forked from Intel-tensorflow/Tensorflow-SSD-VGG
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpreparedata.py
309 lines (231 loc) · 13.7 KB
/
preparedata.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
from datasets import pascalvoc_datasets
import tensorflow as tf
import matplotlib.pyplot as plt
import tensorflow.contrib.slim as slim
# from nets import nets_factory
from preprocessing import preprocessing_factory
import numpy as np
import cv2
from utility import visualization
from nets.ssd import g_ssd_model
from preprocessing.ssd_vgg_preprocessing import np_image_unwhitened
from preprocessing.ssd_vgg_preprocessing import preprocess_for_train
from preprocessing.ssd_vgg_preprocessing import preprocess_for_eval
import tf_utils
import math
class PrepareData():
def __init__(self):
self.batch_size = 8
self.labels_offset = 0
self.matched_thresholds = 0.5 #threshold for anchor matching strategy
return
def __preprocess_data(self, image, labels, bboxes):
out_shape = g_ssd_model.img_shape
if self.is_training_data:
image, labels, bboxes = preprocess_for_train(image, labels, bboxes, out_shape = out_shape)
else:
image, labels, bboxes, _ = preprocess_for_eval(image, labels, bboxes, out_shape = out_shape)
return image, labels, bboxes
def __get_images_labels_bboxes(self,data_sources, num_samples,is_training_data):
self.dataset = pascalvoc_datasets.get_dataset_info(data_sources, num_samples)
self.is_training_data = is_training_data
if self.is_training_data:
shuffle = True
#make sure most samples can be fetched in one epoch
self.num_readers = 2
else:
#make sure data is fetchd in sequence
shuffle = False
self.num_readers = 1
provider = slim.dataset_data_provider.DatasetDataProvider(
self.dataset,
shuffle=shuffle,
num_readers=self.num_readers,
common_queue_capacity=20 * self.batch_size,
common_queue_min=10 * self.batch_size)
# Get for SSD network: image, labels, bboxes.
[image, shape, format, filename, glabels, gbboxes,gdifficults] = provider.get(['image', 'shape', 'format','filename',
'object/label',
'object/bbox',
'object/difficult'])
glabels -= self.labels_offset
# Pre-processing image, labels and bboxes.
image, glabels, gbboxes = self.__preprocess_data(image, glabels, gbboxes)
# Assign groundtruth information for all default/anchor boxes
# gclasses, glocalisations, gscores = g_ssd_model.tf_ssd_bboxes_encode(glabels, gbboxes)
gclasses, glocalisations, gscores = g_ssd_model.match_achors(glabels, gbboxes, matching_threshold=self.matched_thresholds)
return self.__batching_data(image, glabels, format, filename, gbboxes, gdifficults, gclasses, glocalisations, gscores)
def __batching_data(self,image, glabels, format, filename, gbboxes, gdifficults,gclasses, glocalisations, gscores):
#we will want to batch original glabels and gbboxes
#this information is still useful even if they are padded after dequeuing
dynamic_pad = True
batch_shape = [1,1,1,1,1] + [len(gclasses), len(glocalisations), len(gscores)]
tensors = [image, filename,glabels,gbboxes,gdifficults,gclasses, glocalisations, gscores]
#Batch the samples
if self.is_training_data:
self.num_preprocessing_threads = 1
else:
# to make sure data is fectched in sequence during evaluation
self.num_preprocessing_threads = 1
#tf.train.batch accepts only list of tensors, this batch shape can used to
#flatten the list in list, and later on convet it back to list in list.
