-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_2d.py
545 lines (486 loc) · 27.8 KB
/
train_2d.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
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
from __future__ import print_function
from __future__ import absolute_import
from __future__ import division
import tensorflow as tf
from model import net, resnet, inception_v1, inception_v2, vgg , SE, densenet , inception_v3 , inception_resnet_v2
from model import resnet_attention , bn_inception , inception_v1_non_local,vanila_resnet
import os,gc
import time
import numpy as np
import ucf_2d , hmdb_2d
import ucf_2d_val , hmdb_2d_val
gpu_options = tf.GPUOptions (per_process_gpu_memory_fraction=0.95, allow_growth=True)
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
file_save_path = '/mnt/zhujian/action_recognition'
log_dir = 'log_dir/'
tf.app.flags.DEFINE_string ('video_dir', './UCF-101/', 'the orginal and optical video directory')
tf.app.flags.DEFINE_integer ('image_size', 224, 'the uniform input video size')
tf.app.flags.DEFINE_integer ('num_classes', 101, 'the classes number of the video dataset')
tf.app.flags.DEFINE_integer ('frame_counts', 20, 'the input video frame counts')
tf.app.flags.DEFINE_integer ('batch_size', 8, 'the inputed batch size ')
tf.app.flags.DEFINE_integer ('gpu_nums', 2, 'the inputed batch size ')
tf.app.flags.DEFINE_float ('learning_rate', 1e-3, 'the learning rate of optimizer')
tf.app.flags.DEFINE_float ('weight_decay', 5e-4, 'the learning rate of optimizer')
tf.app.flags.DEFINE_integer ('training_step', 100000, 'the training steps of networks')
tf.app.flags.DEFINE_integer ('epoch', 100, 'the training epoches')
tf.app.flags.DEFINE_string ('dataset', 'UCF101', 'the training dataset')
tf.app.flags.DEFINE_string ('eval_type', 'rgb', 'rgb flow or joint')
tf.app.flags.DEFINE_boolean ('reboot', True, 'reboot traing process if True else False')
tf.app.flags.DEFINE_string ('lr_step', '[150,200]', 'epochs to decay learning rate by 10')
tf.app.flags.DEFINE_boolean ('is_training', True, 'is training or not')
tf.app.flags.DEFINE_integer ('split', 0, 'split number ')
tf.app.flags.DEFINE_float ('rgb_dr', 1.0, 'rgb dropout ratio')
tf.app.flags.DEFINE_float ('flow_dr', 1.0, 'flow dropout ratio')
tf.app.flags.DEFINE_float ('clip_grad', 20, 'clip gradient norm')
tf.app.flags.DEFINE_integer ('num_segments', 3, 'num segments for TSN')
tf.app.flags.DEFINE_integer ('threads',8,'data worker threads')
tf.app.flags.DEFINE_integer ('eval_freq',5,'eval frequence')
tf.app.flags.DEFINE_integer ('new_length',10,'length for flow modality')
tf.app.flags.DEFINE_boolean ('use_pbn', False, 'use partial batch norm')
tf.app.flags.DEFINE_string ('model_type', 'inception_v1', 'use which model vgg,inception,resnet and so on')
tf.app.flags.DEFINE_string ('preprocess_name', 'pytorch', 'preprocess type')
tf.app.flags.DEFINE_string ('opt', 'momentum', 'optimizer type')
Flags = tf.app.flags.FLAGS
tf.logging.set_verbosity (tf.logging.INFO)
rgb_dr = Flags.rgb_dr
flow_dr = Flags.flow_dr
split = Flags.split
use_pbn = Flags.use_pbn
num_segments = Flags.num_segments
eval_type = Flags.eval_type
lr_step = eval (Flags.lr_step)
gpu_nums = Flags.gpu_nums
# is_training = Flags.is_training
threads = Flags.threads
model_type = Flags.model_type
eval_freq = Flags.eval_freq
clip_grad = Flags.clip_grad
new_length = Flags.new_length
preprocess_name = Flags.preprocess_name
opt = Flags.opt
class model ():
def __init__(self,
video_dir='./UCF-101/',
image_size=224,
num_classes=101,
frame_counts=64,
batch_size=3,
learning_rate=1e-4,
num_segments=3,
