-
Notifications
You must be signed in to change notification settings - Fork 3
/
inference_pairwise.py
433 lines (363 loc) · 18.2 KB
/
inference_pairwise.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
428
429
430
431
432
433
""" Given a pair of images, generate 'S1 vs. S2' using a trained decerning speaker model.
Given one image, genersate its description using a trained single speaker model.
Output:
html file that collects image pairs and captions
save the dataset with captions and evaluation metrics scores
"""
import os
import sys
import math
import json
import time
import argparse
import numpy as np
import tensorflow as tf
from utils.pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
from utils.pycocoevalcap.bleu.bleu import Bleu
from utils.pycocoevalcap.meteor.meteor import Meteor
from utils.pycocoevalcap.rouge.rouge import Rouge
from utils.pycocoevalcap.cider.cider import Cider
from utils import caption_generator, inference_wrapper_base
from network.speaker_net_discerning import CaptionNet as CaptionNetPair
from network.speaker_net_simple import CaptionNetSingle
from pyVisDifftools.visdiff import VisDiff
import data_loader as data_feeder
html_dir = 'result/inference'
data_dir = 'dataset'
html2img_dir = '../../dataset/images'
# html2img_dir = os.path.join(os.path.abspath(data_dir), 'images')
class InferenceWrapperSingle(inference_wrapper_base.InferenceWrapperBase):
"""Model wrapper class for performing inference with CaptionNet."""
def __init__(self, img_input_mode):
super(InferenceWrapperSingle, self).__init__()
self.img_input_mode = img_input_mode
self.img = None
self.model = None
def build_model(self, args):
model = CaptionNetSingle(args, mode="inference", vocab_size=args.vocab_size, img_h=args.img_h, img_w=args.img_w)
model.build()
print 'vocab_size: %d' % model.vocab_size
self.model = model
return model
def img_initial_state(self, sess, image1, image2=None):
self.img = np.expand_dims(image1, 0)
# print '*** img ***'
# print image1
initial_state = sess.run(fetches=["rnn_1/initial_state:0"],
feed_dict={self.model.img1: self.img})
return initial_state
def inference_step(self, sess, input_feed, state_feed):
sp = np.shape(state_feed)
if np.size(sp) == 3 and sp[1] == 1:
state_feed = state_feed[:, 0, :]
if self.img_input_mode is 'initial':
softmax_output, state_output = sess.run(fetches=["softmax:0", "rnn_1/state:0"],
feed_dict={"text_seq:0": input_feed,
"rnn_1/state_feed:0": state_feed})
else: # self.img_input_mode is 'per_state':
len = np.shape(input_feed)[0]
new_shape = np.ones(np.shape(np.shape(self.img)), dtype=np.int8)
new_shape[0] = len
img = np.tile(self.img, new_shape)
softmax_output, state_output = sess.run(fetches=["softmax:0", "rnn_1/state:0"],
feed_dict={"image1:0": img,
"text_seq:0": input_feed,
"rnn_1/state_feed:0": state_feed})
return softmax_output, state_output, None
class InferenceWrapperPair(inference_wrapper_base.InferenceWrapperBase):
"""Model wrapper class for performing inference with CaptionNet."""
