-
Notifications
You must be signed in to change notification settings - Fork 0
/
Solver.py
432 lines (384 loc) · 23.7 KB
/
Solver.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
__author__ = 'antony'
import config
import tensorflow as tf
import numpy as np
import time,math,utils,json,os
import shutil,sys
from visulize_attention import *
from new_data_loader import VQADataLoader
from pairwaise_dataloader import VQADataLoader_pair
from utils import visualize_failures
sys.path.append(config.VQA_TOOLS_PATH)
sys.path.append(config.VQA_EVAL_TOOLS_PATH)
from vqaTools.vqa import VQA
from vqaEvaluation.vqaEval import VQAEval
from scipy import *
class Solver(object):
def __init__(self,model):
self.model=model
self.global_epoch_step=0
self.lr=tf.placeholder(tf.float32)
self.global_step=tf.Variable(0,name='global_step',trainable=False)
with tf.variable_scope(config.VQA_SCOPE):
self.vqa_optimizer=tf.train.AdamOptimizer(self.lr)
def build_train_ops(self):
learning_rate=config.VQA_LR
def _learning_rate_decay_fn(learning_rate,global_step):
return tf.train.exponential_decay(learning_rate=learning_rate,global_step=global_step,
decay_steps=config.NUM_STEPS_PER_DECAY,
decay_rate=config.DECAY_RATE,
staircase=True)
learning_rate_decay=_learning_rate_decay_fn
decay_lr = tf.train.exponential_decay(learning_rate=learning_rate,global_step=self.global_step,
decay_steps=config.NUM_STEPS_PER_DECAY,
decay_rate=config.DECAY_RATE,
staircase=True)
with tf.variable_scope('optimizer',reuse=tf.AUTO_REUSE):
self.vqa_optimizer=tf.train.AdamOptimizer(
learning_rate=config.VQA_LR
)
train_ops=tf.contrib.layers.optimize_loss(
loss=self.overall_loss,
global_step=self.global_step,
learning_rate=learning_rate,
optimizer=self.vqa_optimizer,
clip_gradients=config.CLIP_GRADIENT,
learning_rate_decay_fn=learning_rate_decay
)
return train_ops,decay_lr
def save_vqa_model(self,sess,scope,step,update=False):
path=config.FOLDER_NAME+'/models/'
saver=tf.train.Saver(tf.get_collection(tf.GraphKeys.VARIABLES,scope=scope))
print 'saving model to %s'%path
if update:
saver.save(sess,path+config.MODEL_SAVE_NAME)
else:
saver.save(sess,path+config.MODEL_SAVE_NAME,step)
def idx2answers(self,predicts):
return [self.dataloader.idx2ans[str(x)] for x in predicts]
def load_vqa_model(self,sess,scope):
path = config.FOLDER_NAME + '/models/'
saver=tf.train.Saver(tf.get_collection(tf.GraphKeys.VARIABLES,scope=scope))
saver.restore(sess,path+config.MODEL_LOAD_NAME)
def train(self):
#################### Basic Operations#########################
if config.POSITIVE_ONLY:
self.overall_loss=self.model.loss1
cross_entrophy=self.model.sigmoid_cross_entrophy1
accuracy=self.model.accuracy1
vqa_predict1=self.model.predict1
af_distance = self.model.distance
else:
self.overall_loss = self.model.overloss
cross_entrophy = self.model.sigmoid_cross_entrophy1+self.model.sigmoid_cross_entrophy2
accuracy = (self.model.accuracy1 + self.model.accuracy2)/2
vqa_predict1 = self.model.predict1
vqa_predict2 = self.model.predict2
af_distance = self.model.distance
vqa_distance = self.model.distance2
distance_loss = self.model.distance_loss
pn_distance = self.model.dpn
apn_distance = self.model.dapn
vqa_train_op,vqa_decay_lr=self.build_train_ops()
#################### Summary Setting #########################
val_average_loss=tf.placeholder('float')
val_average_accuracy=tf.placeholder('float')
train_ol_summary=tf.summary.scalar('train_overall_loss',self.overall_loss)
train_ce_summary=tf.summary.scalar('train_cross_entrophy',cross_entrophy)
train_acc_summary=tf.summary.scalar('accuracy',accuracy)
val_average_l_summary=tf.summary.scalar('val_ave_cross_entrophy',val_average_loss)
val_average_acc_summary=tf.summary.scalar('val_average_accuracy',val_average_accuracy)
