn_vec_g)
        else:
            ###### Multi-head attention with relative positional encoding
            if mems is not None and mems.dim() > 1:
                cat = torch.cat([mems, h], dim=0)
            else:
                cat = h

            # content heads
            q_head_h = torch.einsum('ibh,hnd->ibnd', h, self.q)
            k_head_h = torch.einsum('ibh,hnd->ibnd', cat, self.k)
            v_head_h = torch.einsum('ibh,hnd->ibnd', cat, self.v)

            # positional heads
            k_head_r = torch.einsum('ibh,hnd->ibnd', r, self.r)

            # core attention ops
            attn_vec = self.rel_attn_core(
                q_head_h, k_head_h, v_head_h, k_head_r, seg_mat=seg_mat, attn_mask=attn_mask_h)

            # post processing
thomwolf's avatar
thomwolf committed
564
565
            output_h = self.post_attention(h, attn_vec)
            output_g = None
thomwolf's avatar
thomwolf committed
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584


        # Mask heads if we want to
        # if head_mask is not None:
        #     attention_probs = attention_probs * head_mask

        # context_layer = torch.matmul(attention_probs, value_layer)
        # if self.keep_multihead_output:
        #     self.multihead_output = context_layer
        #     self.multihead_output.retain_grad()

        # context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        # new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        # context_layer = context_layer.view(*new_context_layer_shape)

        # if self.output_attentions:
        #     attentions, self_output = self_output
        # if self.output_attentions:
        #     return attentions, attention_output
thomwolf's avatar
thomwolf committed
585
        return output_h, output_g
thomwolf's avatar
thomwolf committed
586
587
588
589

class XLNetFeedForward(nn.Module):
    def __init__(self, config):
        super(XLNetFeedForward, self).__init__()
thomwolf's avatar
thomwolf committed
590
        self.layer_norm = XLNetLayerNorm(config.d_model, eps=config.layer_norm_eps)
thomwolf's avatar
thomwolf committed
591
592
593
594
595
596
597
598
        self.layer_1 = nn.Linear(config.d_model, config.d_inner)
        self.layer_2 = nn.Linear(config.d_inner, config.d_model)
        self.dropout = nn.Dropout(config.dropout)
        if isinstance(config.ff_activation, str) or (sys.version_info[0] == 2 and isinstance(config.ff_activation, unicode)):
            self.activation_function = ACT2FN[config.ff_activation]
        else:
            self.activation_function = config.ff_activation

thomwolf's avatar
thomwolf committed
599
600
601
602
603
604
605
    def forward(self, inp):
        output = inp
        output = self.layer_1(output)
        output = self.activation_function(output)
        output = self.dropout(output)
        output = self.layer_2(output)
        output = self.dropout(output)
thomwolf's avatar
thomwolf committed
606
        output = self.layer_norm(output + inp)
thomwolf's avatar
thomwolf committed
607
        return output
thomwolf's avatar
thomwolf committed
608
609
610
611
612
613
614
615
616
617
618
619

class XLNetLayer(nn.Module):
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(XLNetLayer, self).__init__()
        self.output_attentions = output_attentions
        self.rel_attn = XLNetRelativeAttention(config, output_attentions=output_attentions,
                                               keep_multihead_output=keep_multihead_output)
        self.ff = XLNetFeedForward(config)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, output_h, output_g,
                attn_mask_h, attn_mask_g,
thomwolf's avatar
thomwolf committed
620
621
                r, seg_mat,
                mems=None, target_mapping=None, head_mask=None):
thomwolf's avatar
thomwolf committed
622
623
624
625
        output_h, output_g = self.rel_attn(output_h, output_g,
                                           attn_mask_h, attn_mask_g,
                                           r, seg_mat,
                                           mems=mems, target_mapping=target_mapping, head_mask=head_mask)
thomwolf's avatar
thomwolf committed
626
        if output_g is not None:
thomwolf's avatar
thomwolf committed
627
628
629
630
631
632
633
            output_g = self.ff(output_g)
        output_h = self.ff(output_h)

        # if self.output_attentions:
        #     return attentions, layer_output
        return output_h, output_g

thomwolf's avatar
thomwolf committed
634
635
636
637
638
639
class XLNetPreTrainedModel(nn.Module):
    """ An abstract class to handle weights initialization and
        a simple interface for dowloading and loading pretrained models.
    """
    def __init__(self, config, *inputs, **kwargs):
        super(XLNetPreTrainedModel, self).__init__()
thomwolf's avatar
thomwolf committed
640
        if not isinstance(config, XLNetBaseConfig):
thomwolf's avatar
thomwolf committed
641
            raise ValueError(
thomwolf's avatar
thomwolf committed
642
                "Parameter config in `{}(config)` should be an instance of class `XLNetBaseConfig`. "
thomwolf's avatar
thomwolf committed
643
644
645
646
647
648
                "To create a model from a Google pretrained model use "
                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
                    self.__class__.__name__, self.__class__.__name__
                ))
        self.config = config

thomwolf's avatar
thomwolf committed
649
    def init_weights(self, module):
thomwolf's avatar
thomwolf committed
650
651
652
653
654
655
656
657
658
        """ Initialize the weights.
        """
        if isinstance(module, (nn.Linear, nn.Embedding)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, XLNetLayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
659
660
661
662
663
        elif isinstance(module, XLNetRelativeAttention):
            for param in [module.q, module.k, module.v, module.o, module.r,
                          module.r_r_bias, module.r_s_bias, module.r_w_bias,
                          module.seg_embed]:
                param.data.normal_(mean=0.0, std=self.config.initializer_range)
thomwolf's avatar
thomwolf committed
664
665
666
667
668
669
670
671
672
673
674
675
676
677
        if isinstance(module, nn.Linear) and module.bias is not None:
            module.bias.data.zero_()

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
        """
        Instantiate a XLNetPreTrainedModel from a pre-trained model file or a pytorch state dict.
        Download and cache the pre-trained model file if needed.