def __init__(self, img_input_mode):
super(InferenceWrapperPair, self).__init__()
self.img_input_mode = img_input_mode
self.img1 = None
self.img2 = None
self.model = None
def build_model(self, args):
model = CaptionNetPair(args, mode="inference", vocab_size=args.vocab_size, img_h=args.img_h, img_w=args.img_w)
model.build()
print 'vocab_size: %d' % model.vocab_size
self.model = model
return model
def img_initial_state(self, sess, image1, image2):
self.img1 = np.expand_dims(image1, 0)
self.img2 = np.expand_dims(image2, 0)
initial_state = sess.run(fetches=["rnn_1/initial_state:0"],
feed_dict={self.model.img1: self.img1, self.model.img2: self.img2})
return initial_state
def inference_step(self, sess, input_feed, state_feed):
sp = np.shape(state_feed)
if np.size(sp) == 3 and sp[1] == 1:
state_feed = state_feed[:, 0, :]
if self.img_input_mode is 'initial':
softmax_output, state_output = sess.run(fetches=["softmax:0", "rnn_1/state:0"],
feed_dict={"text_seq:0": input_feed,
"rnn_1/state_feed:0": state_feed})
else: # self.img_input_mode is 'per_state':
len = np.shape(input_feed)[0]
new_shape = np.ones(np.shape(np.shape(self.img1)), dtype=np.int8)
new_shape[0] = len
img1 = np.tile(self.img1, new_shape)
img2 = np.tile(self.img2, new_shape)
softmax_output, state_output = sess.run(fetches=["softmax:0", "rnn_1/state:0"],
feed_dict={"image1:0": img1,
"image2:0": img2,
"text_seq:0": input_feed,
"rnn_1/state_feed:0": state_feed})
return softmax_output, state_output, None
def captions_id2word(captions, word_list, spliter=' '):
sentences = []
probs = []
for i, caption in enumerate(captions):
sent = ''
for w_id in caption.sentence[1:-1]:
sent += word_list[w_id] + spliter
sentences.append(sent)
probs.append(math.exp(caption.logprob))
return sentences, probs
def _infer_wirte_as_txt(result_f, img_id1, img_id2, true_s1, true_s2, gen_sen_w, gen_probs):
msg = "\n*** image %s vs %s ***" % (img_id1, img_id2)
result_f.write(msg + '\n')
msg = "Ground Truth:"
result_f.write(msg + '\n')
for i in range(5):
msg = ' %s _VS %s' % (true_s1[i], true_s2[i])
result_f.write(msg + '\n')
msg = "Generated:"
result_f.write(msg + '\n')
for i in range(len(gen_probs)):
msg = " %d) %s \t (p=%f)" % (i+1, gen_sen_w[i], gen_probs[i])
result_f.write(msg + '\n')
def _infer_wirte_as_html(result_f, img_id1, img_id2, true_s1, true_s2, gen_sen_w, gen_probs, img_dir=html2img_dir):
msg = "*** image %s vs %s ***" % (img_id1, img_id2)
result_f.write('<tr align=\'center\'><td>' + msg + '</td></tr>\n')
msg = '<img src=\'%s.jpg\' height=150><img src=\'%s.jpg\' height=150>' \
% (os.path.join(img_dir, img_id1), os.path.join(img_dir, img_id2))
result_f.write('<tr align=\'center\'><td>' + msg + '</td></tr>\n')
msg = "Ground Truth:"
result_f.write('<tr align=\'center\'><td>' + msg + '</td></tr>\n')
for i in range(5):
msg = ' %s _VS %s' % (true_s1[i], true_s2[i])
result_f.write('<tr align=\'center\'><td>' + msg + '</td></tr>\n')
msg = "Generated:"
result_f.write('<tr align=\'center\'><td>' + msg + '</td></tr>\n')
for i in range(len(gen_probs)):
msg = " %d) %s \t (p=%f)" % (i+1, gen_sen_w[i], gen_probs[i])
result_f.write('<tr align=\'center\'><td>' + msg + '</td></tr>\n')
result_f.write('<tr> <td colspan="4" bgcolor="#009966" height="3"> </td> </tr>\n')
def inference(args, input_path, model_step, dataset_name='val', case_num=0, output_num=10, beam_size=10,
out_to_html=50, eval_scores=1, save_dataset=False):
if save_dataset:
case_num = 0
model_path = os.path.join(input_path, model_step)
data_json_file = os.path.join(data_dir, 'visdiff_%s.json' % dataset_name)
args.batch_size = 1 # inference for one image at a time
# Load dataset
dataset = VisDiff(data_json_file).dataset['annotations']
# Load imgs_dict
if not args.train_img_model:
imgs_dict_fpath = '../imgs_dict/%s_dict.json' % dataset_name
with open(imgs_dict_fpath) as imgs_dict_file:
imgs_dict = json.load(imgs_dict_file)
else:
imgs_dict = []
