forked from chrischute/squad
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodels.py
224 lines (166 loc) · 10 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
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
"""Top-level model classes.
Author:
Chris Chute ([email protected])
"""
import layers
import torch
import torch.nn as nn
class BiDAFBaseline(nn.Module):
"""Baseline BiDAF model for SQuAD.
Based on the paper:
"Bidirectional Attention Flow for Machine Comprehension"
by Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, Hannaneh Hajishirzi
(https://arxiv.org/abs/1611.01603).
Follows a high-level structure commonly found in SQuAD models:
- Embedding layer: Embed word indices to get word vectors.
- Encoder layer: Encode the embedded sequence.
- Attention layer: Apply an attention mechanism to the encoded sequence.
- Model encoder layer: Encode the sequence again.
- Output layer: Simple layer (e.g., fc + softmax) to get final outputs.
Args:
word_vectors (torch.Tensor): Pre-trained word vectors.
hidden_size (int): Number of features in the hidden state at each layer.
drop_prob (float): Dropout probability.
"""
def __init__(self, word_vectors, char_vectors, hidden_size, drop_prob=0., use_chars=True):
super(BiDAFBaseline, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
use_chars=use_chars,
char_vectors=char_vectors,
num_filters=100)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size,
num_layers=2,
drop_prob=drop_prob)
self.out = layers.SQuADOutput(hidden_size=hidden_size,
drop_prob=drop_prob)
def forward(self, cw_idxs, cc_idxs, qw_idxs, qc_idxs):
c_mask = torch.zeros_like(cw_idxs) != cw_idxs
q_mask = torch.zeros_like(qw_idxs) != qw_idxs
c_len, q_len = c_mask.sum(-1), q_mask.sum(-1)
c_emb = self.emb(cw_idxs, cc_idxs) # (batch_size, c_len, hidden_size)
q_emb = self.emb(qw_idxs, qc_idxs) # (batch_size, q_len, hidden_size)
c_enc = self.enc(c_emb, c_len) # (batch_size, c_len, 2 * hidden_size)
q_enc = self.enc(q_emb, q_len) # (batch_size, q_len, 2 * hidden_size)
att = self.att(c_enc, q_enc,
c_mask, q_mask) # (batch_size, c_len, 8 * hidden_size)
mod = self.mod(att, c_len) # (batch_size, c_len, 2 * hidden_size)
out = self.out(att, mod, c_mask) # 2 tensors, each (batch_size, c_len)
return out
class BiDAFEncoder(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, hidden_size_2, drop_prob=0., use_chars=True):
super(BiDAFEncoder, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
use_chars=use_chars,
char_vectors=char_vectors,
num_filters=100)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttention(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size_2,
num_layers=1,
drop_prob=drop_prob)
def forward(self, cw_idxs, cc_idxs, qw_idxs, qc_idxs):
c_mask = torch.zeros_like(cw_idxs) != cw_idxs
q_mask = torch.zeros_like(qw_idxs) != qw_idxs
c_len, q_len = c_mask.sum(-1), q_mask.sum(-1)
c_emb = self.emb(cw_idxs, cc_idxs) # (batch_size, c_len, hidden_size)
q_emb = self.emb(qw_idxs, qc_idxs) # (batch_size, q_len, hidden_size)
c_enc = self.enc(c_emb, c_len) # (batch_size, c_len, 2 * hidden_size)
q_enc = self.enc(q_emb, q_len) # (batch_size, q_len, 2 * hidden_size)
att = self.att(c_enc, q_enc,
c_mask, q_mask) # (batch_size, c_len, 8 * hidden_size)
mod = self.mod(att, c_len) # (batch_size, c_len, 2 * hidden_size)
return att, mod, c_mask
class BiDAFEncoderMirrored(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, hidden_size_2, drop_prob=0., use_chars=True):
