-
Notifications
You must be signed in to change notification settings - Fork 5
/
model.py
427 lines (354 loc) · 23.7 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
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
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
import tensorflow as tf
import numpy as np
from dnn_library import *
from nets import *
import pdb
from attention import *
from text_encoder import *
class STT(object):
"""
Base class for Cross-Modal Retrieval experiments
"""
def __init__(self, base='inception_v1', vocab_file='/shared/kgcoe-research/mil/peri/mscoco_data/mscoco_1024d_2gru/vocab_mscoco.enc', margin=1., embedding_dim=512,word_dim=1024, vocab_size=26735):
self.base_arch = base
self.margin=margin
self.vocab_file = vocab_file
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.word_dim = word_dim
def process_text(self):
"""
Loads the vocabulary and builds the vocab to idx tables.
Builds the embedding matrix given the vocab and hidden dimension size
"""
self.vocab_table = tf.contrib.lookup.index_table_from_file(self.vocab_file, default_value=0)
self.reverse_vocab_table = tf.contrib.lookup.index_to_string_table_from_file(self.vocab_file)
self.embedding_matrix = tf.get_variable("embeddings/embedding_share", shape=[self.vocab_size, self.word_dim], \
trainable=True, \
initializer=tf.random_uniform_initializer(minval=-0.1, maxval=0.1), \
dtype=tf.float32)
def _feature_extractor(self, image, reuse=None, is_training=True):
"""
Builds the model architecture
"""
# Define the network and pass the input image
with tf.variable_scope('Feature_extractor', reuse=reuse) as scope:
with slim.arg_scope(model[self.base_arch]['scope']):
logits, end_points = model[self.base_arch]['net'](image, num_classes=model[self.base_arch]['num_classes'], is_training=is_training)
# Avg pool features of inception v1 (size: 1024)
feat_anchor = end_points[model[self.base_arch]['end_point']] ## Dropout_0b
feat_anchor = tf.squeeze(end_points[model[self.base_arch]['end_point']])
return feat_anchor
def build_rnn_cell(self, num_units, num_layers, dropout=0.):
"""
Define the RNN cell with num_units, num_layers and dropout (Dropout=0 for inference)
"""
cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.DropoutWrapper(cell=tf.contrib.rnn.GRUCell(num_units, \
kernel_initializer=tf.orthogonal_initializer(),
reuse=tf.AUTO_REUSE), \
input_keep_prob=(1.0 - dropout)) for _ in range(num_layers)])
return cell
def _projection_layer(self):
"""
Builds the projection layer
"""
# Get the projection layer
with tf.variable_scope("build_network"):
with tf.variable_scope("decoder/output_projection"):
projection_layer = tf.layers.Dense(self.vocab_size, use_bias=False, name="output_projection")
return projection_layer
def _build_encoder(self, input_seq, num_units, num_layers, seq_len, dropout=0.):
"""
Builds the encoder part of seq-seq model
"""
# Look up the ids for input sequence
input_word_ids = self.vocab_table.lookup(input_seq)
# Time major
input_word_ids = tf.transpose(input_word_ids)
# Define the encoder cell and build input embeddings
self.cell = self.build_rnn_cell(num_units, num_layers, dropout=dropout)
encoder_emb_inp = tf.nn.embedding_lookup(self.embedding_matrix, input_word_ids)
with tf.variable_scope("dynamic_seq2seq", dtype=tf.float32):
with tf.variable_scope("encoder") as scope:
encoder_outputs, encoder_state = tf.nn.dynamic_rnn(
self.cell,
encoder_emb_inp,
dtype=tf.float32,
sequence_length=seq_len,
time_major=True,
swap_memory=True)
return encoder_outputs, encoder_state
def _build_decoder(self, decoder_initial_state, decoder_input, dec_len, params, phase='train', reuse=tf.AUTO_REUSE, cell=None):
"""
Builds the decoder part of seq-seq model
"""
with tf.variable_scope("dynamic_seq2seq", dtype=tf.float32, reuse=reuse):
with tf.variable_scope("decoder") as decoder_scope:
# Define RNN cell
if cell is not None:
cell=cell
else:
cell = self.build_rnn_cell(params.num_units, params.num_layers, dropout=params.dropout)
if phase=='train':
# Look up the ids for input sequence
decoder_word_ids = self.vocab_table.lookup(decoder_input)
# Time major
decoder_word_ids = tf.transpose(decoder_word_ids)