#################### Session Setting#########################
sess=tf.Session(config=tf.ConfigProto(log_device_placement=False))
sess.run(tf.global_variables_initializer())
train_writer=tf.summary.FileWriter(config.log_path+'train',sess.graph)
valid_writer=tf.summary.FileWriter(config.log_path+'valid')
if config.USE_PRE_VQA_MODEL:
self.load_vqa_model(sess,config.VQA_SCOPE)
#################### Train Step Function#########################
dataloader=VQADataLoader(batchsize=config.BATCH_SIZE,mode='train')
dataloader_pairs = VQADataLoader_pair(batchsize=config.BATCH_SIZE, mode='train')
start_t = time.time()
best_result = [0.0, 0]
val_counter = 0
for i in range(config.MAX_TRAIN_STEP):
if config.POSITIVE_ONLY:
#data_time=time.time()
q_strs, q_word_vec_list, q_len_list, ans_vectors, img_features,a_word_vec,ans_space_score, t_qid_list, img_ids, epoch = dataloader.next_batch(config.BATCH_SIZE)
#print 'data provide time', time.time()-data_time
feed_dict = {self.model.q_input: q_word_vec_list, self.model.a_input: a_word_vec,self.model.ans1: ans_vectors,self.model.ans_space_score:ans_space_score,
self.model.seqlen: q_len_list, self.model.img_vec: img_features,
self.lr: config.VQA_LR, self.model.keep_prob: config.KEEP_PROB, self.model.is_training: True}
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
_,vqadecay_lr,global_step = sess.run([vqa_train_op,vqa_decay_lr,self.global_step], feed_dict=feed_dict)
to_summary, tc_summary, ta_summary, acc, o_loss, c_loss, predict1,af = sess.run(
[train_ol_summary, train_ce_summary, train_acc_summary, accuracy, self.overall_loss, cross_entrophy,
vqa_predict1,af_distance], options=run_options,
run_metadata=run_metadata, feed_dict=feed_dict)
else:
q_strs, q_word_vec_list, q_len_list, ans_vectors, img_features,a_word_vec,ans_space_score, q_strs2, q_word_vec_list2, q_len_list2, ans_vectors2, img_features2, a_word_vec2,t_qid_list, img_ids, epoch = dataloader_pairs.next_batch(config.BATCH_SIZE)
feed_dict = {self.model.q_input: q_word_vec_list, self.model.a_input: a_word_vec, self.model.ans1: ans_vectors,self.model.ans_space_score:ans_space_score,
self.model.seqlen: q_len_list, self.model.img_vec: img_features, self.model.a_input2: a_word_vec2,
self.lr: config.VQA_LR, self.model.keep_prob: config.KEEP_PROB,self.model.is_training: True,
self.model.ans2: ans_vectors2, self.model.img_vec2: img_features2}
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
_, vqadecay_lr, global_step = sess.run([vqa_train_op, vqa_decay_lr, self.global_step],
feed_dict=feed_dict)
to_summary, tc_summary, ta_summary, acc, o_loss, c_loss, predict1, predict2,af,vqad,d_loss,dpn,dapn = sess.run(
[train_ol_summary, train_ce_summary, train_acc_summary, accuracy, self.overall_loss, cross_entrophy,
vqa_predict1, vqa_predict2,af_distance,vqa_distance,distance_loss,pn_distance,apn_distance], options=run_options,
run_metadata=run_metadata, feed_dict=feed_dict)
train_writer.add_summary(to_summary, i)
train_writer.add_summary(tc_summary, i)
train_writer.add_summary(ta_summary, i)
if (i+1)%config.SHOW_STEP==0 or i ==1:
cost_t=time.time()-start_t
print '\n Train Process: Step: %d|%d, Overall Loss: %.7f, Sigmoid Cross Entrophy: %.7f, lr: %.7f, decay_lr: %.7f,accuracy: %.3f, time:%d s, distance1: %.7f, distance2: %.7f, distance_loss: %.7f, pn_distance: %.7f, apn_distance: %.7f' \
% (
global_step + 1, config.MAX_TRAIN_STEP, o_loss, c_loss, config.VQA_LR, vqadecay_lr, acc,
cost_t,
af, vqad, d_loss, dpn, dapn)
print 'Question: %s' % (q_strs[0])
answer = ans_vectors[0].argmax(axis=0)
print 'Answer:', dataloader.vec_to_answer(answer), answer
v_pred1 = predict1[0]
print 'predict:', dataloader.vec_to_answer(v_pred1), v_pred1
answer2 = ans_vectors2[0].argmax(axis=0)
