File size: 8,221 Bytes
2a21e9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import torch.nn as nn
import torch
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss, BCEWithLogitsLoss
import numpy as np
from utils import MyTokenizer
from transformers import (
    RobertaConfig,
    RobertaModel,
    RobertaTokenizer,
    BartConfig,
    BartForConditionalGeneration,
    BartTokenizer,
    T5Config,
    T5ForConditionalGeneration,
    T5Tokenizer,
)
import logging

logger = logging.getLogger(__name__)


class ReviewerModel(T5ForConditionalGeneration):

    def __init__(self, config):
        super().__init__(config)
        self.cls_head = nn.Linear(self.config.d_model, 2, bias=True)
        self.init()

    def init(self):
        nn.init.xavier_uniform_(self.lm_head.weight)
        factor = self.config.initializer_factor
        self.cls_head.weight.data.normal_(mean=0.0, \
            std=factor * ((self.config.d_model) ** -0.5))
        self.cls_head.bias.data.zero_()

    def forward(
        self, *argv, **kwargs
    ):
        r"""
        Doc from Huggingface transformers:
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[-100, 0, ...,
            config.vocab_size - 1]`. All labels set to ``-100`` are ignored (masked), the loss is only computed for
            labels in ``[0, ..., config.vocab_size]``
        Returns:
        Examples::
            >>> from transformers import T5Tokenizer, T5ForConditionalGeneration
            >>> tokenizer = T5Tokenizer.from_pretrained('t5-small')
            >>> model = T5ForConditionalGeneration.from_pretrained('t5-small')
            >>> # training
            >>> input_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids
            >>> labels = tokenizer('<extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids
            >>> outputs = model(input_ids=input_ids, labels=labels)
            >>> loss = outputs.loss
            >>> logits = outputs.logits
            >>> # inference
            >>> input_ids = tokenizer("summarize: studies have shown that owning a dog is good for you", return_tensors="pt").input_ids  # Batch size 1
            >>> outputs = model.generate(input_ids)
            >>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))
            >>> # studies have shown that owning a dog is good for you.
        """
        if "cls" in kwargs:
            assert (
                "input_ids" in kwargs and \
                "labels" in kwargs and \
                "attention_mask" in kwargs
            )
            return self.cls(
                input_ids=kwargs["input_ids"],
                labels=kwargs["labels"],
                attention_mask=kwargs["attention_mask"],
            )
        if "input_labels" in kwargs:
            assert (
                "input_ids" in kwargs and \
                "input_labels" in kwargs and \
                "decoder_input_ids" in kwargs and \
                "attention_mask" in kwargs and \
                "decoder_attention_mask" in kwargs
            ), "Please give these arg keys."
            input_ids = kwargs["input_ids"]
            input_labels = kwargs["input_labels"]
            decoder_input_ids = kwargs["decoder_input_ids"]
            attention_mask = kwargs["attention_mask"]
            decoder_attention_mask = kwargs["decoder_attention_mask"]
            if "encoder_loss" not in kwargs:
                encoder_loss = True
            else:
                encoder_loss = kwargs["encoder_loss"]
            return self.review_forward(input_ids, input_labels, decoder_input_ids, attention_mask, decoder_attention_mask, encoder_loss)
        return super().forward(*argv, **kwargs)

    def cls(
        self,
        input_ids,
        labels,
        attention_mask,
    ):
        encoder_outputs = self.encoder( \
            input_ids=input_ids,
            attention_mask=attention_mask,
            output_attentions=False,
            return_dict=False
        )
        hidden_states = encoder_outputs[0]
        first_hidden = hidden_states[:, 0, :]
        first_hidden = nn.Dropout(0.3)(first_hidden)
        logits = self.cls_head(first_hidden)
        loss_fct = CrossEntropyLoss()
        if labels != None:
            loss = loss_fct(logits, labels)
            return loss
        return logits

    def review_forward(
        self,
        input_ids,
        input_labels,
        decoder_input_ids,
        attention_mask,
        decoder_attention_mask,
        encoder_loss=True
    ):
        encoder_outputs = self.encoder( \
            input_ids=input_ids,
            attention_mask=attention_mask,
            output_attentions=False,
            return_dict=False
        )
        hidden_states = encoder_outputs[0]
        decoder_inputs = self._shift_right(decoder_input_ids)
        # Decode
        decoder_outputs = self.decoder(
            input_ids=decoder_inputs,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=hidden_states,
            encoder_attention_mask=attention_mask,
            output_attentions=False,
            return_dict=False
        )
        sequence_output = decoder_outputs[0]
        if self.config.tie_word_embeddings: # this is True default
            sequence_output = sequence_output * (self.model_dim ** -0.5)
        if encoder_loss:
            # print(self.encoder.get_input_embeddings().weight.shape)
            cls_logits = nn.functional.linear(hidden_states, self.encoder.get_input_embeddings().weight)
            # cls_logits = self.cls_head(hidden_states)
        lm_logits = self.lm_head(sequence_output)
        if decoder_input_ids is not None:
            lm_loss_fct = CrossEntropyLoss(ignore_index=0)      # Warning: PAD_ID should be 0
            loss = lm_loss_fct(lm_logits.view(-1, lm_logits.size(-1)), decoder_input_ids.view(-1))
            if encoder_loss and input_labels is not None:
                cls_loss_fct = CrossEntropyLoss(ignore_index=-100)
                loss += cls_loss_fct(cls_logits.view(-1, cls_logits.size(-1)), input_labels.view(-1))
            return loss
        return cls_logits, lm_logits

def get_model_size(model):
    model_parameters = filter(lambda p: p.requires_grad, model.parameters())
    model_size = sum([np.prod(p.size()) for p in model_parameters])
    return "{}M".format(round(model_size / 1e6))


def build_or_load_gen_model(args):
    config_class, model_class, tokenizer_class = T5Config, ReviewerModel, RobertaTokenizer
    
    config = config_class.from_pretrained(args.model_name_or_path)
    tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
    model = model_class.from_pretrained(args.model_name_or_path, config=config)

    tokenizer.special_dict = {
        f"<e{i}>" : tokenizer.get_vocab()[f"<e{i}>"] for i in range(99, -1, -1)
    }

    tokenizer.mask_id = tokenizer.get_vocab()["<mask>"]
    tokenizer.bos_id = tokenizer.get_vocab()["<s>"]
    tokenizer.pad_id = tokenizer.get_vocab()["<pad>"]
    tokenizer.eos_id = tokenizer.get_vocab()["</s>"]
    tokenizer.msg_id = tokenizer.get_vocab()["<msg>"]
    tokenizer.keep_id = tokenizer.get_vocab()["<keep>"]
    tokenizer.add_id = tokenizer.get_vocab()["<add>"]
    tokenizer.del_id = tokenizer.get_vocab()["<del>"]
    tokenizer.start_id = tokenizer.get_vocab()["<start>"]
    tokenizer.end_id = tokenizer.get_vocab()["<end>"]

    logger.info(
        "Finish loading model [%s] from %s",
        get_model_size(model),
        args.model_name_or_path,
    )

    if args.load_model_path is not None:
        model_path = os.path.join(args.load_model_path, "pytorch_model.bin")
        logger.info("Reload model from {}".format(model_path))
        try:
            model.load_state_dict(torch.load(model_path, map_location="cpu"))
        except RuntimeError:
            saved = model.cls_head
            model.cls_head = None
            model.load_state_dict(torch.load(model_path, map_location="cpu"))
            model.cls_head = saved
        model.to(args.local_rank)

    return config, model, tokenizer