-
Notifications
You must be signed in to change notification settings - Fork 842
/
train_Fuji.py
227 lines (174 loc) · 8.46 KB
/
train_Fuji.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
221
222
223
224
225
226
227
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
input_dir = './dataset/Fuji/short/'
gt_dir = './dataset/Fuji/long/'
checkpoint_dir = './result_Fuji/'
result_dir = './result_Fuji/'
# get train IDs
train_fns = glob.glob(gt_dir + '0*.RAF')
train_ids = [int(os.path.basename(train_fn)[0:5]) for train_fn in train_fns]
ps = 512 # patch size for training
save_freq = 500
def lrelu(x):
return tf.maximum(x * 0.2, x)
def upsample_and_concat(x1, x2, output_channels, in_channels):
pool_size = 2
deconv_filter = tf.Variable(tf.truncated_normal([pool_size, pool_size, output_channels, in_channels], stddev=0.02))
deconv = tf.nn.conv2d_transpose(x1, deconv_filter, tf.shape(x2), strides=[1, pool_size, pool_size, 1])
deconv_output = tf.concat([deconv, x2], 3)
deconv_output.set_shape([None, None, None, output_channels * 2])
return deconv_output
def network(input): # Unet
conv1 = slim.conv2d(input, 32, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv1_1')
conv1 = slim.conv2d(conv1, 32, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv1_2')
pool1 = slim.max_pool2d(conv1, [2, 2], padding='SAME')
conv2 = slim.conv2d(pool1, 64, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv2_1')
conv2 = slim.conv2d(conv2, 64, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv2_2')
pool2 = slim.max_pool2d(conv2, [2, 2], padding='SAME')
conv3 = slim.conv2d(pool2, 128, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv3_1')
conv3 = slim.conv2d(conv3, 128, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv3_2')
pool3 = slim.max_pool2d(conv3, [2, 2], padding='SAME')
conv4 = slim.conv2d(pool3, 256, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv4_1')
conv4 = slim.conv2d(conv4, 256, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv4_2')
pool4 = slim.max_pool2d(conv4, [2, 2], padding='SAME')
conv5 = slim.conv2d(pool4, 512, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv5_1')
conv5 = slim.conv2d(conv5, 512, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv5_2')
up6 = upsample_and_concat(conv5, conv4, 256, 512)
conv6 = slim.conv2d(up6, 256, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv6_1')
conv6 = slim.conv2d(conv6, 256, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv6_2')
up7 = upsample_and_concat(conv6, conv3, 128, 256)
conv7 = slim.conv2d(up7, 128, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv7_1')
conv7 = slim.conv2d(conv7, 128, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv7_2')
up8 = upsample_and_concat(conv7, conv2, 64, 128)
conv8 = slim.conv2d(up8, 64, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv8_1')
conv8 = slim.conv2d(conv8, 64, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv8_2')
up9 = upsample_and_concat(conv8, conv1, 32, 64)
conv9 = slim.conv2d(up9, 32, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv9_1')
conv9 = slim.conv2d(conv9, 32, [3, 3], rate=1, activation_fn=lrelu, scope='g_conv9_2')
conv10 = slim.conv2d(conv9, 27, [1, 1], rate=1, activation_fn=None, scope='g_conv10')
out = tf.depth_to_space(conv10, 3)
return out
def pack_raw(raw):
# pack X-Trans image to 9 channels
im = raw.raw_image_visible.astype(np.float32)
im = np.maximum(im - 1024, 0) / (16383 - 1024) # subtract the black level
img_shape = im.shape
H = (img_shape[0] // 6) * 6
W = (img_shape[1] // 6) * 6
out = np.zeros((H // 3, W // 3, 9))
# 0 R
out[0::2, 0::2, 0] = im[0:H:6, 0:W:6]
out[0::2, 1::2, 0] = im[0:H:6, 4:W:6]
out[1::2, 0::2, 0] = im[3:H:6, 1:W:6]
out[1::2, 1::2, 0] = im[3:H:6, 3:W:6]
# 1 G
out[0::2, 0::2, 1] = im[0:H:6, 2:W:6]
out[0::2, 1::2, 1] = im[0:H:6, 5:W:6]
out[1::2, 0::2, 1] = im[3:H:6, 2:W:6]
out[1::2, 1::2, 1] = im[3:H:6, 5:W:6]
# 1 B
out[0::2, 0::2, 2] = im[0:H:6, 1:W:6]
out[0::2, 1::2, 2] = im[0:H:6, 3:W:6]
out[1::2, 0::2, 2] = im[3:H:6, 0:W:6]
out[1::2, 1::2, 2] = im[3:H:6, 4:W:6]
# 4 R
out[0::2, 0::2, 3] = im[1:H:6, 2:W:6]
out[0::2, 1::2, 3] = im[2:H:6, 5:W:6]
out[1::2, 0::2, 3] = im[5:H:6, 2:W:6]
out[1::2, 1::2, 3] = im[4:H:6, 5:W:6]
# 5 B
out[0::2, 0::2, 4] = im[2:H:6, 2:W:6]
out[0::2, 1::2, 4] = im[1:H:6, 5:W:6]
out[1::2, 0::2, 4] = im[4:H:6, 2:W:6]
out[1::2, 1::2, 4] = im[5:H:6, 5:W:6]
out[:, :, 5] = im[1:H:3, 0:W:3]
out[:, :, 6] = im[1:H:3, 1:W:3]
out[:, :, 7] = im[2:H:3, 0:W:3]
out[:, :, 8] = im[2:H:3, 1:W:3]
return out
sess = tf.Session()
in_image = tf.placeholder(tf.float32, [None, None, None, 9])
gt_image = tf.placeholder(tf.float32, [None, None, None, 3])
out_image = network(in_image)
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
in_images = {}
in_images['300'] = [None] * len(train_ids)
in_images['250'] = [None] * len(train_ids)
in_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*.RAF' % train_id)
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*.RAF' % 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 in_images[str(ratio)[0:3]][ind] is None:
raw = rawpy.imread(in_path)
in_images[str(ratio)[0:3]][ind] = np.expand_dims(pack_raw(raw), axis=0) * ratio
gt_raw = rawpy.imread(gt_path)
im = gt_raw.postprocess(use_camera_wb=True, half_size=False, no_auto_bright=True, output_bps=16)
gt_images[ind] = np.expand_dims(np.float32(im / 65535.0), axis=0)
# crop
H = in_images[str(ratio)[0:3]][ind].shape[1]
W = in_images[str(ratio)[0:3]][ind].shape[2]
xx = np.random.randint(0, W - ps)
yy = np.random.randint(0, H - ps)
input_patch = in_images[str(ratio)[0:3]][ind][:, yy:yy + ps, xx:xx + ps, :]
gt_patch = gt_images[ind][:, yy * 3:yy * 3 + ps * 3, xx * 3:xx * 3 + ps * 3, :]
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')