-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathbase_model.py
301 lines (262 loc) · 13.1 KB
/
base_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
import os
import numpy as np
import pandas as pd
import tensorflow as tf
import pickle as pickle
import copy
import json
# from tqdm import tqdm
from utils.nn import NN
# from utils.coco.coco import COCO
# from utils.coco.pycocoevalcap.eval import COCOEvalCap
from utils.misc import CaptionData, TopN
slim = tf.contrib.slim
class BaseModel(object):
def __init__(self, config, mode="train"):
self.config = config
self.mode= mode
self.is_train = True if mode == 'train' else False
self.train_cnn = False
# self.image_loader = ImageLoader('./utils/ilsvrc_2012_mean.npy')
# self.image_shape = [224, 224, 3]
self.nn = NN(config, mode)
self.global_step = tf.Variable(0,
name = 'global_step',
trainable = False)
self.reader = tf.TFRecordReader()
self.rnn_attend_state = None
# self.build()
def build(self):
raise NotImplementedError()
# def train(self):
# """ Train the model using the COCO train2014 data. """
# print("Training the model...")
# config = self.config
# if not os.path.exists(config.summary_dir):
# os.mkdir(config.summary_dir)
# train_writer = tf.summary.FileWriter(config.summary_dir,
# sess.graph)
# for _ in tqdm(list(range(config.num_epochs)), desc='epoch'):
# for _ in tqdm(list(range(train_data.num_batches)), desc='batch'):
# batch = train_data.next_batch()
# image_files, sentences, masks = batch
# images = self.image_loader.load_images(image_files)
# feed_dict = {self.images: images,
# self.sentences: sentences,
# self.masks: masks}
# _, summary, global_step = sess.run([self.opt_op,
# self.summary,
# self.global_step],
# feed_dict=feed_dict)
# if (global_step + 1) % config.save_period == 0:
# self.save()
# train_writer.add_summary(summary, global_step)
# train_data.reset()
# self.save()
# train_writer.close()
# print("Training complete.")
# def eval(self, sess, eval_gt_coco, eval_data, vocabulary):
# """ Evaluate the model using the COCO val2014 data. """
# print("Evaluating the model ...")
# config = self.config
# results = []
# if not os.path.exists(config.eval_result_dir):
# os.mkdir(config.eval_result_dir)
# # Generate the captions for the images
# idx = 0
# for k in tqdm(list(range(eval_data.num_batches)), desc='batch'):
# batch = eval_data.next_batch()
# caption_data = self.beam_search(sess, batch, vocabulary)
# fake_cnt = 0 if k<eval_data.num_batches-1 \
# else eval_data.fake_count
# for l in range(eval_data.batch_size-fake_cnt):
# word_idxs = caption_data[l][0].sentence
# score = caption_data[l][0].score
# caption = vocabulary.get_sentence(word_idxs)
# results.append({'image_id': eval_data.image_ids[idx].item(),
# 'caption': caption})
# idx += 1
# # Save the result in an image file, if requested
# if config.save_eval_result_as_image:
# image_file = batch[l]
# image_name = image_file.split(os.sep)[-1]
# image_name = os.path.splitext(image_name)[0]
# img = plt.imread(image_file)
# plt.imshow(img)
# plt.axis('off')
# plt.title(caption)
# plt.savefig(os.path.join(config.eval_result_dir,
# image_name+'_result.jpg'))
# fp = open(config.eval_result_file, 'w')
# json.dump(results, fp)
# fp.close()
# # Evaluate these captions
# eval_result_coco = eval_gt_coco.loadRes(config.eval_result_file)
# scorer = COCOEvalCap(eval_gt_coco, eval_result_coco)
# scorer.evaluate()
# print("Evaluation complete.")
# def test(self, sess, test_data, vocabulary):
# """ Test the model using any given images. """
# print("Testing the model ...")
# config = self.config
# if not os.path.exists(config.test_result_dir):
# os.mkdir(config.test_result_dir)
# captions = []
# scores = []
# # Generate the captions for the images
# for k in tqdm(list(range(test_data.num_batches)), desc='path'):
# batch = test_data.next_batch()
# caption_data = self.beam_search(sess, batch, vocabulary)
# fake_cnt = 0 if k<test_data.num_batches-1 \
# else test_data.fake_count
# for l in range(test_data.batch_size-fake_cnt):
# word_idxs = caption_data[l][0].sentence
# score = caption_data[l][0].score
# caption = vocabulary.get_sentence(word_idxs)
# captions.append(caption)
# scores.append(score)
# # Save the result in an image file
# image_file = batch[l]
# image_name = image_file.split(os.sep)[-1]
# image_name = os.path.splitext(image_name)[0]
# img = plt.imread(image_file)
# plt.imshow(img)
# plt.axis('off')
# plt.title(caption)
# plt.savefig(os.path.join(config.test_result_dir,
# image_name+'_result.jpg'))
# # Save the captions to a file
# results = pd.DataFrame({'image_files':test_data.image_files,
# 'caption':captions,
# 'prob':scores})
# results.to_csv(config.test_result_file)
# print("Testing complete.")
def beam_search(self, sess, vocabulary):
"""Use beam search to generate the captions for a batch of images."""
