File size: 10,201 Bytes
629db7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import sys
sys.path.append('..')
from os.path import join
import torch
import numpy as np
import scipy.sparse as sp
from torch.utils.data import Dataset
from torch.nn.utils.rnn import pad_sequence
from transformers import BertTokenizer
from config import *
from utils.utils import *


torch.manual_seed(TORCH_SEED)
torch.cuda.manual_seed_all(TORCH_SEED)
torch.backends.cudnn.deterministic = True


def build_train_data(configs, fold_id, shuffle=True):
    train_dataset = MyDataset(configs, fold_id, data_type='train')
    train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=configs.batch_size,
                                               shuffle=shuffle, collate_fn=bert_batch_preprocessing)
    return train_loader


def build_inference_data(configs, fold_id, data_type):
    dataset = MyDataset(configs, fold_id, data_type)
    data_loader = torch.utils.data.DataLoader(dataset=dataset, batch_size=configs.batch_size,
                                              shuffle=False, collate_fn=bert_batch_preprocessing)
    return data_loader


class MyDataset(Dataset):
    def __init__(self, configs, fold_id, data_type, data_dir=DATA_DIR):
        self.data_dir = data_dir
        self.split = configs.split

        self.data_type = data_type
        self.train_file = join(data_dir, self.split, TRAIN_FILE % fold_id)
        self.valid_file = join(data_dir, self.split, VALID_FILE % fold_id)
        self.test_file = join(data_dir, self.split, TEST_FILE % fold_id)

        self.batch_size = configs.batch_size
        self.epochs = configs.epochs

        self.bert_tokenizer = BertTokenizer.from_pretrained(configs.bert_cache_path)

        self.doc_couples_list, self.y_emotions_list, self.y_causes_list, \
        self.doc_len_list, self.doc_id_list, \
        self.bert_token_idx_list, self.bert_clause_idx_list, self.bert_segments_idx_list, \
        self.bert_token_lens_list = self.read_data_file(self.data_type)

    def __len__(self):
        return len(self.y_emotions_list)

    def __getitem__(self, idx):
        doc_couples, y_emotions, y_causes = self.doc_couples_list[idx], self.y_emotions_list[idx], self.y_causes_list[idx]
        doc_len, doc_id = self.doc_len_list[idx], self.doc_id_list[idx]
        bert_token_idx, bert_clause_idx = self.bert_token_idx_list[idx], self.bert_clause_idx_list[idx]
        bert_segments_idx, bert_token_lens = self.bert_segments_idx_list[idx], self.bert_token_lens_list[idx]

        if bert_token_lens > 512:
            bert_token_idx, bert_clause_idx, \
            bert_segments_idx, bert_token_lens, \
            doc_couples, y_emotions, y_causes, doc_len = self.token_trunk(bert_token_idx, bert_clause_idx,
                                                                          bert_segments_idx, bert_token_lens,
                                                                          doc_couples, y_emotions, y_causes, doc_len)

        bert_token_idx = torch.LongTensor(bert_token_idx)
        bert_segments_idx = torch.LongTensor(bert_segments_idx)
        bert_clause_idx = torch.LongTensor(bert_clause_idx)

        assert doc_len == len(y_emotions)
        return doc_couples, y_emotions, y_causes, doc_len, doc_id, \
               bert_token_idx, bert_segments_idx, bert_clause_idx, bert_token_lens

    def read_data_file(self, data_type):
        if data_type == 'train':
            data_file = self.train_file
        elif data_type == 'valid':
            data_file = self.valid_file
        elif data_type == 'test':
            data_file = self.test_file

        doc_id_list = []
        doc_len_list = []
        doc_couples_list = []
        y_emotions_list, y_causes_list = [], []
        bert_token_idx_list = []
        bert_clause_idx_list = []
        bert_segments_idx_list = []
        bert_token_lens_list = []

        data_list = read_json(data_file)
        for doc in data_list:
            doc_id = doc['doc_id']
            doc_len = doc['doc_len']
            doc_couples = doc['pairs']
            doc_emotions, doc_causes = zip(*doc_couples)
            doc_id_list.append(doc_id)
            doc_len_list.append(doc_len)
            doc_couples = list(map(lambda x: list(x), doc_couples))
            doc_couples_list.append(doc_couples)

            y_emotions, y_causes = [], []
            doc_clauses = doc['clauses']
            doc_str = ''
            for i in range(doc_len):
                emotion_label = int(i + 1 in doc_emotions)
                cause_label = int(i + 1 in doc_causes)
                y_emotions.append(emotion_label)
                y_causes.append(cause_label)

                clause = doc_clauses[i]
                clause_id = clause['clause_id']
                assert int(clause_id) == i + 1
                doc_str += '[CLS] ' + clause['clause'] + ' [SEP] '

            indexed_tokens = self.bert_tokenizer.encode(doc_str.strip(), add_special_tokens=False)

            clause_indices = [i for i, x in enumerate(indexed_tokens) if x == 101]
            doc_token_len = len(indexed_tokens)

            segments_ids = []
            segments_indices = [i for i, x in enumerate(indexed_tokens) if x == 101]
            segments_indices.append(len(indexed_tokens))
            for i in range(len(segments_indices)-1):
                semgent_len = segments_indices[i+1] - segments_indices[i]
                if i % 2 == 0:
                    segments_ids.extend([0] * semgent_len)
                else:
                    segments_ids.extend([1] * semgent_len)

