-
Notifications
You must be signed in to change notification settings - Fork 14
/
train.py
220 lines (187 loc) · 10.2 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
import sys, os
#sys.path.insert(0, '/home/eecs568/miniconda3/envs/tensorflow/lib/python3.5/site-packages')
import data_handler
from datetime import datetime
import numpy as np
import random
import tensorflow as tf
from tqdm import tqdm
import cv2, imutils
import vgg
import gen_data_cam as gen_data
import pdb
def delete_network_backups(filename_prefix):
try:
os.remove(filename_prefix+str(".index"))
except OSError:
print(OSError)
pass
try:
os.remove(filename_prefix+str(".meta"))
except OSError:
pass
try:
os.remove(filename_prefix+str(".data-00000-of-00001"))
except OSError:
pass
class trainer():
def __init__(self,path_to_weight, path_to_data, beta, use_quaternion=True, resume_training=False, just_test=False):
self.network_input_size = 224
self.output_dim = 7 if use_quaternion else 6
self.sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
if resume_training:
self.restore_network(path_to_weight)
else:
self.image_inputs = tf.placeholder(tf.float32, [None, self.network_input_size, self.network_input_size, 3])
self.label_inputs = tf.placeholder(tf.float32, [None, self.output_dim]) # [ X Y Z W P Q R]
self.network = vgg.VGG16({'data': self.image_inputs})
self.regen_regression_network()
self.build_loss(beta)
self.saver = tf.train.Saver()
self.merged_summary = tf.summary.merge_all()
now = datetime.now()
self.summary_now = now.strftime("%Y%m%d-%H%M%S")
#self.train_writer = tf.summary.FileWriter('./summary/train/', self.sess.graph)
self.train_writer = tf.summary.FileWriter('./summary/train'+ self.summary_now + "/", self.sess.graph)
self.test_writer = tf.summary.FileWriter( './summary/test'+ now.strftime("%Y%m%d-%H%M%S") + "/")
# initialize
if just_test == False:
self.init_data_handler(path_to_data)
self.init_op = tf.global_variables_initializer() # tf.variables_initializer(self.init_vars )
if not resume_training:
self.sess.run(self.init_op)
self.load_weight(path_to_weight)
print("Model initialized")
def init_data_handler(self,path_to_data):
#self.data_handler = data_handler.Process(path_to_data, 'dataset_train.txt', False)
self.data_handler = gen_data.get_data()
def load_weight(self,path_to_weight):
self.network.load(path_to_weight, self.sess)
#self.saver.restore(self.sess, path_to_weight)
print("Model Restored")
def regen_regression_network(self):
graph = tf.get_default_graph()
fc7 = graph.get_tensor_by_name("fc7/fc7:0")
with tf.variable_scope("fc8", reuse=tf.AUTO_REUSE) as scope:
fc8_input_shape = fc7.get_shape()
feed_in, dim = (fc7, fc8_input_shape[-1].value)
fc8_weights = tf.get_variable('weights', [dim, 2048])
self.network.variable_summaries( fc8_weights, "_weights_fc" )
fc8_biases = tf.get_variable('biases', [2048])
op = tf.nn.xw_plus_b
fc8 = op(feed_in, fc8_weights, fc8_biases, name=scope.name)
with tf.variable_scope("fc9", reuse=tf.AUTO_REUSE) as scope:
fc9_input_shape = fc8.get_shape()
feed_in, dim = (fc8, fc9_input_shape[-1].value)
fc9_weights = tf.get_variable('weights', [dim, self.output_dim])
self.network.variable_summaries( fc9_weights, "_weights_fc" )
fc9_biases = tf.get_variable('biases', [self.output_dim])
op = tf.nn.xw_plus_b
fc9 = op(feed_in, fc9_weights, fc9_biases, name=scope.name)
self.init_vars = [fc8_weights, fc8_biases , fc9_weights, fc9_biases]
self.regression_out = fc9
tf.identity(self.regression_out, name="regression_output")
self.network.variable_summaries(self.regression_out, "regression_output_")
def restore_network(self, path_to_weight):
self.saver = tf.train.import_meta_graph(path_to_weight + ".meta" )
graph = tf.get_default_graph()
self.regression_out = tf.get_default_graph().get_tensor_by_name('fc9/fc9:0')
self.loss = graph.get_tensor_by_name("add:0")
self.train_op = tf.get_default_graph().get_operation_by_name("Adam_minimizer")
self.saver.restore(self.sess, path_to_weight)#tf.train.latest_checkpoint('./'))
self.image_inputs = tf.get_default_graph().get_tensor_by_name('Placeholder:0')
self.label_inputs = tf.get_default_graph().get_tensor_by_name('Placeholder_1:0')
print("Model restored.")
