forked from bupt-mmai/CNN-Caption
-
Notifications
You must be signed in to change notification settings - Fork 0
/
DualCNN.py
executable file
·361 lines (284 loc) · 15.3 KB
/
DualCNN.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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
# -*- coding: utf-8 -*-
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import numpy as np
def Conv1d(in_channels, out_channels, kernel_size, padding):
m = nn.Conv1d(in_channels, out_channels, kernel_size, padding=padding)
# Xavier Init
std = math.sqrt((4 / (kernel_size * in_channels)))
m.weight.data.normal_(mean=0, std=std)
m.bias.data.zero_()
# return m with weight normalization
return nn.utils.weight_norm(m)
def Embedding(num_embeddings, embedding_dim, padding_idx, word2idx):
m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx)
m.weight.data.normal_(0, 0.1)
# Embedding of <padding> is set to 0
m.weight.data[padding_idx].fill_(0)
return m
def Linear(in_features, out_features, activation=None):
m = nn.Linear(in_features, out_features)
# Xavier Init
m.weight.data.normal_(mean=0, std=math.sqrt((2 / (in_features + out_features))))
m.bias.data.zero_()
# return m with weight normalization
return nn.utils.weight_norm(m)
class DualCNN(nn.Module):
def __init__(self, args):
super(DualCNN, self).__init__()
self.emb_size = args.emb_size
self.s_kernel_size = args.s_kernel_size
self.s_padding = self.s_kernel_size - 1
self.w_kernel_size = args.w_kernel_size
self.w_padding = self.w_kernel_size - 1
self.s_num_layers = args.s_num_layers
self.w_num_layers = args.w_num_layers
self.s_max = args.s_max
self.w_max = args.w_max
self.feat_size = args.feat_size
self.proj_size = args.proj_size
self.num_boxes = args.num_boxes
self.idx2word = args.idx2word
self.word2idx = args.word2idx
self.vocab_len = len(self.idx2word)
self.padding_idx = args.pad_idx
self.softmax = nn.Softmax(-1)
self.project = nn.Sequential(
Linear(self.feat_size, self.proj_size),
nn.ReLU()
)
self.sCNN = nn.ModuleList()
self.wCNN = nn.ModuleList()
self.device = args.device
for i in range(self.s_num_layers):
i_size = self.proj_size + self.emb_size if i == 0 else self.emb_size
o_size = 2 * self.emb_size
self.sCNN.append(Conv1d(i_size, o_size, self.s_kernel_size, self.s_padding))
for i in range(self.w_num_layers):
self.wCNN.append(Conv1d(self.emb_size, 2 * self.emb_size, self.w_kernel_size, self.w_padding))
self.gen_topic = nn.Sequential(
Linear(self.emb_size, self.emb_size, 'relu'),
nn.ReLU(),
Linear(self.emb_size, self.emb_size, 'relu'),
nn.ReLU()
)
self.w_res_proj = Linear(2 * self.emb_size, self.emb_size)
self.s_res_proj = Linear(self.emb_size + self.proj_size, self.emb_size)
# self.sent_proj = nn.Sequential(
# Linear(self.emb_size, self.emb_size, 'relu'),
# nn.ReLU()
# )
self.stop_classifier = Linear(self.emb_size, 2)
self.classifier = Linear(self.emb_size, self.vocab_len)
self.embedding = Embedding(self.vocab_len, self.emb_size, self.padding_idx, self.word2idx)
# self.ln2 = nn.LayerNorm([self.s_max, self.emb_size])
# self.ln3 = nn.LayerNorm([self.emb_size])
# self.ln4 = nn.LayerNorm([self.s_max, self.emb_size])
def forward(self, img_feats, para_words, words_mask, fake_words, fake_words_mask):
