-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodels.py
198 lines (187 loc) · 9.78 KB
/
models.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
import torch
import torch.nn as nn
# def mean_pooling(batch, lengths):
# '''
# avg-pooling for pytorch 1.1.0 or beyond, where pack_padded_sequence is available
# '''
# return (torch.sum(batch, dim=1).transpose(0, 1)/torch.tensor(lengths, dtype=torch.float).to('cuda')).transpose(0, 1)
'''
pytorch 1.0.1 avg-pooling
'''
def mean_pooling(batch, lengths):
'''
- to apply mean_pooling for RNN-modules
- input:
batch: [batch_size, seq_len, hid_size]
lengths: the true length of each sequence
- output: [batch_size, hid_size]
'''
for i, _ in enumerate(batch):
if int(lengths[i])==0:
lengths[i]+=1
if i==0:
tmp_vec=nn.functional.avg_pool1d(_.transpose(0, 1).unsqueeze(0), int(lengths[i]))[:, :, 0]
else:
tmp_vec=torch.cat([tmp_vec, nn.functional.avg_pool1d(_.transpose(0, 1).unsqueeze(0), int(lengths[i]))[:, :, 0]])
return tmp_vec
class LayerNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-12):
'''
Construct a layernorm module in the TF style (epsilon inside the square root).
'''
super(LayerNorm, self).__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.bias = nn.Parameter(torch.zeros(hidden_size))
self.variance_epsilon = eps
def forward(self, x):
u = x.mean(-1, keepdim=True)
s = (x - u).pow(2).mean(-1, keepdim=True)
x = (x - u) / torch.sqrt(s + self.variance_epsilon)
return self.weight * x + self.bias
class BiLSTM_centric_layer(nn.Module):
'''
basic BiLSTM review-centric layer
input:
in_size - input_size for LSTM
hidden_size - hidden_size for LSTM
in_raw - raw_text, [batch_size, seq_len, vec_dim]
in_sum - summary_text, [batch_size, seq_len, vec_dim]
len_raw/len_sum - list of original (unpadded) lengths of the current batch of raw/summary text
'''
def __init__(self, in_size, hidden_size, num_heads=4, dropout_rate=0.2):
super(BiLSTM_centric_layer, self).__init__()
self.in_size=in_size
self.hidden_size=hidden_size
self.num_heads=num_heads
self.dropout_rate=dropout_rate
self.dropout_layer=nn.Dropout(p=self.dropout_rate)
self.lstm_raw=nn.LSTM(self.in_size, self.hidden_size, bidirectional=True, batch_first=True)
self.lstm_sum=nn.LSTM(self.in_size, self.hidden_size, bidirectional=True, batch_first=True)
self.proj_layers_query=nn.ModuleList([nn.Linear(self.hidden_size*2, int(self.hidden_size*2/self.num_heads), bias=False) for i in range(self.num_heads)])
self.proj_layers_key=nn.ModuleList([nn.Linear(self.hidden_size*2, int(self.hidden_size*2/self.num_heads), bias=False) for i in range(self.num_heads)])
self.proj_layers_value=nn.ModuleList([nn.Linear(self.hidden_size*2, int(self.hidden_size*2/self.num_heads), bias=False) for i in range(self.num_heads)])
# self.projection_layer=nn.Linear(self.hidden_size*4*self.num_heads, self.hidden_size*4)
def forward(self,
in_raw,
in_sum,
len_raw,
len_sum,
use_sum_lstm=False,
output_sum=None,
use_residual=True,
use_sum_avg_pooling=True,
use_concate_raw=False,
use_concate_sum=False,
use_divide_dk=False):
output_raw, (h_n_raw, c_n_raw)=self.lstm_raw(in_raw)
output_raw=self.dropout_layer(output_raw)
if use_sum_lstm:
output_sum, (h_n_sum, c_n_sum)=self.lstm_sum(in_sum)
output_sum=self.dropout_layer(output_sum)
sum_vec=mean_pooling(output_sum, len_sum)
sum_vec=self.dropout_layer(sum_vec)
rst=[]
for j in range(self.num_heads):
if use_sum_avg_pooling:
tmp_query=self.proj_layers_query[j](output_raw)
tmp_key=self.proj_layers_key[j](sum_vec)
tmp_value=self.proj_layers_value[j](sum_vec)
if use_divide_dk:
tmp_rst=torch.bmm((nn.Softmax(dim=-1)(torch.bmm(tmp_query, tmp_key.unsqueeze(-1)).squeeze(-1)/((self.hidden_size*2/self.num_heads)**0.5))).unsqueeze(-1), tmp_value.unsqueeze(-2))
else:
tmp_rst=torch.bmm(nn.Softmax(dim=-1)(torch.bmm(tmp_query, tmp_key.unsqueeze(-1)).squeeze(-1)).unsqueeze(-1), tmp_value.unsqueeze(-2))
else:
tmp_query=self.proj_layers_query[j](output_raw)
tmp_key=self.proj_layers_key[j](output_sum)
tmp_value=self.proj_layers_value[j](output_sum)
if use_divide_dk:
tmp_rst=torch.bmm(nn.Softmax(dim=-1)(torch.bmm(tmp_query, tmp_key.transpose(-2, -1))/((self.hidden_size*2/self.num_heads)**0.5)), tmp_value)
else:
tmp_rst=torch.bmm(nn.Softmax(dim=-1)(torch.bmm(tmp_query, tmp_key.transpose(-2, -1))), tmp_value)
if use_residual:
if use_concate_sum:
rst.append(torch.cat([tmp_rst+tmp_query, tmp_value], -2))
else:
rst.append(tmp_rst+tmp_query) ########### +tmp_query
else:
if use_concate_sum:
rst.append(torch.cat([tmp_rst, tmp_value], -2))
else:
rst.append(tmp_rst)
rst=torch.cat(rst, -1)
#if use_residual:
# # rst=rst+sum_vec.unsqueeze(-2)
# rst+=output_raw
if use_concate_raw and use_concate_sum:
raise KeyError('cannot concatenate output_raw and output_sum at the same time.')
