-
Notifications
You must be signed in to change notification settings - Fork 0
/
seq2seq.py
203 lines (158 loc) · 8.21 KB
/
seq2seq.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
import torch
import torch.nn as nn
from torchseq.utils.functional import softmax_mask
class Encoder(nn.Module):
def __init__(self, n_tokens, embed_dim, hidden_dim, n_layers, dropout):
super(Encoder, self).__init__()
self.n_tokens = n_tokens
self.embed_dim = embed_dim
self.hidden_dim = hidden_dim
self.n_layers = n_layers
self.embed = nn.Embedding(n_tokens, embed_dim, padding_idx=0)
self.encoder = nn.GRU(embed_dim, hidden_size=hidden_dim,
num_layers=n_layers, bidirectional=True)
def initializeWeights(self):
nn.init.uniform_(self.embed.weight, -0.1, 0.1)
for name, w in self.encoder.named_parameters():
if name.startswith("weight"):
nn.init.orthogonal_(w)
elif name.startswith("bias"):
nn.init.constant_(w, 0)
def getOuputDim(self):
return self.hidden_dim * 2
def flatten_parameters(self):
self.encoder.flatten_parameters()
def forward(self, x):
# x: seq_len, N
embed = self.embed(x) # seq_len, N, embed_dim
encoded, hidden = self.encoder(embed, None) # seq_len, N, hidden_dim * 2 ; n_layers * 2, N, hidden_dim
return encoded, hidden
class Attention(nn.Module):
def __init__(self, source_dim, target_dim, score_method, concat_hidden_dim=None):
super(Attention, self).__init__()
self.source_dim = source_dim
self.target_dim = target_dim
self.score_method = score_method
if score_method == "dot":
assert(source_dim == target_dim), \
"source_dim must equal to target_dim when use dot score_method"
elif score_method == "general":
self.atten = nn.Linear(source_dim, target_dim)
elif score_method == "concat":
if concat_hidden_dim == None:
concat_hidden_dim = source_dim + target_dim
self.atten = nn.Sequential(
nn.Linear(source_dim + target_dim, concat_hidden_dim),
nn.Tanh()
)
self.v = nn.Parameter(torch.empty(concat_hidden_dim, dtype=torch.float))
def initializeWeights(self):
if self.score_method == "general":
nn.init.normal_(self.atten.weight, 0, 0.1)
nn.init.constant_(self.atten.bias, 0)
elif self.score_method == "concat":
nn.init.normal_(self.atten.weight[0], 0, 0.1)
nn.init.constant_(self.atten.bias[0], 0)
nn.init.normal_(self.v, 0, 0.1)
def forward(self, target_h, source_hs, source_mask):
# target_h: N, target_dim
# source_hs: seq_len, N, source_dim
# source_mask: seq_len, N
a = self.score(target_h, source_hs, source_mask) # seq_len, N
a = a.unsqueeze(2).expand_as(source_hs) # seq_len, N, source_dim
att = torch.mul(a, source_hs) # seq_len, N, source_dim
att = torch.sum(att, dim=0) # N, source_dim
return att
def score(self, target_h, source_hs, source_mask):
# target_h: N, target_dim
# source_hs: seq_len, N, source_dim
# source_mask: seq_len, N
target_h = target_h.unsqueeze(0) # 1, N, target_dim
target_size = (source_hs.size(0), -1, -1) # seq_len, N, target_dim
if self.score_method == "dot":
att = torch.mul(target_h.expand(target_size), source_hs) # seq_len, N, target_dim
att = torch.sum(att, dim=2) # seq_len, N
elif self.score_method == "general":
att = self.atten(source_hs) # seq_len, N, target_dim
att = torch.mul(target_h.expand(target_size), att) # seq_len, N, target_dim
att = torch.sum(att, dim=2) # seq_len, N
elif self.score_method == "concat":
target_h = target_h.expand(target_size) # seq_len, N, target_dim
concat = torch.concat((source_hs, target_h), dim=2) # seq_len, N, source_dim + target_dim
att = self.atten(concat) # seq_len, N, concat_hidden_dim
v = self.v.unsqueeze(0).unsqueeze(0) # 1, 1, concat_hidden_dim
