correct/model.py at main · cezhang01/correct · GitHub

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
import torch
import torch.nn as nn
import numpy as np
import sklearn
import transformers
from transformers import BertConfig, BertModel
from encoder import Encoder
from decoder import Classifier


class Model(nn.Module):

    def __init__(self, args, data):

        super(Model, self).__init__()
        self.data = data
        self.parse_args(args)
        if args.local_rank in [-1, 0]:
            self.show_config()
        self.generate_modules(args, data)

    def parse_args(self, args):

        self.dataset_name = args.dataset_name
        self.current_device = args.device
        self.mode = args.mode
        self.num_shots = args.num_shots
        self.ddp = args.distributed_training
        if self.ddp:
            self.world_size = args.world_size
        self.evidence_setting = args.evidence_setting
        self.num_prompt_embs = args.num_prompt_embs
        self.random_prompt_init = args.random_prompt_init
        self.language_model = args.language_model
        self.num_claims = self.data.num_claims
        self.num_training_claims = len(self.data.training_claim_ids)
        self.num_evidences = self.data.num_evidences
        self.has_contexts = self.data.has_contexts
        args.has_contexts = self.data.has_contexts
        if self.has_contexts:
            self.num_contexts = self.data.num_contexts
        self.has_references = self.data.has_references
        args.has_references = self.data.has_references
        if self.has_references:
            self.num_references = self.data.num_references
        self.num_labels = self.data.num_labels
        args.num_labels = self.data.num_labels
        self.max_text_length = args.max_text_length
        self.num_sampled_evidence = args.num_sampled_evidence
        self.num_sampled_references = args.num_sampled_references
        self.num_epochs = args.num_epochs
        self.learning_rate_for_lm = args.learning_rate_for_lm
        self.learning_rate_for_prompt_embs = args.learning_rate_for_prompt_embs
        self.minibatch_size = args.minibatch_size

    def show_config(self):

        print('******************************************************')
        print('dataset name:', self.dataset_name)
        print('torch version:', torch.__version__)
        print('np version:', np.__version__)
        print('sklearn version:', sklearn.__version__)
        print('transformers version:', transformers.__version__)
        print('device:', self.current_device)
        print('distributed training:', self.ddp)
        if self.ddp:
            print('world size:', self.world_size)
        print('mode:', self.mode)
        if self.mode == 'few_shot':
            print('#shots:', self.num_shots)
        print('evidence setting:', self.evidence_setting)
        print('#prompt embeddings:', self.num_prompt_embs)
        print('random prompt init:', self.random_prompt_init)
        print('language model:', self.language_model)
        print('max text length:', self.max_text_length)
        print('#claims:', self.num_claims)
        print('#training claims:', self.num_training_claims)
        print('#evidence:', self.num_evidences)
        if self.has_contexts:
            print('#contextual documents:', self.num_contexts)
        if self.has_references:
            print('#referential documents:', self.num_references)
        print('#labels:', self.num_labels)
        print('#sampled evidence:', self.num_sampled_evidence)
        print('#sampled references:', self.num_sampled_references)
        print('#epochs:', self.num_epochs)
        print('learning rate for lm:', self.learning_rate_for_lm)
        print('learning rate for prompt embeddings:', self.learning_rate_for_prompt_embs)
        print('minibatch size:', self.minibatch_size)
        print('******************************************************')

    def generate_modules(self, args, data):

        config = BertConfig.from_pretrained(args.pretrained_model_name_or_path)
        self.lm = BertModel.from_pretrained(args.pretrained_model_name_or_path, config=config)
        args.num_hidden_layers = config.num_hidden_layers
        args.hidden_size = config.hidden_size

        self.encoder = Encoder(self.lm, config, args, data)
        self.classifier = Classifier(config, args)

    def preprocess_data(self, claim_ids, data):

        claim_ids = claim_ids.detach().cpu().numpy()
        evid_ids = np.array([data.sampled_evid_ids[claim_id] for claim_id in claim_ids])
        evid_ids = np.reshape(evid_ids, [-1])

        claim_input_ids = np.array([data.claim_input_ids[claim_id] for claim_id in claim_ids])
        claim_attention_mask = np.array([data.claim_attention_mask[claim_id] for claim_id in claim_ids])
        claim_input_ids = np.reshape(claim_input_ids, [-1, self.max_text_length])
        claim_input_ids = torch.LongTensor(claim_input_ids).to(self.current_device)
        claim_attention_mask = np.reshape(claim_attention_mask, [-1, self.max_text_length])
        claim_attention_mask = torch.LongTensor(claim_attention_mask).to(self.current_device)

