-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathparall_training_laplace.py
225 lines (205 loc) · 11.1 KB
/
parall_training_laplace.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
import tensorflow as tf
from tensorflow.python.ops import variable_scope as vs
import numpy as np
from laplace_temporal_net import Dataloader,Dataset,create_r3d,SummaryWriter
def averge(tower_grads):
averg_grads_vars = []
for var_and_grad in zip(*tower_grads):
grads = []
for g, _ in var_and_grad:
grad = tf.expand_dims(g, 0)
grads.append(grad)
grad = tf.concat(grads, 0)
grad = tf.reduce_mean(grad, 0)
v = var_and_grad[0][1]
grad_var = (grad, v)
averg_grads_vars.append(grad_var)
return averg_grads_vars
FLAGS = tf.flags.FLAGS
tf.flags.DEFINE_integer('batch_size', 9, 'batch_size to specify the input_placeholder')
tf.flags.DEFINE_boolean('checkpoint', False, "is there snapshot to load to continue training")
tf.flags.DEFINE_boolean('pretrained', False, 'use pretrained models for hot starting')
if __name__ == '__main__':
tf.logging.set_verbosity(tf.logging.INFO)
_SAMPLE_VIDEO_FRAMES = 16
_IMAGE_SIZE = 112
NUM_CLASS = 101
gpus=[0]
_CHECKPOINT_PATHS = {
'pretrained_model': 'pretrained-models/r2plus1-18/Caffe2TfR2.5d.ckpt',
'snapshots': 'saved_models/model.ckpt'
}
is_checkpoint = FLAGS.checkpoint
batch_size = FLAGS.batch_size * len(gpus)
use_pretrained = FLAGS.pretrained
train_set = Dataset.DataSet(clip_length=_SAMPLE_VIDEO_FRAMES,
sample_step=2,
data_root='/home/pr606/Pictures/UCF101DATASET/ucf101',
annotation_path='/home/pr606/Pictures/dataset_annotations/ucf101_json_file/ucf101_01.json',
spatial_transform=None,
mode='train')
validate_set = Dataset.DataSet(clip_length=_SAMPLE_VIDEO_FRAMES,
sample_step=2,
data_root='/home/pr606/Pictures/UCF101DATASET/ucf101',
annotation_path='/home/pr606/Pictures/dataset_annotations/ucf101_json_file/ucf101_01.json',
spatial_transform=None,
mode='validation')
train_generator = Dataloader.DataGenerator(train_set, batch_size=batch_size)
validate_generator = Dataloader.DataGenerator(validate_set, batch_size=batch_size)
num_train = train_generator.__len__()
num_validate = validate_generator.__len__()
print("training data num is %d" % num_train) # 733
print("validation data num is %d" % num_validate) # 291
data = tf.placeholder(shape=(batch_size, _SAMPLE_VIDEO_FRAMES, _IMAGE_SIZE, _IMAGE_SIZE, 1), dtype=tf.float32,name='clips')
label = tf.placeholder(shape=(batch_size, NUM_CLASS), dtype=tf.int32,name='labels')
data_list = tf.split(data,len(gpus),axis=0,name='sep_0_for_%d'%len(gpus))
label_list = tf.split(label,len(gpus),axis=0,name='sep_1_for_%d'%len(gpus))
global_step = tf.Variable(name='global_step', initial_value=0, trainable=False)
# global_step = tf.get_variable("global_step", shape=(),trainable=False,initializer=tf.constant_initializer(0), dtype=tf.int64)
graph = tf.get_default_graph()
learning_rate = tf.train.exponential_decay(learning_rate=0.005, global_step=global_step, decay_steps=1200,
decay_rate=0.98)
opt = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
# opt = tf.train.AdamOptimizer(learning_rate=1e-3)
# train_step = opt.minimize(loss, global_step=global_step)
grads_vars = []
variable_map = {}
regular_map = {}
extra_map = {}
for i in gpus:
with tf.device("/gpu:%d" % i):
# tf.name_scope()不会为变量自动添加前缀，
# 可以保证重用的变量是之前唯一创建的,这与tf.variable_scope()不同
with tf.name_scope('model_%d' % i):
result = create_r3d(
data=data_list[i],
model_depth=10,
num_labels=NUM_CLASS,
num_input_channels=1,
is_decomposed=True,
no_bias=1,
spatial_bn_mom=0.9,
final_temporal_kernel=2,
)
if i == 0:
labels = tf.stop_gradient(label_list[i],
name='disallow_grad_labels%d' % i)
entropy_loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(logits=result, labels=labels),
axis=0)
top1_prediction = tf.math.in_top_k(result, tf.argmax(labels, 1), 1)
top5_prediction = tf.math.in_top_k(result, tf.argmax(labels, 1), 5)
acc_top1 = tf.reduce_mean(tf.cast(top1_prediction, tf.float32))
acc_top5 = tf.reduce_mean(tf.cast(top5_prediction, tf.float32))
grad_var = opt.compute_gradients(entropy_loss,
var_list=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='parall_model'))
grads_vars.append(grad_var)
tf.get_variable_scope().reuse_variables()