batch = tf.train.batch(
tf_utils.reshape_list(tensors),
batch_size=self.batch_size,
num_threads=self.num_preprocessing_threads,
dynamic_pad=dynamic_pad,
capacity=5 * self.batch_size)
#convert it back to the list in list format which allows us to easily use later on
batch= tf_utils.reshape_list(batch, batch_shape)
return batch
def __disp_image(self, img, classes, bboxes):
bvalid = (classes !=0)
classes = classes[bvalid]
bboxes = bboxes[bvalid]
scores =np.full(classes.shape, 1.0)
visualization.plt_bboxes(img, classes, scores, bboxes,title='Ground Truth')
return
def __disp_matched_anchors(self,img, target_labels_data, target_localizations_data, target_scores_data):
found_matched = False
all_anchors = g_ssd_model.get_allanchors()
for i, target_score_data in enumerate(target_scores_data):
num_pos = (target_labels_data[i] != 0).sum()
if (num_pos == 0):
continue
print('Found {} matched default boxes in layer {}'.format(num_pos,g_ssd_model.feat_layers[i]))
# pos_sample_inds = ((target_labels_data[i] != 0) & (target_score_data <=self.matched_thresholds)).nonzero()
pos_sample_inds = (target_labels_data[i] != 0).nonzero()
classes = target_labels_data[i][pos_sample_inds]
scores = target_scores_data[i][pos_sample_inds]
print("matched scores :{}".format(scores))
print("matched labels: {}".format(classes))
bboxes_default= g_ssd_model.get_allanchors(minmaxformat=True)[i][pos_sample_inds]
bboxes_gt = g_ssd_model.decode_bboxes_layer(target_localizations_data[i][pos_sample_inds],
all_anchors[i][pos_sample_inds])
# print("default box minimum, {} gt box minimum, {}".format(bboxes_default.min(), bboxes_gt.min()))
marks_default = np.full(classes.shape, True)
marks_gt = np.full(classes.shape, False)
scores_gt = np.full(scores.shape, 1.0)
bboxes = bboxes_default
neg_marks = marks_default
add_gt = True
if add_gt :
bboxes = np.vstack([bboxes_default,bboxes_gt])
neg_marks = np.hstack([marks_default,marks_gt])
classes = np.tile(classes, 2)
scores = np.hstack([scores, scores_gt])
title = "Default boxes: Layer {}".format(g_ssd_model.feat_layers[i])
visualization.plt_bboxes(img, classes, scores, bboxes,neg_marks=neg_marks,title=title)
found_matched = True
return found_matched
def get_voc_2007_train_data(self,is_training_data=True):
data_sources = "../data/voc/tfrecords/voc_train_2007*.tfrecord"
num_samples = pascalvoc_datasets.DATASET_SIZE['2007_train']
return self.__get_images_labels_bboxes(data_sources, num_samples, is_training_data)
def get_voc_2012_train_data(self,is_training_data=True):
data_sources = "../data/voc/tfrecords/voc_train_2012*.tfrecord"
num_samples = pascalvoc_datasets.DATASET_SIZE['2012_train']
return self.__get_images_labels_bboxes(data_sources, num_samples, is_training_data)
def get_voc_2007_2012_train_data(self,is_training_data=True):
data_sources = "../data/voc/tfrecords/voc_train*.tfrecord"
num_samples = pascalvoc_datasets.DATASET_SIZE['2007_train'] + pascalvoc_datasets.DATASET_SIZE['2012_train']
return self.__get_images_labels_bboxes(data_sources, num_samples, is_training_data)
def get_voc_2007_test_data(self):
data_sources = "../data/voc/tfrecords/voc_test_2007*.tfrecord"
num_samples = pascalvoc_datasets.DATASET_SIZE['2007_test']
return self.__get_images_labels_bboxes(data_sources, num_samples, False)
def get_gtsdb_train_data(self,is_training_data=True):
data_sources = "../GTSDB/tfrecords/gtsdb_train*.tfrecord"
num_samples = 588
return self.__get_images_labels_bboxes(data_sources, num_samples, is_training_data)
def get_gtsdb_test_data(self,is_training_data=False):
data_sources = "../GTSDB/tfrecords/gtsdb_test*.tfrecord"
num_samples = 153
