/
model.py
228 lines (183 loc) · 9.4 KB
/
model.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
import torch.nn as nn
import torch
import torch.nn.functional as F
from torch.nn.utils import spectral_norm
import torch.distributions.binomial as binomial
class SelectItem(nn.Module):
def __init__(self, item_index):
super(SelectItem, self).__init__()
self._name = 'selectitem'
self.item_index = item_index
def forward(self, inputs):
return inputs[self.item_index]
class Embedding(nn.Module):
def __init__(self, vocab_size, embedding_dim, pretrained_embedding=None):
super(Embedding, self).__init__()
self.embedding = nn.Embedding(vocab_size, embedding_dim)
if pretrained_embedding is not None:
self.embedding.weight.data = torch.from_numpy(pretrained_embedding)
self.embedding.weight.requires_grad = False
def forward(self, x):
"""
Inputs:
x -- (batch_size, seq_length)
Outputs
shape -- (batch_size, seq_length, embedding_dim)
"""
return self.embedding(x)
class Multi_gen(nn.Module):
def __init__(self,args):
super(Multi_gen, self).__init__()
self.args = args
self.embedding_layer = Embedding(args.vocab_size,
args.embedding_dim,
args.pretrained_embedding)
self.cls = nn.GRU(input_size=args.embedding_dim,
hidden_size=args.hidden_dim // 2,
num_layers=args.num_layers,
batch_first=True,
bidirectional=True)
self.cls_fc = nn.Linear(args.hidden_dim, args.num_class)
self.z_dim = 2
self.dropout = nn.Dropout(args.dropout)
self.gen_list=[]
if args.share==0:
for idx in range(args.num_gen):
# temp=nn.Sequential(nn.GRU(input_size=args.embedding_dim,
# hidden_size=args.hidden_dim // 2,
# num_layers=args.num_layers,
# batch_first=True,
# bidirectional=True),
# SelectItem(0),
# nn.LayerNorm(args.hidden_dim),
# self.dropout,
# nn.Linear(args.hidden_dim, self.z_dim)).to('cuda:{}'.format(args.gpu))
self.gen_list.append(nn.Sequential(nn.GRU(input_size=args.embedding_dim,
hidden_size=args.hidden_dim // 2,
num_layers=args.num_layers,
batch_first=True,
bidirectional=True),
SelectItem(0),
nn.LayerNorm(args.hidden_dim),
self.dropout,
nn.Linear(args.hidden_dim, self.z_dim)).to('cuda:{}'.format(args.gpu)))
elif args.share==1:
self.gen = nn.GRU(input_size=args.embedding_dim,
hidden_size=args.hidden_dim // 2,
num_layers=args.num_layers,
batch_first=True,
bidirectional=True)
for idx in range(args.num_gen):
self.gen_list.append(nn.Sequential(self.gen,
SelectItem(0),
nn.LayerNorm(args.hidden_dim),
self.dropout,
nn.Linear(args.hidden_dim, self.z_dim)).to('cuda:{}'.format(args.gpu)))
def _independent_soft_sampling(self, rationale_logits):
"""
Use the hidden states at all time to sample whether each word is a rationale or not.
No dependency between actions. Return the sampled (soft) rationale mask.
Outputs:
z -- (batch_size, sequence_length, 2)
"""
z = torch.softmax(rationale_logits, dim=-1)
return z
def independent_straight_through_sampling(self, rationale_logits):
"""
Straight through sampling.