# Note: key of variable_map should be specific enough to match the unique variable
for var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
print(var.name, var.shape)
if 'L101' in var.name:
extra_map[var.name.replace(":0", '')] = var
continue
elif 'conv1_middle/kernel' in var.name or 'conv1/kernel' in var.name:
extra_map[var.name.replace(":0", '')] = var
continue
elif 'conv1_middle_spatbn_relu' in var.name:
extra_map[var.name.replace(":0", '')] = var
continue
elif 'Lp_conv_' in var.name or 'global_step' in var.name:
continue
elif 'gaussian_filter' in var.name:
continue
elif 'kernel:0' in var.name:
regular_map[var.name.replace(":0", '')] = var
continue
else:
variable_map[var.name.replace(":0", '')] = var
# weight_deacy_loss = 0.0
weight_deacy_loss = tf.reduce_mean(
[0.005 * tf.nn.l2_loss(var) for var in list(regular_map.values())[-6:]])
else:
labels = tf.stop_gradient(label_list[i], name='disallow_grad_labels%d' % i)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(logits=result, labels=labels), axis=0)
grad_var = opt.compute_gradients(loss,
var_list=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='parall_model'))
grads_vars.append(grad_var)
averge_grads_vars = averge(grads_vars)
# rgb_loader = tf.train.Saver(var_list=dict(variable_map,**regular_map), reshape=True)
rgb_resaver = tf.train.Saver(var_list=dict(variable_map, **regular_map, **extra_map), reshape=True)
"""
with tf.Session() as sess:
global_init = tf.global_variables_initializer()
sess.run(global_init)
rgb_loader.restore(sess, _CHECKPOINT_PATHS['pretrained_model'])
rgb_resaver.save(sess, _CHECKPOINT_PATHS['snapshots'])
"""
train_step = opt.apply_gradients(averge_grads_vars, global_step=global_step)
with tf.name_scope("train_step"):
tf.summary.scalar("accuracy_top1", acc_top1)
tf.summary.scalar("accuracy_top5", acc_top5)
tf.summary.scalar("entropy_loss", entropy_loss)
update_ops = tf.group(*tf.get_collection(tf.GraphKeys.UPDATE_OPS))
with tf.control_dependencies([train_step, update_ops]):
train_op = tf.no_op(name='train')
merge1 = tf.summary.merge(tf.get_collection(tf.GraphKeys.SUMMARIES, scope="train_step"))
train_writer = tf.summary.FileWriter("./logs/train", graph)
validate_writer = SummaryWriter("./logs/validate", graph) # tensorboardX
global_init = tf.global_variables_initializer()
config = tf.ConfigProto(allow_soft_placement=True)
# config.gpu_options.allow_growth=True
config.gpu_options.per_process_gpu_memory_fraction = 0.85
with tf.Session(config=config) as sess:
sess.run(global_init)
feed_dict = {}
if is_checkpoint:
rgb_resaver.restore(sess, _CHECKPOINT_PATHS['snapshots'])
tf.logging.info('RGB checkpoint restored')
'''according to the value of global_step of saved graph if necessary:sess.run(tf.assign(global_step, 500))'''
# sess.run(tf.assign(global_step, 13767))
epoch = 0
while epoch <= 8:
for i, (datas, labells) in enumerate(train_generator):
if i > num_train - 1:
break
feed_dict[data] = datas[:, :, :, :, np.newaxis]
feed_dict[label] = labells
_, Entropyloss, top1, top5, summaries = sess.run([train_op, entropy_loss, acc_top1, acc_top5, merge1],
feed_dict=feed_dict)
step = global_step.eval()
train_writer.add_summary(summaries, step)
print('steps/epochs:{}/{}===> train_loss: {:.6f}, acc_top1: {:.4f}, acc_top5: {:.4f}'
.format(step, epoch, Entropyloss, top1, top5))
top1 = 0.0
top5 = 0.0
Entropyloss = 0.0
k = 0
for i, (datas, labells) in enumerate(validate_generator):
feed_dict[data] = datas[:, :, :, :, np.newaxis]
feed_dict[label] = labells
_Entropyloss, _top1, _top5 = sess.run([entropy_loss, acc_top1, acc_top5],
feed_dict=feed_dict)
Entropyloss += _Entropyloss
top1 += _top1
top5 += _top5
k += 1
print("steps/epochs:{}/{}===>validate_loss:{:.6f},acc_top1:{:.4f},acc_top5: {:.4f}"
.format(k, epoch, _Entropyloss, _top1, _top5))
if i >= num_validate - 1:
break
if k >= 300:
break
top1 = float(top1) / k
top5 = float(top5) / k
Entropyloss = float(Entropyloss) / k
validate_writer.add_scalar("validate/mean_loss", Entropyloss, epoch)
validate_writer.add_scalar("validate/mean_accuracy_top1", top1, epoch)
validate_writer.add_scalar("validate/mean_accuracy_top5", top5, epoch)
epoch += 1
rgb_resaver.save(sess, _CHECKPOINT_PATHS['snapshots'])
train_writer.close()
validate_writer.close()