print 'Answer2:', dataloader.vec_to_answer(answer2), answer2
v_pred2 = predict2[0]
print 'predict2:', dataloader.vec_to_answer(v_pred2), v_pred2
start_t = time.time()
fshow = file('show.txt', 'a+')
fshow.write(
'\n Train Process: Step: %d|%d, Overall Loss: %.7f, Sigmoid Cross Entrophy: %.7f, lr: %.7f, decay_lr: %.7f,accuracy: %.3f, time:%d s, distance1: %.7f, distance2: %.7f, distance_loss: %.7f, pn_distance: %.7f, apn_distance: %.7f' \
% (
global_step + 1, config.MAX_TRAIN_STEP, o_loss, c_loss, config.VQA_LR, vqadecay_lr, acc,
cost_t,
af, vqad, d_loss, dpn, dapn))
fshow.close()
if (i+1)%config.VAL_STEP==0:
val_ave_l,val_ave_acc,acc_per_question_type,acc_per_answer_type=self.exec_validation(sess,mode='val',it=i,folder=config.FOLDER_NAME)
vl_summary, va_summary = sess.run([val_average_l_summary, val_average_acc_summary],
feed_dict={val_average_loss: val_ave_l,
val_average_accuracy: val_ave_acc})
print 'the average validation losses is %.7f, average accuracy is %.5f' % \
(val_ave_l, val_ave_acc)
print 'per answer:',acc_per_answer_type
print 'per quetion type:',acc_per_question_type
valid_writer.add_summary(vl_summary, i)
valid_writer.add_summary(va_summary, i)
if val_ave_acc>best_result[0]:
print 'the previous max val acc is', best_result[0], 'at iter',best_result[1]
best_result[0]=val_ave_acc
best_result[1]=i
print 'now the best result is',best_result[0]
else:
print 'the best result is',best_result[0],'at iter',best_result[1]
val_counter+=1
f = file('result.txt','a+')
f.write('\n' + 'the step is %d,the average validation losses is %.7f, average accuracy is %.5f' % \
(i+1,val_ave_l, val_ave_acc)+'\n'+'per quetion type: %s'%(acc_per_answer_type))
f.close()
if val_counter>5:
self.save_vqa_model(sess,config.VQA_SCOPE,i+1)
break
if (i+1)%config.SAVE_STEP==0:
print 'saving model .......'
self.save_vqa_model(sess,config.VQA_SCOPE,i+1) # Save VQA Model
def exec_validation(self,sess,mode,folder, it=0, visualize=False):
dp = VQADataLoader(mode=mode, batchsize=config.VAL_BATCH_SIZE, folder=folder)
total_questions = len(dp.getQuesIds())
epoch = 0
pred_list = []
testloss_list = []
stat_list = []
while epoch == 0:
q_strs, q_word_vec_list, q_len_list, ans_vectors, img_features,a_word_vec,ans_score,ans_space_score, t_qid_list, img_ids, epoch = dp.next_batch(config.BATCH_SIZE)
feed_dict = {self.model.q_input: q_word_vec_list, self.model.ans1: ans_vectors,
self.model.seqlen: q_len_list, self.model.img_vec: img_features,
self.lr: config.VQA_LR, self.model.keep_prob: 1.0, self.model.is_training: False}
t_predict_list,predict_loss=sess.run([self.model.predict1, self.model.softmax_cross_entrophy1], feed_dict=feed_dict)
t_pred_str = [dp.vec_to_answer(pred_symbol) for pred_symbol in t_predict_list]
testloss_list.append(predict_loss)
ans_vectors=np.asarray(ans_vectors).argmax(1)
for qid, iid, ans, pred in zip(t_qid_list, img_ids, ans_vectors, t_pred_str):
# pred_list.append({u'answer':pred, u'question_id': int(dp.getStrippedQuesId(qid))})
pred_list.append((pred, int(dp.getStrippedQuesId(qid))))
if visualize:
q_list = dp.seq_to_list(dp.getQuesStr(qid))
if mode == 'test-dev' or 'test':
ans_str = ''
ans_list = [''] * 10
else:
ans_str = dp.vec_to_answer(ans)
ans_list = [dp.getAnsObj(qid)[i]['answer'] for i in xrange(10)]
stat_list.append({ \
'qid': qid,
'q_list': q_list,
'iid': iid,
'answer': ans_str,
'ans_list': ans_list,
'pred': pred})
percent = 100 * float(len(pred_list)) / total_questions
sys.stdout.write('\r' + ('%.2f' % percent) + '%')
sys.stdout.flush()
print 'Deduping arr of len', len(pred_list)
deduped = []
seen = set()
for ans, qid in pred_list:
if qid not in seen:
seen.add(qid)
deduped.append((ans, qid))
print 'New len', len(deduped)
final_list = []
for ans, qid in deduped:
