-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
166 lines (124 loc) · 6.37 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
from __future__ import absolute_import
from __future__ import print_function
from __future__ import division
import os
import numpy as np
import tensorflow as tf
from data.dataset_factory import get_dataset
from model.config import cfg
from model.bilinear_cnn import bilinear_cnn
def _preprocess_for_training(input_image, input_height, input_width, image_name, image_label, label_desc):
input_image = tf.expand_dims(input_image, 0)
input_image = tf.cast(input_image, dtype=tf.float32)
_R_MEAN, _G_MEAN, _B_MEAN = cfg._RGB_MEAN
rgb_mean = tf.reshape(np.array([_R_MEAN, _G_MEAN, _B_MEAN]).astype(np.float32), [1,1,1,3])
input_image = input_image - rgb_mean
resized = tf.image.resize_images(input_image, (488, 488))
crop_fn = lambda x: tf.random_crop(x, [448, 448, 3])
processed = tf.map_fn(crop_fn, resized)
flip_fn = lambda x: tf.image.random_flip_left_right(x)
processed = tf.map_fn(flip_fn, processed)
brightness_fn = lambda x: tf.image.random_brightness(x, max_delta=0.2)
processed = tf.map_fn(brightness_fn, processed)
processed = processed[0]
return processed, input_height, input_width, image_name, image_label, label_desc,
def input_pipeline(num_epochs=10):
dataset = get_dataset(dataset_name=cfg.current_dataset, split_name='train')
dataset = dataset.map(_preprocess_for_training)
dataset = dataset.shuffle(buffer_size=500).repeat(num_epochs).batch(cfg.BATCH_SIZE)
iterator = dataset.make_one_shot_iterator()
input_image, input_height, input_width, image_name, image_label, label_desc = iterator.get_next()
return input_image, image_label
def get_init_fn_for_train(vgg_pretrained_path, model_dir, exclude_vars=[]):
if tf.train.latest_checkpoint(model_dir):
tf.logging.info("Ignore pretrained vgg model because a checkpoint file already exists")
return None
variables_to_restore = []
for var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
if 'vgg_16' in var.name:
variables_to_restore.append(var)
if tf.gfile.IsDirectory(vgg_pretrained_path):
vgg_pretrained_path = tf.train.latest_checkpoint(vgg_pretrained_path)
tf.logging.info("Fine tuning from %s" %(vgg_pretrained_path))
if not variables_to_restore:
raise ValueError("variables to restore cannot be empty.")
saver = tf.train.Saver(variables_to_restore, reshape=False)
saver.build()
def callback(scaffold, session):
saver.restore(session, vgg_pretrained_path)
return callback
def get_init_fn_for_finetune(train_dir, finetune_dir):
if tf.train.latest_checkpoint(finetune_dir):
tf.logging.info("Ignore pretrained model because a checkpoint file already exists")
return None
variables_to_restore = []
for var in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
if not 'Momentum' in var.name:
variables_to_restore.append(var)
saver = tf.train.Saver(variables_to_restore)
saver.build()
if tf.gfile.IsDirectory(train_dir):
train_dir = tf.train.latest_checkpoint(train_dir)
def callback(scaffold, sess):
saver.restore(sess, train_dir)
return callback
def bcnn_train_model(features, labels, mode, params):
"""
Args:
features: input image batch
labels: image label
mode: 'TRAIN' | 'EVAL' | 'PREDICT'
params: additional params
"""
logits = bilinear_cnn(features, is_training=True, fine_tuning=False, num_class=cfg.num_classes)
optimizer = tf.train.MomentumOptimizer(learning_rate=cfg.train_base_lr, momentum=cfg.momentum)
loss = tf.losses.softmax_cross_entropy(onehot_labels=tf.one_hot(labels, depth=cfg.num_classes), logits=logits)
tf.summary.scalar('softmax_loss', loss)
global_step = tf.train.get_or_create_global_step()
train_op = optimizer.minimize(loss, global_step)
return tf.estimator.EstimatorSpec(mode=mode,
predictions=logits,
loss=loss,
train_op=train_op,
scaffold=tf.train.Scaffold(init_fn=get_init_fn_for_train(cfg.vgg_pretrained_path, cfg.train_dir)))
def bcnn_finetune_model(features, labels, mode, params):
'''
features: input image batch
labels: image label
mode: 'TRAIN' | 'EVAL' | 'PREDICT'
params: additional params
'''
logits = bilinear_cnn(features, is_training=True, fine_tuning=True, num_class=cfg.num_classes)
optimizer = tf.train.MomentumOptimizer(learning_rate=cfg.finetune_base_lr, momentum=cfg.momentum)
loss = tf.losses.softmax_cross_entropy(onehot_labels=tf.one_hot(labels, depth=cfg.num_classes), logits=logits)
tf.summary.scalar('softmax_loss', loss)
global_step = tf.train.get_or_create_global_step()
train_op = optimizer.minimize(loss, global_step)
return tf.estimator.EstimatorSpec(mode=mode,
predictions=logits,
loss=loss,
train_op=train_op,
scaffold=tf.train.Scaffold(init_fn=get_init_fn_for_finetune(cfg.train_dir, cfg.finetune_dir)))
def main(unused_argv):
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
run_config = tf.estimator.RunConfig()\
.replace(save_summary_steps=2500)\
.replace(log_step_count_steps=10)
train_dir = cfg.train_dir
model = tf.estimator.Estimator(model_fn=bcnn_train_model,
model_dir=train_dir,
config=run_config,
params={})
tf.logging.info('start training model')
model.train(input_fn=lambda :input_pipeline(num_epochs=45), hooks=None, max_steps=25000)
tf.logging.info('Finish training model')
finetune_model = tf.estimator.Estimator(model_fn=bcnn_finetune_model,
model_dir=cfg.finetune_dir,
config=run_config,
params={})
tf.logging.info('start finetuning model')
finetune_model.train(input_fn=lambda :input_pipeline(num_epochs=20), hooks=None)
tf.logging.info('finish finetuning model.')
if __name__ == "__main__":
tf.logging.set_verbosity(tf.logging.INFO)
tf.app.run()