-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_Sony.py
200 lines (160 loc) · 6.89 KB
/
train_Sony.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
# uniform content loss + adaptive threshold + per_class_input + recursive G
# improvement upon cqf37
from __future__ import division
import os, time, scipy.io
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
import rawpy
import glob
from model import *
import os
import scipy.misc
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
#/workspace/sony_data/Sony/long
input_dir = '/workspace/sony_data/Sony/short/'
gt_dir = '/workspace/sony_data/Sony/long/'
checkpoint_dir = './result_Sony/'
result_dir = './result_Sony/'
# get train IDs
train_fns = glob.glob(gt_dir + '0*.ARW')
print("train_fns = glob.glob(gt_dir + '0*.ARW')")
print(train_fns)
train_ids = [int(os.path.basename(train_fn)[0:5]) for train_fn in train_fns]
print("train_ids = [int(os.path.basename(train_fn)[0:5]) for train_fn in train_fns]")
print(train_ids)
ps = 512 # patch size for training
save_freq = 500
DEBUG = 0
if DEBUG == 1:
save_freq = 2
train_ids = train_ids[0:5]
def pack_raw(raw):
# pack Bayer image to 4 channels，SID数据集是14位，所以这里是16383
im = raw.raw_image_visible.astype(np.float32)
im = np.maximum(im - 512, 0) / (16383 - 512) # subtract the black level
im = np.expand_dims(im, axis=2)
img_shape = im.shape
H = img_shape[0]
W = img_shape[1]
##
out = np.concatenate((im[0:H:2, 0:W:2, :],
im[0:H:2, 1:W:2, :],
im[1:H:2, 1:W:2, :],
im[1:H:2, 0:W:2, :]), axis=2)
return out
def pack_raw_gt(raw):
# pack Bayer image to 4 channels，SID数据集是14位，所以这里是16383
im = raw.raw_image_visible.astype(np.float32)
im = np.maximum(im-512, 0) / (16383-512) # subtract the black level
im = np.expand_dims(im, axis=2)
img_shape = im.shape
H = img_shape[0]
W = img_shape[1]
##
out = np.concatenate((im[0:H:2, 0:W:2, :],
im[0:H:2, 1:W:2, :],
im[1:H:2, 1:W:2, :],
im[1:H:2, 0:W:2, :]), axis=2)
print(out.shape)
return out
sess = tf.Session()
in_image = tf.placeholder(tf.float32, [None, None, None, 4])
gt_image = tf.placeholder(tf.float32, [None, None, None, 4])
out_image = denoise_net(in_image)
### 这个loss应该对应的是网络输出out_image和gt_image的L1范数
#这里是不是要确认一下这点
G_loss = tf.reduce_mean(tf.abs(out_image - gt_image))
t_vars = tf.trainable_variables()
lr = tf.placeholder(tf.float32)
G_opt = tf.train.AdamOptimizer(learning_rate=lr).minimize(G_loss)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt:
print('loaded ' + ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
# Raw data takes long time to load. Keep them in memory after loaded.
gt_images = [None] * 6000
input_images = {}
input_images['300'] = [None] * len(train_ids)
input_images['250'] = [None] * len(train_ids)
input_images['100'] = [None] * len(train_ids)
g_loss = np.zeros((5000, 1))
allfolders = glob.glob(result_dir + '*0')
lastepoch = 0
for folder in allfolders:
lastepoch = np.maximum(lastepoch, int(folder[-4:]))
learning_rate = 1e-4
for epoch in range(lastepoch, 4001):
if os.path.isdir(result_dir + '%04d' % epoch):
continue
cnt = 0
if epoch > 2000:
learning_rate = 1e-5
for ind in np.random.permutation(len(train_ids)):
# get the path from image id
train_id = train_ids[ind]
in_files = glob.glob(input_dir + '%05d_00*.ARW' % train_id)
print("in_files = glob.glob(input_dir + '%05d_00*.ARW' % train_id)")
print(train_id)
print(in_files)
if len(in_files) == 0:
continue
in_path = in_files[np.random.random_integers(0, len(in_files) - 1)]
in_fn = os.path.basename(in_path)
gt_files = glob.glob(gt_dir + '%05d_00*.ARW' % train_id)
gt_path = gt_files[0]
gt_fn = os.path.basename(gt_path)
in_exposure = float(in_fn[9:-5])
gt_exposure = float(gt_fn[9:-5])
ratio = min(gt_exposure / in_exposure, 300)
st = time.time()
cnt += 1
if input_images[str(ratio)[0:3]][ind] is None:
raw = rawpy.imread(in_path)
input_images[str(ratio)[0:3]][ind] = np.expand_dims(pack_raw(raw), axis=0) * ratio
temp = input_images[str(ratio)[0:3]][ind]
print("temp = input_images[str(ratio)[0:3]][ind]")
print(temp.shape)
gt_raw = rawpy.imread(gt_path)
#im = gt_raw
im=pack_raw_gt(gt_raw)
gt_images[ind] = np.expand_dims(pack_raw_gt(gt_raw), axis=0)
temp_gt = gt_images[ind]
print("gt_images[ind].shape")
print(temp_gt.shape)
# TODO pack the im(ground truth) to RGGB tensor
#im = gt_raw.postprocess(use_camera_wb=True, half_size=False, no_auto_bright=True, output_bps=16)
# TODO the next line code below may be wrong
####gt_images[ind] = np.expand_dims(np.float32(im / 16383.0), axis=0)
# crop
H = input_images[str(ratio)[0:3]][ind].shape[1]
W = input_images[str(ratio)[0:3]][ind].shape[2]
xx = np.random.randint(0, W - ps)
yy = np.random.randint(0, H - ps)
input_patch = input_images[str(ratio)[0:3]][ind][:, yy:yy + ps, xx:xx + ps, :]
gt_patch = gt_images[ind][:, yy:yy + ps, xx:xx + ps, :]
##gt_patch = gt_images[ind][:, yy * 2:yy * 2 + ps * 2, xx * 2:xx * 2 + ps * 2, :]
if np.random.randint(2, size=1)[0] == 1: # random flip
input_patch = np.flip(input_patch, axis=1)
gt_patch = np.flip(gt_patch, axis=1)
if np.random.randint(2, size=1)[0] == 1:
input_patch = np.flip(input_patch, axis=2)
gt_patch = np.flip(gt_patch, axis=2)
if np.random.randint(2, size=1)[0] == 1: # random transpose
input_patch = np.transpose(input_patch, (0, 2, 1, 3))
gt_patch = np.transpose(gt_patch, (0, 2, 1, 3))
input_patch = np.minimum(input_patch, 1.0)
_, G_current, output = sess.run([G_opt, G_loss, out_image],
feed_dict={in_image: input_patch, gt_image: gt_patch, lr: learning_rate})
output = np.minimum(np.maximum(output, 0), 1)
g_loss[ind] = G_current
print("%d %d Loss=%.3f Time=%.3f" % (epoch, cnt, np.mean(g_loss[np.where(g_loss)]), time.time() - st))
## if epoch % save_freq == 0:
## if not os.path.isdir(result_dir + '%04d' % epoch):
## os.makedirs(result_dir + '%04d' % epoch)
## temp = np.concatenate((gt_patch[0, :, :, :], output[0, :, :, :]), axis=1)
## scipy.misc.toimage(temp * 255, high=255, low=0, cmin=0, cmax=255).save(
## result_dir + '%04d/%05d_00_train_%d.jpg' % (epoch, train_id, ratio))
saver.save(sess, checkpoint_dir + 'model.ckpt')