# Look up embedding vectors for decoder inputs
decoder_emb_inp = tf.nn.embedding_lookup(self.embedding_matrix, decoder_word_ids)
# Helper
helper = tf.contrib.seq2seq.TrainingHelper(decoder_emb_inp, dec_len, time_major=True)
# Decoder
my_decoder = tf.contrib.seq2seq.BasicDecoder(cell, helper, decoder_initial_state)
# Dynamic decoding
outputs, final_context_state, _ = tf.contrib.seq2seq.dynamic_decode(my_decoder, output_time_major=True, swap_memory=True, scope=decoder_scope)
# Get the logits from the projection layer
projection_layer = self._projection_layer()
logits = projection_layer(outputs.rnn_output)
sample_id = outputs.sample_id
elif phase=='inference':
# Look up for start and end tokens
start_token_id = tf.cast(self.vocab_table.lookup(tf.constant('<s>')), tf.int32)
end_token = tf.cast(self.vocab_table.lookup(tf.constant('</s>')), tf.int32)
start_tokens = tf.fill([100], start_token_id)
# Greedy decoding
helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(self.embedding_matrix, start_tokens, end_token)
# Decoder
projection_layer=self._projection_layer()
my_decoder = tf.contrib.seq2seq.BasicDecoder(cell, helper, decoder_initial_state, output_layer=projection_layer)
# Dynamic decoding
outputs, final_context_state, _ = tf.contrib.seq2seq.dynamic_decode(my_decoder,
maximum_iterations=20,
output_time_major=True,
swap_memory=True,
scope=decoder_scope)
logits = outputs.rnn_output
sample_id = outputs.sample_id
return logits, sample_id, final_context_state
def _build_embedding(self, feat_anchor, embedding_dim=512, scope_name="embedding", act_fn=None, reuse=None):
"""
Build the embedding network
"""
with slim.arg_scope([slim.fully_connected],
activation_fn=act_fn,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.0002),
reuse=reuse):
embedding = slim.fully_connected(feat_anchor, embedding_dim, activation_fn=act_fn, scope=scope_name)
return embedding
def build_stt_model(self, images, encoder_captions, decoder_captions, enc_len, dec_len, params, reuse=None):
"""
Builds the Show, Translate and Tell model
"""
# Build the embeddings for images.
if not params.precompute:
image_features = self._feature_extractor(images, is_training=params.finetune_with_cnn, reuse=reuse)
else:
image_features = images
# Build the embeddings for images, encoder_captions
image_embeddings = self._build_embedding(image_features, self.embedding_dim, act_fn=None, reuse=reuse, scope_name='image_embedding')
# Build the encoder and its embeddings
self.process_text()
encoder_outputs, encoder_state = self._build_encoder(encoder_captions, params.num_units, params.num_layers, enc_len, dropout=params.dropout)
encoder_features = tf.concat(encoder_state, axis=1, name='encoder_features')
text_embeddings = self._build_embedding(encoder_features, self.embedding_dim, act_fn=None, reuse=reuse, scope_name='text_embedding')
# L2 normalize the embeddings
image_embeddings = tf.nn.l2_normalize(image_embeddings, axis=1, name='norm_image_embeddings')
text_embeddings = tf.nn.l2_normalize(text_embeddings, axis=1, name='norm_text_embeddings')
dec_image_logits, dec_sent_logits= None, None
if params.mode =='val':
phase = 'inference'
else:
phase='train'
# Build the decoder with encoder text embeddings
text_embed_split = tuple(tf.split(text_embeddings, num_or_size_splits=params.num_layers, axis=1))
dec_sent_logits, dec_sent_sample_id, sent_context_state = self._build_decoder(text_embed_split, decoder_captions, dec_len, params, phase=phase)
# Build the decoder with image embeddings
image_embed_split = tuple(tf.split(image_embeddings, num_or_size_splits=params.num_layers, axis=1))
dec_image_logits, dec_im_sample_id, image_context_state = self._build_decoder(image_embed_split, decoder_captions, dec_len, params, phase=phase, reuse=True)
if params.mode=='val':
im_pred_words = self.reverse_vocab_table.lookup(tf.cast(dec_im_sample_id, tf.int64))
sent_pred_words = self.reverse_vocab_table.lookup(tf.cast(dec_sent_sample_id, tf.int64))
return image_embeddings, text_embeddings, im_pred_words, sent_pred_words
return image_embeddings, text_embeddings, dec_image_logits, dec_sent_logits
def build_stt_attention_model(self, images, encoder_captions, decoder_captions, enc_len, dec_len, params, reuse=None):
"""
Builds the Show, Translate and Tell model
"""