        Params:
            pretrained_model_name_or_path: either:
                - a str with the name of a pre-trained model to load selected in the list of:
                    . `xlnet-large-cased`
                - a path or url to a pretrained model archive containing:
678
                    . `config.json` a configuration file for the model
thomwolf's avatar
thomwolf committed
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
                    . `pytorch_model.bin` a PyTorch dump of a XLNetForPreTraining instance
                - a path or url to a pretrained model archive containing:
                    . `xlnet_config.json` a configuration file for the model
                    . `model.chkpt` a TensorFlow checkpoint
            from_tf: should we load the weights from a locally saved TensorFlow checkpoint
            cache_dir: an optional path to a folder in which the pre-trained models will be cached.
            state_dict: an optional state dictionnary (collections.OrderedDict object) to use instead of Google pre-trained models
            *inputs, **kwargs: additional input for the specific XLNet class
                (ex: num_labels for XLNetForSequenceClassification)
        """
        state_dict = kwargs.get('state_dict', None)
        kwargs.pop('state_dict', None)
        cache_dir = kwargs.get('cache_dir', None)
        kwargs.pop('cache_dir', None)
        from_tf = kwargs.get('from_tf', False)
        kwargs.pop('from_tf', None)

        if pretrained_model_name_or_path in PRETRAINED_MODEL_ARCHIVE_MAP:
            archive_file = PRETRAINED_MODEL_ARCHIVE_MAP[pretrained_model_name_or_path]
            config_file = PRETRAINED_CONFIG_ARCHIVE_MAP[pretrained_model_name_or_path]
        else:
            if from_tf:
                # Directly load from a TensorFlow checkpoint
                archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME)
                config_file = os.path.join(pretrained_model_name_or_path, XLNET_CONFIG_NAME)
            else:
                archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
                config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
        # redirect to the cache, if necessary
        try:
            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir)
        except EnvironmentError:
            if pretrained_model_name_or_path in PRETRAINED_MODEL_ARCHIVE_MAP:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained weights.".format(
                        archive_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find any file "
                    "associated to this path or url.".format(
                        pretrained_model_name_or_path,
                        ', '.join(PRETRAINED_MODEL_ARCHIVE_MAP.keys()),
                        archive_file))
            return None
        try:
            resolved_config_file = cached_path(config_file, cache_dir=cache_dir)
        except EnvironmentError:
            if pretrained_model_name_or_path in PRETRAINED_CONFIG_ARCHIVE_MAP:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
                        config_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find any file "
                    "associated to this path or url.".format(
                        pretrained_model_name_or_path,
                        ', '.join(PRETRAINED_CONFIG_ARCHIVE_MAP.keys()),
                        config_file))
            return None
        if resolved_archive_file == archive_file and resolved_config_file == config_file:
            logger.info("loading weights file {}".format(archive_file))
            logger.info("loading configuration file {}".format(config_file))
        else:
            logger.info("loading weights file {} from cache at {}".format(
                archive_file, resolved_archive_file))
            logger.info("loading configuration file {} from cache at {}".format(
                config_file, resolved_config_file))
748

thomwolf's avatar
thomwolf committed
749
750
        # Load config
        config = XLNetConfig.from_json_file(resolved_config_file)
751
752

        # Update config with kwargs if needed
753
754
        to_remove = []
        for key, value in kwargs.items():
755
756
            if hasattr(config, key):
                setattr(config, key, value)
757
758
759
760
761
                to_remove.append(key)
        for key in to_remove:
            kwargs.pop(key, None)

        logger.info("Model config {}".format(config))
762

thomwolf's avatar
thomwolf committed
763
764
765
766
767
768
        # Instantiate model.
        model = cls(config, *inputs, **kwargs)
        if state_dict is None and not from_tf:
            state_dict = torch.load(resolved_archive_file, map_location='cpu')
        if from_tf:
            # Directly load from a TensorFlow checkpoint
769
770
            return load_tf_weights_in_xlnet(model, config, resolved_archive_file)

thomwolf's avatar
thomwolf committed
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
        # Load from a PyTorch state_dict
        missing_keys = []
        unexpected_keys = []
        error_msgs = []
        # copy state_dict so _load_from_state_dict can modify it
        metadata = getattr(state_dict, '_metadata', None)
        state_dict = state_dict.copy()
        if metadata is not None:
            state_dict._metadata = metadata

        def load(module, prefix=''):
            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
            module._load_from_state_dict(
                state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)
            for name, child in module._modules.items():
                if child is not None:
                    load(child, prefix + name + '.')
        start_prefix = ''
789
790
        if not hasattr(model, 'transformer') and any(s.startswith('transformer') for s in state_dict.keys()):
            start_prefix = 'transformer.'
thomwolf's avatar
thomwolf committed
791
792
793
794
795
796
797
798
799
800
        load(model, prefix=start_prefix)
        if len(missing_keys) > 0:
            logger.info("Weights of {} not initialized from pretrained model: {}".format(
                model.__class__.__name__, missing_keys))
        if len(unexpected_keys) > 0:
            logger.info("Weights from pretrained model not used in {}: {}".format(
                model.__class__.__name__, unexpected_keys))
        if len(error_msgs) > 0:
            raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
                               model.__class__.__name__, "\n\t".join(error_msgs)))
801
802
        if isinstance(model, XLNetLMHeadModel):
            model.tie_weights()  # make sure word embedding weights are still tied
thomwolf's avatar
thomwolf committed
803
804
805
806
        return model


class XLNetModel(XLNetPreTrainedModel):
thomwolf's avatar
thomwolf committed
807
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
thomwolf's avatar
thomwolf committed
808
        super(XLNetModel, self).__init__(config)
thomwolf's avatar
thomwolf committed
809
810
811
        self.output_attentions = output_attentions
        self.mem_len = config.mem_len
        self.reuse_len = config.reuse_len
thomwolf's avatar
thomwolf committed
812
813
814
815
816
        self.d_model = config.d_model
        self.same_length = config.same_length
        self.attn_type = config.attn_type
        self.bi_data = config.bi_data
        self.clamp_len = config.clamp_len
thomwolf's avatar
thomwolf committed
817