        evid_input_ids = np.array([data.evid_input_ids[evid_id] for evid_id in evid_ids])
        evid_attention_mask = np.array([data.evid_attention_mask[evid_id] for evid_id in evid_ids])
        evid_input_ids = np.reshape(evid_input_ids, [-1, self.max_text_length])
        evid_input_ids = torch.LongTensor(evid_input_ids).to(self.current_device)
        evid_attention_mask = np.reshape(evid_attention_mask, [-1, self.max_text_length])
        evid_attention_mask = torch.LongTensor(evid_attention_mask).to(self.current_device)

        ctx_input_ids, ctx_attention_mask = None, None
        if data.has_contexts:
            ctx_ids = [data.evidences[evid_id]['ctx_id'][0] for evid_id in evid_ids]
            ctx_input_ids = np.array([data.ctx_input_ids[ctx_id] for ctx_id in ctx_ids])
            ctx_attention_mask = np.array([data.ctx_attention_mask[ctx_id] for ctx_id in ctx_ids])
            ctx_input_ids = np.reshape(ctx_input_ids, [-1, self.max_text_length])
            ctx_input_ids = torch.LongTensor(ctx_input_ids).to(self.current_device)
            ctx_attention_mask = np.reshape(ctx_attention_mask, [-1, self.max_text_length])
            ctx_attention_mask = torch.LongTensor(ctx_attention_mask).to(self.current_device)

        ref_input_ids, ref_attention_mask = None, None
        if data.has_references:
            ref_input_ids, ref_attention_mask = [], []
            for evid_id in evid_ids:
                ref_ids_one_evid = data.sampled_ref_ids[evid_id]
                if ref_ids_one_evid[0] == -1:
                    ref_input_ids_one_evid = np.tile(np.reshape(np.array(data.evid_input_ids[evid_id]), [1, -1]), [self.num_sampled_references, 1])
                    ref_attention_mask_one_evid = np.tile(np.reshape(np.array(data.evid_attention_mask[evid_id]), [1, -1]), [self.num_sampled_references, 1])
                else:
                    ref_input_ids_one_evid = np.array([data.ref_input_ids[ref_id] for ref_id in ref_ids_one_evid])
                    ref_attention_mask_one_evid = np.array([data.ref_attention_mask[ref_id] for ref_id in ref_ids_one_evid])
                ref_input_ids.append(ref_input_ids_one_evid)
                ref_attention_mask.append(ref_attention_mask_one_evid)
            ref_input_ids = np.concatenate(ref_input_ids, axis=0)
            ref_attention_mask = np.concatenate(ref_attention_mask, axis=0)
            ref_input_ids = np.reshape(ref_input_ids, [-1, self.max_text_length])
            ref_input_ids = torch.LongTensor(ref_input_ids).to(self.current_device)
            ref_attention_mask = np.reshape(ref_attention_mask, [-1, self.max_text_length])
            ref_attention_mask = torch.LongTensor(ref_attention_mask).to(self.current_device)

        return (claim_input_ids,
                claim_attention_mask,
                evid_input_ids,
                evid_attention_mask,
                ctx_input_ids,
                ctx_attention_mask,
                ref_input_ids,
                ref_attention_mask)

    def forward(self, claim_ids, labels, data, mode):

        # preprocess data
        (claim_input_ids,
         claim_attention_mask,
         evid_input_ids,
         evid_attention_mask,
         ctx_input_ids,
         ctx_attention_mask,
         ref_input_ids,
         ref_attention_mask) = self.preprocess_data(claim_ids, data)

        # evidence encoder
        evid_emb = self.encoder(lm=self.lm,
                                input_ids=evid_input_ids,
                                attention_mask=evid_attention_mask,
                                ctx_input_ids=ctx_input_ids,
                                ctx_attention_mask=ctx_attention_mask,
                                ref_input_ids=ref_input_ids,
                                ref_attention_mask=ref_attention_mask,
                                claim_or_evid='evid',
                                mode=mode)

        # claim encoder
        claim_emb_list = self.encoder(lm=self.lm,
                                      input_ids=claim_input_ids,
                                      attention_mask=claim_attention_mask,
                                      evid_emb=evid_emb,
                                      claim_or_evid='claim',
                                      mode=mode)

        # classifier
        loss, y_pred = self.classifier(claim_emb_list, evid_emb, labels)

        return [loss, y_pred]