super(BiDAFEncoderMirrored, self).__init__()
self.emb = layers.Embedding(word_vectors=word_vectors,
hidden_size=hidden_size,
drop_prob=drop_prob,
use_chars=use_chars,
char_vectors=char_vectors,
num_filters=100)
self.enc = layers.RNNEncoder(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
drop_prob=drop_prob)
self.att = layers.BiDAFAttentionMirrored(hidden_size=2 * hidden_size,
drop_prob=drop_prob)
self.mod = layers.RNNEncoder(input_size=8 * hidden_size,
hidden_size=hidden_size_2,
num_layers=1,
drop_prob=drop_prob)
def forward(self, cw_idxs, cc_idxs, qw_idxs, qc_idxs):
c_mask = torch.zeros_like(cw_idxs) != cw_idxs
q_mask = torch.zeros_like(qw_idxs) != qw_idxs
c_len, q_len = c_mask.sum(-1), q_mask.sum(-1)
c_emb = self.emb(cw_idxs, cc_idxs) # (batch_size, c_len, hidden_size)
q_emb = self.emb(qw_idxs, qc_idxs) # (batch_size, q_len, hidden_size)
c_enc = self.enc(c_emb, c_len) # (batch_size, c_len, 2 * hidden_size)
q_enc = self.enc(q_emb, q_len) # (batch_size, q_len, 2 * hidden_size)
c_att, q_att = self.att(c_enc, q_enc,
c_mask, q_mask) # (batch_size, c_len, 8 * hidden_size)
c_mod = self.mod(c_att, c_len) # (batch_size, c_len, 2 * hidden_size)
q_mod = self.mod(q_att, q_len) # (batch_size, c_len, 2 * hidden_size)
return c_att, c_mod, c_mask, q_att, q_mod, q_mask
class BiDAFSQuAD(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, hidden_size_2, drop_prob=0., use_chars=True):
super(BiDAFSQuAD, self).__init__()
self.encoder = BiDAFEncoder(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
hidden_size_2=hidden_size_2,
drop_prob=drop_prob,
use_chars=use_chars)
self.output_layer = layers.SQuADOutput(hidden_size=hidden_size,
hidden_size_2=hidden_size_2,
drop_prob=drop_prob)
def forward(self, cw_idxs, cc_idxs, qw_idxs, qc_idxs):
att, mod, c_mask = self.encoder(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
out = self.output_layer(att, mod, c_mask)
return out
class BiDAFGT(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, hidden_size_2, drop_prob=0., use_chars=True):
super(BiDAFGT, self).__init__()
self.encoder = BiDAFEncoder(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
hidden_size_2=hidden_size_2,
drop_prob=drop_prob,
use_chars=use_chars)
self.output_layer = layers.GTOutputDoubleWindowPooling(hidden_size=hidden_size, hidden_size_2=hidden_size_2)
def forward(self, cw_idxs, cc_idxs, qw_idxs, qc_idxs, gap_indices):
att, mod, c_mask = self.encoder(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
out = self.output_layer(att, mod, gap_indices, c_mask)
return out
class BiDAFMLMNSP(nn.Module):
def __init__(self, word_vectors, char_vectors, hidden_size, hidden_size_2, drop_prob=0., use_chars=True):
super(BiDAFMLMNSP, self).__init__()
self.encoder = BiDAFEncoderMirrored(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=hidden_size,
hidden_size_2=hidden_size_2,
drop_prob=drop_prob,
use_chars=use_chars)
self.output_layer = layers.MLMNSPOutput(hidden_size=hidden_size,
hidden_size_2=hidden_size_2,
vocab_size=word_vectors.shape[0])
def forward(self, cw_idxs, cc_idxs, qw_idxs, qc_idxs, mask_1, mask_2, texts_1_words_unmasked, texts_2_words_unmasked):
c_att, c_mod, c_mask, q_att, q_mod, q_mask = self.encoder(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
MLM_logits, masked_words, NSP_logits = self.output_layer(c_att, c_mod, q_att, q_mod, mask_1, mask_2, texts_1_words_unmasked, texts_2_words_unmasked)
return MLM_logits, masked_words, NSP_logits