thomwolf's avatar
thomwolf committed
818
819
        self.word_embedding = nn.Embedding(config.n_token, config.d_model)
        self.mask_emb = nn.Parameter(torch.Tensor(1, 1, config.d_model))
thomwolf's avatar
thomwolf committed
820
821
822
823
        layer = XLNetLayer(config, output_attentions=output_attentions,
                                   keep_multihead_output=keep_multihead_output)
        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.n_layer)])
        self.dropout = nn.Dropout(config.dropout)
thomwolf's avatar
thomwolf committed
824

thomwolf's avatar
thomwolf committed
825
826
827
828
829
830
831
832
833
834
835
836
837
    def prune_heads(self, heads_to_prune):
        """ Prunes heads of the model.
            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
        """
        for layer, heads in heads_to_prune.items():
            self.layer[layer].attention.prune_heads(heads)

    def get_multihead_outputs(self):
        """ Gather all multi-head outputs.
            Return: list (layers) of multihead module outputs with gradients
        """
        return [layer.attention.self.multihead_output for layer in self.layer]

thomwolf's avatar
thomwolf committed
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
    def create_mask(self, qlen, mlen):
        """ create causal attention mask.
            float mask where 1.0 indicate masked, 0.0 indicated not-masked.
             same_length=False:      same_length=True:
             <mlen > <  qlen >       <mlen > <  qlen >
          ^ [0 0 0 0 0 1 1 1 1]     [0 0 0 0 0 1 1 1 1]
            [0 0 0 0 0 0 1 1 1]     [1 0 0 0 0 0 1 1 1]
       qlen [0 0 0 0 0 0 0 1 1]     [1 1 0 0 0 0 0 1 1]
            [0 0 0 0 0 0 0 0 1]     [1 1 1 0 0 0 0 0 1]
          v [0 0 0 0 0 0 0 0 0]     [1 1 1 1 0 0 0 0 0]
        """
        attn_mask = torch.ones([qlen, qlen])
        mask_up = torch.triu(attn_mask, diagonal=1)
        attn_mask_pad = torch.zeros([qlen, mlen])
        ret = torch.cat([attn_mask_pad, mask_up], dim=1)
        if self.same_length:
            mask_lo = torch.tril(attn_mask, diagonal=-1)
            ret = torch.cat([ret[:, :qlen] + mask_lo, ret[:, qlen:]], dim=1)

        ret = ret.to(next(self.parameters()))
thomwolf's avatar
thomwolf committed
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
        return ret

    def cache_mem(self, curr_out, prev_mem):
        """cache hidden states into memory."""
        if self.mem_len is None or self.mem_len == 0:
            return None
        else:
            if self.reuse_len is not None and self.reuse_len > 0:
                curr_out = curr_out[:self.reuse_len]

            if prev_mem is None:
                new_mem = curr_out[-self.mem_len:]
            else:
                new_mem = torch.cat([prev_mem, curr_out], dim=0)[-self.mem_len:]

        return new_mem.detach()

thomwolf's avatar
thomwolf committed
875
876
877
878
879
880
881
882
883
884
885
886
    @staticmethod
    def positional_embedding(pos_seq, inv_freq, bsz=None):
        sinusoid_inp = torch.einsum('i,d->id', pos_seq, inv_freq)
        pos_emb = torch.cat([torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)], dim=-1)
        pos_emb = pos_emb[:, None, :]

        if bsz is not None:
            pos_emb = pos_emb.expand(-1, bsz, -1)

        return pos_emb

    def relative_positional_encoding(self, qlen, klen, bsz=None):
thomwolf's avatar
thomwolf committed
887
        """create relative positional encoding."""
thomwolf's avatar
thomwolf committed
888
889
        freq_seq = torch.arange(0, self.d_model, 2.0, dtype=torch.float)
        inv_freq = 1 / (10000 ** (freq_seq / self.d_model))
thomwolf's avatar
thomwolf committed
890
891
892
893
894
895
896
897
898
899
900

        if self.attn_type == 'bi':
            # beg, end = klen - 1, -qlen
            beg, end = klen, -qlen
        elif self.attn_type == 'uni':
            # beg, end = klen - 1, -1
            beg, end = klen, -1
        else:
            raise ValueError('Unknown `attn_type` {}.'.format(self.attn_type))

        if self.bi_data:
thomwolf's avatar
thomwolf committed
901
902
            fwd_pos_seq = torch.arange(beg, end, -1.0, dtype=torch.float)
            bwd_pos_seq = torch.arange(-beg, -end, 1.0, dtype=torch.float)
thomwolf's avatar
thomwolf committed
903
904
905
906
907
908

            if self.clamp_len > 0:
                fwd_pos_seq = fwd_pos_seq.clamp(-self.clamp_len, self.clamp_len)
                bwd_pos_seq = bwd_pos_seq.clamp(-self.clamp_len, self.clamp_len)

            if bsz is not None:
thomwolf's avatar
thomwolf committed
909
910
                fwd_pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz//2)
                bwd_pos_emb = self.positional_embedding(bwd_pos_seq, inv_freq, bsz//2)
thomwolf's avatar
thomwolf committed
911
            else:
thomwolf's avatar
thomwolf committed
912
913
                fwd_pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq)
                bwd_pos_emb = self.positional_embedding(bwd_pos_seq, inv_freq)
thomwolf's avatar
thomwolf committed
914
915
916

            pos_emb = torch.cat([fwd_pos_emb, bwd_pos_emb], dim=1)
        else:
thomwolf's avatar
thomwolf committed
917
            fwd_pos_seq = torch.arange(beg, end, -1.0)
thomwolf's avatar
thomwolf committed
918
919
            if self.clamp_len > 0:
                fwd_pos_seq = fwd_pos_seq.clamp(-self.clamp_len, self.clamp_len)
thomwolf's avatar
thomwolf committed
920
            pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz)
thomwolf's avatar
thomwolf committed
921

thomwolf's avatar
thomwolf committed
922
        pos_emb = pos_emb.to(next(self.parameters()))
thomwolf's avatar
thomwolf committed
923
924
        return pos_emb

