forked from RubanSeven/CRAFT_keras
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
400 lines (337 loc) · 18.3 KB
/
train.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
# -*- coding: utf-8 -*-
# @Author: Ruban
# @License: Apache Licence
# @File: train.py
import io
import os
import cv2
import keras
import argparse
import numpy as np
from PIL import Image
from keras import Input
import tensorflow as tf
import keras.backend as K
from keras.models import Model
from keras.layers import Lambda
from keras.optimizers import Adam
from keras.callbacks import TensorBoard
from module.fake import Fake
from net.vgg16 import VGG16_UNet
from utils.gaussian import GaussianGenerator
from utils.data_util import load_data
from utils.box_util import reorder_points
from utils.img_util import load_sample, img_unnormalize, load_image, img_normalize
from utils.fake_util import crop_image, watershed, find_box, un_warping, cal_confidence, divide_region, \
enlarge_char_boxes
from module.loss import craft_mse_loss, craft_mae_loss, craft_huber_loss
parser = argparse.ArgumentParser()
parser.add_argument('--learning_rate', type=float, default=0.0001) # initial learning rate
parser.add_argument('--batch_size', type=int, default=5) # batch size for training
parser.add_argument('--img_size', type=int, default=600) # batch size for training
parser.add_argument('--max_epochs', type=int, default=800) # maximum number of epochs
parser.add_argument('--gpu_list', type=str, default='0') # list of gpus to use
parser.add_argument('--use_fake', type=bool, default=True) # list of gpus to use
# path to training data
parser.add_argument('--truth_data_path', type=str, default=r'D:\data\synthText\SynthText')
parser.add_argument('--pseudo_data_path', type=str, default=r'D:\data\ICDAR2013')
parser.add_argument('--val_data_path', type=str, default='D:\data\ICDAR2013') # path to validation data
parser.add_argument('--max_image_size', type=int, default=1280)
FLAGS = parser.parse_args()
class SampleGenerator(keras.callbacks.Callback):
def __init__(self, base_model, train_sample_lists, train_sample_probs, fakes, img_size, batch_size):
super().__init__()
assert len(train_sample_lists) == len(train_sample_probs)
assert len(train_sample_lists) == len(fakes)
self.base_model = base_model
self.train_sample_lists = train_sample_lists
self.fakes = fakes
self.train_sample_probs = np.array(train_sample_probs) / np.sum(train_sample_probs)
self.sample_count_list = [len(sample_list) for sample_list in train_sample_lists]
self.sample_idx_list = [0] * len(train_sample_lists)
self.sample_mark_list = list(range(len(train_sample_lists)))
self.img_size = img_size
self.batch_size = batch_size
def get_batch(self, size, is_true=True):
images = list()
word_boxes_list = list()
word_lengths_list = list()
region_scores = list()
affinity_scores = list()
confidence_score_list = list()
fg_masks = list()
bg_masks = list()
gaussian_generator = GaussianGenerator()
word_count_list = list()
for i in range(size):
if is_true:
sample_mark = np.random.choice(self.sample_mark_list, p=self.train_sample_probs)
else:
while 1:
sample_mark = np.random.choice(self.sample_mark_list, p=self.train_sample_probs)
if self.fakes[sample_mark]:
break
img_path, word_boxes, words, char_boxes_list, confidence_list = \
self.train_sample_lists[sample_mark][self.sample_idx_list[sample_mark]]
self.sample_idx_list[sample_mark] += 1