            assert len(clause_indices) == doc_len
            assert len(segments_ids) == len(indexed_tokens)
            bert_token_idx_list.append(indexed_tokens)
            bert_clause_idx_list.append(clause_indices)
            bert_segments_idx_list.append(segments_ids)
            bert_token_lens_list.append(doc_token_len)

            y_emotions_list.append(y_emotions)
            y_causes_list.append(y_causes)

        return doc_couples_list, y_emotions_list, y_causes_list, doc_len_list, doc_id_list, \
               bert_token_idx_list, bert_clause_idx_list, bert_segments_idx_list, bert_token_lens_list

    def token_trunk(self, bert_token_idx, bert_clause_idx, bert_segments_idx, bert_token_lens,
                    doc_couples, y_emotions, y_causes, doc_len):
        # TODO: cannot handle some extreme cases now
        emotion, cause = doc_couples[0]
        if emotion > doc_len / 2 and cause > doc_len / 2:
            i = 0
            while True:
                temp_bert_token_idx = bert_token_idx[bert_clause_idx[i]:]
                if len(temp_bert_token_idx) <= 512:
                    cls_idx = bert_clause_idx[i]
                    bert_token_idx = bert_token_idx[cls_idx:]
                    bert_segments_idx = bert_segments_idx[cls_idx:]
                    bert_clause_idx = [p - cls_idx for p in bert_clause_idx[i:]]
                    doc_couples = [[emotion - i, cause - i]]
                    y_emotions = y_emotions[i:]
                    y_causes = y_causes[i:]
                    doc_len = doc_len - i
                    break
                i = i + 1
        if emotion < doc_len / 2 and cause < doc_len / 2:
            i = doc_len - 1
            while True:
                temp_bert_token_idx = bert_token_idx[:bert_clause_idx[i]]
                if len(temp_bert_token_idx) <= 512:
                    cls_idx = bert_clause_idx[i]
                    bert_token_idx = bert_token_idx[:cls_idx]
                    bert_segments_idx = bert_segments_idx[:cls_idx]
                    bert_clause_idx = bert_clause_idx[:i]
                    y_emotions = y_emotions[:i]
                    y_causes = y_causes[:i]
                    doc_len = i
                    break
                i = i - 1
        return bert_token_idx, bert_clause_idx, bert_segments_idx, bert_token_lens, \
               doc_couples, y_emotions, y_causes, doc_len


def bert_batch_preprocessing(batch):
    doc_couples_b, y_emotions_b, y_causes_b, doc_len_b, doc_id_b, \
    bert_token_b, bert_segment_b, bert_clause_b, bert_token_lens_b = zip(*batch)

    y_mask_b, y_emotions_b, y_causes_b = pad_docs(doc_len_b, y_emotions_b, y_causes_b)
    adj_b = pad_matrices(doc_len_b)
    bert_token_b = pad_sequence(bert_token_b, batch_first=True, padding_value=0)
    bert_segment_b = pad_sequence(bert_segment_b, batch_first=True, padding_value=0)
    bert_clause_b = pad_sequence(bert_clause_b, batch_first=True, padding_value=0)

    bsz, max_len = bert_token_b.size()
    bert_masks_b = np.zeros([bsz, max_len], dtype=np.float)
    for index, seq_len in enumerate(bert_token_lens_b):
        bert_masks_b[index][:seq_len] = 1

    bert_masks_b = torch.FloatTensor(bert_masks_b)
    assert bert_segment_b.shape == bert_token_b.shape
    assert bert_segment_b.shape == bert_masks_b.shape

    return np.array(doc_len_b), np.array(adj_b), \
           np.array(y_emotions_b), np.array(y_causes_b), np.array(y_mask_b), doc_couples_b, doc_id_b, \
           bert_token_b, bert_segment_b, bert_masks_b, bert_clause_b


def pad_docs(doc_len_b, y_emotions_b, y_causes_b):
    max_doc_len = max(doc_len_b)

    y_mask_b, y_emotions_b_, y_causes_b_ = [], [], []
    for y_emotions, y_causes in zip(y_emotions_b, y_causes_b):
        y_emotions_ = pad_list(y_emotions, max_doc_len, -1)
        y_causes_ = pad_list(y_causes, max_doc_len, -1)
        y_mask = list(map(lambda x: 0 if x == -1 else 1, y_emotions_))

        y_mask_b.append(y_mask)
        y_emotions_b_.append(y_emotions_)
        y_causes_b_.append(y_causes_)

    return y_mask_b, y_emotions_b_, y_causes_b_


def pad_matrices(doc_len_b):
    N = max(doc_len_b)
    adj_b = []
    for doc_len in doc_len_b:
        adj = np.ones((doc_len, doc_len))
        adj = sp.coo_matrix(adj)
        adj = sp.coo_matrix((adj.data, (adj.row, adj.col)),
                            shape=(N, N), dtype=np.float32)
        adj_b.append(adj.toarray())
    return adj_b


def pad_list(element_list, max_len, pad_mark):
    element_list_pad = element_list[:]
    pad_mark_list = [pad_mark] * (max_len - len(element_list))
    element_list_pad.extend(pad_mark_list)
    return element_list_pad