925
    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
thomwolf's avatar
thomwolf committed
926
927
928
929
                mems=None, perm_mask=None, target_mapping=None, inp_q=None,
                output_all_encoded_layers=True, head_mask=None):
        """
        Args:
930
            inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
thomwolf's avatar
thomwolf committed
931
            token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
932
            input_mask: [optional] float32 Tensor in shape [bsz, len], the input mask.
thomwolf's avatar
thomwolf committed
933
                0 for real tokens and 1 for padding.
934
935
936
937
            attention_mask: [optional] float32 Tensor, SAME FUNCTION as `input_mask`
                but with 1 for real tokens and 0 for padding.
                Added for easy compatibility with the BERT model (which uses this negative masking).
                You can only uses one among `input_mask` and `attention_mask`
thomwolf's avatar
thomwolf committed
938
            mems: [optional] a list of float32 Tensors in shape [mem_len, bsz, d_model], memory
thomwolf's avatar
thomwolf committed
939
940
                from previous batches. The length of the list equals n_layer.
                If None, no memory is used.
941
942
943
            perm_mask: [optional] float32 Tensor in shape [bsz, len, len].
                If perm_mask[k, i, j] = 0, i attend to j in batch k;
                if perm_mask[k, i, j] = 1, i does not attend to j in batch k.
thomwolf's avatar
thomwolf committed
944
                If None, each position attends to all the others.
945
946
            target_mapping: [optional] float32 Tensor in shape [bsz, num_predict, len].
                If target_mapping[k, i, j] = 1, the i-th predict in batch k is
thomwolf's avatar
thomwolf committed
947
948
949
                on the j-th token.
                Only used during pretraining for partial prediction.
                Set to None during finetuning.
950
            inp_q: [optional] float32 Tensor in shape [bsz, len].
thomwolf's avatar
thomwolf committed
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
                1 for tokens with losses and 0 for tokens without losses.
                Only used during pretraining for two-stream attention.
                Set to None during finetuning.

            mem_len: int, the number of tokens to cache.
            reuse_len: int, the number of tokens in the currect batch to be cached
                and reused in the future.
            bi_data: bool, whether to use bidirectional input pipeline.
                Usually set to True during pretraining and False during finetuning.
            clamp_len: int, clamp all relative distances larger than clamp_len.
                -1 means no clamping.
            same_length: bool, whether to use the same attention length for each token.
            summary_type: str, "last", "first", "mean", or "attn". The method
                to pool the input to get a vector representation.
        """
966
967
968
969
        # the original code for XLNet uses shapes [len, bsz] with the batch dimension at the end
        # but we want a unified interface in the library with the batch size on the first dimension
        # so we move here the first dimension (batch) to the end
        inp_k = inp_k.transpose(0, 1).contiguous()
thomwolf's avatar
thomwolf committed
970
        token_type_ids = token_type_ids.transpose(0, 1).contiguous() if token_type_ids is not None else None
971
        input_mask = input_mask.transpose(0, 1).contiguous() if input_mask is not None else None
thomwolf's avatar
thomwolf committed
972
        attention_mask = attention_mask.transpose(0, 1).contiguous() if attention_mask is not None else None
973
974
975
976
        perm_mask = perm_mask.permute(1, 2, 0).contiguous() if perm_mask is not None else None
        target_mapping = target_mapping.permute(1, 2, 0).contiguous() if target_mapping is not None else None
        inp_q = inp_q.transpose(0, 1).contiguous() if inp_q is not None else None

thomwolf's avatar
thomwolf committed
977
        qlen, bsz = inp_k.shape[0], inp_k.shape[1]
thomwolf's avatar
thomwolf committed
978
979
        mlen = mems[0].shape[0] if mems is not None else 0
        klen = mlen + qlen
thomwolf's avatar
thomwolf committed
980
981
982

        dtype_float = next(self.parameters()).dtype
        device = next(self.parameters()).device
thomwolf's avatar
thomwolf committed
983
984
985
986

        ##### Attention mask
        # causal attention mask
        if self.attn_type == 'uni':
thomwolf's avatar
thomwolf committed
987
            attn_mask = self.create_mask(qlen, mlen)
thomwolf's avatar
thomwolf committed
988
989
990
991
992
993
994
            attn_mask = attn_mask[:, :, None, None]
        elif self.attn_type == 'bi':
            attn_mask = None
        else:
            raise ValueError('Unsupported attention type: {}'.format(self.attn_type))

        # data mask: input mask & perm mask
995
996
997
998
999
1000
1001
1002
1003
        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) "
        "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
        if input_mask is None and attention_mask is not None:
            input_mask = 1.0 - attention_mask
        if input_mask is not None and perm_mask is not None:
            data_mask = input_mask[None] + perm_mask
        elif input_mask is not None and perm_mask is None:
            data_mask = input_mask[None]
        elif input_mask is None and perm_mask is not None:
thomwolf's avatar
thomwolf committed
1004
1005
1006
1007
1008
1009
            data_mask = perm_mask
        else:
            data_mask = None

        if data_mask is not None:
            # all mems can be attended to
thomwolf's avatar
thomwolf committed
1010
            mems_mask = torch.zeros([data_mask.shape[0], mlen, bsz]).to(data_mask)
thomwolf's avatar
thomwolf committed
1011
1012
1013
1014
1015
1016
1017
            data_mask = torch.cat([mems_mask, data_mask], dim=1)
            if attn_mask is None:
                attn_mask = data_mask[:, :, :, None]
            else:
                attn_mask += data_mask[:, :, :, None]

        if attn_mask is not None:
thomwolf's avatar
thomwolf committed
1018
            attn_mask = (attn_mask > 0).to(dtype_float)
thomwolf's avatar
thomwolf committed
1019
1020

        if attn_mask is not None:
thomwolf's avatar
thomwolf committed
1021
1022
1023
            non_tgt_mask = -torch.eye(qlen).to(attn_mask)
            non_tgt_mask = torch.cat([torch.zeros([qlen, mlen]).to(attn_mask), non_tgt_mask], dim=-1)
            non_tgt_mask = ((attn_mask + non_tgt_mask[:, :, None, None]) > 0).to(attn_mask)
thomwolf's avatar
thomwolf committed
1024
1025
1026
        else:
            non_tgt_mask = None

