/
dataset2.py
250 lines (227 loc) · 11.7 KB
/
dataset2.py
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
import torch
import pickle
from dataset import BaseDataset
from torch.utils.data import Dataset, DataLoader
from torch.nn.utils.rnn import pad_sequence
import os
import numpy as np
class SkaigDataset(Dataset):
def __init__(self, model_size='base', phase='train', window=10):
super(SkaigDataset, self).__init__()
self.base_dataset = BaseDataset(model_size, phase)
self.window = window * 2
knowledge_path = 'skaig_data/dailydialog_' + phase + '_know_processed.pkl'
data = pickle.load(open(knowledge_path, 'rb'), encoding='latin1')
knowledge, knowledge_adj = data[0], data[1]
self.knowledge = knowledge
self.knowledge_adj = knowledge_adj
def __getitem__(self, item):
utter, token_id, att_mask, label, adj_idx, evid, msk, clen, act = self.base_dataset[item]
edge_mask = msk.clone()
knowledge = torch.tensor(self.knowledge[item], dtype=torch.float)
knowledge_adj = torch.tensor(self.knowledge_adj[item], dtype=torch.long)
knowledge_adj = knowledge_adj.triu(-self.window)
edge_mask = edge_mask.triu(-self.window)
adj_idx = adj_idx.triu(-self.window)
return utter, token_id, att_mask, label, adj_idx, evid, msk, \
clen, act, knowledge, knowledge_adj, edge_mask
def __len__(self):
return len(self.base_dataset)
def collate_fn_know(data):
token_ids_1 = []
attention_mask_1 = []
label = []
adj_index = []
ece_pair = []
mask = []
edge_mask = []
clen = []
act_label = []
knowledge = []
knowledge_adj = []
for i, d in enumerate(data):
token_ids_1.append(d[1])
attention_mask_1.append(d[2])
label.append(d[3])
# [2, conv_len, conv_len]
adj_index.append(d[4])
ece_pair.append(d[5])
mask.append(d[6])
clen.append(d[7])
act_label.append(d[8])
knowledge.append(d[9])
knowledge_adj.append(d[10])
edge_mask.append(d[11])
label = pad_sequence(label, batch_first=True, padding_value=-1)
act_label = pad_sequence(act_label, batch_first=True, padding_value=-1)
max_len = max(clen)
mask = [torch.cat([torch.cat([m, torch.zeros(max_len - m.shape[0], m.shape[1])], dim=0), torch.zeros(max_len, max_len - m.shape[1])], dim=1) for m in mask]
mask = torch.stack(mask, dim=0)
edge_mask = [torch.cat([torch.cat([m, torch.zeros(max_len - m.shape[0], m.shape[1])], dim=0), torch.zeros(max_len, max_len - m.shape[1])], dim=1) for m in edge_mask]
edge_mask = torch.stack(edge_mask, dim=0)
ece_pair = [torch.cat([torch.cat([ep, torch.zeros(max_len - ep.shape[0], ep.shape[1])], dim=0), torch.zeros(max_len, max_len - ep.shape[1])], dim=1) for ep in ece_pair]
ece_pair = torch.stack(ece_pair, dim=0)
adj_index = [torch.cat([torch.cat([a, torch.zeros(2, max_len - a.shape[1], a.shape[2])], dim=1), torch.zeros(2, max_len, max_len - a.shape[2])], dim=2) for a in adj_index]
# [batch_size, 2, conv_len, conv_len]
adj_index = torch.stack(adj_index, dim=0)
num_knowledge = [k.shape[0] for k in knowledge]
knowledge = torch.cat(knowledge, dim=0)
accu_know = 0
for i in range(len(num_knowledge)):
knowledge_adj[i] = knowledge_adj[i] + accu_know
accu_know += num_knowledge[i]
knowledge_adj[i] = torch.cat([torch.cat([knowledge_adj[i], torch.zeros(max_len - knowledge_adj[i].shape[0], knowledge_adj[i].shape[1])], dim=0), torch.zeros(max_len, max_len-knowledge_adj[i].shape[1])], dim=1)
knowledge_adj = torch.stack(knowledge_adj, dim=0)
