File size: 12,959 Bytes
eac156b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import types
import torch
import transformers
import torch.nn.functional as F
from torch import nn
from torch.nn import CrossEntropyLoss
import numpy as np

class FiDT5(transformers.T5ForConditionalGeneration):
    def __init__(self, config):
        super().__init__(config)
        self.wrap_encoder()

    def forward_(self, **kwargs):
        if 'input_ids' in kwargs:
            kwargs['input_ids'] = kwargs['input_ids'].view(kwargs['input_ids'].size(0), -1)
        if 'attention_mask' in kwargs:
            kwargs['attention_mask'] = kwargs['attention_mask'].view(kwargs['attention_mask'].size(0), -1)

        return super(FiDT5, self).forward(
            **kwargs
        )

    # We need to resize as B x (N * L) instead of (B * N) x L here
    # because the T5 forward method uses the input tensors to infer
    # dimensions used in the decoder.
    # EncoderWrapper resizes the inputs as (B * N) x L.
    def forward(self, input_ids=None, attention_mask=None, **kwargs):
        if input_ids != None:
            # inputs might have already be resized in the generate method
            if input_ids.dim() == 3:
                self.encoder.n_passages = input_ids.size(1)
            input_ids = input_ids.view(input_ids.size(0), -1)
        if attention_mask != None:
            attention_mask = attention_mask.view(attention_mask.size(0), -1)
        return super().forward(
            input_ids=input_ids,
            attention_mask=attention_mask,
            **kwargs
        )

    # We need to resize the inputs here, as the generate method expect 2D tensors
    def generate(self, input_ids, attention_mask, max_length):
        self.encoder.n_passages = input_ids.size(1)
        return super().generate(
            input_ids=input_ids.view(input_ids.size(0), -1),
            attention_mask=attention_mask.view(attention_mask.size(0), -1),
            max_length=max_length
        )

    def wrap_encoder(self, use_checkpoint=False):
        """
        Wrap T5 encoder to obtain a Fusion-in-Decoder model.
        """
        self.encoder = EncoderWrapper(self.encoder, use_checkpoint=use_checkpoint)

    def unwrap_encoder(self):
        """
        Unwrap Fusion-in-Decoder encoder, useful to load T5 weights.
        """
        self.encoder = self.encoder.encoder
        block = []
        for mod in self.encoder.block:
            block.append(mod.module)
        block = nn.ModuleList(block)
        self.encoder.block = block

    def load_t5(self, state_dict):
        self.unwrap_encoder()
        self.load_state_dict(state_dict)
        self.wrap_encoder()

    def set_checkpoint(self, use_checkpoint):
        """
        Enable or disable checkpointing in the encoder.
        See https://pytorch.org/docs/stable/checkpoint.html
        """
        for mod in self.encoder.encoder.block:
            mod.use_checkpoint = use_checkpoint

    def reset_score_storage(self):
        """
        Reset score storage, only used when cross-attention scores are saved
        to train a retriever.
        """
        for mod in self.decoder.block:
            mod.layer[1].EncDecAttention.score_storage = None

    def get_crossattention_scores(self, context_mask):
        """
        Cross-attention scores are aggregated to obtain a single scalar per
        passage. This scalar can be seen as a similarity score between the
        question and the input passage. It is obtained by averaging the
        cross-attention scores obtained on the first decoded token over heads,
        layers, and tokens of the input passage.
        More details in Distilling Knowledge from Reader to Retriever:
        https://arxiv.org/abs/2012.04584.
        """
        scores = []
        n_passages = context_mask.size(1)
        for mod in self.decoder.block:
            scores.append(mod.layer[1].EncDecAttention.score_storage)
        scores = torch.cat(scores, dim=2)
        bsz, n_heads, n_layers, _ = scores.size()
        # batch_size, n_head, n_layers, n_passages, text_maxlength
        scores = scores.view(bsz, n_heads, n_layers, n_passages, -1)
        scores = scores.masked_fill(~context_mask[:, None, None], 0.)
        scores = scores.sum(dim=[1, 2, 4])
        ntokens = context_mask.sum(dim=[2]) * n_layers * n_heads
        scores = scores/ntokens
        return scores