thomwolf's avatar
thomwolf committed
1027
1028
        ##### Word embeddings and prepare h & g hidden states
        word_emb_k = self.word_embedding(inp_k)
thomwolf's avatar
thomwolf committed
1029
1030
1031
        output_h = self.dropout(word_emb_k)
        if inp_q is not None:
            if target_mapping is not None:
1032
                word_emb_q = self.mask_emb.expand(target_mapping.shape[0], bsz, -1)
thomwolf's avatar
thomwolf committed
1033
1034
            else:
                inp_q_ext = inp_q[:, :, None]
1035
                word_emb_q = inp_q_ext * self.mask_emb + (1 - inp_q_ext) * word_emb_k
thomwolf's avatar
thomwolf committed
1036
1037
1038
1039
1040
            output_g = self.dropout(word_emb_q)
        else:
            output_g = None

        ##### Segment embedding
thomwolf's avatar
thomwolf committed
1041
1042
        if token_type_ids is not None:
            # Convert `token_type_ids` to one-hot `seg_mat`
thomwolf's avatar
thomwolf committed
1043
            mem_pad = torch.zeros([mlen, bsz], dtype=torch.long, device=device)
thomwolf's avatar
thomwolf committed
1044
            cat_ids = torch.cat([mem_pad, token_type_ids], dim=0)
thomwolf's avatar
thomwolf committed
1045
1046

            # `1` indicates not in the same segment [qlen x klen x bsz]
thomwolf's avatar
thomwolf committed
1047
            seg_mat = (token_type_ids[:, None] != cat_ids[None, :]).long()
thomwolf's avatar
thomwolf committed
1048
            seg_mat = F.one_hot(seg_mat, num_classes=2).to(dtype_float)
thomwolf's avatar
thomwolf committed
1049
1050
1051
1052
        else:
            seg_mat = None

        ##### Positional encoding
thomwolf's avatar
thomwolf committed
1053
        pos_emb = self.relative_positional_encoding(qlen, klen, bsz=bsz)
thomwolf's avatar
thomwolf committed
1054
1055
1056
1057
1058
        pos_emb = self.dropout(pos_emb)

        ##### Head mask if needed (for bertology/pruning)
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
thomwolf's avatar
thomwolf committed
1059
1060
        # input head_mask has shape [num_heads] or [n_layer x num_heads]
        # and head_mask is converted to shape [n_layer x batch x num_heads x seq_length x seq_length]
thomwolf's avatar
thomwolf committed
1061
1062
1063
        if head_mask is not None:
            if head_mask.dim() == 1:
                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
thomwolf's avatar
thomwolf committed
1064
                head_mask = head_mask.expand(self.config.n_layer, -1, -1, -1, -1)
thomwolf's avatar
thomwolf committed
1065
1066
1067
1068
            elif head_mask.dim() == 2:
                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
        else:
thomwolf's avatar
thomwolf committed
1069
            head_mask = [None] * self.config.n_layer
thomwolf's avatar
thomwolf committed
1070
1071
1072
1073
1074

        new_mems = []
        if mems is None:
            mems = [None] * len(self.layer)

1075
        hidden_states = []
thomwolf's avatar
thomwolf committed
1076
1077
1078
1079
1080
        for i, layer_module in enumerate(self.layer):
            # cache new mems
            new_mems.append(self.cache_mem(output_h, mems[i]))

            output_h, output_g = layer_module(output_h, output_g,
thomwolf's avatar
thomwolf committed
1081
1082
                                              attn_mask_h=non_tgt_mask, attn_mask_g=attn_mask,
                                              r=pos_emb, seg_mat=seg_mat,
thomwolf's avatar
thomwolf committed
1083
1084
                                              mems=mems[i], target_mapping=target_mapping,
                                              head_mask=head_mask)
1085
            hidden_states.append(output_h)
thomwolf's avatar
thomwolf committed
1086
1087
        output = self.dropout(output_g if output_g is not None else output_h)

1088
1089
1090
1091
1092
        # We transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
        output = output.permute(1, 0, 2).contiguous()
        hidden_states = [hs.permute(1, 0, 2).contiguous() for hs in hidden_states]

        return output, hidden_states, new_mems
thomwolf's avatar
thomwolf committed
1093
1094
1095


class XLNetLMHeadModel(XLNetPreTrainedModel):
thomwolf's avatar
thomwolf committed
1096
1097
1098
1099
1100
1101
1102
1103
1104
    """XLNet model ("XLNet: Generalized Autoregressive Pretraining for Language Understanding").

    Params:
        `config`: a XLNetConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
1105
        inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
thomwolf's avatar
thomwolf committed
1106
        token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
1107
        input_mask: [optional] float32 Tensor in shape [bsz, len], the input mask.
1108
            0 for real tokens and 1 for padding.
1109
1110
1111
1112
        attention_mask: [optional] float32 Tensor, SAME FUNCTION as `input_mask`
            but with 1 for real tokens and 0 for padding.
            Added for easy compatibility with the BERT model (which uses this negative masking).
            You can only uses one among `input_mask` and `attention_mask`
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
        mems: [optional] a list of float32 Tensors in shape [mem_len, bsz, d_model], memory
            from previous batches. The length of the list equals n_layer.
            If None, no memory is used.
        perm_mask: [optional] float32 Tensor in shape [bsz, len, len].
            If perm_mask[k, i, j] = 0, i attend to j in batch k;
            if perm_mask[k, i, j] = 1, i does not attend to j in batch k.
            If None, each position attends to all the others.
        target_mapping: [optional] float32 Tensor in shape [bsz, num_predict, len].
            If target_mapping[k, i, j] = 1, the i-th predict in batch k is
            on the j-th token.
            Only used during pretraining for partial prediction.
            Set to None during finetuning.
        inp_q: [optional] float32 Tensor in shape [bsz, len].
            1 for tokens with losses and 0 for tokens without losses.
            Only used during pretraining for two-stream attention.
            Set to None during finetuning.
thomwolf's avatar
thomwolf committed
1129
1130
1131
1132
1133
1134