return token_ids_1, attention_mask_1, clen, mask, edge_mask, adj_index, \
label, ece_pair, act_label, knowledge, knowledge_adj, num_knowledge
def get_dataloaders(model_size, batch_size, valid_shuffle, window=10):
train_set = SkaigDataset(model_size, 'train', window)
dev_set = SkaigDataset(model_size, 'dev', window)
test_set = SkaigDataset(model_size, 'test', window)
train_loader = DataLoader(train_set, batch_size, True, collate_fn=collate_fn_know)
dev_loader = DataLoader(dev_set, batch_size, valid_shuffle, collate_fn=collate_fn_know)
test_loader = DataLoader(test_set, batch_size, valid_shuffle, collate_fn=collate_fn_know)
return train_loader, dev_loader, test_loader
class KagDataset(Dataset):
def __init__(self, model_size='base', phase='train', unidirection=False):
super(KagDataset, self).__init__()
self.base_dataset = BaseDataset(model_size, phase)
self.unidirection = unidirection
know_adj_path = 'kag_data/' + phase + '_conceptnet_adj_1.pkl'
know_adj_processed_path = 'kag_data/' + phase + '_conceptnet_adj_processed.pkl'
if os.path.exists(know_adj_processed_path):
know_data = pickle.load(open(know_adj_processed_path, 'rb'), encoding='latin1')
self.knowledge_text = know_data[0]
self.knowledge_attention_mask = know_data[1]
self.knowledge_adj = know_data[2]
self.knowledge_len = know_data[3]
self.sequence_adj = know_data[4]
else:
self.knowledge_text = []
self.knowledge_attention_mask = []
self.knowledge_adj = []
self.knowledge_len = []
self.sequence_adj = []
know_data = pickle.load(open(know_adj_path, 'rb'), encoding='latin1')
for conv_know, label in zip(know_data, self.base_dataset.emotion):
num_utter = len(label)
conv_adj = np.zeros((num_utter, num_utter), dtype=np.int)
know_len = []
kw_text = []
kw_att_mask = []
index_count = 1
for k, v in conv_know.items():
source_node = int(k.split('-')[0])
target_node = int(k.split('-')[1])
conv_adj[target_node, source_node] = index_count
index_count += 1
know_len.append(len(v))
for kw in v:
processed_kw = self.base_dataset.tokenizer(kw)
kw_text.append(torch.tensor(processed_kw.input_ids, dtype=torch.long))
kw_att_mask.append(torch.tensor(processed_kw.attention_mask))
kw_text = pad_sequence(kw_text, batch_first=True, padding_value=1)
kw_att_mask = pad_sequence(kw_att_mask, batch_first=True, padding_value=0)
conv_adj = torch.tensor(conv_adj, dtype=torch.long)
seq_adj = torch.zeros_like(conv_adj, dtype=torch.long)
for i in range(num_utter):
if i == 0:
seq_adj[i+1, i] = 1
elif 0 < i < num_utter - 1:
seq_adj[i+1, i] = 1
seq_adj[i-1, i] = 1
else:
seq_adj[i-1, i] = 1
self.knowledge_text.append(kw_text)
self.knowledge_attention_mask.append(kw_att_mask)
self.knowledge_adj.append(conv_adj)
self.knowledge_len.append(know_len)
self.sequence_adj.append(seq_adj)
pickle.dump([self.knowledge_text, self.knowledge_attention_mask, self.knowledge_adj,
self.knowledge_len, self.sequence_adj], open(know_adj_processed_path, 'wb'))
def __getitem__(self, item):
utter, token_id, att_mask, label, adj_idx, evid, msk, clen, act = self.base_dataset[item]
knowledge = self.knowledge_text[item]
know_attn_mask = self.knowledge_attention_mask[item]
know_adj = self.knowledge_adj[item]
seq_adj = self.sequence_adj[item]
if self.unidirection:
seq_adj = torch.tril(seq_adj, 0)
know_len = self.knowledge_len[item]
return token_id, att_mask, label, evid, msk, clen, \