TRAINING_STEP=10000,
epoch=10,
dataset='UCF101',
reboot=False):
self._video_dir = video_dir
self._IMAGE_SIZE = image_size
self._NUM_CLASSES = num_classes
self._FRAME_COUNTS = frame_counts
self.batch_size = batch_size
self.learning_rate = learning_rate
self.TRAINING_STEP = TRAINING_STEP
self.num_segments = num_segments
self.epoch = epoch
self.dataset_name = dataset
self.reboot = reboot
self.init_config ()
self.init_type ()
self.init_dataset ()
self.init_model ()
self.train ()
self._init_sess ()
self._restore_model ()
def init_config(self):
if eval_type in ['rgb', 'joint']:
self.rgb_train = tf.placeholder (tf.float32,
[None, self._IMAGE_SIZE, self._IMAGE_SIZE, 3])
else:
self.rgb_train = None
if eval_type in ['flow', 'joint']:
self.flow_train = tf.placeholder (tf.float32,
[None, self._IMAGE_SIZE, self._IMAGE_SIZE, new_length * 2])
else:
self.flow_train = None
# init learning rate
self.lr = tf.placeholder (tf.float32, [])
# init label
self.y_ = tf.placeholder (tf.float32, [None, self._NUM_CLASSES])
# init optimizer
if opt == 'momentum':
self.opt = tf.train.MomentumOptimizer (self.lr, 0.9)
elif opt == 'adam':
self.opt = tf.train.AdamOptimizer (self.lr)
self.opt = tf.contrib.estimator.clip_gradients_by_norm(self.opt,clip_grad)
self.is_training = tf.placeholder(tf.bool)
def init_dataset(self):
dataset = Flags.dataset
if dataset == 'UCF101':
self.dataset = ucf_2d.ucf_dataset (split_number=split, is_training_split=True,
batch_size=self.batch_size, epoch=1,new_length=new_length,
eval_type=eval_type,preprocess_name=preprocess_name,
image_size=self._IMAGE_SIZE,
prefetch_buffer_size=self.batch_size).dataset ()
self.val_dataset = ucf_2d_val.ucf_dataset (split_number=split, is_training_split=False,
batch_size=self.batch_size, epoch=1,new_length=new_length,
eval_type=eval_type,preprocess_name=preprocess_name,
image_size=self._IMAGE_SIZE,
prefetch_buffer_size=self.batch_size).dataset ()
elif dataset == 'hmdb51':
self.dataset = hmdb_2d.hmdb_dataset (split_number=split, is_training_split=True,
batch_size=self.batch_size, epoch=1,
eval_type=eval_type,
image_size=self._IMAGE_SIZE,
prefetch_buffer_size=self.batch_size).dataset ()
self.val_dataset = hmdb_2d_val.hmdb_dataset (split_number=split, is_training_split=False,
batch_size=self.batch_size, epoch=1,
eval_type=eval_type,
image_size=self._IMAGE_SIZE,
prefetch_buffer_size=self.batch_size).dataset ()
iter = tf.data.Iterator.from_structure (self.dataset.output_types, self.dataset.output_shapes)
self.next_element = iter.get_next ()
self.training_init_op = iter.make_initializer (self.dataset)
self.validation_init_op = iter.make_initializer (self.val_dataset)
def init_type(self):
if model_type == 'resnet18':
self.base_net = 'resnet_v1_18'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_resnet18/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_resnet18/model.ckpt'
if model_type == 'resnet50':
self.base_net = 'resnet_v1_50'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_snt_resnetV1_50/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_snt_resnetV1_50/model.ckpt'
if model_type == 'resnet50_attention':
self.base_net = 'resnet_v1_50_attention'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_snt_resnetV1_50/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_snt_resnetV1_50/model.ckpt'
elif model_type == 'se_resnet50':
self.base_net = 'se_resnet50'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_se_resnet50/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/rgb_se_resnet50/model.ckpt'
elif model_type == 'resnet101':
