/
Train.py
208 lines (160 loc) · 7.61 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
# Copyright (C) 2019 * Ltd. All rights reserved.
# author : SangHyeon Jo <josanghyeokn@gmail.com>
import os
import cv2
import sys
import glob
import time
import random
import numpy as np
import tensorflow as tf
from Define import *
from Utils import *
from Teacher import *
from FCOS import *
from FCOS_Loss import *
from FCOS_Utils import *
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
# 1. dataset
train_data_list = np.load('./dataset/train_detection.npy', allow_pickle = True)
valid_data_list = np.load('./dataset/validation_detection.npy', allow_pickle = True)
valid_count = len(valid_data_list)
open('log.txt', 'w')
log_print('[i] Train : {}'.format(len(train_data_list)))
log_print('[i] Valid : {}'.format(len(valid_data_list)))
# 2. build
input_var = tf.placeholder(tf.float32, [BATCH_SIZE, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNEL])
is_training = tf.placeholder(tf.bool)
fcos_dic, fcos_sizes = FCOS(input_var, is_training)
fcos_utils = FCOS_Utils(fcos_sizes)
pred_bboxes_op = fcos_dic['pred_bboxes']
pred_centers_op = fcos_dic['pred_centers']
pred_classes_op = fcos_dic['pred_classes']
log_print('[i] pred_bboxes_op : {}'.format(pred_bboxes_op))
log_print('[i] pred_centers_op : {}'.format(pred_centers_op))
log_print('[i] pred_classes_op : {}'.format(pred_classes_op))
_, fcos_size, _ = pred_bboxes_op.shape.as_list()
gt_bboxes_var = tf.placeholder(tf.float32, [BATCH_SIZE, fcos_size, 4])
gt_centers_var = tf.placeholder(tf.float32, [BATCH_SIZE, fcos_size, 1])
gt_classes_var = tf.placeholder(tf.float32, [BATCH_SIZE, fcos_size, CLASSES])
log_print('[i] gt_bboxes_var : {}'.format(gt_bboxes_var))
log_print('[i] gt_centers_var : {}'.format(gt_centers_var))
log_print('[i] gt_classes_var : {}'.format(gt_classes_var))
pred_ops = [pred_bboxes_op, pred_centers_op, pred_classes_op]
gt_ops = [gt_bboxes_var, gt_centers_var, gt_classes_var]
loss_op, focal_loss_op, center_loss_op, giou_loss_op = FCOS_Loss(pred_ops, gt_ops)
vars = tf.trainable_variables()
l2_reg_loss_op = tf.add_n([tf.nn.l2_loss(var) for var in vars]) * WEIGHT_DECAY
loss_op += l2_reg_loss_op
learning_rate_var = tf.placeholder(tf.float32)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
# train_op = tf.train.AdamOptimizer(learning_rate_var).minimize(loss_op)
train_op = tf.train.MomentumOptimizer(learning_rate_var, momentum = 0.9).minimize(loss_op)
train_summary_dic = {
'Loss/Total_Loss' : loss_op,
'Loss/Focal_Loss' : focal_loss_op,
'Loss/Center_Loss' : center_loss_op,
'Loss/GIoU_Loss' : giou_loss_op,
'Loss/L2_Regularization_Loss' : l2_reg_loss_op,
'Learning_rate' : learning_rate_var,
}
train_summary_list = []
for name in train_summary_dic.keys():
value = train_summary_dic[name]
train_summary_list.append(tf.summary.scalar(name, value))
train_summary_op = tf.summary.merge(train_summary_list)
log_image_var = tf.placeholder(tf.float32, [None, SAMPLE_IMAGE_HEIGHT, SAMPLE_IMAGE_WIDTH, IMAGE_CHANNEL])
log_image_op = tf.summary.image('Image/Train', log_image_var, SAMPLES)
# 3. train
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# '''
pretrained_vars = []
for var in vars:
if 'resnet_v1_50' in var.name:
pretrained_vars.append(var)
pretrained_saver = tf.train.Saver(var_list = pretrained_vars)
pretrained_saver.restore(sess, './resnet_v1_model/resnet_v1_50.ckpt')
# '''
saver = tf.train.Saver(max_to_keep = 100)
saver.restore(sess, './model/FCOS_{}.ckpt'.format(115000))
learning_rate = INIT_LEARNING_RATE
log_print('[i] max_iteration : {}'.format(MAX_ITERATION))