# Feed in the images to get the contexts and the initial LSTM states
config = self.config
contexts, initial_memory, initial_output = sess.run(
[self.conv_feats, self.initial_memory, self.initial_output])
partial_caption_data = []
complete_caption_data = []
for k in range(config.batch_size):
initial_beam = CaptionData(sentence = [],
memory = initial_memory[k],
output = initial_output[k],
score = 0)
partial_caption_data.append(TopN(config.beam_size))
partial_caption_data[-1].push(initial_beam)
complete_caption_data.append(TopN(config.beam_size))
# Run beam search
for idx in range(config.max_caption_length):
partial_caption_data_lists = []
for k in range(config.batch_size):
data = partial_caption_data[k].extract()
partial_caption_data_lists.append(data)
partial_caption_data[k].reset()
num_steps = 1 if idx == 0 else config.beam_size
for b in range(num_steps):
if idx == 0:
last_word = vocabulary.start_id * np.ones((config.batch_size), np.int32)
else:
last_word = np.array([pcl[b].sentence[-1]
for pcl in partial_caption_data_lists],
np.int32)
last_memory = np.array([pcl[b].memory
for pcl in partial_caption_data_lists],
np.float32)
last_output = np.array([pcl[b].output
for pcl in partial_caption_data_lists],
np.float32)
memory, output, scores = sess.run(
[self.memory, self.output, self.probs],
feed_dict = {self.contexts: contexts,
self.last_word: last_word,
self.last_memory: last_memory,
self.last_output: last_output})
# Find the beam_size most probable next words
for k in range(config.batch_size):
caption_data = partial_caption_data_lists[k][b]
words_and_scores = list(enumerate(scores[k]))
words_and_scores.sort(key=lambda x: -x[1])
words_and_scores = words_and_scores[0:config.beam_size+1]
# Append each of these words to the current partial caption
for w, s in words_and_scores:
sentence = caption_data.sentence + [w]
score = caption_data.score + np.log2(s)
beam = CaptionData(sentence,
memory[k],
output[k],
score)
if w == vocabulary.end_id:
complete_caption_data[k].push(beam)
else:
partial_caption_data[k].push(beam)
results = []
for k in range(config.batch_size):
if complete_caption_data[k].size() == 0:
complete_caption_data[k] = partial_caption_data[k]
results.append(complete_caption_data[k].extract(sort=True))
return results
def save(self):
""" Save the model. """
config = self.config
data = {v.name: v.eval() for v in tf.global_variables()}
save_path = os.path.join(config.save_dir, str(self.global_step.eval()))
print((" Saving the model to %s..." % (save_path+".npy")))
np.save(save_path, data)
info_file = open(os.path.join(config.save_dir, "config.pickle"), "wb")
config_ = copy.copy(config)
config_.global_step = self.global_step.eval()
pickle.dump(config_, info_file)
info_file.close()
print("Model saved.")
# def load(self, sess, model_file=None):
# """ Load the model. """
# config = self.config
# if model_file is not None:
# save_path = model_file
# else:
# info_path = os.path.join(config.save_dir, "config.pickle")
# info_file = open(info_path, "rb")
# config = pickle.load(info_file)
# global_step = config.global_step
# info_file.close()
# save_path = os.path.join(config.save_dir,
# str(global_step)+".npy")
# print("Loading the model from %s..." %save_path)
# data_dict = np.load(save_path).item()
# count = 0
# for v in tqdm(tf.global_variables()):
# if v.name in data_dict.keys():
# sess.run(v.assign(data_dict[v.name]))
# count += 1
# print("%d tensors loaded." %count)
# def load_cnn(self, session, data_path, ignore_missing=True):
# """ Load a pretrained CNN model. """
# print("Loading the CNN from %s..." %data_path)
# data_dict = np.load(data_path,encoding="latin1").item()
# count = 0
# for op_name in tqdm(data_dict):
# with tf.variable_scope(op_name, reuse = True):
# for param_name, data in data_dict[op_name].items():
# try:
# var = tf.get_variable(param_name)
# session.run(var.assign(data))
# count += 1
# except ValueError:
# pass
# print("%d tensors loaded." %count)
def load_faster_rcnn_feature_extractor(self, model_file):
variables_to_restore = slim.get_variables_to_restore(
include=["SecondStageFeatureExtractor","SecondStageBoxPredictor"])
init_assign_op, input_feed_dict = slim.assign_from_checkpoint(
model_file, variables_to_restore)
return init_assign_op, input_feed_dict
def load_model_except_faster_rcnn(self, model_file):
variables_to_restore = slim.get_variables_to_restore(
exclude=["SecondStageFeatureExtractor","SecondStageBoxPredictor",
"optimizer/OptimizeLoss/SecondStageFeatureExtractor/"])
init_assign_op, input_feed_dict = slim.assign_from_checkpoint(
model_file, variables_to_restore)
return init_assign_op, input_feed_dict