Outputs:
z -- shape (batch_size, sequence_length, 2)
"""
z = self._independent_soft_sampling(rationale_logits)
z = F.gumbel_softmax(rationale_logits, tau=1, hard=True)
return z
def forward(self, inputs, masks):
masks_ = masks.unsqueeze(-1)
########## Genetator ##########
embedding = masks_ * self.embedding_layer(inputs) # (batch_size, seq_length, embedding_dim)
# gen_output, _ = self.gen(embedding) # (batch_size, seq_length, hidden_dim)
# if self.lay:
# gen_output = self.layernorm1(gen_output)
# gen_logits = self.gen_fc(self.dropout(gen_output)) # (batch_size, seq_length, 2)
gen_logits=[gen(embedding) for gen in self.gen_list ]
# gen_logits=self.generator(embedding)
########## Sample ##########
# z = self.independent_straight_through_sampling(gen_logits) # (batch_size, seq_length, 2)
z_list=[self.independent_straight_through_sampling(logit) for logit in gen_logits]
########## Classifier ##########
cls_logits_list=[]
for z in z_list:
cls_embedding = embedding * (z[:, :, 1].unsqueeze(-1)) # (batch_size, seq_length, embedding_dim)
cls_outputs, _ = self.cls(cls_embedding) # (batch_size, seq_length, hidden_dim)
cls_outputs = cls_outputs * masks_ + (1. -
masks_) * (-1e6)
# (batch_size, hidden_dim, seq_length)
cls_outputs = torch.transpose(cls_outputs, 1, 2)
cls_outputs, _ = torch.max(cls_outputs, axis=2)
# shape -- (batch_size, num_classes)
cls_logits = self.cls_fc(self.dropout(cls_outputs))
cls_logits_list.append(cls_logits)
return z_list, cls_logits_list
def test(self,inputs, masks):
masks_ = masks.unsqueeze(-1)
########## Genetator ##########
embedding = masks_ * self.embedding_layer(inputs) # (batch_size, seq_length, embedding_dim)
# gen_output, _ = self.gen(embedding) # (batch_size, seq_length, hidden_dim)
# if self.lay:
# gen_output = self.layernorm1(gen_output)
# gen_logits = self.gen_fc(self.dropout(gen_output)) # (batch_size, seq_length, 2)
gen_logits = [torch.softmax(gen(embedding),dim=-1) for gen in self.gen_list]
mean_logits=sum(gen_logits)/len(gen_logits)
########## Sample ##########
# z = self.independent_straight_through_sampling(gen_logits) # (batch_size, seq_length, 2)
# z_list = [self.independent_straight_through_sampling(logit) for logit in gen_logits]
# z=self.independent_straight_through_sampling(mean_logits)
z_distribution=binomial.Binomial(1,mean_logits)
z=z_distribution.sample()
########## Classifier ##########
cls_embedding = embedding * (z[:, :, 1].unsqueeze(-1)) # (batch_size, seq_length, embedding_dim)
cls_outputs, _ = self.cls(cls_embedding) # (batch_size, seq_length, hidden_dim)
cls_outputs = cls_outputs * masks_ + (1. -
masks_) * (-1e6)
# (batch_size, hidden_dim, seq_length)
cls_outputs = torch.transpose(cls_outputs, 1, 2)
cls_outputs, _ = torch.max(cls_outputs, axis=2)
# shape -- (batch_size, num_classes)
cls_logits = self.cls_fc(self.dropout(cls_outputs))
return z, cls_logits
def test_one_head(self,inputs, masks):
head_1=self.gen_list[0]
masks_ = masks.unsqueeze(-1)
embedding = masks_ * self.embedding_layer(inputs)
gen_logits=head_1(embedding)
z = self.independent_straight_through_sampling(gen_logits) # (batch_size, seq_length, 2)
########## Classifier ##########
cls_embedding = embedding * (z[:, :, 1].unsqueeze(-1)) # (batch_size, seq_length, embedding_dim)
cls_outputs, _ = self.cls(cls_embedding) # (batch_size, seq_length, hidden_dim)
cls_outputs = cls_outputs * masks_ + (1. -
masks_) * (-1e6)
# (batch_size, hidden_dim, seq_length)
cls_outputs = torch.transpose(cls_outputs, 1, 2)
cls_outputs, _ = torch.max(cls_outputs, axis=2)
# shape -- (batch_size, num_classes)
cls_logits = self.cls_fc(self.dropout(cls_outputs))
return z, cls_logits
def get_cls_param(self):
layers = [self.cls, self.cls_fc]
params = []
for layer in layers:
params.extend([param for param in layer.parameters() if param.requires_grad])
return params
def train_one_step(self, inputs, masks):
masks_ = masks.unsqueeze(-1)
# (batch_size, seq_length, embedding_dim)
embedding = masks_ * self.embedding_layer(inputs)
outputs, _ = self.cls(embedding) # (batch_size, seq_length, hidden_dim)
outputs = outputs * masks_ + (1. -
masks_) * (-1e6)
# (batch_size, hidden_dim, seq_length)
outputs = torch.transpose(outputs, 1, 2)
outputs, _ = torch.max(outputs, axis=2)
# shape -- (batch_size, num_classes)
logits = self.cls_fc(self.dropout(outputs))
return logits