-
Notifications
You must be signed in to change notification settings - Fork 2
/
models.py
206 lines (159 loc) · 7.49 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models import resnet152, resnet101
from torch.nn.utils.rnn import pack_padded_sequence
from torch.nn.utils.rnn import pad_packed_sequence
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class Encoder(nn.Module):
def __init__(self, architecture='resnet152'):
super(Encoder, self).__init__()
self.architecture = architecture
if architecture == 'resnet152':
self.net = resnet152(pretrained=True)
self.net = nn.Sequential(*list(self.net.children())[:-2])
self.dim = 2048
else:
self.net = resnet101(pretrained=True)
self.net = nn.Sequential(*list(self.net.children())[:-2])
self.dim = 2048
self.fine_tune()
def forward(self, img):
feats = self.net(img)
feats = feats.permute(0, 2, 3, 1)
feats = feats.view(feats.size(0), -1, feats.size(-1))
return feats
def fine_tune(self, fine_tune=False):
if not fine_tune:
for param in self.net.parameters():
param.requires_grad = False
class Attention(nn.Module):
def __init__(self, encoder_dim, decoder_dim, attention_dim):
super(Attention, self).__init__()
self.W1 = nn.Linear(encoder_dim, attention_dim)
self.W2 = nn.Linear(decoder_dim, attention_dim)
self.V = nn.Linear(attention_dim, 1)
self.tanh = nn.Tanh()
self.softmax = nn.Softmax(dim=1)
def forward(self, img_feats, hidden):
x = self.W1(img_feats)
y = self.W2(hidden)
x = self.V(self.tanh(x + y.unsqueeze(1))).squeeze(2)
alphas = self.softmax(x)
weighted_feats = (img_feats * alphas.unsqueeze(2)).sum(dim=1)
return weighted_feats, alphas
class Generator(nn.Module):
def __init__(self,
attention_dim,
embedding_dim,
gru_units,
vocab_size,
encoder_dim=2048,
dropout=0.5
):
super(Generator, self).__init__()
self.encoder_dim = encoder_dim
self.attention_dim = attention_dim
self.embedding_dim = embedding_dim
self.gru_units = gru_units
self.vocab_size = vocab_size
self.dropout = dropout
self.attention_net = Attention(encoder_dim, gru_units, attention_dim)
self.embedding = nn.Embedding(vocab_size, embedding_dim)
self.dropout = nn.Dropout(p=self.dropout)
self.gru = nn.GRUCell(embedding_dim + encoder_dim, gru_units, bias=True)
self.init_h = nn.Linear(encoder_dim, gru_units)
self.f_beta = nn.Linear(gru_units, encoder_dim)
self.sigmoid = nn.Sigmoid()
self.fc = nn.Linear(gru_units, vocab_size)
self.softmax = nn.Softmax(dim=1)
self.relu = nn.ReLU()
def init_hidden_state(self, img_feats):
mean_img_feats = img_feats.mean(dim=1)
hidden = self.init_h(mean_img_feats)
hidden = self.relu(hidden)
return hidden
def forward(self, img_feats, caps, cap_lens):
batch_size = img_feats.size(0)
vocab_size = self.vocab_size
num_pixels = img_feats.size(1)
cap_lens, indices = cap_lens.sort(dim=0, descending=True)
img_feats = img_feats[indices]
caps = caps[indices]
embeddings = self.embedding(caps)
hidden_state = self.init_hidden_state(img_feats)
output_lens = (cap_lens - 1).tolist()
preds = torch.zeros(batch_size, caps.shape[1] - 1, vocab_size).to(device)
alphas = torch.zeros(batch_size, caps.shape[1] - 1, num_pixels).to(device)
for t in range(max(output_lens)):
context_vec, alpha = self.attention_net(img_feats, hidden_state)
gate = self.sigmoid(self.f_beta(hidden_state))
context_vec = gate * context_vec
hidden_state = self.gru(torch.cat([embeddings[:, t],
context_vec], dim=1), hidden_state)
preds[:, t] = self.fc(self.dropout(hidden_state))
alphas[:, t] = alpha
return preds, caps, output_lens, alphas, indices
def step(self, input_word, hidden_state, img_feats):
embeddings = self.embedding(input_word)
context_vec, alpha = self.attention_net(img_feats, hidden_state)
gate = self.sigmoid(self.f_beta(hidden_state))
context_vec = gate * context_vec
hidden_state = self.gru(torch.cat([embeddings, context_vec], dim=1), hidden_state)
preds = self.softmax(self.fc(hidden_state))
return preds, hidden_state
def sample(self, cap_len, col_shape, img_feats, input_word, sampling_method='multinomial', hidden_state=None):
samples = torch.zeros(input_word.shape[0], col_shape).long().to(device)
if hidden_state is None:
hidden_states = torch.zeros(input_word.shape[0], col_shape, self.gru_units).to(device)
hidden_state = self.init_hidden_state(img_feats)
samples[:, 0] = input_word
for i in range(cap_len):
preds, hidden_state = self.step(input_word, hidden_state, img_feats)
if sampling_method == 'multinomial':
input_word = torch.multinomial(preds, 1)
input_word = input_word.squeeze(-1)
else:
input_word = torch.argmax(preds, 1)
samples[:, i + 1] = input_word
hidden_states[:, i] = hidden_state
return samples, hidden_states
else:
for i in range(cap_len):
preds, hidden_state = self.step(input_word, hidden_state, img_feats)
if sampling_method == 'multinomial':
input_word = torch.multinomial(preds, 1)
input_word = input_word.squeeze(-1)
else:
input_word = torch.argmax(preds, 1)
samples[:, i] = input_word
return samples
class GRUDiscriminator(nn.Module):
def __init__(self, embedding_dim, encoder_dim, gru_units, vocab_size):
super(GRUDiscriminator, self).__init__()
self.encoder_dim = encoder_dim
self.embedding_dim = embedding_dim
self.gru_units = gru_units
self.vocab_size = vocab_size
self.embedding = nn.Embedding(vocab_size, embedding_dim)
self.gru = nn.GRU(input_size=embedding_dim, hidden_size=gru_units, batch_first=True)
self.fc1 = nn.Linear(encoder_dim, embedding_dim)
self.fc2 = nn.Linear(gru_units, 1)
self.sigmoid = nn.Sigmoid()
def forward(self, img_feats, caps, cap_lens):
img_feats = img_feats.permute(0, 2, 1)
img_feats = F.avg_pool1d(img_feats, img_feats.shape[-1]).squeeze(-1)
img_feats = self.fc1(img_feats)
embeddings = self.embedding(caps)
inputs = torch.cat((img_feats.unsqueeze(1), embeddings), 1)
inputs_packed = pack_padded_sequence(inputs, cap_lens + 1, batch_first=True, enforce_sorted=False)
outputs, _ = self.gru(inputs_packed)
try:
outputs = pad_packed_sequence(outputs, batch_first=True)[0]
except:
print(outputs)
print(outputs.shape)
row_indices = torch.arange(0, caps.size(0)).long()
last_hidden = outputs[row_indices, cap_lens, :]
pred = self.sigmoid(self.fc2(last_hidden))
return pred.squeeze(-1)