final_list.append({u'answer': ans, u'question_id': qid})
mean_testloss = np.array(testloss_list).mean()
if mode == 'val':
valFile = './%s/val2015_resfile_%d' % (folder,it)
with open(valFile, 'w') as f:
json.dump(final_list, f)
if visualize:
visualize_failures(stat_list, mode)
annFile = config.DATA_PATHS['val']['ans_file']
quesFile = config.DATA_PATHS['val']['ques_file']
vqa = VQA(annFile, quesFile)
vqaRes = vqa.loadRes(valFile, quesFile)
vqaEval = VQAEval(vqa, vqaRes, n=2)
vqaEval.evaluate()
acc_overall = vqaEval.accuracy['overall']
acc_perQuestionType = vqaEval.accuracy['perQuestionType']
acc_perAnswerType = vqaEval.accuracy['perAnswerType']
return mean_testloss, acc_overall, acc_perQuestionType, acc_perAnswerType
elif mode == 'test-dev':
filename = './%s/vqa_OpenEnded_mscoco_test-dev2015_%s-%d-' % (folder, folder,it) + str(it).zfill(8) + '_results'
with open(filename + '.json', 'w') as f:
json.dump(final_list, f)
if visualize:
visualize_failures(stat_list, mode)
elif mode == 'test':
filename = './%s/vqa_OpenEnded_mscoco_test2015_%s-%d-' % (folder, folder,it) + str(it).zfill(8) + '_results'
with open(filename + '.json', 'w') as f:
json.dump(final_list, f)
if visualize:
visualize_failures(stat_list, mode)
def eval(self):
#################### Session Setting#########################
sess = tf.Session(config=tf.ConfigProto(log_device_placement=False))
sess.run(tf.global_variables_initializer())
self.load_vqa_model(sess, config.VQA_SCOPE)
self.exec_validation(sess, mode='test-dev', folder=config.FOLDER_NAME)
def train_eval(self):
#################### Basic Operations#########################
if config.POSITIVE_ONLY:
self.overall_loss = self.model.loss1
cross_entrophy = self.model.sigmoid_cross_entrophy1
accuracy = self.model.accuracy1
vqa_predict1 = self.model.predict1
af_distance = self.model.distance
else:
self.overall_loss = self.model.overloss
cross_entrophy = self.model.sigmoid_cross_entrophy1 + self.model.sigmoid_cross_entrophy2
accuracy = (self.model.accuracy1 + self.model.accuracy2) / 2
vqa_predict1 = self.model.predict1
vqa_predict2 = self.model.predict2
af_distance = self.model.distance
vqa_distance = self.model.distance2
distance_loss = self.model.distance_loss
pn_distance = self.model.dpn
apn_distance = self.model.dapn
vqa_train_op, vqa_decay_lr = self.build_train_ops()
#################### Summary Setting #########################
val_average_loss = tf.placeholder('float')
val_average_accuracy = tf.placeholder('float')
train_ol_summary = tf.summary.scalar('train_overall_loss', self.overall_loss)
train_ce_summary = tf.summary.scalar('train_cross_entrophy', cross_entrophy)
train_acc_summary = tf.summary.scalar('accuracy', accuracy)
val_average_l_summary = tf.summary.scalar('val_ave_cross_entrophy', val_average_loss)
val_average_acc_summary = tf.summary.scalar('val_average_accuracy', val_average_accuracy)
#################### Session Setting#########################
sess = tf.Session(config=tf.ConfigProto(log_device_placement=False))
sess.run(tf.global_variables_initializer())
train_writer = tf.summary.FileWriter(config.log_path + 'train', sess.graph)
valid_writer = tf.summary.FileWriter(config.log_path + 'valid')
if config.USE_PRE_VQA_MODEL:
self.load_vqa_model(sess, config.VQA_SCOPE)
#################### Train Step Function#########################
dataloader = VQADataLoader(batchsize=config.BATCH_SIZE, mode='train')
dataloader_pairs = VQADataLoader_pair(batchsize=config.BATCH_SIZE, mode='train')
start_t = time.time()
best_result = [0.0, 0]
val_counter = 0
for i in range(config.MAX_TRAIN_STEP):
if config.POSITIVE_ONLY:
# data_time=time.time()
q_strs, q_word_vec_list, q_len_list, ans_vectors, img_features, a_word_vec, ans_space_score, t_qid_list, img_ids, epoch = dataloader.next_batch(
