-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
173 lines (145 loc) · 9.98 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
import os
import os.path as ops
import tensorflow as tf
import numpy as np
import utils
import sys
from counting_model import counting_net
import cv2
import matplotlib
matplotlib.use('Agg') # not show up just write into disk
from matplotlib import pyplot as plt
import train_data_provider
import validate_data_provider
os.environ['CUDA_VISIBLE_DEVICES'] = "0"
dataset_dir = "./ShanghaiTech_Crowd_Counting_Dataset/part_A_final/"
result_output = open("./result_A", "w")
best_result_output = open("./best_MAE_MSE.txt", "w")
output_density_map = "./output"
batch_size = 1
epoch = 600
loss_c_weight = 0.001
loss_weight = 0.5
if __name__ == "__main__":
train_dataset = train_data_provider.DataSet(ops.join(dataset_dir, 'train.txt'))
validate_dataset = validate_data_provider.DataSet(ops.join(dataset_dir, 'test.txt'))
# declare tensor
x = tf.placeholder(tf.float32, shape=[None, None, None, 3], name="input")
x_2 = tf.placeholder(tf.float32, shape=[None, None, None, 3], name="input_2")
x_4 = tf.placeholder(tf.float32, shape=[None, None, None, 3], name="input_4")
x_8 = tf.placeholder(tf.float32, shape=[None, None, None, 3], name="input_8")
y = tf.placeholder(tf.float32, shape=[None, None, None, 1], name="label")
phase_tensor = tf.placeholder(tf.string, shape=[], name='phase')
MAE = sys.maxsize
MSE = sys.maxsize
# 声明网络，计算loss
density_map_list_typelist = counting_net.CountingNet(phase_tensor).build(x, x_2, x_4, x_8, name='density_map_ret')
density_map_list = tf.convert_to_tensor(density_map_list_typelist)
estimated_counting_2 = tf.reduce_sum(density_map_list[2], reduction_indices=[1, 2, 3], name="crowd_counting_2")
ground_truth_counting = tf.cast(tf.reduce_sum(y, reduction_indices=[1, 2, 3]), tf.float32)
eval_metric_ops = {
'MAE': tf.reduce_mean(tf.abs(tf.subtract(estimated_counting_2, ground_truth_counting)), axis=0, name="MAE"),
'MSE': tf.reduce_mean(tf.square(tf.subtract(ground_truth_counting, estimated_counting_2)), axis=0, name="MSE"),
}
loss_e = tf.constant(0.0, tf.float32)
loss_c = tf.constant(0.0, tf.float32)
n = density_map_list.get_shape().as_list()[0]
for index, density_map in enumerate(density_map_list_typelist):
mse_loss = tf.multiply(loss_weight, tf.losses.mean_squared_error(labels=y, predictions=density_map_list[index]))
ssim_loss = tf.multiply(loss_weight, utils.structural_similarity_index_metric(density_map_list[index], y))
loss_e = tf.add(loss_e, mse_loss)
loss_c = tf.add(loss_c, ssim_loss)
loss = tf.add(loss_e, tf.multiply(loss_c_weight, loss_c))
train_op = tf.train.AdamOptimizer(learning_rate=1e-5).minimize(loss=loss, global_step=tf.train.get_global_step())
# for visulization, 3600 and 182 are consistent with the training and validating numbers of batches in each epoch
z1 = tf.placeholder(tf.float32, shape=[3600, 1], name="loss_train")
h1 = tf.reduce_mean(z1, name="loss_train_mean")
z2 = tf.placeholder(tf.float32, shape=[3600], name="MAE_train")
h2 = tf.reduce_mean(z2, axis=0, name="MAE_train_mean")
z3 = tf.placeholder(tf.float32, shape=[3600], name="MSE_train")
h3 = tf.sqrt(tf.reduce_mean(z3, axis=0), name="RMSE_train")
z4 = tf.placeholder(tf.float32, shape=[182, 1], name="loss_validate")
h4 = tf.reduce_mean(z4, name="loss_validate_mean")
z5 = tf.placeholder(tf.float32, shape=[182], name="MAE_validate")
h5 = tf.reduce_mean(z5, axis=0, name="MAE_validate_mean")
z6 = tf.placeholder(tf.float32, shape=[182], name="MSE_validate")
h6 = tf.sqrt(tf.reduce_mean(z6, axis=0), name="RMSE_validate")
# set tf saver
saver = tf.train.Saver()
# set tf summary
tboard_save_path = './tboard'
loss_train_mean_scalar = tf.summary.scalar(name='train_loss', tensor=h1)
MAE_train_mean_scalar = tf.summary.scalar(name='train_MAE', tensor=h2)
RMSE_train_scalar = tf.summary.scalar(name='train_MSE', tensor=h3)
loss_validate_mean_scalar = tf.summary.scalar(name='validate_loss', tensor=h4)
MAE_validate_mean_scalar = tf.summary.scalar(name='validate_MAE', tensor=h5)
RMSE_validate_scalar = tf.summary.scalar(name='validate_MSE', tensor=h6)
summary_ops = tf.summary.merge_all()
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
# # model restoration
# weights_path = './checkpoint_dir/counting_epoch256.ckpt'
# saver.restore(sess=sess, save_path=weights_path)
summary_writer = tf.summary.FileWriter(tboard_save_path, sess.graph)
for i in range(epoch):
# train
loss_train_ = []
MAE_train_ = []
MSE_train_ = []
image_train_num = 3600
for j in range(image_train_num):