# Get the vocabulary.
if args.word_embed_type == 'char':
vocab = data_feeder.build_vocabulary('%s/visdiff_trainval.json'
% data_dir, args.word_count_thresh, args.word_embed_type)
else:
vocab = np.load(args.vocab_file).item()
args.vocab_size = len(vocab)
# Build reverse vocabulary
word_list = [''] * len(vocab)
for w, case_i in vocab.iteritems():
word_list[case_i] = w
# Set img size
if args.img_model == 'alexnet' or args.img_model == 'vgg_16':
img_w = img_h = 224
elif args.img_model == 'inception_v3':
img_w = img_h = 299
else:
raise Exception("model type not supported: {}".format(args.img_model))
args.img_h = img_h
args.img_w = img_w
# build data_feeder
my_feeder = data_feeder.DataFeeder(dataset, vocabulary=vocab, train_img_model=args.train_img_model,
img_dir='../../../dataset/images_224/', img_dict=imgs_dict,
max_length=args.max_text_length, img_w=img_w, img_h=img_h)
cases = dataset
if case_num > 0:
cases = dataset[0: case_num]
# Build the inference graph.
g = tf.Graph()
with g.as_default():
if args.speaker_mode == 'S':
infer_model = InferenceWrapperSingle(args.img_input_mode)
else: # mode 'DS'
infer_model = InferenceWrapperPair(args.img_input_mode)
restore_fn = infer_model.build_graph_from_config(args, model_path)
# g.finalize()
tf.reset_default_graph()
# gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
# with tf.Session(graph=g, config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
with tf.Session(graph=g) as sess:
init = tf.global_variables_initializer()
sess.run(init)
# Load the model from checkpoint.
restore_fn(sess)
g.finalize()
# Prepare the caption generator. Here we are implicitly using the default
# beam search parameters. See caption_generator.py for a description of the
# available beam search parameters.
if eval_scores:
true_text_dict = {}
gen_text_collect = []
for sent_j in range(output_num):
gen_text_collect.append({})
generator = caption_generator.CaptionGenerator(infer_model, vocab, beam_size=beam_size)
spliter = ' '
if args.word_embed_type == 'char':
spliter = ''
if out_to_html:
result_path = os.path.join(html_dir, 'infer_%s_%s_%d.html' % (dataset_name, model_step, beam_size))
result_f = open(result_path, 'w')
result_f.write('<table align=\'center\'>\n')
for case_i, case in enumerate(cases):
img_id1 = case['img1_id']
img_id2 = case['img2_id']
img1 = my_feeder.get_img(img_id1, args.img_w, args.img_h)
img2 = my_feeder.get_img(img_id2, args.img_w, args.img_h)
true_s1 = case['sentences1']
true_s2 = case['sentences2']
gen_sen_id = generator.beam_search(sess, img1, img2)
gen_sen_w, gen_probs = captions_id2word(gen_sen_id[0: output_num], word_list, spliter)
if out_to_html > 0 and case_i < out_to_html:
_infer_wirte_as_html(result_f, img_id1, img_id2, true_s1, true_s2, gen_sen_w, gen_probs)
if eval_scores:
true_text_dict[case_i*10] = []
true_text_dict[case_i*10+1] = []
for idx in range(5):
# true_text_dict[case_i].append(true_s1[idx] + ' _VS ' + true_s2[idx])
true_text_dict[case_i*10].append(true_s1[idx])
true_text_dict[case_i*10 + 1].append(true_s2[idx])
for sent_j in range(output_num):
ss = gen_sen_w[sent_j].split('_VS')
left_s = ss[0].strip()
if len(ss) > 1:
right_s = ss[1].strip()
else:
right_s = ''
gen_text_collect[sent_j][case_i*10] = [left_s]
gen_text_collect[sent_j][case_i*10+1] = [right_s]
if save_dataset:
case['gen_texts'] = gen_sen_w
case['gen_probs'] = gen_probs
if eval_scores:
case['scores'] = {}
print '%s: %d / %d' % (time.strftime('[%H:%M:%S]'), case_i, len(cases))
if out_to_html:
result_f.write('</table>\n')
if eval_scores:
score_record_list, scores_record_list = _eval_infer(true_text_dict, gen_text_collect)
if out_to_html:
msg = '>> ' + ' '.join(score_record_list[0].iterkeys())
result_f.write(msg + '<br/>\n')
score_mean = {}
out_num = len(score_record_list)
for sent_j, score_record in enumerate(score_record_list):
if out_to_html:
msg = '%d)\t' % sent_j + '\t'.join('%.3f' % sc for sc in score_record.itervalues())