self.base_net = 'resnet_v1_101'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_resnet_v1_101/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_resnet_v1_101/model.ckpt'
elif model_type == 'resnet152':
self.base_net = 'resnet_v1_152'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_resnet_v1_152/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_resnet_v1_152/model.ckpt'
elif model_type == 'inception_v1':
self.base_net = 'inceptionv1'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_inception_v1/rgb_inception_v1.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_inception_v1/flow_inception_v1.ckpt'
elif model_type == 'inception_v1_non_local':
self.base_net = 'inceptionv1_non_local'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_inception_v1/rgb_inception_v1.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_inception_v1/flow_inception_v1.ckpt'
elif model_type == 'inception_v2':
self.base_net = 'inceptionv2'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_snt_inception_v2/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_snt_inception_v2/model.ckpt'
elif model_type == 'bn_inception':
self.base_net = 'bn_inception'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_snt_bn_inception/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_snt_bn_inception/model.ckpt'
elif model_type == 'inception_v3':
self.base_net = 'inceptionv3'
self._rgb_reboot_path = '/mnt/zhujian/ckpt/rgb_snt_inception_v3/model.ckpt'
self._flow_reboot_path = '/mnt/zhujian/ckpt/flow_snt_inception_v3/model.ckpt'
elif model_type == 'inception_resnet_v2':
self.base_net = 'inception_resnet_v2'
self._rgb_reboot_path = "/mnt/zhujian/ckpt/rgb_snt_inception_resnet_v2/model.ckpt"
self._flow_reboot_path = "/mnt/zhujian/ckpt/flow_snt_inception_resnet_v2/model.ckpt"
def init_model(self):
if eval_type in ['rgb', 'joint']:
with tf.variable_scope ('RGB', reuse=tf.AUTO_REUSE):
if model_type == 'resnet18':
self.rgb_model = vanila_resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_18')
if model_type == 'resnet50':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'resnet50_attention':
self.rgb_model = resnet_attention.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'se_resnet50':
self.rgb_model = SE.SE_Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'inception_v1':
self.rgb_model = inception_v1.InceptionV1 (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_v1_non_local':
self.rgb_model = inception_v1_non_local.InceptionV1 (num_classes=self._NUM_CLASSES)
elif model_type == 'vgg_16':
self.rgb_model = vgg.vgg (num_classes=self._NUM_CLASSES)
elif model_type == 'resnet101':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_101',
unit_num=[3, 4, 23, 3])
elif model_type == 'resnet152':
self.rgb_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_152',
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v2':
self.rgb_model = inception_v2.InceptionV2 (num_classes=self._NUM_CLASSES)
elif model_type == 'bn_inception':
self.rgb_model = bn_inception.BNInception (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_v3':
self.rgb_model = inception_v3.InceptionV3 (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_resnet_v2':
self.rgb_model = inception_resnet_v2.InceptionResnetV2(num_classes=self._NUM_CLASSES)
if eval_type in ['flow', 'joint']:
with tf.variable_scope ('Flow', reuse=tf.AUTO_REUSE):
if model_type == 'resnet18':
self.flow_model = vanila_resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_18')
if model_type == 'resnet50':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'resnet50_attention':