    Outputs: Tuple of (encoded_layers, pooled_output)
        `encoded_layers`: controled by `output_all_encoded_layers` argument:
            - `output_all_encoded_layers=True`: outputs a list of the full sequences of encoded-hidden-states at the end
                of each attention block (i.e. 12 full sequences for XLNet-base, 24 for XLNet-large), each
thomwolf's avatar
thomwolf committed
1135
                encoded-hidden-state is a torch.FloatTensor of size [batch_size, sequence_length, d_model],
thomwolf's avatar
thomwolf committed
1136
            - `output_all_encoded_layers=False`: outputs only the full sequence of hidden-states corresponding
thomwolf's avatar
thomwolf committed
1137
1138
                to the last attention block of shape [batch_size, sequence_length, d_model],
        `pooled_output`: a torch.FloatTensor of size [batch_size, d_model] which is the output of a
thomwolf's avatar
thomwolf committed
1139
1140
1141
1142
1143
1144
1145
            classifier pretrained on top of the hidden state associated to the first character of the
            input (`CLS`) to train on the Next-Sentence task (see XLNet's paper).

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
1146
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
thomwolf's avatar
thomwolf committed
1147
1148
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

thomwolf's avatar
thomwolf committed
1149
    config = modeling.XLNetConfig(vocab_size_or_config_json_file=32000, d_model=768,
thomwolf's avatar
thomwolf committed
1150
        n_layer=12, num_attention_heads=12, intermediate_size=3072)
thomwolf's avatar
thomwolf committed
1151
1152

    model = modeling.XLNetModel(config=config)
1153
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
thomwolf's avatar
thomwolf committed
1154
1155
    ```
    """
thomwolf's avatar
thomwolf committed
1156
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
thomwolf's avatar
thomwolf committed
1157
        super(XLNetLMHeadModel, self).__init__(config)
thomwolf's avatar
thomwolf committed
1158
        self.output_attentions = output_attentions
thomwolf's avatar
thomwolf committed
1159
1160
        self.attn_type = config.attn_type
        self.same_length = config.same_length
thomwolf's avatar
thomwolf committed
1161

thomwolf's avatar
thomwolf committed
1162
1163
        self.transformer = XLNetModel(config, output_attentions=output_attentions,
                                              keep_multihead_output=keep_multihead_output)
thomwolf's avatar
thomwolf committed
1164
        self.lm_loss = nn.Linear(config.d_model, config.n_token, bias=True)
thomwolf's avatar
thomwolf committed
1165

thomwolf's avatar
thomwolf committed
1166
1167
        # Tie weights

thomwolf's avatar
thomwolf committed
1168
        self.apply(self.init_weights)
thomwolf's avatar
thomwolf committed
1169
        self.tie_weights()
thomwolf's avatar
thomwolf committed
1170

thomwolf's avatar
thomwolf committed
1171
1172
    def tie_weights(self):
        """ Make sure we are sharing the embeddings
thomwolf's avatar
thomwolf committed
1173
        """
thomwolf's avatar
thomwolf committed
1174
        self.lm_loss.weight = self.transformer.word_embedding.weight
thomwolf's avatar
thomwolf committed
1175

1176
    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
thomwolf's avatar
thomwolf committed
1177
                mems=None, perm_mask=None, target_mapping=None, inp_q=None,
thomwolf's avatar
thomwolf committed
1178
                target=None, output_all_encoded_layers=True, head_mask=None):
thomwolf's avatar
thomwolf committed
1179
1180
        """
        Args:
1181
            inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
thomwolf's avatar
thomwolf committed
1182
            token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
1183
            input_mask: float32 Tensor in shape [bsz, len], the input mask.
thomwolf's avatar
thomwolf committed
1184
                0 for real tokens and 1 for padding.
1185
1186
1187
1188
            attention_mask: [optional] float32 Tensor, SAME FUNCTION as `input_mask`
                but with 1 for real tokens and 0 for padding.
                Added for easy compatibility with the BERT model (which uses this negative masking).
                You can only uses one among `input_mask` and `attention_mask`
thomwolf's avatar
thomwolf committed
1189
1190
1191
            mems: a list of float32 Tensors in shape [mem_len, bsz, d_model], memory
                from previous batches. The length of the list equals n_layer.
                If None, no memory is used.
1192
1193
1194
            perm_mask: float32 Tensor in shape [bsz, len, len].
                If perm_mask[k, i, j] = 0, i attend to j in batch k;
                if perm_mask[k, i, j] = 1, i does not attend to j in batch k.
thomwolf's avatar
thomwolf committed
1195
                If None, each position attends to all the others.
1196
1197
            target_mapping: float32 Tensor in shape [bsz, num_predict, len].
                If target_mapping[k, i, j] = 1, the i-th predict in batch k is
thomwolf's avatar
thomwolf committed
1198
1199
1200
                on the j-th token.
                Only used during pretraining for partial prediction.
                Set to None during finetuning.
1201
            inp_q: float32 Tensor in shape [bsz, len].
thomwolf's avatar
thomwolf committed
1202
1203
1204
1205
1206
1207
1208
                1 for tokens with losses and 0 for tokens without losses.
                Only used during pretraining for two-stream attention.
                Set to None during finetuning.

            summary_type: str, "last", "first", "mean", or "attn". The method
                to pool the input to get a vector representation.
        """
1209
        output, hidden_states, new_mems = self.transformer(inp_k, token_type_ids, input_mask, attention_mask,
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
                                            mems, perm_mask, target_mapping, inp_q,
                                            output_all_encoded_layers, head_mask)

        logits = self.lm_loss(output)

        if target is not None:
            # Flatten the tokens
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            loss = loss_fct(logits.view(-1, logits.size(-1)),
                            target.view(-1))
            return loss, new_mems