real_bsz = para_words.size(0)
proj_feats = self.project(img_feats)
pool_feats = torch.max(proj_feats, 1)[0].unsqueeze(1) / math.sqrt(2)
label_words_embed = self.embedding(para_words)
# ignore these, just for try
# fake_words_embed = self.embedding(fake_words)[:, :-1, 1:, :]
# fake_sents = torch.max(fake_words_embed, 2)[0]
pred_words = label_words_embed.new_zeros(real_bsz, self.s_max, self.w_max, self.vocab_len)
# <start> token embedding
start_words = torch.zeros((real_bsz, 1), dtype=torch.long, device=self.device)
start_words_embed = self.embedding(start_words)
sents_embed = torch.max(label_words_embed[:, :-1, 1:, :], 2)[0]
sents_embed = torch.cat([start_words_embed, sents_embed], 1) * math.sqrt(2)
# sents_embed = self.ln2(sents_embed)
s_inputs = pool_feats.expand(real_bsz, self.s_max, self.proj_size)
s_inputs = torch.cat([s_inputs, s_inputs.new_zeros(real_bsz, self.s_max, self.emb_size)], 2)
s_inputs[:, :, self.proj_size:] = sents_embed
s_inputs = s_inputs.transpose(2, 1)
for i, conv in enumerate(self.sCNN):
res = s_inputs if i != 0 else self.s_res_proj(s_inputs.transpose(2, 1)).transpose(2, 1)
s_outputs = conv(s_inputs)
s_outputs = s_outputs[:, :, :-self.s_padding]
s_outputs = F.glu(s_outputs, dim=1)
s_outputs = self.ln3(s_outputs.transpose(2, 1)).transpose(2, 1)
att_res = s_outputs
att_weight = torch.bmm(proj_feats, s_outputs)
att_weight = att_weight / math.sqrt(self.emb_size)
att_weight = nn.Softmax(1)(att_weight)
s_outputs = torch.bmm(proj_feats.permute(0, 2, 1), att_weight)
s_outputs = (att_res + s_outputs) * math.sqrt(.5)
s_outputs = (s_outputs + res) * math.sqrt(.5)
s_inputs = s_outputs
s_outputs = s_outputs.transpose(2, 1)
sent_topics = self.gen_topic(s_outputs)
pred_stop = self.stop_classifier(sent_topics)
# wCNN
for sent_id in range(self.s_max):
topic = sent_topics[:, sent_id:sent_id + 1, :]
words = label_words_embed[:, sent_id, :-1, :]
w_inputs = torch.cat([topic, words], 1).transpose(2, 1)
for i, conv in enumerate(self.wCNN):
res = w_inputs
w_outputs = conv(w_inputs)
w_outputs = w_outputs[:, :, :-self.w_padding]
w_outputs = F.glu(w_outputs, dim=1)
w_outputs = (w_outputs + res) * math.sqrt(.5)
w_inputs = w_outputs
w_outputs = w_outputs.transpose(2, 1)
pred_words[:, sent_id, :, :] = self.classifier(w_outputs[:, 1:, :])
return pred_words, pred_stop
def beam_search(self, img_feats, beam_size=2):
img_feats = img_feats.view(-1, self.num_boxes, img_feats.size(-1))
real_bts = img_feats.size(0)
proj_feats = self.project(img_feats)
pool_feats = torch.max(proj_feats, 1)[0].unsqueeze(1) / math.sqrt(2)
result = ['' for _ in range(real_bts)]
global_s_inputs = pool_feats.expand(real_bts, self.s_max, self.proj_size)
global_s_inputs = torch.cat([global_s_inputs, global_s_inputs.new_zeros(real_bts, self.s_max, self.emb_size)],
2)
global_s_inputs = global_s_inputs.transpose(2, 1)
start_words = torch.zeros((real_bts, 1), dtype=torch.long, device=self.device)
start_words = self.embedding(start_words)
for sent_id in range(self.s_max):
if sent_id == 0:
prev_sents_embed = start_words.squeeze(1)
