-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_octunet_sid.py
222 lines (185 loc) · 8.62 KB
/
train_octunet_sid.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
# uniform content loss + adaptive threshold + per_class_input + recursive G
# improvement upon cqf37
from __future__ import division
import os, time, scipy.io, imageio
import rawpy
import PIL
import cv2
import tensorflow as tf
tf.set_random_seed(819)
import tensorflow.contrib.slim as slim
import numpy as np
np.random.seed(819)
# import rawpy
import glob
from loss import *
from octconv_unet import oct_unet
input_dir = '../../datasets/SID/Sony/short/'
gt_dir = '../../datasets/SID/Sony/gt/' # use rgb gt instead of raw
checkpoint_dir = './checkpoint/Sony_oct_ssim/'
result_dir = './result_Sony_oct_ssim/'
if not os.path.exists(checkpoint_dir):
os.mkdir(checkpoint_dir)
if not os.path.exists(result_dir):
os.mkdir(result_dir)
# get train IDs
train_fns = glob.glob(gt_dir + '0*.png')
train_ids = [int(os.path.basename(train_fn)[0:5]) for train_fn in train_fns]
alpha = 0.25 # octave conv 'alpha' param
ps = 512 # patch size for training
lmd = 0.5 # l1 and perceptual loss weight
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
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
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=1)
config = tf.ConfigProto(allow_soft_placement=True, gpu_options=gpu_options)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
in_image = tf.placeholder(tf.float32, [None, None, None, 4])
gt_image = tf.placeholder(tf.float32, [None, None, None, 3])
out_image = oct_unet(in_image, alpha)
# 测试的时候才看metric，训练的时候没有意义
# psnr = tf.reduce_mean(tf.image.psnr(out_image, gt_image, max_val=1.0), axis=0)
# ssim = tf.reduce_mean(tf.image.ssim(out_image, gt_image, max_val=1.0), axis=0)
# tf.summary.image('psnr', psnr)
# tf.summary.image('ssim', ssim)
G_l1loss = tf.reduce_mean(tf.abs(out_image - gt_image))
G_msssimloss = tf.reduce_mean(1 - tf.image.ssim_multiscale(out_image, gt_image, 1.0))
# G_l1loss = tf.reduce_mean(compute_l1_loss(out_image, gt_image))
# features = ["conv1_2", "conv2_2", "conv3_2"]
# G_perceploss = tf.reduce_mean(compute_percep_loss(gt_image, out_image, features, withl1=False))
G_loss = lmd * G_l1loss + (1 - lmd) * G_msssimloss
tf.summary.scalar('l1loss', G_l1loss)
tf.summary.scalar('msssimloss', G_msssimloss)
tf.summary.scalar('sum_loss', G_loss)
t_vars = tf.trainable_variables()
# global_step = tf.Variable(tf.constant(0), trainable=False)
lr = tf.placeholder(tf.float32)
tf.summary.scalar('lr', lr)
# start_lr = 5e-4
# lr = tf.train.exponential_decay(start_lr, global_step, 1000, 0.9, staircase=True)
# tf.summary.scalar('lr', lr)
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
G_opt = tf.train.AdamOptimizer(learning_rate=lr).minimize(G_loss)
merged = tf.summary.merge_all()
log_dir = result_dir + 'logs'
if not os.path.exists(log_dir):
os.mkdir(log_dir)
train_writer = tf.summary.FileWriter(log_dir, sess.graph)
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] * len(train_ids)
input_images = {}
input_images['300'] = [None] * len(train_ids)
input_images['250'] = [None] * len(train_ids)
input_images['100'] = [None] * len(train_ids)
# input_images['300'] = None
# input_images['250'] = None
# input_images['100'] = None
g_loss = np.zeros((5000, 1))
allfolders = glob.glob(result_dir + '*0')
lastepoch = 0
cnt = 0
for folder in allfolders:
lastepoch = np.maximum(lastepoch, int(folder[-4:]))
cnt = lastepoch * len(train_ids)
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)
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*.png' % 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
with tf.device('/cpu:0'):
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
# 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)
# raw = rawpy.imread(in_path)
# input_images[str(ratio)[0:3]] = np.expand_dims(pack_raw(raw), axis=0) * ratio
#
# gt_image_rgb = np.expand_dims(
# np.float32(np.array(PIL.Image.open(gt_path)) / 65535.), axis=0)
# TODO: 目前gt rgb图有问题，导致loss收敛但结果有问题，
# 修改需要读取16bit png 用cv2尝试
print("time raw: ", time.time() - st)
if gt_images[ind] is None:
gt_images[ind] = np.expand_dims(
np.float32(cv2.imread(gt_path, cv2.IMREAD_UNCHANGED)[..., ::-1] / 65535.), axis=0)
print("time gt: ", time.time() - st)
# 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 * 2:yy * 2 + ps * 2, xx * 2:xx * 2 + ps * 2, :]
# gt_patch = gt_image_rgb[:, yy * 2:yy * 2 + ps * 2, xx * 2:xx * 2 + ps * 2, :]
# gt_patch = gt_image_rgb[:, yy * 2:yy * 2 + ps, xx * 2:xx * 2 + ps, :]
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))
# gt_patch = np.transpose(gt_patch, (0, 2, 1))
input_patch = np.minimum(input_patch, 1.0)
# print('**debug shape: ', sess.run(tf.shape(input_patch)),
# sess.run(tf.shape(gt_patch)),
# sess.run(tf.shape(out_image)))
summary, _, G_current, output = sess.run([merged, 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
# if cnt % 20 == 0:
train_writer.add_summary(summary, cnt)
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)
# PIL.Image.fromarray((temp * 255).astype('uint8')).convert('RGB').save(
# result_dir + '%04d/%05d_00_train_%d_pil.jpg' % (epoch, train_id, ratio))
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')