return self.__get_images_labels_bboxes(data_sources, num_samples, is_training_data)
def get_gtsdb_original_data(self,is_training_data=False):
data_sources = "../GTSDB/tfrecords_old/gtsdb*.tfrecord"
num_samples = 506
return self.__get_images_labels_bboxes(data_sources, num_samples, is_training_data)
def iterate_file_name(self, batch_data):
num_batches = 1*math.ceil(self.dataset.num_samples / float(self.batch_size))
all_files = []
with tf.Session('') as sess:
init = tf.global_variables_initializer()
sess.run(init)
with slim.queues.QueueRunners(sess):
for i in range(num_batches):
image, filename,glabels,gbboxes,gdifficults,gclasses, glocalisations, gscores = sess.run(list(batch_data))
print(filename)
all_files.append(filename)
all_files = np.concatenate(all_files)
all_files_unique = np.unique(all_files)
print(len(all_files_unique))
return
def check_match_statistics(self,filename,gclasses, gscores):
#flatten the array into Batch_num x bbox_num
gt_anchor_labels = []
gt_anchor_scores = []
for i in range(len(gclasses)):
gt_anchor_labels.append(np.reshape(gclasses[i], [self.batch_size, -1]))
gt_anchor_scores.append(np.reshape(gscores[i], [self.batch_size,-1]))
gt_anchor_labels = np.concatenate(gt_anchor_labels, axis = 1)
gt_anchor_scores = np.concatenate(gt_anchor_scores, axis = 1)
#find out missed match
inds = (gt_anchor_scores <= self.matched_thresholds) & (gt_anchor_labels != 0)
real_inds = inds.nonzero()
print("missed match: {}".format(filename[real_inds[0]]))
print("missed match scores: {}".format(gt_anchor_scores[real_inds]))
print("missed match labels: {}".format(gt_anchor_labels[real_inds]))
return
def run(self):
with tf.Graph().as_default():
# batch_data= self.get_voc_2007_train_data(is_training_data=True)
# batch_data = self.get_voc_2007_test_data()
# batch_data = self.get_voc_2012_train_data()
# batch_data = self.get_voc_2007_2012_train_data(is_training_data = True)
batch_data = self.get_gtsdb_train_data(is_training_data = True)
return self.iterate_file_name(batch_data)
with tf.Session('') as sess:
init = tf.global_variables_initializer()
sess.run(init)
with slim.queues.QueueRunners(sess):
while True:
image, filename,glabels,gbboxes,gdifficults,gclasses, glocalisations, gscores = sess.run(list(batch_data))
# print("min: {}, max: {}".format(gbboxes.min(), gbboxes.max()))
# return
# print(glabels)
# print("number of zero label patch {}".format((glabels.sum(axis=1) == 0).sum()))
# return
#
print(filename)
selected_file = b'000050'
picked_inds = None
#selet the first image in the batch
if selected_file is None:
picked_inds = 0
else:
picked_inds = (selected_file == filename).nonzero()
if len(picked_inds[0]) == 0:
picked_inds = None
else:
picked_inds = picked_inds[0][0]
if picked_inds is None:
continue
self.check_match_statistics(filename, gclasses, gscores)
target_labels_data = [item[picked_inds] for item in gclasses]
target_localizations_data = [item[picked_inds] for item in glocalisations]
target_scores_data = [item[picked_inds] for item in gscores]
image_data = image[picked_inds]
print("picked file {}".format(filename[picked_inds]))
image_data = np_image_unwhitened(image_data)
self.__disp_image(image_data, glabels[picked_inds], gbboxes[picked_inds])
found_matched = self.__disp_matched_anchors(image_data,target_labels_data, target_localizations_data, target_scores_data)
plt.show()
break;
#exit the batch data testing right after a successful match have been found
# if found_matched:
#this could be a potential issue to be solved since sometime not all grouth truth bboxes are encoded.
return
if __name__ == "__main__":
obj= PrepareData()
obj.run()