        # if self.output_attentions:
        #     all_attentions, encoded_layers = encoded_layers
        # sequence_output = encoded_layers[-1]
        # pooled_output = self.pooler(sequence_output)
        # if not output_all_encoded_layers:
        #     encoded_layers = encoded_layers[-1]
        # if self.output_attentions:
        return logits, new_mems
        #     return all_attentions, encoded_layers, pooled_output

1232
1233
1234
1235
1236
1237
class XLNetSequenceSummary(nn.Module):
    def __init__(self, config, summary_type="last", use_proj=True,
                 output_attentions=False, keep_multihead_output=False):
        super(XLNetSequenceSummary, self).__init__()
        self.summary_type = summary_type
        if use_proj:
thomwolf's avatar
thomwolf committed
1238
            self.summary = nn.Linear(config.d_model, config.d_model)
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
        else:
            self.summary = None
        if summary_type == 'attn':
            # 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
        self.dropout = nn.Dropout(config.dropout)
        self.activation = nn.Tanh()

thomwolf's avatar
thomwolf committed
1249
1250
    def forward(self, hidden_states):
        """ hidden_states: float Tensor in shape [bsz, seq_len, d_model], the hidden-states of the last layer."""
1251
        if self.summary_type == 'last':
thomwolf's avatar
thomwolf committed
1252
            output = hidden_states[:, -1]
1253
        elif self.summary_type == 'first':
thomwolf's avatar
thomwolf committed
1254
            output = hidden_states[:, 0]
1255
        elif self.summary_type == 'mean':
thomwolf's avatar
thomwolf committed
1256
            output = hidden_states.mean(dim=1)
1257
1258
1259
1260
1261
        elif summary_type == 'attn':
            raise NotImplementedError

        output = self.summary(output)
        output = self.activation(output)
thomwolf's avatar
thomwolf committed
1262
        output = self.dropout(output)
1263
1264
        return output

1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278

class XLNetForSequenceClassification(XLNetPreTrainedModel):
    """XLNet model ("XLNet: Generalized Autoregressive Pretraining for Language Understanding").

    Params:
        `config`: a XLNetConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False
        `summary_type`: str, "last", "first", "mean", or "attn". The method
            to pool the input to get a vector representation. Default: last

    Inputs:
        inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
thomwolf's avatar
thomwolf committed
1279
        token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
1280
        input_mask: float32 Tensor in shape [bsz, len], the input mask.
1281
            0 for real tokens and 1 for padding.
1282
1283
1284
1285
        attention_mask: [optional] float32 Tensor, SAME FUNCTION as `input_mask`
            but with 1 for real tokens and 0 for padding.
            Added for easy compatibility with the BERT model (which uses this negative masking).
            You can only uses one among `input_mask` and `attention_mask`
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
        mems: a list of float32 Tensors in shape [mem_len, bsz, d_model], memory
            from previous batches. The length of the list equals n_layer.
            If None, no memory is used.
        perm_mask: float32 Tensor in shape [bsz, len, len].
            If perm_mask[k, i, j] = 0, i attend to j in batch k;
            if perm_mask[k, i, j] = 1, i does not attend to j in batch k.
            If None, each position attends to all the others.
        target_mapping: float32 Tensor in shape [bsz, num_predict, len].
            If target_mapping[k, i, j] = 1, the i-th predict in batch k is
            on the j-th token.
            Only used during pretraining for partial prediction.
            Set to None during finetuning.
        inp_q: float32 Tensor in shape [bsz, len].
            1 for tokens with losses and 0 for tokens without losses.
            Only used during pretraining for two-stream attention.
            Set to None during finetuning.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.


    Outputs: Tuple of (logits or loss, mems)
        `logits or loss`:
            if target is None:
                Token logits with shape [batch_size, sequence_length] 
            else:
                CrossEntropy loss with the targets
        `new_mems`: list (num layers) of updated mem states at the entry of each layer
            each mem state is a torch.FloatTensor of size [self.config.mem_len, batch_size, self.config.d_model]
            Note that the first two dimensions are transposed in `mems` with regards to `input_ids` and `target`

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
1320
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
1321
1322
1323
1324
1325
1326
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = modeling.XLNetConfig(vocab_size_or_config_json_file=32000, d_model=768,
        n_layer=12, num_attention_heads=12, intermediate_size=3072)

    model = modeling.XLNetModel(config=config)
1327
    all_encoder_layers, pooled_output = model(input_ids, token_type_ids, input_mask)
1328
1329
    ```
    """
1330
1331
    def __init__(self, config, summary_type="last", use_proj=True, num_labels=2,
                 is_regression=False, output_attentions=False, keep_multihead_output=False):
1332
1333
1334
1335
1336
        super(XLNetForSequenceClassification, self).__init__(config)
        self.output_attentions = output_attentions
        self.attn_type = config.attn_type
        self.same_length = config.same_length
        self.summary_type = summary_type
1337
        self.is_regression = is_regression
1338
1339
1340
1341

        self.transformer = XLNetModel(config, output_attentions=output_attentions,
                                              keep_multihead_output=keep_multihead_output)

1342
1343
1344
        self.sequence_summary = XLNetSequenceSummary(config, summary_type=summary_type,
                                                     use_proj=use_proj, output_attentions=output_attentions,
                                                     keep_multihead_output=keep_multihead_output)
1345
        self.logits_proj = nn.Linear(config.d_model, num_labels if not is_regression else 1)
thomwolf's avatar
thomwolf committed
1346
        self.apply(self.init_weights)
1347