self.flow_model = resnet_attention.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'se_resnet50':
self.flow_model = SE.SE_Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_50')
elif model_type == 'inception_v1':
self.flow_model = inception_v1.InceptionV1 (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_v1_non_local':
self.flow_model = inception_v1_non_local.InceptionV1 (num_classes=self._NUM_CLASSES)
elif model_type == 'vgg_16':
self.flow_model = vgg.vgg (num_classes=self._NUM_CLASSES)
elif model_type == 'resnet101':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_101',
unit_num=[3, 4, 23, 3])
elif model_type == 'resnet152':
self.flow_model = resnet.Resnet (num_classes=self._NUM_CLASSES, name='resnet_v1_152',
unit_num=[3, 8, 36, 3])
elif model_type == 'inception_v2':
self.flow_model = inception_v2.InceptionV2 (num_classes=self._NUM_CLASSES,use_pbn=use_pbn)
elif model_type == 'inception_v3':
self.flow_model = inception_v3.InceptionV3 (num_classes=self._NUM_CLASSES)
elif model_type == 'inception_resnet_v2':
self.flow_model = inception_resnet_v2.InceptionResnetV2(num_classes=self._NUM_CLASSES)
@staticmethod
def Multigpu_train(model_fn,num_gpus,rgb_input,flow_input):
in_splits = {}
in_splits['rgb'] = tf.split(rgb_input,num_gpus) if rgb_input is not None else None
in_splits['flow'] = tf.split(flow_input,num_gpus) if flow_input is not None else None
out_split = []
for i in range(num_gpus):
if tf.test.is_built_with_cuda():
device_type = 'GPU'
else:
device_type = 'CPU'
with tf.device(tf.DeviceSpec(device_type=device_type,device_index=i)):
with tf.variable_scope(tf.get_variable_scope(),reuse=tf.AUTO_REUSE):
if in_splits['flow'] is None:
out_split.append(model_fn(in_splits['rgb'][i],None))
elif in_splits['rgb'] is None:
out_split.append(model_fn(None,in_splits['flow'][i]))
else:
out_split.append (model_fn (in_splits['rgb'][i], in_splits['flow'][i]))
tf.get_variable_scope().reuse_variables()
out = tf.concat(out_split,axis=0)
return out
def train_model(self,rgb_input,flow_input):
if eval_type in ['rgb', 'joint'] and rgb_input is not None:
rgb_logits, rgb_endpoints = self.rgb_model (rgb_input, is_training=self.is_training,
dropout_keep_prob=rgb_dr)
if eval_type in ['flow', 'joint'] and flow_input is not None:
flow_logits, flow_endpoints = self.flow_model (flow_input, is_training=self.is_training,
dropout_keep_prob=flow_dr)
if eval_type == 'rgb':
model_logits = rgb_logits
elif eval_type == 'flow':
model_logits = flow_logits
elif eval_type == 'joint':
model_logits = rgb_logits + flow_logits
return model_logits
def get_saver(self):
if eval_type in ['rgb', 'joint'] :
rgb_var_map = {}
global_rgb_var_map = {}
for var in tf.global_variables ():
if var.name.split ('/')[0] == 'RGB':
global_rgb_var_map[var.name.replace (':0', '')] = var
if 'logits' not in var.name and 'Logits' not in var.name and 'attention' not in var.name:
rgb_var_map[var.name.replace (':0', '')] = var
self.rgb_saver = tf.train.Saver (var_list=rgb_var_map, reshape=True)
self.global_rgb_saver = tf.train.Saver (var_list=global_rgb_var_map, reshape=True)
if eval_type in ['flow', 'joint'] :
flow_var_map = {}
global_flow_var_map = {}
for var in tf.global_variables ():
if var.name.split ('/')[0] == 'Flow':
global_flow_var_map[var.name.replace (':0', '')] = var
if 'logits' not in var.name and 'Logits' not in var.name and 'attention' not in var.name:
flow_var_map[var.name.replace (':0', '')] = var
self.flow_saver = tf.train.Saver (var_list=flow_var_map, reshape=True)
self.global_flow_saver = tf.train.Saver (var_list=global_flow_var_map, reshape=True)
self.saver = tf.train.Saver (max_to_keep=15, reshape=True)