def build_loss(self, beta=100):
self.translation_loss = tf.sqrt(tf.nn.l2_loss(self.regression_out[0:3] - self.label_inputs[0:3]))
self.rotation_loss = tf.sqrt(tf.nn.l2_loss( self.regression_out[3:] - self.label_inputs[3:] ))
self.loss = self.translation_loss + beta * self.rotation_loss
tf.identity(self.loss, name="final_loss")
self.optimizer = tf.train.AdamOptimizer( learning_rate=0.00001, beta1=0.9, beta2=0.999, epsilon=0.00000001, use_locking=False)
slot_var_names = self.optimizer.get_slot_names()
for v in tf.trainable_variables():
for i in slot_var_names:
self.init_vars.append(self.optimizer.get_slot(v, i))
train_vars = tf.trainable_variables()
del train_vars[-6]
del train_vars[-5]
self.compute_gradients = self.optimizer.compute_gradients (self.loss, train_vars)
self.train_op = self.optimizer.apply_gradients(self.compute_gradients , name='Adam_minimizer')
#self.train_op = self.optimizer.minimize(self.loss, name='Adam_minimizer')
grad_summ_op = tf.summary.merge([tf.summary.histogram("%s_grad" % g[1].name, g[0]) for g in self.compute_gradients ])
self.network.variable_summaries(self.translation_loss, "translation_loss_")
self.network.variable_summaries(self.rotation_loss, "rotation_loss_")
self.network.variable_summaries(self.loss, "final_weighted_loss_")
def test(self, img, need_rotate_angle=270, num_random_crops=20):
if img.shape[2] != 3:
print ("We only accept 3-dimensional rgb images")
if img.shape[0] > img.shape[1]:
img = imutils.rotate(img, need_rotate_angle)
img = imutils.resize(img , height=256)
input_size = self.network_input_size # 224 here
input_batch = np.zeros((num_random_crops,input_size,input_size,3))
if num_random_crops == 1:
rand_range = [img.shape[0]-input_size, img.shape[1]-input_size] # height, width
for index in range(num_random_crops):
h = np.random.randint(rand_range[0])
w = np.random.randint(rand_range[1])
input_batch[index, :] = img[h:h+input_size, w:w+input_size, :]
t_r_output = self.sess.run([self.regression_out],
feed_dict={self.image_inputs: input_batch})
return np.mean(t_r_output, axis=0)
else:
tf_output = self.sess.run([self.regression_out],
feed_dict={self.image_inputs: gen_data.centeredCrop(img, input_size)} )
return tf_output
def train(self, batch_size, epochs):
total_loss = 0
total_batch = 125 #int(self.data_handler.numimages() * self.data_handler.genNum * 1.0 / batch_size) #100
if total_batch==0:
pdb.set_trace()
#print("[trainer] Start Training, size of dataset is " +str(self.data_handler.numimages() * self.data_handler.num_crops ))
#pdb.set_trace()
min_loss = 100000000000.0
for epoch in range(epochs):
#self.data_handler.reset()
#self.data_handler.generateData(500)
data_gen = gen_data.gen_data_batch(self.data_handler )
for i in range(total_batch):
#data_runout_flag, one_batch_image , one_batch_label = self.data_handler.fetch(batch_size)
'''
if data_runout_flag == False:
if self.data_handler.remimages() > 0:
self.data_handler.generateData(500)
else:
self.data_handler.reset()
self.data_handler.generateData(500)
data_runout_flag, one_batch_image , one_batch_label = self.data_handler.fetch(batch_size)
'''
one_batch_image, np_poses_x, np_poses_q = next(data_gen)
one_batch_label = np.hstack((np_poses_x, np_poses_q))
feeds ={self.image_inputs: one_batch_image, self.label_inputs: one_batch_label }
#summary, loss, gradients = self.sess.run([self.merged_summary, self.loss, self.compute_gradients ], feed_dict=feeds)
summary, loss, _= self.sess.run([self.merged_summary, self.loss, self.train_op ], feed_dict=feeds)
#self.sess.run([self.train_op], feed_dict=feeds )
print("[Epoch "+str(epoch)+" trainer] Train one batch of size "+str(batch_size)+", loss is "+str(loss))
total_loss += loss
self.train_writer.add_summary(summary, epoch * total_batch + i)
avg_loss = (total_loss)/total_batch
if avg_loss < min_loss:
min_loss = avg_loss
self.saver.save(self.sess, "./"+self.summary_now+"model_epoch_"+str(epoch)+".ckpt")
#if epoch > 0: delete_network_backups("./"+self.summary_now+"model_epoch_"+str(epoch-1)+".ckpt" )
print("[trainer] Epoch " + str(epoch )+ " ends, avg loss =" + "{:.3f}".format(avg_loss))
total_loss = 0
if __name__ == "__main__":
argv = sys.argv
if len(sys.argv) < 5:
argv = ['' for _ in range(6)]
argv[1] = './vgg.data'
argv[2] = './ShopFacade/'
argv[3] = 100
argv[4] = True
argv[5] = bool(int(False))
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
#os.environ['CUDA_VISIBLE_DEVICES'] = '1'
train_thread1 = trainer(argv[1], argv[2], 100, use_quaternion=argv[4], resume_training=False )
train_thread1.train(32, 10)