if self.sample_idx_list[sample_mark] >= self.sample_count_list[sample_mark]:
self.sample_idx_list[sample_mark] = 0
np.random.shuffle(self.train_sample_lists[sample_mark])
img, word_boxes, char_boxes_list, region_box_list, affinity_box_list, img_shape = \
load_sample(img_path, self.img_size, word_boxes, char_boxes_list)
images.append(img)
word_count = min(len(word_boxes), len(words))
word_boxes = np.array(word_boxes[:word_count], dtype=np.int32) // 2
word_boxes_list.append(word_boxes)
word_count_list.append(word_count)
word_lengths = [len(words[j]) if len(char_boxes_list[j]) == 0 else 0 for j in range(word_count)]
word_lengths_list.append(word_lengths)
height, width = img.shape[:2]
heat_map_size = (height // 2, width // 2)
mask_shape = (img_shape[1] // 2, img_shape[0] // 2)
confidence_score = np.ones(heat_map_size, dtype=np.float32)
for word_box, confidence_value in zip(word_boxes, confidence_list):
if confidence_value == 1:
continue
tmp_confidence_score = np.zeros(heat_map_size, dtype=np.uint8)
cv2.fillPoly(tmp_confidence_score, [np.array(word_box)], 1)
tmp_confidence_score = np.float32(tmp_confidence_score) * confidence_value
confidence_score = \
np.where(tmp_confidence_score > confidence_score, tmp_confidence_score, confidence_score)
confidence_score_list.append(confidence_score)
fg_mask = np.zeros(heat_map_size, dtype=np.uint8)
cv2.fillPoly(fg_mask, [np.array(word_box) for word_box in word_boxes], 1)
fg_masks.append(fg_mask)
bg_mask = np.zeros(heat_map_size, dtype=np.float32)
bg_mask[:mask_shape[0], :mask_shape[1]] = 1
bg_mask = bg_mask - fg_mask
bg_mask = np.clip(bg_mask, 0, 1)
bg_masks.append(bg_mask)
region_score = gaussian_generator.gen(heat_map_size, np.array(region_box_list) // 2)
region_scores.append(region_score)
affinity_score = gaussian_generator.gen(heat_map_size, np.array(affinity_box_list) // 2)
affinity_scores.append(affinity_score)
max_word_count = np.max(word_count_list)
max_word_count = max(1, max_word_count)
new_word_boxes_list = np.zeros((size, max_word_count, 4, 2), dtype=np.int32)
new_word_lengths_list = np.zeros((size, max_word_count), dtype=np.int32)
for i in range(size):
if word_count_list[i] > 0:
new_word_boxes_list[i, :word_count_list[i]] = np.array(word_boxes_list[i])
new_word_lengths_list[i, :word_count_list[i]] = np.array(word_lengths_list[i])
images = np.array(images)
region_scores = np.array(region_scores, dtype=np.float32)
affinity_scores = np.array(affinity_scores, dtype=np.float32)
confidence_scores = np.array(confidence_score_list, dtype=np.float32)
fg_masks = np.array(fg_masks, dtype=np.float32)
bg_masks = np.array(bg_masks, dtype=np.float32)
inputs = {
'image': images,
'word_box': new_word_boxes_list,
'word_length': new_word_lengths_list,
'region': region_scores,
'affinity': affinity_scores,
'confidence': confidence_scores,
'fg_mask': fg_masks,
'bg_mask': bg_masks,
}
outputs = {
'craft': np.zeros([size])
}
return inputs, outputs
def fake_char_boxes(self, src, word_box, word_length):
img, src_points, crop_points = crop_image(src, word_box, dst_height=64.)
h, w = img.shape[:2]
if min(h, w) == 0:
confidence = 0.5
region_boxes = divide_region(word_box, word_length)
region_boxes = [reorder_points(region_box) for region_box in region_boxes]
return region_boxes, confidence
img = img_normalize(img)
# print(img.shape)
region_score, _ = self.base_model.predict(np.array([img]))
heat_map = region_score[0] * 255.