global_s_inputs[:, self.proj_size:, sent_id] = prev_sents_embed
s_inputs = global_s_inputs.clone()
for i, conv in enumerate(self.sCNN):
res = s_inputs if i != 0 else self.s_res_proj(s_inputs.transpose(2, 1)).transpose(2, 1)
s_outputs = conv(s_inputs)
s_outputs = s_outputs[:, :, :-self.s_padding]
s_outputs = F.glu(s_outputs, dim=1)
s_outputs = self.ln3(s_outputs.transpose(2, 1)).transpose(2, 1)
att_res = s_outputs
att_weight = torch.bmm(proj_feats, s_outputs)
att_weight = att_weight / math.sqrt(self.emb_size)
att_weight = nn.Softmax(1)(att_weight)
s_outputs = torch.bmm(proj_feats.permute(0, 2, 1), att_weight)
s_outputs = (s_outputs + att_res) * math.sqrt(.5)
s_outputs = (s_outputs + res) * math.sqrt(.5)
s_inputs = s_outputs
s_outputs = s_outputs.transpose(2, 1)
sent_topics = self.gen_topic(s_outputs)
# for i in range(5):
# print (torch.dist(sent_topics[0, i+1, :], sent_topics[0, 0, :]))
pred_stops = self.stop_classifier(sent_topics)
topics = sent_topics[:, sent_id:sent_id + 1, :]
global_inputs = torch.cat([topics, start_words,
start_words.new_zeros((real_bts, self.w_max - 1, self.emb_size))], 1)
global_inputs = global_inputs.transpose(2, 1)
res_words_idx = torch.ones((real_bts, self.w_max), dtype=torch.int)
state = [[res_words_idx.clone(), global_inputs.clone(), [0] * real_bts] for _ in range(beam_size)]
for j in range(1, self.w_max + 1):
tmp = [[{'log_prob': 0, 'cur_idx': 0, 'prev_idx': 0} for _ in range(beam_size * beam_size)] for _ in
range(real_bts)]
for beam_i in range(beam_size):
w_inputs = state[beam_i][1]
for i, conv in enumerate(self.wCNN):
res = w_inputs
w_outputs = conv(w_inputs)
w_outputs = w_outputs[:, :, :-self.w_padding]
w_outputs = F.glu(w_outputs, dim=1)
w_outputs = (w_outputs + res) * math.sqrt(.5)
w_inputs = w_outputs
now_words = self.classifier(w_outputs[:, :, j])
now_words = self.softmax(now_words).cpu()
for beam_j in range(beam_size - 1, -1, -1):
j_th_max = torch.kthvalue(-now_words, beam_size - beam_j)
jth_prob = -j_th_max[0]
jth_idx = j_th_max[1]
for k in range(real_bts):
tmp[k][beam_i + beam_j * beam_size]['log_prob'] = state[beam_i][2][k] + math.log(
float(jth_prob[k]) + 1e-9)
tmp[k][beam_i + beam_j * beam_size]['cur_idx'] = int(jth_idx[k])
tmp[k][beam_i + beam_j * beam_size]['prev_idx'] = beam_i
new_state = [[res_words_idx.clone(), global_inputs.clone(), [0] * real_bts] for _ in range(beam_size)]
for k in range(real_bts):
tmp[k].sort(key=lambda p: p['log_prob'], reverse=True)
if j == 1:
for o in range(1, beam_size):
tmp[k][o], tmp[k][o + beam_size * o] = tmp[k][o + beam_size * o], tmp[k][o]
for beam_i in range(beam_size):
prev_beam_idx = tmp[k][beam_i]['prev_idx']
new_state[beam_i][0][k] = state[prev_beam_idx][0][k].clone()
new_state[beam_i][0][k, j - 1] = tmp[k][beam_i]['cur_idx']
new_state[beam_i][2][k] = tmp[k][beam_i]['log_prob']
if j != self.w_max:
new_state[beam_i][1][k] = state[prev_beam_idx][1][k].clone()
new_state[beam_i][1][k, :, j + 1] = self.embedding(
new_state[beam_i][0][k, j - 1].long().to(self.device))
state = new_state
res_words_idx = [list(i) for i in list(state[0][0].numpy())]