if use_concate_raw:
rst=torch.cat([output_raw, rst], 2)
# elif use_concate_sum:
# rst=torch.cat([output_sum, rst], 2)
return rst
class BiLSTM_centric_model(nn.Module):
'''
BiLSTM review-centric model
'''
def __init__(self,
in_size,
hid_size,
out_classes,
num_heads=2,
num_layers=2,
dropout_rate=0.15,
use_residual=False,
use_concate_raw=False,
use_concate_sum=False,
use_layer_norm=False,
use_divide_dk=False):
super(BiLSTM_centric_model, self).__init__()
self.in_size=in_size
self.hid_size=hid_size
self.out_classes=out_classes
self.num_layers=num_layers
self.dropout_rate=dropout_rate
self.dropout_layer=nn.Dropout(p=self.dropout_rate)
self.use_residual=use_residual
self.use_concate_raw=use_concate_raw
self.use_concate_sum=use_concate_sum
self.use_layer_norm=use_layer_norm
self.num_heads=num_heads
self.use_divide_dk=use_divide_dk
self.lstm_sum=nn.LSTM(self.in_size, self.hid_size, batch_first=True, bidirectional=True)
self.attention_layers=nn.ModuleList([BiLSTM_centric_layer(self.in_size, self.hid_size, dropout_rate=self.dropout_rate, num_heads=self.num_heads) for i in range(self.num_layers)])
if self.use_concate_raw:
self.proj_layers=nn.ModuleList([nn.Linear(self.hid_size*4, self.in_size) for i in range(self.num_layers-1)])
self.norm_layers=nn.ModuleList([LayerNorm(self.hid_size*4) for i in range(self.num_layers)])
self.classifier=nn.Linear(self.hid_size*4, self.out_classes)
elif self.use_concate_sum:
self.proj_layers=nn.ModuleList([nn.Linear(self.hid_size*2, self.in_size) for i in range(self.num_layers-1)])
self.norm_layers=nn.ModuleList([LayerNorm(self.hid_size*2) for i in range(self.num_layers)])
self.classifier=nn.Linear(self.hid_size*2, self.out_classes)
else:
self.proj_layers=nn.ModuleList([nn.Linear(self.hid_size*2, self.in_size) for i in range(self.num_layers-1)])
self.norm_layers=nn.ModuleList([LayerNorm(self.hid_size*2) for i in range(self.num_layers)])
self.classifier=nn.Linear(self.hid_size*2, self.out_classes)
def forward(self, in_vec_raw, in_vec_sum, len_raw, len_sum, use_norm=False, use_sum_avg_pooling=True):
output_sum, (h_n, c_n)=self.lstm_sum(in_vec_sum)
output_sum=self.dropout_layer(output_sum)
for i in range(self.num_layers):
in_vec_raw=self.dropout_layer(self.attention_layers[i](in_vec_raw,
in_vec_sum,
len_raw,
len_sum,
use_sum_lstm=False,
output_sum=output_sum,
use_residual=self.use_residual,
use_sum_avg_pooling=use_sum_avg_pooling,
use_concate_raw=self.use_concate_raw,
use_concate_sum=self.use_concate_sum,
use_divide_dk=self.use_divide_dk))
if self.use_layer_norm:
in_vec_raw=self.dropout_layer(self.norm_layers[i](in_vec_raw))
if i!=self.num_layers-1:
in_vec_raw=self.dropout_layer(self.proj_layers[i](in_vec_raw))
rst=self.dropout_layer(self.classifier(mean_pooling(in_vec_raw, len_raw)))
if use_norm:
rst=nn.Softmax(dim=-1)(rst)
return rst