# Build the embeddings for images.
if not params.precompute:
image_features = self._feature_extractor(images, is_training=params.finetune_with_cnn, reuse=reuse)
else:
image_features = images
# Build the embeddings for images, encoder_captions
image_embeddings = self._build_embedding(image_features, self.embedding_dim, act_fn=None, reuse=reuse, scope_name='image_embedding')
# Avg pool the regions
mean_image_embeddings = tf.reduce_mean(image_embeddings, axis=1)
mean_norm_image_embeddings = tf.nn.l2_normalize(mean_image_embeddings, axis=1)
image_embeddings = tf.nn.l2_normalize(image_embeddings, axis=2, name='norm_image_reg_embeddings')
#Build the embeddings for captions
self.process_text()
# Look up the ids for input sequence
input_word_ids = self.vocab_table.lookup(encoder_captions)
encoder_embeddings = tf.nn.embedding_lookup(self.embedding_matrix, input_word_ids)
scan = SCAN()
word_features, text_features = scan._build_text_encoder(encoder_embeddings, params, enc_len)
word_features = tf.nn.l2_normalize(word_features, axis=2)
sim_matrix = t2i_attention(image_embeddings, word_features, enc_len, params)
text_features = tf.nn.l2_normalize(text_features, axis=1)
input_word_ids = self.vocab_table.lookup(encoder_captions)
encoder_embeddings = tf.nn.embedding_lookup(self.embedding_matrix, input_word_ids)
scan = SCAN()
word_features, text_features = scan._build_text_encoder(encoder_embeddings, params, enc_len)
dec_image_logits, dec_sent_logits= None, None
if params.mode =='val':
phase = 'inference'
else:
phase='train'
# Build the decoder with encoder text embeddings
text_embed_split = tuple(tf.split(text_features, num_or_size_splits=params.num_layers, axis=1))
dec_sent_logits, dec_sent_sample_id, sent_context_state = self._build_decoder(text_embed_split, decoder_captions, dec_len, params, phase=phase)
# Build the decoder with image embeddings
image_embed_split = tuple(tf.split(mean_norm_image_embeddings, num_or_size_splits=params.num_layers, axis=1))
dec_image_logits, dec_im_sample_id, image_context_state = self._build_decoder(image_embed_split, decoder_captions, dec_len, params, phase=phase)
if params.mode=='val':
im_pred_words = self.reverse_vocab_table.lookup(tf.cast(dec_im_sample_id, tf.int64))
sent_pred_words = self.reverse_vocab_table.lookup(tf.cast(dec_sent_sample_id, tf.int64))
return image_embeddings, word_features, im_pred_words, sent_pred_words
return mean_norm_image_embeddings, text_features, dec_image_logits, dec_sent_logits, sim_matrix
def build_stt_att_t2t_model(self, images, encoder_captions, decoder_captions, enc_len, dec_len, params, reuse=None):
"""
Builds the Show, Translate and Tell model
"""