prev_sents = torch.LongTensor(res_words_idx).to(self.device)
res_words = [[self.idx2word[int(j)] for j in i] for i in res_words_idx]
prev_mask_sum = []
for k in range(real_bts):
end_pos = res_words[k].index('<eos>') if '<eos>' in res_words[k] else self.w_max + 1
prev_mask_sum.append(end_pos + 1)
prev_sents[k][end_pos + 1:] = 2
result[k] += ' '.join(res_words[k][:end_pos] + ['. '])
prev_sents_embed = self.embedding(prev_sents)
prev_mask_sum = torch.Tensor(prev_mask_sum).to(self.device).unsqueeze(1)
prev_sents_embed = torch.max(prev_sents_embed, 1)[0]
stop_flag = self.softmax(self.stop_classifier(sRNN_output.squeeze(0)))
if float(stop_flag[0, 1]) >= 0.5:
break
return [r.strip() for r in result]
def sample(self, img_feats):
img_feats = img_feats.view(-1, self.num_boxes, img_feats.size(-1))
real_bts = img_feats.size(0)
proj_feats = self.project(img_feats)
pool_feats = torch.max(proj_feats, 1)[0].view(real_bts, 1, self.proj_size)
result = ['' for _ in range(real_bts)]
s_inputs = pool_feats.expand(real_bts, self.s_max, self.emb_size)
s_inputs = s_inputs.transpose(2, 1)
for i, conv in enumerate(self.sCNN):
if i != 0:
att_weight = torch.bmm(proj_feats, self.att_proj(s_inputs.transpose(2, 1)).transpose(2, 1))
att_weight = nn.Softmax(1)(att_weight)
s_inputs = torch.bmm(proj_feats.permute(0, 2, 1), att_weight) + s_inputs
res = s_inputs
s_outputs = conv(s_inputs)
s_outputs = s_outputs[:, :, :-self.s_padding]
s_outputs = F.glu(s_outputs, dim=1)
s_outputs = (s_outputs + res) * math.sqrt(0.5)
s_inputs = s_outputs
s_outputs = s_outputs.transpose(2, 1)
sent_topics = self.gen_topic(s_outputs)
pred_stops = self.stop_classifier(sent_topics)
for sent_id in range(self.s_max):
topics = sent_topics[:, sent_id:sent_id + 1, :]
start_words = torch.zeros((real_bts, 1), dtype=torch.long, device=self.device)
start_words = self.embedding(start_words)
global_inputs = torch.cat([topics, start_words,
start_words.new_zeros(real_bts, self.w_max - 1, self.emb_size)], 1)
global_inputs = global_inputs.transpose(2, 1)
res_words_idx = torch.zeros((real_bts, self.w_max), dtype=torch.int)
for j in range(1, self.w_max + 1):
w_inputs = global_inputs
for i, conv in enumerate(self.wCNN):
res = w_inputs
w_outputs = conv(w_inputs)
w_outputs = w_outputs[:, :, :-self.w_padding]
w_outputs = F.glu(w_outputs, dim=1)
w_outputs = (w_outputs + res) * math.sqrt(.5)
w_inputs = w_outputs
now_words = self.classifier(w_outputs[:, :, j])
now_words = torch.max(now_words, 1)[1]
res_words_idx[:, j - 1] = now_words
now_words = self.embedding(now_words)
if j != self.w_max:
global_inputs[:, :, j + 1] = now_words
res_words_idx = [list(i) for i in list(res_words_idx.numpy())]
res_words = [[self.idx2word[int(j)] for j in i] for i in res_words_idx]
for k in range(real_bts):
end_pos = res_words[k].index('<eos>') if '<eos>' in res_words[k] else self.w_max + 1
result[k] += ' '.join(res_words[k][:end_pos] + ['. '])
stop_flag = self.softmax(self.stop_classifier(sRNN_output.squeeze(0)))
if float(stop_flag[0, 1]) >= 0.5:
break
return [r.strip() for r in result]
net = DualCNN()
param = net.parameters()
for i in param:
print (i.size())