forked from luopeixiang/im2latex
-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
178 lines (151 loc) · 6.31 KB
/
model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
INIT = 1e-2
class Im2LatexModel(nn.Module):
def __init__(self, out_size, emb_size,
enc_rnn_h, dec_rnn_h, n_layer=1):
super(Im2LatexModel, self).__init__()
# follow the original paper's table2: CNN specification
self.cnn_encoder = nn.Sequential(
nn.Conv2d(3, 512, 3, 1, 0),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.Conv2d(512, 512, 3, 1, 1),
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d((1, 2), (1, 2), (0, 0)),
nn.Conv2d(512, 256, 3, 1, 1),
nn.ReLU(),
nn.MaxPool2d((2, 1), (2, 1), (0, 0)),
nn.Conv2d(256, 256, 3, 1, 1),
nn.BatchNorm2d(256),
nn.ReLU(),
nn.Conv2d(256, 128, 3, 1, 1),
nn.ReLU(),
nn.MaxPool2d((2, 2), (2, 2), (0, 0)),
nn.Conv2d(128, 64, 3, 1, 1),
nn.ReLU(),
nn.MaxPool2d((2, 2), (2, 2), (1, 1))
)
self.rnn_encoder = nn.LSTM(64, enc_rnn_h,
bidirectional=True,
batch_first=True)
self.rnn_decoder = nn.LSTMCell(enc_rnn_h+emb_size, dec_rnn_h)
self.embedding = nn.Embedding(out_size, emb_size)
# enc_rnn_h*2 is the dimension of context
self.W_c = nn.Linear(dec_rnn_h+2*enc_rnn_h, enc_rnn_h)
self.W_out = nn.Linear(enc_rnn_h, out_size)
# a trainable initial hidden state V_h_0 for each row
self.V_h_0 = nn.Parameter(torch.Tensor(n_layer*2, enc_rnn_h))
self.V_c_0 = nn.Parameter(torch.Tensor(n_layer*2, enc_rnn_h))
init.uniform_(self.V_h_0, -INIT, INIT)
init.uniform_(self.V_c_0, -INIT, INIT)
# Attention mechanism
self.beta = nn.Parameter(torch.Tensor(dec_rnn_h))
init.uniform_(self.beta, -INIT, INIT)
self.W_h = nn.Linear(dec_rnn_h, dec_rnn_h)
self.W_v = nn.Linear(enc_rnn_h*2, dec_rnn_h)
def forward(self, imgs, formulas):
"""args:
imgs: [B, C, H, W]
formulas: [B, MAX_LEN]
return:
logits: [B, MAX_LEN, VOCAB_SIZE]
"""
# encoding
row_enc_out, hiddens = self.encode(imgs)
# init decoder's states
dec_states, O_t = self.init_decoder(row_enc_out, hiddens)
max_len = formulas.size(1)
logits = []
for t in range(max_len):
tgt = formulas[:, t:t+1]
# ont step decoding
dec_states, O_t, logit = self.step_decoding(
dec_states, O_t, row_enc_out, tgt)
logits.append(logit)
logits = torch.stack(logits, dim=1) # [B, MAX_LEN, out_size]
return logits
def encode(self, imgs):
encoded_imgs = self.cnn_encoder(imgs) # [B, 64, H', W']
encoded_imgs = encoded_imgs.permute(0, 2, 3, 1) # [B, H', W', 64]
# Prepare data for Row Encoder
# poccess data like a new big batch
B, H, W, out_channels = encoded_imgs.size()
encoded_imgs = encoded_imgs.contiguous().view(B*H, W, out_channels)
# prepare init hidden for each row
init_hidden_h = self.V_h_0.unsqueeze(
1).expand(-1, B*H, -1).contiguous()
init_hidden_c = self.V_c_0.unsqueeze(
1).expand(-1, B*H, -1).contiguous()
init_hidden = (init_hidden_h, init_hidden_c)
# Row Encoder
row_enc_out, (h, c) = self.rnn_encoder(encoded_imgs, init_hidden)
# row_enc_out [B*H, W, enc_rnn_h]
# hidden: [2, B*H, enc_rnn_h]
row_enc_out = row_enc_out.view(B, H, W, -1) # [B, H, W, enc_rnn_h]
h, c = h.view(2, B, H, -1), c.view(2, B, H, -1)
return row_enc_out, (h, c)
def step_decoding(self, dec_states, O_t, enc_out, tgt):
"""Runing one step decoding"""
prev_y = self.embedding(tgt).squeeze(1) # [B, emb_size]
inp = torch.cat([prev_y, O_t], dim=1) # [B, emb_size+enc_rnn_h]
h_t, c_t = self.rnn_decoder(inp, dec_states)
context_t, attn_scores = self._get_attn(enc_out, dec_states[0])
# [B, enc_rnn_h]
O_t = self.W_c(torch.cat([h_t, context_t], dim=1)).tanh()
# calculate logit
logit = F.softmax(self.W_out(O_t), dim=1) # [B, out_size]
return (h_t, c_t), O_t, logit
def _get_attn(self, enc_out, prev_h):
"""Attention mechanism
args:
enc_out: row encoder's output [B, H, W, enc_rnn_h]
prev_h: the previous time step hidden state [B, dec_rnn_h]
return:
context: this time step context [B, enc_rnn_h]
attn_scores: Attention scores
"""
# self.W_v(enc_out) [B, H, W, enc_rnn_h]
# self.W_h(prev_h) [B, enc_rnn_h]
B, H, W, _ = enc_out.size()
linear_prev_h = self.W_h(prev_h).view(B, 1, 1, -1)
linear_prev_h = linear_prev_h.expand(-1, H, W, -1)
e = torch.sum(
self.beta * torch.tanh(
linear_prev_h +
self.W_v(enc_out)
),
dim=-1
) # [B, H, W]
alpha = F.softmax(e.view(B, -1), dim=-1).view(B, H, W)
attn_scores = alpha.unsqueeze(-1)
context = torch.sum(attn_scores * enc_out,
dim=[1, 2]) # [B, enc_rnn_h]
return context, attn_scores
def init_decoder(self, enc_out, hiddens):
"""args:
enc_out: the output of row encoder [B, H, W, enc_rnn_h]
hidden: the last step hidden of row encoder [2, B, H, enc_rnn_h]
return:
h_0, c_0 h_0 and c_0's shape: [B, dec_rnn_h]
init_O : the average of enc_out [B, enc_rnn_h]
for decoder
"""
h, c = hiddens
h, c = self._convert_hidden(h), self._convert_hidden(c)
context_0 = enc_out.mean(dim=[1, 2])
init_O = torch.tanh(
self.W_c(torch.cat([h, context_0], dim=1))
)
return (h, c), init_O
def _convert_hidden(self, hidden):
"""convert row encoder hidden to decoder initial hidden"""
hidden = hidden.permute(1, 2, 0, 3).contiguous()
# Note that 2*enc_rnn_h = dec_rnn_h
hidden = hidden.view(hidden.size(
0), hidden.size(1), -1) # [B, H, dec_rnn_h]
hidden = hidden.mean(dim=1) # [B, dec_rnn_h]
return hidden