-
Notifications
You must be signed in to change notification settings - Fork 0
/
EncoderDecoder.py
158 lines (118 loc) · 4.37 KB
/
EncoderDecoder.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
import torch
import torch.nn as nn
from vocabulary import max_len
import torchvision.models as models
class CNNEncoder(nn.Module):
def __init__(
self,
device,
):
super(CNNEncoder, self).__init__()
vgg16 = models.vgg16(pretrained=True).to(device)
for param in vgg16.parameters():
param.requires_grad_(False)
# Remove linear and pool layers (since we're not doing classification)
modules = list(vgg16.children())[:-1]
self.vgg16 = nn.Sequential(*modules)
def forward(
self,
images,
):
# Feed images through VGG16
features = self.vgg16(images)
features = features.permute(0, 2,3,1)
features = features.view(features.size(0), -1, features.size(-1)).squeeze()
return features
def fine_tune(
self,
fine_tune=True
):
"""
Allow or prevent the computation of gradients for convolutional blocks 2 through 4 of the encoder.
:param fine_tune: Allow?
"""
for p in self.vgg16.parameters():
p.requires_grad = False
# If fine-tuning, only fine-tune convolutional blocks 2 through 4
for c in list(self.vgg16.children())[5:]:
for p in c.parameters():
p.requires_grad = fine_tune
class Attention(nn.Module):
def __init__(
self, decoder_out_dim, attention_dim, encoder_dim=512
):
super(Attention, self).__init__()
self.attention_dim = attention_dim
self.hidden_map = nn.Linear(decoder_out_dim, attention_dim)
self.feature_map = nn.Linear(encoder_dim, attention_dim)
self.attention_map = nn.Linear(attention_dim, 1)
self.softmax = nn.Softmax(dim = 1)
def forward(
self,
features,
hidden,
):
mapped_features = self.feature_map(features)
mapped_hidden = self.hidden_map(hidden)
combined_states = torch.tanh(mapped_features + mapped_hidden.unsqueeze(1))
attn_scores = self.attention_map(combined_states).squeeze(2)
alphas = self.softmax(attn_scores)
attn_weights = features * alphas.unsqueeze(2)
attn_weights = attn_weights.sum(dim=1)
return alphas, attn_weights
class DecoderWithAttention(nn.Module):
def __init__(
self,
vocab_size,
embedding_dim,
attention_dim,
encoder_dim,
decoder_dim,
device,
drop_prob=0.1,
pad_token=None,
):
super(DecoderWithAttention, self).__init__()
self.device = device
self.decoder_dim = decoder_dim
self.embedding = nn.Embedding(
vocab_size,
embedding_dim,
padding_idx=pad_token,
)
self.dropout = nn.Dropout(drop_prob)
self.attention = Attention(decoder_dim, attention_dim).to(device)
self.lstm_cell = nn.LSTMCell(embedding_dim + encoder_dim, decoder_dim)
self.LinearMap = nn.Linear(
in_features=decoder_dim,
out_features=vocab_size,
)
def forward(
self,
sentences, # (batch_size)
encoder_features, # (batch_size, number_of_layers=49, encoder_output_size=512)
):
global max_len
# Embedding
embedded = self.embedding(sentences)
embedded = self.dropout(embedded) # (batch_size, embedding_dim)
batch_size = sentences.shape[0]
hidden, cell = self.getInitialHidden(batch_size)
outputs = []
alphas = []
for s in range(max_len-1):
# Attention
alpha, attn_weight = self.attention(encoder_features, hidden)
lstm_input = torch.cat((embedded[:, s], attn_weight) , -1)
hidden, cell = self.lstm_cell(lstm_input, (hidden, cell))
hidden = self.dropout(hidden)
output = self.LinearMap(hidden)
outputs.append(output)
alphas.append(alpha)
outputs = torch.stack(outputs, dim = 1)
alphas = torch.stack(alphas, dim = 1)
return outputs, alphas
def getInitialHidden(self, batch_size):
hidden = torch.zeros((batch_size, self.decoder_dim)).to(self.device)
cell = torch.zeros((batch_size, self.decoder_dim)).to(self.device)
return hidden, cell