heat_map = heat_map.astype(np.uint8)
marker_map = watershed(heat_map)
region_boxes = find_box(marker_map)
confidence = cal_confidence(region_boxes, word_length)
if confidence <= 0.5:
confidence = 0.5
region_boxes = divide_region(word_box, word_length)
region_boxes = [reorder_points(region_box) for region_box in region_boxes]
else:
region_boxes = np.array(region_boxes) * 2
region_boxes = enlarge_char_boxes(region_boxes, crop_points)
region_boxes = [un_warping(region_box, src_points, crop_points) for region_box in region_boxes]
# print(word_box, region_boxes)
return region_boxes, confidence
def init_sample(self, flag=False):
for sample_mark in self.sample_mark_list:
if self.fakes[sample_mark]:
sample_list = self.train_sample_lists[sample_mark]
new_sample_list = list()
for sample in sample_list:
if len(sample) == 5:
img_path, word_boxes, words, _, _ = sample
else:
img_path, word_boxes, words, _ = sample
img = load_image(img_path)
char_boxes_list = list()
confidence_list = list()
for word_box, word in zip(word_boxes, words):
char_boxes, confidence = self.fake_char_boxes(img, word_box, len(word))
char_boxes_list.append(char_boxes)
confidence_list.append(confidence)
new_sample_list.append([img_path, word_boxes, words, char_boxes_list, confidence_list])
self.train_sample_lists[sample_mark] = new_sample_list
elif flag:
sample_list = self.train_sample_lists[sample_mark]
new_sample_list = list()
for sample in sample_list:
if len(sample) == 5:
img_path, word_boxes, words, char_boxes_list, _ = sample
else:
img_path, word_boxes, words, char_boxes_list = sample
confidence_list = [1] * len(word_boxes)
new_sample_list.append([img_path, word_boxes, words, char_boxes_list, confidence_list])
self.train_sample_lists[sample_mark] = new_sample_list
def on_epoch_end(self, epoch, logs=None):
self.init_sample()
def on_train_begin(self, logs=None):
self.init_sample(True)
def next_train(self):
while 1:
ret = self.get_batch(self.batch_size)
yield ret
def next_val(self):
while 1:
ret = self.get_batch(self.batch_size, False)
yield ret
def make_image_summary(tensor):
"""
Convert an numpy representation image to Image protobuf.
Copied from https://github.com/lanpa/tensorboard-pytorch/
"""
if len(tensor.shape) == 2:
height, width = tensor.shape
channel = 1
else:
height, width, channel = tensor.shape
if channel == 1:
tensor = tensor[:, :, 0]
image = Image.fromarray(tensor)
output = io.BytesIO()
image.save(output, format='PNG')
image_string = output.getvalue()
output.close()
return tf.Summary.Image(height=height,
width=width,
colorspace=channel,
encoded_image_string=image_string)
class CraftTensorBoard(TensorBoard):
def __init__(self, log_dir, write_graph, test_model, callback_model, data_generator):
self.test_model = test_model
self.callback_model = callback_model
self.data_generator = data_generator
super(CraftTensorBoard, self).__init__(log_dir=log_dir, write_graph=write_graph)
def on_epoch_end(self, epoch, logs=None):
logs.update({'learning_rate': K.eval(self.model.optimizer.lr)})
# self.data_generator.init_sample()
data = next(self.data_generator.next_val())
images = data[0]['image']
word_boxes = data[0]['word_box']
word_lengths = data[0]['word_length']
target_region = data[0]['region']
target_affinity = data[0]['affinity']
confidence_scores = data[0]['confidence']
region, affinity, region_gt, affinity_gt = self.callback_model.predict([images, word_boxes, word_lengths,
target_region, target_affinity,
confidence_scores])
img_summaries = []
for i in range(3):
input_image_summary = make_image_summary(img_unnormalize(images[i]))
pred_region_summary = make_image_summary((region[i] * 255).astype('uint8'))
pred_affinity_summary = make_image_summary((affinity[i] * 255).astype('uint8'))
gt_region_summary = make_image_summary((region_gt[i] * 255).astype('uint8'))
gt_affinity_summary = make_image_summary((affinity_gt[i] * 255).astype('uint8'))
img_summaries.append(tf.Summary.Value(tag='input_image/%d' % i, image=input_image_summary))
img_summaries.append(tf.Summary.Value(tag='region_pred/%d' % i, image=pred_region_summary))