result_f.write(msg + '<br/>\n')
for method in score_record.iterkeys():
if sent_j == 0:
score_mean[method] = score_record[method] / out_num
else:
score_mean[method] += score_record[method] / out_num
# save overall scores to json
to_save = {'input_path': input_path, 'model_step': model_step, 'score_record_list': score_record_list,
'score_mean': score_mean}
score_f = open(os.path.join(input_path, 'infer_scores_%s_%s_case%d_beam%d_sent%d.json'
% (dataset_name, model_step, case_num, beam_size, output_num)), 'w')
json.dump(to_save, score_f)
score_f.close()
if save_dataset:
if eval_scores:
for case_i, case in enumerate(cases):
for method in score_record_list[0].iterkeys():
case['scores'][method] = [[0.0, 0.0]] * output_num
for sent_i in range(output_num):
scores = scores_record_list[sent_i][method]
case['scores'][method][sent_i] = [scores[case_i*2], scores[case_i*2 + 1]]
dataset_f = open(os.path.join(input_path, 'infer_annotations_%s_%s_case%d_beam%d_sent%d.json'
% (dataset_name, model_step, case_num, beam_size, output_num)), 'w')
json.dump(cases, dataset_f)
dataset_f.close()
if out_to_html:
msg = 'ave\t' + '\t'.join('%.3f' % sc for sc in score_mean.itervalues())
result_f.write(msg + '<br/>\n')
if out_to_html:
result_f.close()
def _eval_infer(true_text_dict, gen_text_collect):
score_record_list = []
scores_record_list = []
for i, gen_txt_dict in enumerate(gen_text_collect):
score_record = {}
scores_record = {}
tokenizer = PTBTokenizer()
gts = tokenizer.tokenize(true_text_dict)
res = tokenizer.tokenize(gen_txt_dict)
# print 'setting up scorers...'
scorers = [
(Bleu(4), ["Bleu_1", "Bleu_2", "Bleu_3", "Bleu_4"]),
(Meteor(), "METEOR"),
(Rouge(), "ROUGE_L"),
(Cider(), "CIDEr")
]
for scorer, method in scorers:
score, scores = scorer.compute_score(gts, res)
if type(method) == list:
for sc, ss, m in zip(score, scores, method):
score_record[m] = sc
scores_record[m] = ss
else:
score_record[method] = score
scores_record[method] = scores
score_record_list.append(score_record)
scores_record_list.append(scores_record)
return score_record_list, scores_record_list
def main(_):
parser = argparse.ArgumentParser()
parser.add_argument('--case_num', type=int, default=0,
help='Number of examples to run, <1 means whole dataset')
parser.add_argument('--out_to_html', type=int, default=50,
help='Number of examples to be shown in html. 0 means no html file')
parser.add_argument('--output_num', type=int, default=10,
help='Number of sentences for each example')
parser.add_argument('--beam_size', type=int, default=10,
help='For beam search. Must guarantee beam_size >= output_num')
parser.add_argument('--input_path', type=str,
default='result/speaker/DS_fixCNN__2017-03-06-23',
help='Directory with model and config.json inside')
parser.add_argument('--model_step', type=str, default='model-50000',
help='The model to use inside the input_path')
parser.add_argument('--dataset_name', type=str, default='val',
help='The dataset_name used for inference, train / val / test')
parser.add_argument('--eval_scores', type=int, default=1,
help='Whether or not compute evaluation scores.')
parser.add_argument('--save_dataset', type=int, default=1,
help='Whether or not save the dataset with inference and scores added.')
args = parser.parse_args()
load_arguments(parser, os.path.join(args.input_path, 'config.json'))
args = parser.parse_args()
args.train_set_name =args.dataset_name
# print vars(args)
inference(args=args, input_path=args.input_path, model_step=args.model_step,
dataset_name=args.dataset_name, case_num=args.case_num, output_num=args.output_num,
beam_size=args.beam_size, out_to_html=args.out_to_html, eval_scores=args.eval_scores,
save_dataset=args.save_dataset)
def load_arguments(parser, json_fpath):
with open(json_fpath) as data_file:
arg_dict = json.load(data_file)
for (key, value) in arg_dict.items():
parser.add_argument('--' + key, default=value)
if __name__ == "__main__":
tf.app.run()