1348
    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
1349
1350
1351
1352
1353
                mems=None, perm_mask=None, target_mapping=None, inp_q=None,
                target=None, output_all_encoded_layers=True, head_mask=None):
        """
        Args:
            inp_k: int32 Tensor in shape [bsz, len], the input token IDs.
thomwolf's avatar
thomwolf committed
1354
            token_type_ids: int32 Tensor in shape [bsz, len], the input segment IDs.
1355
            input_mask: float32 Tensor in shape [bsz, len], the input mask.
1356
                0 for real tokens and 1 for padding.
1357
1358
1359
1360
            attention_mask: [optional] float32 Tensor, SAME FUNCTION as `input_mask`
                but with 1 for real tokens and 0 for padding.
                Added for easy compatibility with the BERT model (which uses this negative masking).
                You can only uses one among `input_mask` and `attention_mask`
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
            mems: a list of float32 Tensors in shape [mem_len, bsz, d_model], memory
                from previous batches. The length of the list equals n_layer.
                If None, no memory is used.
            perm_mask: float32 Tensor in shape [bsz, len, len].
                If perm_mask[k, i, j] = 0, i attend to j in batch k;
                if perm_mask[k, i, j] = 1, i does not attend to j in batch k.
                If None, each position attends to all the others.
            target_mapping: float32 Tensor in shape [bsz, num_predict, len].
                If target_mapping[k, i, j] = 1, the i-th predict in batch k is
                on the j-th token.
                Only used during pretraining for partial prediction.
                Set to None during finetuning.
            inp_q: float32 Tensor in shape [bsz, len].
                1 for tokens with losses and 0 for tokens without losses.
                Only used during pretraining for two-stream attention.
                Set to None during finetuning.
        """
1378
        output, _, new_mems = self.transformer(inp_k, token_type_ids, input_mask, attention_mask,
thomwolf's avatar
thomwolf committed
1379
1380
                                            mems, perm_mask, target_mapping, inp_q,
                                            output_all_encoded_layers, head_mask)
thomwolf's avatar
thomwolf committed
1381

1382
        output = self.sequence_summary(output)
1383
        logits = self.logits_proj(output)
thomwolf's avatar
thomwolf committed
1384

thomwolf's avatar
thomwolf committed
1385
        if target is not None:
1386
1387
1388
1389
1390
1391
            if self.is_regression:
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), target.view(-1))
            else:
                loss_fct = CrossEntropyLoss(ignore_index=-1)
                loss = loss_fct(logits.view(-1, logits.size(-1)), target.view(-1))
thomwolf's avatar
thomwolf committed
1392
1393
            return loss, new_mems

thomwolf's avatar
thomwolf committed
1394
1395
1396
1397
1398
1399
1400
        # if self.output_attentions:
        #     all_attentions, encoded_layers = encoded_layers
        # sequence_output = encoded_layers[-1]
        # pooled_output = self.pooler(sequence_output)
        # if not output_all_encoded_layers:
        #     encoded_layers = encoded_layers[-1]
        # if self.output_attentions:
thomwolf's avatar
thomwolf committed
1401
        return logits, new_mems
thomwolf's avatar
thomwolf committed
1402
        #     return all_attentions, encoded_layers, pooled_output
thomwolf's avatar
thomwolf committed
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417

class XLNetForQuestionAnswering(XLNetPreTrainedModel):
    """XLNet model for Question Answering (span extraction).
    This module is composed of the XLNet model with a linear layer on top of
    the sequence output that computes start_logits and end_logits

    Params:
        `config`: a XLNetConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Inputs:
        `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length]
            with the word token indices in the vocabulary(see the tokens preprocessing logic in the scripts
1418
            `run_bert_extract_features.py`, `run_bert_classifier.py` and `run_bert_squad.py`)
thomwolf's avatar
thomwolf committed
1419
1420
1421
        `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token
            types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to
            a `sentence B` token (see XLNet paper for more details).
1422
1423
1424
1425
1426
        `attention_mask`: [optional] float32 Tensor, SAME FUNCTION as `input_mask`
            but with 1 for real tokens and 0 for padding.
            Added for easy compatibility with the BERT model (which uses this negative masking).
            You can only uses one among `input_mask` and `attention_mask`
        `input_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices
thomwolf's avatar
thomwolf committed
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
            selected in [0, 1]. It's a mask to be used if the input sequence length is smaller than the max
            input sequence length in the current batch. It's the mask that we typically use for attention when
            a batch has varying length sentences.
        `start_positions`: position of the first token for the labeled span: torch.LongTensor of shape [batch_size].
            Positions are clamped to the length of the sequence and position outside of the sequence are not taken
            into account for computing the loss.
        `end_positions`: position of the last token for the labeled span: torch.LongTensor of shape [batch_size].
            Positions are clamped to the length of the sequence and position outside of the sequence are not taken
            into account for computing the loss.
        `head_mask`: an optional torch.Tensor of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
            It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.

    Outputs:
        if `start_positions` and `end_positions` are not `None`:
            Outputs the total_loss which is the sum of the CrossEntropy loss for the start and end token positions.
        if `start_positions` or `end_positions` is `None`:
            Outputs a tuple of start_logits, end_logits which are the logits respectively for the start and end
            position tokens of shape [batch_size, sequence_length].

    Example usage:
    ```python
    # Already been converted into WordPiece token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
1450
    input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
thomwolf's avatar
thomwolf committed
1451
1452
1453
1454
1455
1456
    token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])

    config = XLNetConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)

    model = XLNetForQuestionAnswering(config)
1457
    start_logits, end_logits = model(input_ids, token_type_ids, input_mask)
thomwolf's avatar
thomwolf committed
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
    ```
    """
    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(XLNetForQuestionAnswering, self).__init__(config)
        self.output_attentions = output_attentions
        self.transformer = XLNetModel(config, output_attentions=output_attentions,
                                      keep_multihead_output=keep_multihead_output)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)
        self.apply(self.init_weights)

1468
    def forward(self, inp_k, token_type_ids=None, input_mask=None, attention_mask=None,
thomwolf's avatar
thomwolf committed
1469
1470
1471
                mems=None, perm_mask=None, target_mapping=None, inp_q=None,
                start_positions=None, end_positions=None,
                output_all_encoded_layers=True, head_mask=None):
1472
        output, _, new_mems = self.transformer(inp_k, token_type_ids, input_mask, attention_mask,
thomwolf's avatar
thomwolf committed
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
                                            mems, perm_mask, target_mapping, inp_q,
                                            output_all_encoded_layers, head_mask)

        logits = self.qa_outputs(output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions.clamp_(0, ignored_index)
            end_positions.clamp_(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2
            return total_loss
        elif self.output_attentions:
            return all_attentions, start_logits, end_logits
        return start_logits, end_logits