img_summaries.append(tf.Summary.Value(tag='affinity_pred/%d' % i, image=pred_affinity_summary))
img_summaries.append(tf.Summary.Value(tag='region_gt/%d' % i, image=gt_region_summary))
img_summaries.append(tf.Summary.Value(tag='affinity_gt/%d' % i, image=gt_affinity_summary))
tf_summary = tf.Summary(value=img_summaries)
self.writer.add_summary(tf_summary, epoch + 1)
super(CraftTensorBoard, self).on_epoch_end(epoch + 1, logs)
self.test_model.save_weights(r'weights/weight.h5'.format(epoch))
def train():
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu_list
input_image = Input(shape=(None, None, 3), name='image', dtype=tf.float32)
input_box = Input(shape=(None, 4, 2), name='word_box', dtype=tf.int32)
input_word_length = Input(shape=(None,), name='word_length', dtype=tf.int32)
input_region = Input(shape=(None, None), name='region', dtype=tf.float32)
input_affinity = Input(shape=(None, None), name='affinity', dtype=tf.float32)
input_confidence = Input(shape=(None, None), name='confidence', dtype=tf.float32)
input_fg_mask = Input(shape=(None, None), name='fg_mask', dtype=tf.float32)
input_bg_mask = Input(shape=(None, None), name='bg_mask', dtype=tf.float32)
region, affinity = VGG16_UNet(input_tensor=input_image, weights='imagenet')
# if FLAGS.use_fake:
# region_gt, affinity_gt, confidence_gt = \
# Fake(input_box, input_word_length, input_region, input_affinity, input_confidence, name='fake')(region)
# else:
# region_gt = Lambda(lambda x: x)(input_region)
# affinity_gt = Lambda(lambda x: x)(input_affinity)
# confidence_gt = Lambda(lambda x: x)(input_confidence)
region_gt = Lambda(lambda x: x)(input_region)
affinity_gt = Lambda(lambda x: x)(input_affinity)
confidence_gt = Lambda(lambda x: x)(input_confidence)
loss_funs = [craft_mse_loss, craft_mae_loss, craft_huber_loss]
loss_out = Lambda(loss_funs[2], output_shape=(1,), name='craft')(
[region_gt, affinity_gt, region, affinity, confidence_gt, input_fg_mask, input_bg_mask])
model = Model(inputs=[input_image, input_box, input_word_length, input_region,
input_affinity, input_confidence, input_fg_mask, input_bg_mask],
outputs=loss_out)
callback_model = Model(inputs=[input_image, input_box, input_word_length, input_region,
input_affinity, input_confidence],
outputs=[region, affinity, region_gt, affinity_gt])
test_model = Model(inputs=[input_image], outputs=[region, affinity])
test_model.summary()
weight_path = r'weights/weight.h5'
if os.path.exists(weight_path):
test_model.load_weights(weight_path)
# optimizer = SGD(lr=FLAGS.learning_rate, decay=1e-6, momentum=0.9, nesterov=True, clipnorm=5)
optimizer = Adam(lr=FLAGS.learning_rate)
model.compile(loss={'craft': lambda y_true, y_pred: y_pred}, optimizer=optimizer)
true_sample_list = load_data(os.path.join(FLAGS.truth_data_path, r'gt.pkl'))
train_sample_list = true_sample_list
np.random.shuffle(train_sample_list)
if FLAGS.use_fake:
pseudo_sample_list = load_data(os.path.join(FLAGS.pseudo_data_path, r'gt.pkl'))
np.random.shuffle(pseudo_sample_list)
train_generator = SampleGenerator(test_model, [train_sample_list, pseudo_sample_list], [5, 1], [False, True],
FLAGS.img_size, FLAGS.batch_size)
# tensor_board_data_generator = SampleGenerator(test_model, [pseudo_sample_list], [1], [True],
# FLAGS.img_size, FLAGS.batch_size)
else:
train_generator = SampleGenerator(test_model, [train_sample_list], [1], [False],
FLAGS.img_size, FLAGS.batch_size)
# tensor_board_data_generator = SampleGenerator(test_model, [train_sample_list], [1], [False],
# FLAGS.img_size, FLAGS.batch_size)
# train_generator.init_sample(True)
# val_pkl_path = os.path.join(FLAGS.val_data_path, r'gt.pkl')
# if os.path.exists(val_pkl_path):
# val_sample_list = load_data(val_pkl_path)
steps_per_epoch = 1000
tensor_board = CraftTensorBoard(log_dir=r'logs',
write_graph=False,
test_model=test_model,
callback_model=callback_model,
data_generator=train_generator,
)
model.fit_generator(generator=train_generator.next_train(),
steps_per_epoch=steps_per_epoch,
initial_epoch=0,
epochs=FLAGS.max_epochs,
callbacks=[train_generator, tensor_board]
)
if __name__ == '__main__':
train()