knowledge, know_attn_mask, know_adj, seq_adj, know_len
def __len__(self):
return len(self.base_dataset.emotion)
def collate_fn_kag(data):
token_ids = []
attention_mask = []
label = []
ece_pair = []
mask = []
clen = []
knowledge = []
knowledge_attention_mask = []
knowledge_adj = []
sequence_adj = []
knowledge_len = []
batch_knowledge_len = []
for d in data:
tk_ids = d[0]
at_msk = d[1]
lbl = d[2]
ece_p = d[3]
msk = d[4]
cl = d[5]
know = d[6]
know_att_msk = d[7]
know_adj = d[8]
seq_adj = d[9]
know_len = d[10]
token_ids.append(tk_ids)
attention_mask.append(at_msk)
label.append(lbl)
ece_pair.append(ece_p)
mask.append(msk)
clen.append(cl)
knowledge.append(know)
knowledge_adj.append(know_adj)
knowledge_attention_mask.append(know_att_msk)
sequence_adj.append(seq_adj)
knowledge_len.append(know_len)
batch_knowledge_len.append(sum(know_len))
label = pad_sequence(label, batch_first=True, padding_value=-1)
max_len = max(clen)
mask = [torch.cat([torch.cat([m, torch.zeros(max_len - m.shape[0], m.shape[1])], dim=0), torch.zeros(max_len, max_len - m.shape[1])], dim=1) for m in mask]
mask = torch.stack(mask, dim=0)
ece_pair = [torch.cat([torch.cat([ep, torch.zeros(max_len - ep.shape[0], ep.shape[1])], dim=0), torch.zeros(max_len, max_len - ep.shape[1])], dim=1) for ep in ece_pair]
ece_pair = torch.stack(ece_pair, dim=0)
sequence_adj = [torch.cat([torch.cat([sa, torch.zeros(max_len - sa.shape[0], sa.shape[1])], dim=0), torch.zeros(max_len, max_len - sa.shape[1])], dim=1) for sa in sequence_adj]
sequence_adj = torch.stack(sequence_adj, dim=0)
max_knowledge_len = max([k.shape[1] if k is not None else 0 for k in knowledge])
num_knowledge = [len(k) for k in knowledge_len]
knowledge_ = []
knowledge_attention_mask_ = []
for k, km in zip(knowledge, knowledge_attention_mask):
knowledge_.append(torch.cat([k, torch.ones(k.shape[0], max_knowledge_len - k.shape[1])], dim=1))
knowledge_attention_mask_.append(torch.cat([km, torch.zeros(km.shape[0], max_knowledge_len - km.shape[1])], dim=1))
knowledge = torch.cat(knowledge_, dim=0)
knowledge_attention_mask = torch.cat(knowledge_attention_mask_, dim=0)
accu_know = 0
for i in range(len(num_knowledge)):
mmask = knowledge_adj[i].eq(0)
knowledge_adj[i] = knowledge_adj[i] + accu_know
knowledge_adj[i] = torch.masked_fill(knowledge_adj[i], mmask, 0)
accu_know += num_knowledge[i]
knowledge_adj[i] = torch.cat([torch.cat([knowledge_adj[i], torch.zeros(max_len - knowledge_adj[i].shape[0], knowledge_adj[i].shape[1])], dim=0), torch.zeros(max_len, max_len - knowledge_adj[i].shape[1])], dim=1)
knowledge_adj = torch.stack(knowledge_adj, dim=0)
return token_ids, attention_mask, clen, \
label, mask, ece_pair, knowledge, knowledge_attention_mask, \
knowledge_adj, sequence_adj, knowledge_len, batch_knowledge_len
def get_kag_dataloaders(model_size, batch_size, valid_shuffle, unidirection=False):
train_set = KagDataset(model_size, 'train', unidirection)
dev_set = KagDataset(model_size, 'dev', unidirection)
test_set = KagDataset(model_size, 'test', unidirection)
train_loader = DataLoader(train_set, batch_size, True, collate_fn=collate_fn_kag)
dev_loader = DataLoader(dev_set, batch_size, valid_shuffle, collate_fn=collate_fn_kag)
test_loader = DataLoader(test_set, batch_size, valid_shuffle, collate_fn=collate_fn_kag)
return train_loader, dev_loader, test_loader