log_print('[i] decay_iteration : {}'.format(DECAY_ITERATIONS))
loss_list = []
focal_loss_list = []
center_loss_list = []
giou_loss_list = []
l2_reg_loss_list = []
train_time = time.time()
train_writer = tf.summary.FileWriter('./logs/train')
train_threads = []
for i in range(NUM_THREADS):
train_thread = Teacher('./dataset/train_detection.npy', fcos_sizes, debug = False)
train_thread.start()
train_threads.append(train_thread)
sample_data_list = train_data_list[:SAMPLES]
for iter in range(1, MAX_ITERATION + 1):
if iter in DECAY_ITERATIONS:
learning_rate /= 10
log_print('[i] learning rate decay : {} -> {}'.format(learning_rate * 10, learning_rate))
# Thread
find = False
while not find:
for train_thread in train_threads:
if train_thread.ready:
find = True
batch_image_data, batch_encode_bboxes, batch_encode_centers, batch_encode_classes = train_thread.get_batch_data()
break
_feed_dict = {input_var : batch_image_data, gt_bboxes_var : batch_encode_bboxes, gt_centers_var : batch_encode_centers, gt_classes_var : batch_encode_classes,
is_training : True, learning_rate_var : learning_rate}
log = sess.run([train_op, loss_op, focal_loss_op, center_loss_op, giou_loss_op, l2_reg_loss_op, train_summary_op], feed_dict = _feed_dict)
# print(log[1:-1])
if np.isnan(log[1]):
print('[!]', log[1:-1])
input()
loss_list.append(log[1])
focal_loss_list.append(log[2])
center_loss_list.append(log[3])
giou_loss_list.append(log[4])
l2_reg_loss_list.append(log[5])
train_writer.add_summary(log[6], iter)
if iter % LOG_ITERATION == 0:
loss = np.mean(loss_list)
focal_loss = np.mean(focal_loss_list)
center_loss = np.mean(center_loss_list)
giou_loss = np.mean(giou_loss_list)
l2_reg_loss = np.mean(l2_reg_loss_list)
train_time = int(time.time() - train_time)
log_print('[i] iter : {}, loss : {:.4f}, focal_loss : {:.4f}, center_loss : {:.4f}, giou_loss : {:.4f}, l2_reg_loss : {:.4f}, train_time : {}sec'.format(iter, loss, focal_loss, center_loss, giou_loss, l2_reg_loss, train_time))
loss_list = []
focal_loss_list = []
center_loss_list = []
giou_loss_list = []
l2_reg_loss_list = []
train_time = time.time()
if iter % SAMPLE_ITERATION == 0:
sample_images = []
batch_image_data = np.zeros((BATCH_SIZE, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNEL), dtype = np.float32)
for i, data in enumerate(sample_data_list):
image_name, gt_bboxes, gt_classes = data
image = cv2.imread(TRAIN_DIR + image_name)
tf_image = cv2.resize(image, (IMAGE_WIDTH, IMAGE_HEIGHT), interpolation = cv2.INTER_CUBIC)
batch_image_data[i] = tf_image.copy()
total_pred_bboxes, total_pred_centers, total_pred_classes = sess.run([pred_bboxes_op, pred_centers_op, pred_classes_op], feed_dict = {input_var : batch_image_data, is_training : False})
for i in range(BATCH_SIZE):
image = batch_image_data[i]
pred_bboxes, pred_classes = fcos_utils.Decode(total_pred_bboxes[i], total_pred_centers[i], total_pred_classes[i], [IMAGE_WIDTH, IMAGE_HEIGHT], detect_threshold = 0.20)
for bbox, class_index in zip(pred_bboxes, pred_classes):
xmin, ymin, xmax, ymax = bbox[:4].astype(np.int32)
conf = bbox[4]
class_name = CLASS_NAMES[class_index]
string = "{} : {:.2f}%".format(class_name, conf * 100)
cv2.putText(image, string, (xmin, ymin - 10), 1, 1, (0, 255, 0))
cv2.rectangle(image, (xmin, ymin), (xmax, ymax), (0, 255, 0), 2)
image = cv2.resize(image, (SAMPLE_IMAGE_WIDTH, SAMPLE_IMAGE_HEIGHT))
sample_images.append(image.copy())
image_summary = sess.run(log_image_op, feed_dict = {log_image_var : sample_images})
train_writer.add_summary(image_summary, iter)
if iter % SAVE_ITERATION == 0:
saver.save(sess, './model/FCOS_{}.ckpt'.format(iter))