config.BATCH_SIZE)
# print 'data provide time', time.time()-data_time
feed_dict = {self.model.q_input: q_word_vec_list, self.model.a_input: a_word_vec,
self.model.ans1: ans_vectors, self.model.ans_space_score: ans_space_score,
self.model.seqlen: q_len_list, self.model.img_vec: img_features,
self.lr: config.VQA_LR, self.model.keep_prob: config.KEEP_PROB,
self.model.is_training: True}
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
_, vqadecay_lr, global_step = sess.run([vqa_train_op, vqa_decay_lr, self.global_step],
feed_dict=feed_dict)
to_summary, tc_summary, ta_summary, acc, o_loss, c_loss, predict1, af = sess.run(
[train_ol_summary, train_ce_summary, train_acc_summary, accuracy, self.overall_loss, cross_entrophy,
vqa_predict1, af_distance], options=run_options,
run_metadata=run_metadata, feed_dict=feed_dict)
else:
q_strs, q_word_vec_list, q_len_list, ans_vectors, img_features, a_word_vec, ans_space_score, q_strs2, q_word_vec_list2, q_len_list2, ans_vectors2, img_features2, a_word_vec2, t_qid_list, img_ids, epoch = dataloader_pairs.next_batch(
config.BATCH_SIZE)
feed_dict = {self.model.q_input: q_word_vec_list, self.model.a_input: a_word_vec,
self.model.ans1: ans_vectors, self.model.ans_space_score: ans_space_score,
self.model.seqlen: q_len_list, self.model.img_vec: img_features,
self.model.a_input2: a_word_vec2,
self.lr: config.VQA_LR, self.model.keep_prob: config.KEEP_PROB,
self.model.is_training: True,
self.model.ans2: ans_vectors2, self.model.img_vec2: img_features2}
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
_, vqadecay_lr, global_step = sess.run([vqa_train_op, vqa_decay_lr, self.global_step],
feed_dict=feed_dict)
to_summary, tc_summary, ta_summary, acc, o_loss, c_loss, predict1, predict2, af, vqad, d_loss, dpn, dapn = sess.run(
[train_ol_summary, train_ce_summary, train_acc_summary, accuracy, self.overall_loss, cross_entrophy,
vqa_predict1, vqa_predict2, af_distance, vqa_distance, distance_loss, pn_distance, apn_distance],
options=run_options,
run_metadata=run_metadata, feed_dict=feed_dict)
train_writer.add_summary(to_summary, i)
train_writer.add_summary(tc_summary, i)
train_writer.add_summary(ta_summary, i)
if (i + 1) % config.SHOW_STEP == 0 or i == 1:
cost_t = time.time() - start_t
print '\n Train Process: Step: %d|%d, Overall Loss: %.7f, Sigmoid Cross Entrophy: %.7f, lr: %.7f, decay_lr: %.7f,accuracy: %.3f, time:%d s, distance1: %.7f, distance2: %.7f, distance_loss: %.7f, pn_distance: %.7f, apn_distance: %.7f' \
% (
global_step + 1, config.MAX_TRAIN_STEP, o_loss, c_loss, config.VQA_LR, vqadecay_lr, acc,
cost_t,
af, vqad, d_loss, dpn, dapn)
print 'Question: %s' % (q_strs[0])
answer = ans_vectors[0].argmax(axis=0)
print 'Answer:', dataloader.vec_to_answer(answer), answer
v_pred1 = predict1[0]
print 'predict:', dataloader.vec_to_answer(v_pred1), v_pred1
answer2 = ans_vectors2[0].argmax(axis=0)
print 'Answer2:', dataloader.vec_to_answer(answer2), answer2
v_pred2 = predict2[0]
print 'predict2:', dataloader.vec_to_answer(v_pred2), v_pred2
start_t = time.time()
fshow = file('show.txt', 'a+')
fshow.write(
'\n Train Process: Step: %d|%d, Overall Loss: %.7f, Sigmoid Cross Entrophy: %.7f, lr: %.7f, decay_lr: %.7f,accuracy: %.3f, time:%d s, distance1: %.7f, distance2: %.7f, distance_loss: %.7f, pn_distance: %.7f, apn_distance: %.7f' \
% (
global_step + 1, config.MAX_TRAIN_STEP, o_loss, c_loss, config.VQA_LR, vqadecay_lr, acc,
cost_t,
af, vqad, d_loss, dpn, dapn))
fshow.close()
if (i + 1) % config.VAL_STEP == 0:
self.exec_validation(sess, mode='test',it=i,folder=config.FOLDER_NAME)
if (i+1)%config.SAVE_STEP==0:
print 'saving model .......'
self.save_vqa_model(sess,config.VQA_SCOPE,i+1) # Save VQA Model