# Build the embeddings for images.
if not params.precompute:
image_features = self._feature_extractor(images, is_training=params.finetune_with_cnn, reuse=reuse)
else:
image_features = images
# Build the embeddings for images, encoder_captions
image_embeddings = self._build_embedding(image_features, self.embedding_dim, act_fn=None, reuse=reuse, scope_name='image_embedding')
# Avg pool the regions
mean_image_embeddings = tf.reduce_mean(image_embeddings, axis=1)
mean_norm_image_embeddings = tf.nn.l2_normalize(mean_image_embeddings, axis=1)
image_embeddings = tf.nn.l2_normalize(image_embeddings, axis=2, name='norm_image_reg_embeddings')
#Build the embeddings for captions
self.process_text()
# Look up the ids for input sequence
input_word_ids = self.vocab_table.lookup(encoder_captions)
encoder_embeddings = tf.nn.embedding_lookup(self.embedding_matrix, input_word_ids)
scan = SCAN()
word_features, text_features = scan._build_text_encoder(encoder_embeddings, params, enc_len)
word_features = tf.nn.l2_normalize(word_features, axis=2)
sim_matrix = t2i_attention(image_embeddings, word_features, enc_len, params)
text_features = tf.nn.l2_normalize(text_features, axis=1)
# For paraphrases
paraphrase_word_ids = self.vocab_table.lookup(decoder_captions)
paraphrase_embeddings = tf.nn.embedding_lookup(self.embedding_matrix, paraphrase_word_ids)
para_word_features, para_text_features = scan._build_text_encoder(paraphrase_embeddings, params, dec_len)
para_text_features = tf.nn.l2_normalize(para_text_features, axis=1)
# para_word_features = tf.nn.l2_normalize(para_word_features, axis=2)
# sim_para_matrix = t2i_attention(image_embeddings, para_word_features, dec_len, params)
sim_para_matrix=None
# Text to text similarity
t2t_matrix = tf.matmul(text_features, para_text_features, transpose_b=True)
dec_image_logits, dec_sent_logits= None, None
if params.mode =='val':
phase = 'inference'
else:
phase='train'
# Build the decoder with encoder text embeddings
text_embed_split = tuple(tf.split(text_features, num_or_size_splits=params.num_layers, axis=1))
dec_sent_logits, dec_sent_sample_id, sent_context_state = self._build_decoder(text_embed_split, decoder_captions, dec_len, params, phase=phase)
# Build the decoder with image embeddings
image_embed_split = tuple(tf.split(mean_norm_image_embeddings, num_or_size_splits=params.num_layers, axis=1))
dec_image_logits, dec_im_sample_id, image_context_state = self._build_decoder(image_embed_split, decoder_captions, dec_len, params, phase=phase)
if params.mode=='val':
im_pred_words = self.reverse_vocab_table.lookup(tf.cast(dec_im_sample_id, tf.int64))
sent_pred_words = self.reverse_vocab_table.lookup(tf.cast(dec_sent_sample_id, tf.int64))
return image_embeddings, word_features, text_features, para_word_features, para_text_features, im_pred_words, sent_pred_words
return mean_norm_image_embeddings, text_features, para_text_features, dec_image_logits, dec_sent_logits, sim_matrix, sim_para_matrix, t2t_matrix
def get_max_time(self, tensor):
return tensor.shape[0].value or tf.shape(tensor)[0]
def stt_loss(self, image_embeddings, text_embeddings, dec_image_logits, dec_sent_logits, decoder_target_caption, dec_len, args, sim_scores=None, sim_para_scores=None, t2t_scores=None):
"""
Loss for Show, Translate and Tell model
"""
# Look up the ids for target sequence
output_word_ids = self.vocab_table.lookup(decoder_target_caption)
# Time major
output_word_ids = tf.transpose(output_word_ids)
sim_loss, sent_loss, image_loss = self.sim_loss(image_embeddings, text_embeddings, args, sim_scores=sim_scores)
t2t_loss=None
sim_para_loss=None
if sim_para_scores is not None:
sim_para_loss, _, _ = self.sim_loss(image_embeddings, text_embeddings, args, sim_scores=sim_para_scores)
if t2t_scores is not None:
t2t_loss, _, _ = self.sim_loss(image_embeddings, text_embeddings, args, sim_scores=t2t_scores)