self.best_saver = tf.train.Saver (reshape=True)
def train(self):
if gpu_nums == 1:
model_logits = self.train_model (self.rgb_train,self.flow_train)
else:
model_logits = self.Multigpu_train(self.train_model,gpu_nums,self.rgb_train,self.flow_train)
self.get_saver()
model_predictions = tf.nn.softmax (model_logits)
self.cross_entropy = tf.reduce_mean (-tf.reduce_sum (self.y_ * tf.log (model_predictions + 1e-10),reduction_indices=[1]))
self.regular_loss = tf.losses.get_regularization_loss ()
self.total_loss = self.cross_entropy + self.regular_loss
self.global_step = tf.Variable (0, trainable=False)
self.train_op = self.opt.minimize (self.total_loss, global_step=self.global_step,colocate_gradients_with_ops=True)
update_ops = tf.group(*tf.get_collection(tf.GraphKeys.UPDATE_OPS))
self.train_op = tf.group(self.train_op, update_ops)
correct_prediction = tf.equal (tf.argmax (model_predictions, 1), tf.argmax (self.y_, 1))
self.train_accuracy = tf.reduce_mean (tf.cast (correct_prediction, tf.float32))
tf.summary.scalar (name='loss', tensor=self.cross_entropy)
tf.summary.scalar (name='train_regular_loss', tensor=self.regular_loss)
tf.summary.scalar (name='train_accuarcy', tensor=self.train_accuracy)
def _init_sess(self):
self.init_op = tf.global_variables_initializer ()
self.merge_summary = tf.summary.merge_all ()
self.sess = tf.Session (config=tf.ConfigProto (gpu_options=gpu_options,allow_soft_placement=True))
self.sess.run (self.init_op)
log_path = os.path.join (file_save_path, log_dir, self.dataset_name, str (split), self.base_net,
eval_type)
if os.path.exists (log_path) is False:
os.makedirs (log_path)
self.train_writer = tf.summary.FileWriter (
log_path, self.sess.graph)
self.train_log_path = os.path.join (log_dir, self.dataset_name, str (split), self.base_net, eval_type)
if os.path.exists (self.train_log_path) is False:
os.makedirs (self.train_log_path)
self.best_val_accuracy = 0
self.best_saver_path = os.path.join ('/mnt/zhujian/action_recognition/two_stream/', self.dataset_name,
str (split), self.base_net, eval_type)
if os.path.exists (self.best_saver_path) is False:
os.makedirs (self.best_saver_path)
def _restore_model(self):
if self.reboot:
if eval_type in ['rgb']:
self.rgb_saver.restore (self.sess, self._rgb_reboot_path)
if eval_type in ['flow']:
self.flow_saver.restore (self.sess, self._flow_reboot_path)
if eval_type in ['joint']:
rgb_path = os.path.join ('/mnt/zhujian/action_recognition/two_stream/', self.dataset_name, str (split),
self.base_net,
'rgb')
rgb_ckpt = tf.train.get_checkpoint_state (rgb_path)
self.global_rgb_saver.restore (self.sess, rgb_ckpt.model_checkpoint_path)
flow_path = os.path.join ('/mnt/zhujian/action_recognition/two_stream/', self.dataset_name, str (split),
self.base_net,
'flow')
flow_ckpt = tf.train.get_checkpoint_state (flow_path)
self.global_flow_saver.restore (self.sess, flow_ckpt.model_checkpoint_path)
else:
path = os.path.join ('/mnt/zhujian/action_recognition/two_stream/', self.dataset_name, str (split),
self.base_net,eval_type,'final')
ckpt = tf.train.get_checkpoint_state (path)
if ckpt is not None:
self.saver.restore (self.sess, ckpt.model_checkpoint_path)
def one_epoch_train(self, epoch):
self.sess.run (self.training_init_op)
one_epoch_loss = []
one_epoch_accuracy = []
feed_dict = {}
gc.collect()
while 1:
try:
read_time = time.time ()
if eval_type == 'rgb':
rgb_file, label = self.sess.run (self.next_element)
elif eval_type == 'flow':
flow_file, label = self.sess.run (self.next_element)
else:
rgb_file, flow_file, label = self.sess.run (self.next_element)