    def overwrite_forward_crossattention(self):
        """
        Replace cross-attention forward function, only used to save
        cross-attention scores.
        """
        for mod in self.decoder.block:
            attn = mod.layer[1].EncDecAttention
            attn.forward = types.MethodType(cross_attention_forward, attn)

class EncoderWrapper(torch.nn.Module):
    """
    Encoder Wrapper for T5 Wrapper to obtain a Fusion-in-Decoder model.
    """
    def __init__(self, encoder, use_checkpoint=False):
        super().__init__()

        self.encoder = encoder
        apply_checkpoint_wrapper(self.encoder, use_checkpoint)

    def forward(self, input_ids=None, attention_mask=None, **kwargs,):
        # total_length = n_passages * passage_length
        bsz, total_length = input_ids.shape
        passage_length = total_length // self.n_passages
        input_ids = input_ids.view(bsz*self.n_passages, passage_length)
        attention_mask = attention_mask.view(bsz*self.n_passages, passage_length)
        outputs = self.encoder(input_ids, attention_mask, **kwargs)
        outputs = (outputs[0].view(bsz, self.n_passages*passage_length, -1), ) + outputs[1:]
        return outputs

class CheckpointWrapper(torch.nn.Module):
    """
    Wrapper replacing None outputs by empty tensors, which allows the use of
    checkpointing.
    """
    def __init__(self, module, use_checkpoint=False):
        super().__init__()
        self.module = module
        self.use_checkpoint = use_checkpoint

    def forward(self, hidden_states, attention_mask, position_bias, **kwargs):
        if self.use_checkpoint and self.training:
            kwargs = {k: v for k, v in kwargs.items() if v is not None}
            def custom_forward(*inputs):
                output = self.module(*inputs, **kwargs)
                empty = torch.tensor(
                    [],
                    dtype=torch.float,
                    device=output[0].device,
                    requires_grad=True)
                output = tuple(x if x is not None else empty for x in output)
                return output

            output = torch.utils.checkpoint.checkpoint(
                custom_forward,
                hidden_states,
                attention_mask,
                position_bias
            )
            output = tuple(x if x.size() != 0 else None for x in output)
        else:
            output = self.module(hidden_states, attention_mask, position_bias, **kwargs)
        return output

def apply_checkpoint_wrapper(t5stack, use_checkpoint):
    """
    Wrap each block of the encoder to enable checkpointing.
    """
    block = []
    for mod in t5stack.block:
        wrapped_mod = CheckpointWrapper(mod, use_checkpoint)
        block.append(wrapped_mod)
    block = nn.ModuleList(block)
    t5stack.block = block

def cross_attention_forward(
        self,
        input,
        mask=None,
        kv=None,
        position_bias=None,
        past_key_value_state=None,
        head_mask=None,
        query_length=None,
        use_cache=False,
        output_attentions=False,
    ):
    """
    This only works for computing cross attention over the input
    """
    assert(kv != None)
    assert(head_mask == None)
    assert(position_bias != None or self.has_relative_attention_bias)

    bsz, qlen, dim = input.size()
    n_heads, d_heads = self.n_heads, self.d_kv
    klen = kv.size(1)

    q = self.q(input).view(bsz, -1, n_heads, d_heads).transpose(1, 2)
    if past_key_value_state == None:
        k = self.k(kv).view(bsz, -1, n_heads, d_heads).transpose(1, 2)
        v = self.v(kv).view(bsz, -1, n_heads, d_heads).transpose(1, 2)
    else:
        k, v = past_key_value_state

    scores = torch.einsum("bnqd,bnkd->bnqk", q, k)

    if mask is not None:
       scores += mask

    if position_bias is None:
        position_bias = self.compute_bias(qlen, klen)
    scores += position_bias