att = torch.mul(v.expand_as(att), att) # seq_len, N, concat_hidden_dim
att = torch.sum(att, dim=2) # seq_len, N
else:
raise RuntimeError("Unsupported score_method")
return softmax_mask(att, source_mask, dim=0) # seq_len, N
class Decoder(nn.Module):
def __init__(self, n_tokens, embed_dim, hidden_dim, n_layers, dropout,
source_dim, atten_method=None):
super(Decoder, self).__init__()
self.n_tokens = n_tokens
self.embed_dim = embed_dim
self.hidden_dim = hidden_dim
self.n_layers = n_layers
self.source_dim = source_dim
self.embed = nn.Embedding(n_tokens, embed_dim, padding_idx=0)
self.decoder = nn.GRU(embed_dim + source_dim, hidden_size=hidden_dim,
num_layers=n_layers, dropout=dropout, bidirectional=False)
self.classifier = nn.Linear(hidden_dim + source_dim, n_tokens)
self.atten = Attention(source_dim, hidden_dim, score_method=atten_method)
def initializeWeights(self):
nn.init.uniform_(self.embed.weight, -0.1, 0.1)
for name, w in self.decoder.named_parameters():
if name.startswith("weight"):
nn.init.orthogonal_(w)
elif name.startswith("bias"):
nn.init.constant_(w, 0)
nn.init.normal_(self.classifier.weight, 0, 0.1)
nn.init.constant_(self.classifier.bias, 0)
self.atten.initializeWeights()
def flatten_parameters(self):
self.decoder.flatten_parameters()
def forward(self, x, hidden, ctx, source_hs, source_mask):
# x: N
# ctx: N, source_dim
# source_hs: seq_len, N, source_dim
# source_mask: seq_len, N
embed = self.embed(x) # N, embed_dim
if ctx is None:
ctx = embed.new_zeros(x.size(0), self.source_dim, requires_grad=False)
decoder_input = torch.cat((embed, ctx), dim=1) # N, embed_dim + source_dim
decoder_input = decoder_input.unsqueeze(0) # 1, N, embed_dim + source_dim
decoder_output, hidden = self.decoder(decoder_input, hidden) # decoder_output: 1, N, hidden_dim
decoder_output = decoder_output.squeeze(0) # N, hidden_dim
ctx = self.atten(decoder_output, source_hs, source_mask) # N, source_dim
output = self.classifier(torch.cat((decoder_output, ctx), dim=1)) # N, n_tokens
return output, hidden, ctx
class Seq2seq(nn.Module):
def __init__(self, n_tokens: list, embed_dim: list, hidden_dim: list, n_layers: list, dropout: list=[0.1, 0.1], atten_method="general"):
super(Seq2seq, self).__init__()
self.n_tokens = n_tokens
self.embed_dim = embed_dim
self.hidden_dim = hidden_dim
self.n_layers = n_layers
self.encoder = Encoder(n_tokens[0], embed_dim[0], hidden_dim[0], n_layers[0], dropout[0])
self.hiddenTransformer = nn.Linear(n_layers[0] * 2 * hidden_dim[0], n_layers[1] * hidden_dim[1])
self.decoder = Decoder(n_tokens[1], embed_dim[1], hidden_dim[1], n_layers[1], dropout[0], self.encoder.getOuputDim(), atten_method)
def initializeWeights(self):
self.encoder.initializeWeights()
self.decoder.initializeWeights()
nn.init.normal_(self.hiddenTransformer.weight, 0, 0.1)
nn.init.constant_(self.hiddenTransformer.bias, 0)
def flatten_parameters(self):
self.encoder.flatten_parameters()
self.decoder.flatten_parameters()
def transformHidden(self, hidden):
# hidden: n_layers[0] * 2, N, hidden_dim[0]
# target: n_layers[1], N, hidden_dim[1]
hidden = hidden.transpose(0, 1).contiguous() # N, n_layers[0] * 2, hidden_dim[0]
hidden = hidden.view(hidden.size(0), -1) # N, n_layers[0] * 2 * hidden_dim[0]
hidden = self.hiddenTransformer(hidden) # N, n_layers[1] * hidden_dim[1]
hidden = hidden.view(hidden.size(0), self.n_layers[1], self.hidden_dim[1]) # N, n_layers[1], hidden_dim[1]
hidden = hidden.transpose(0, 1).contiguous() # n_layers[1], N, hidden_dim[1]
return hidden
def forward(self, *args, **kwargs):
raise NotImplementedError("Use encoder and decoder directly instead of Seq2seq itself")