duration = time.time () - read_time
if eval_type in ['rgb', 'joint']:
feed_dict[self.rgb_train] = rgb_file
if eval_type in ['flow', 'joint']:
feed_dict[self.flow_train] = flow_file
feed_dict[self.y_] = label
feed_dict[self.lr] = self.learning_rate
feed_dict[self.is_training] = True
start_time = time.time ()
step, loss, _, out_predictions, r_loss = self.sess.run (
[self.global_step, self.cross_entropy, self.train_op, self.train_accuracy,
self.regular_loss], feed_dict=feed_dict)
train_time = time.time () - start_time
one_epoch_loss.append (loss)
one_epoch_accuracy.append (out_predictions)
print (
'the training step is %d , the out prediction is %.4f,duratio is %.3f,'
'train time is %.2f,the train loss is %.4f,the regular loss is %.4f'
%
(step, out_predictions, duration, train_time, loss, r_loss))
# self.train_writer.add_summary (summary, step)
if eval_type in ['rgb', 'joint']:
del rgb_file
if eval_type in ['flow', 'joint']:
del flow_file
del label
except tf.errors.OutOfRangeError:
if epoch in lr_step:
self.learning_rate /= 10
train_loss = np.mean (one_epoch_loss)
train_accuracy = np.mean (one_epoch_accuracy)
log_file = os.path.join (self.train_log_path, 'train_log.txt')
with open (log_file, 'a') as f:
f.writelines ('At %d epoch , loss is %.3f , accuracy is %.3f\n' %
(epoch, train_loss, train_accuracy))
break
def one_epoch_eval(self, epoch):
self.sess.run (self.validation_init_op)
valid_accuracy = []
valid_loss = []
feed_dict = {}
while 1:
try:
t = time.time ()
if eval_type == 'rgb':
test_rgb, test_label = self.sess.run (self.next_element)
feed_dict[self.rgb_train] = test_rgb
elif eval_type == 'flow':
test_flow, test_label = self.sess.run (self.next_element)
feed_dict[self.flow_train] = test_flow
else:
test_rgb, test_flow, test_label = self.sess.run (self.next_element)
feed_dict[self.flow_train] = test_flow
feed_dict[self.rgb_train] = test_rgb
feed_dict[self.y_] = test_label
feed_dict[self.lr] = self.learning_rate
feed_dict[self.is_training] = False
test_accuracy, test_loss = self.sess.run ([self.train_accuracy, self.cross_entropy],
feed_dict=feed_dict)
valid_loss.append (test_loss)
valid_accuracy.append (test_accuracy)
print ('epoch %d,the test accuracy is %f,the test loss is %.4f ,use time %.2f ' %
(epoch, np.mean (np.array (valid_accuracy)),
np.mean (np.array (valid_loss)), time.time () - t))
if eval_type in ['rgb', 'joint']:
del test_rgb
if eval_type in ['flow', 'joint']:
del test_flow
del test_label
except tf.errors.OutOfRangeError:
log_file = os.path.join (self.train_log_path, 'val_log.txt')
with open (log_file, 'a') as f:
f.writelines ('At %d epoch,the test accuracy is %f,the test loss is %.4f\n' %
(epoch,np.mean (np.array (valid_accuracy)),
np.mean (np.array (valid_loss))))
try:
self.best_saver.save (self.sess, os.path.join (self.best_saver_path, 'final/model.ckpt'))
except:
print('saver last ckpt failed')
if np.mean (np.array (valid_accuracy)) >= self.best_val_accuracy:
self.best_val_accuracy = np.mean (np.array (valid_accuracy))
self.best_saver.save (self.sess, os.path.join (self.best_saver_path, 'model.ckpt'))
break
def main(self):
for i in range (self.epoch):
self.one_epoch_train (i)
if (i+1) % eval_freq == 0 or i == self.epoch - 1:
self.one_epoch_eval (i)
def main(argv):
m = model (video_dir=Flags.video_dir,
image_size=Flags.image_size,
num_classes=Flags.num_classes,
frame_counts=Flags.frame_counts,
batch_size=Flags.batch_size,
learning_rate=Flags.learning_rate,
TRAINING_STEP=Flags.training_step,
epoch=Flags.epoch,
dataset=Flags.dataset,
reboot=Flags.reboot)
m.main ()
if __name__ == '__main__':
tf.app.run (main)