    if self.score_storage is None:
        self.score_storage = scores

    attn = F.softmax(scores.float(), dim=-1).type_as(scores)
    attn = F.dropout(attn, p=self.dropout, training=self.training)

    output = torch.matmul(attn, v)
    output = output.transpose(1, 2).contiguous().view(bsz, -1, self.inner_dim)
    output = self.o(output)

    if use_cache:
        output = (output,) + ((k, v),)
    else:
        output = (output,) + (None,)

    if output_attentions:
        output = output + (attn,)

    if self.has_relative_attention_bias:
        output = output + (position_bias,)

    return output

class RetrieverConfig(transformers.BertConfig):

    def __init__(self,
                 indexing_dimension=768,
                 apply_question_mask=False,
                 apply_passage_mask=False,
                 extract_cls=False,
                 passage_maxlength=200,
                 question_maxlength=40,
                 projection=True,
                 **kwargs):
        super().__init__(**kwargs)
        self.indexing_dimension = indexing_dimension
        self.apply_question_mask = apply_question_mask
        self.apply_passage_mask = apply_passage_mask
        self.extract_cls=extract_cls
        self.passage_maxlength = passage_maxlength
        self.question_maxlength = question_maxlength
        self.projection = projection

class Retriever(transformers.PreTrainedModel):

    config_class = RetrieverConfig
    base_model_prefix = "retriever"

    def __init__(self, config, initialize_wBERT=False):
        super().__init__(config)
        assert config.projection or config.indexing_dimension == 768, \
            'If no projection then indexing dimension must be equal to 768'
        self.config = config
        if initialize_wBERT:
            self.model = transformers.BertModel.from_pretrained('bert-base-uncased')
        else:
            self.model = transformers.BertModel(config)
        if self.config.projection:
            self.proj = nn.Linear(
                self.model.config.hidden_size,
                self.config.indexing_dimension
            )
            self.norm = nn.LayerNorm(self.config.indexing_dimension)
        self.loss_fct = torch.nn.KLDivLoss()

    def forward(self,
                question_ids,
                question_mask,
                passage_ids,
                passage_mask,
                gold_score=None):
        question_output = self.embed_text(
            text_ids=question_ids,
            text_mask=question_mask,
            apply_mask=self.config.apply_question_mask,
            extract_cls=self.config.extract_cls,
        )
        bsz, n_passages, plen = passage_ids.size()
        passage_ids = passage_ids.view(bsz * n_passages, plen)
        passage_mask = passage_mask.view(bsz * n_passages, plen)
        passage_output = self.embed_text(
            text_ids=passage_ids,
            text_mask=passage_mask,
            apply_mask=self.config.apply_passage_mask,
            extract_cls=self.config.extract_cls,
        )

        score = torch.einsum(
            'bd,bid->bi',
            question_output,
            passage_output.view(bsz, n_passages, -1)
        )
        score = score / np.sqrt(question_output.size(-1))
        if gold_score is not None:
            loss = self.kldivloss(score, gold_score)
        else:
            loss = None

        return question_output, passage_output, score, loss

    def embed_text(self, text_ids, text_mask, apply_mask=False, extract_cls=False):
        text_output = self.model(
            input_ids=text_ids,
            attention_mask=text_mask if apply_mask else None
        )
        if type(text_output) is not tuple:
            text_output.to_tuple()
        text_output = text_output[0]
        if self.config.projection:
            text_output = self.proj(text_output)
            text_output = self.norm(text_output)

        if extract_cls:
            text_output = text_output[:, 0]
        else:
            if apply_mask:
                text_output = text_output.masked_fill(~text_mask[:, :, None], 0.)
                text_output = torch.sum(text_output, dim=1) / torch.sum(text_mask, dim=1)[:, None]
            else:
                text_output = torch.mean(text_output, dim=1)
        return text_output

    def kldivloss(self, score, gold_score):
        gold_score = torch.softmax(gold_score, dim=-1)
        score = torch.nn.functional.log_softmax(score, dim=-1)
        return self.loss_fct(score, gold_score)