image_captioning_ce = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=output_word_ids, logits=dec_image_logits)
seq2seq_ce = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=output_word_ids, logits=dec_sent_logits)
# Use target sequence mask to filter out outputs of padded words
max_time = self.get_max_time(output_word_ids)
target_weights = tf.sequence_mask(dec_len, max_time, dtype=dec_image_logits.dtype)
target_weights = tf.transpose(target_weights)
batch_size=image_embeddings.shape.as_list()[0]
image_captioning_loss = tf.reduce_sum(image_captioning_ce * target_weights) / tf.to_float(batch_size)
seq2seq_loss = tf.reduce_sum(seq2seq_ce * target_weights) / tf.to_float(batch_size)
return seq2seq_loss, image_captioning_loss, sim_loss, sim_para_loss, t2t_loss
def sim_loss(self, image_embeddings, text_embeddings, args, sim_scores=None):
"""
Order violation or cosine similarity loss for image and text embeddings
"""
with tf.name_scope('Sim_Loss') as scope:
if sim_scores is None:
if args.use_abs:
image_embeddings = tf.abs(image_embeddings)
text_embeddings = tf.abs(text_embeddings)
if args.measure=='cosine':
sim_scores = tf.matmul(image_embeddings, text_embeddings, transpose_b=True, name='sim_score')
elif args.measure=='order':
# refer to eqn in paper or code of http://openaccess.thecvf.com/content_cvpr_2018/papers/Wehrmann_Bidirectional_Retrieval_Made_CVPR_2018_paper.pdf
im_emb = tf.expand_dims(image_embeddings, 0) # 1x128x2048
text_emb = tf.expand_dims(text_embeddings, 1) # 128x1x2048
im_emb = tf.tile(im_emb, [image_embeddings.shape.as_list()[0], 1, 1]) # 128x128x2048 (Each row has 128x2048 im_emb)
text_emb = tf.tile(text_emb, [1, text_embeddings.shape.as_list()[0], 1]) # 128x128x2048 (Each row has its text emb replicated 128 times)
sqr_diff = tf.square(tf.maximum(text_emb - im_emb, 0.))
sqr_diff_sum = tf.squeeze(tf.reduce_sum(sqr_diff, 2))
sim_scores = -tf.transpose(tf.sqrt(sqr_diff_sum), name='order_sim_scores') # Note the negative sign as this is a distance metric
# Get the diagonal of the matrix
sim_diag = tf.expand_dims(tf.diag_part(sim_scores), 0, name='sim_diag')
sim_diag_tile = tf.tile(sim_diag, multiples=[sim_diag.shape.as_list()[1], 1], name='sim_diag_tile')
sim_diag_transpose = tf.transpose(sim_diag, name='sim_diag_transpose')
sim_diag_tile_transpose = tf.tile(sim_diag_transpose, multiples=[1, sim_diag.shape.as_list()[1]], name='sim_diag_tile_transpose')
# compare every diagonal score to scores in its column
# caption retrieval
loss_s = tf.maximum(args.margin + sim_scores - sim_diag_tile_transpose, 0.)
# compare every diagonal score to scores in its row
# image retrieval
loss_im = tf.maximum(args.margin + sim_scores - sim_diag_tile, 0.)
# clear the costs for diagonal elements
mask = tf.eye(loss_s.shape.as_list()[0], dtype=tf.bool, name='Mask')
mask_not = tf.cast(tf.logical_not(mask, name='mask_not'), tf.float32)
neg_s_loss = tf.multiply(loss_s, mask_not, name='neg_s_loss')
neg_im_loss = tf.multiply(loss_im, mask_not, name='neg_im_loss')
# Mining the hardest negative for each sample
if args.mine_n_hard>0:
if args.mine_n_hard==1:
loss_s = tf.reduce_max(neg_s_loss, axis=1)
loss_im = tf.reduce_max(neg_im_loss, axis=0)
else:
loss_s = tf.contrib.framework.sort(neg_s_loss, axis=1, direction='DESCENDING')
loss_im = tf.contrib.framework.sort(neg_im_loss, axis=0, direction='DESCENDING')
# Build the index matrix to gather_nd
batch_size=loss_s.shape.as_list()[0]
indices= np.zeros((batch_size, mine_n_hard, 2))
for it in range(batch_size):
for m in range(mine_n_hard):
indices[it][m][0] = it
indices[it][m][1] = m
# Get the top N distances and reduce them
loss_s = tf.gather_nd(loss_s, indices.astype(np.int32))
loss_im = tf.gather_nd(loss_im, indices.astype(np.int32))
loss_s = tf.reduce_sum(loss_s, name='loss_s')
loss_im = tf.reduce_sum(loss_im, name='loss_im')
total_loss = loss_s + loss_im
return total_loss, loss_s, loss_im