gt_imgs, gt_imgs_2, gt_imgs_4, gt_imgs_8, label_imgs = train_dataset.next_batch(batch_size)
sess.run(train_op, feed_dict={x: gt_imgs, x_2:gt_imgs_2, x_4:gt_imgs_4, x_8:gt_imgs_8, y: label_imgs, phase_tensor: 'train'})
loss_train, metric_train = sess.run([loss, eval_metric_ops], feed_dict={x: gt_imgs, x_2: gt_imgs_2, x_4:gt_imgs_4, x_8: gt_imgs_8, y: label_imgs, phase_tensor: 'train'})
loss_train_.append(loss_train)
MAE_train_.append(metric_train['MAE'])
MSE_train_.append(metric_train['MSE'])
# validate
loss_validate_ = []
MAE_validate_ = []
MSE_validate_ = []
image_validate_num = 182
for k in range(image_validate_num):
gt_imgs_9patches, height, weight, gt_imgs_9patches_2, gt_imgs_9patches_4, gt_imgs_9patches_8, label_imgs, gt_imgs = validate_dataset.next_batch(batch_size)
density_map_list_9patches = sess.run(density_map_list_typelist, feed_dict={x: gt_imgs_9patches, x_2: gt_imgs_9patches_2, x_4: gt_imgs_9patches_4, x_8: gt_imgs_9patches_8, phase_tensor: 'test'})
n = len(density_map_list_9patches) # how many steps
density_map_full_list = np.zeros((n, batch_size, height, weight, 1))
for l in range(n):
density_map_full_list[l][0][0:(height//4+height//8), 0:(weight//4+weight//8), :] = density_map_list_9patches[l][0][0:(height//4+height//8), 0:(weight//4+weight//8), :]
density_map_full_list[l][0][0:(height//4+height//8), (weight//4+weight//8):(weight//2+weight//8), :] = density_map_list_9patches[l][1][0:(height//4+height//8), weight//8:(weight//2-weight//4+weight//8), :]
density_map_full_list[l][0][0:(height//4+height//8), (weight//2+weight//8):weight, :] = density_map_list_9patches[l][2][0:(height//4+height//8), weight//8:(weight-weight//2), :]
density_map_full_list[l][0][(height//4+height//8):(height//2+height//8), 0:(weight//4+weight//8), :] = density_map_list_9patches[l][3][height//8:(height//2-height//4+height//8), 0:(weight//4+weight//8), :]
density_map_full_list[l][0][(height//4+height//8):(height//2+height//8), (weight//4+weight//8):(weight//2+weight//8), :] = density_map_list_9patches[l][4][height//8:(height//2-height//4+height//8), weight//8:(weight//2-weight//4+weight//8), :]
density_map_full_list[l][0][(height//4+height//8):(height//2+height//8), (weight//2+weight//8):weight, :] = density_map_list_9patches[l][5][height//8:(height//2-height//4+height//8), weight//8:(weight-weight//2), :]
density_map_full_list[l][0][(height//2+height//8):height, 0:(weight//4+weight//8), :] = density_map_list_9patches[l][6][height//8:(height-height//2), 0:(weight//4+weight//8), :]
density_map_full_list[l][0][(height//2+height//8):height, (weight//4+weight//8):(weight//2+weight//8), :] = density_map_list_9patches[l][7][height//8:(height-height//2), weight//8:(weight//2-weight//4+weight//8), :]
density_map_full_list[l][0][(height//2+height//8):height, (weight//2+weight//8):weight, :] = density_map_list_9patches[l][8][height//8:(height-height//2), weight//8:(weight-weight//2), :]
loss_validate, metric_validate = sess.run([loss, eval_metric_ops], feed_dict={density_map_list: density_map_full_list, y: label_imgs, phase_tensor: 'test'})
loss_validate_.append(loss_validate)
MAE_validate_.append(metric_validate['MAE'])
MSE_validate_.append(metric_validate['MSE'])
loss_validate_mean = sess.run(h4, feed_dict={z4: loss_validate_})
MAE_validate_mean = sess.run(h5, feed_dict={z5: MAE_validate_})
RMSE_validate = sess.run(h6, feed_dict={z6: MSE_validate_})
# visualize metrics
print('In epoch {}, with loss {}, MAE = {}, MSE = {}\r'.format(i, loss_validate_mean, MAE_validate_mean, RMSE_validate))
result_output.write("epoch: " + str(i) + " loss: " + str(loss_validate_mean) + " MAE: " + str(MAE_validate_mean) + " MSE: " + str(RMSE_validate) + "\r\n")
result_output.flush()
summary_result = sess.run(summary_ops, feed_dict={z1: loss_train_, z2: MAE_train_, z3: MSE_train_, z4: loss_validate_, z5: MAE_validate_, z6: MSE_validate_})
summary_writer.add_summary(summary_result, global_step=i)
# save model
if MAE > MAE_validate_mean and MSE > RMSE_validate:
MAE = MAE_validate_mean
MSE = RMSE_validate
best_result_output.write("epoch: " + str(i) + " loss: " + str(loss_validate_mean) + " MAE: " + str(MAE_validate_mean) + " MSE: " + str(RMSE_validate) + "\r\n")
best_result_output.flush()
model_save_dir = './checkpoint_dir'
if not ops.exists(model_save_dir):
os.makedirs(model_save_dir)
model_name = 'counting_epoch{}.ckpt'.format(i)
model_save_path = ops.join(model_save_dir, model_name)
saver.save(sess, model_save_path)
best_result_output.flush()
result_output.close()
best_result_output.close()