/
train.py
executable file
·335 lines (292 loc) · 14.2 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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
import argparse
import json
import os
import torch
import torch.nn as nn
import numpy as np
from PIL import Image
from torch.utils.data import DataLoader
import torch.distributions as tdist
from torch.optim.lr_scheduler import MultiStepLR
from unet_model_ori import UNet
from unet_attention_decouple import AttenUnet_style
from data_utils import autoTrainSetRaw2jpgProcessed
from exposure_module import ExposureNet
from isp import isp
os.environ["CUDA_VISIBLE_DEVICES"]="7"
def load_checkpoint(checkpoint_path, model, optimizer):
assert os.path.isfile(checkpoint_path)
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
iteration = checkpoint_dict['iteration']
optimizer.load_state_dict(checkpoint_dict['optimizer'])
model_for_loading = checkpoint_dict['model']
model.load_state_dict(model_for_loading.state_dict())
print("Loaded checkpoint '{}' (iteration {})" .format(
checkpoint_path, iteration))
return model, optimizer, iteration
def save_checkpoint(model, net_type, optimizer, learning_rate, iteration, filepath, parallel):
print("Saving model and optimizer state at iteration {} to {}".format(
iteration, filepath))
if net_type == "exposure":
model_for_saving = ExposureNet().cuda()
if net_type == "u_net":
model_for_saving = UNet(**network_config).cuda()
if net_type == "unet_att_style":
model_for_saving = AttenUnet_style(**network_config2).cuda()
if parallel:
model_for_saving = nn.DataParallel(model_for_saving)
model_for_saving.load_state_dict(model.state_dict())
torch.save({'model': model_for_saving,
'iteration': iteration,
'optimizer': optimizer.state_dict(),
'learning_rate': learning_rate}, filepath)
def save_training_images_raw(img_list, image_path, img_name, alpha):
print("Saving output images")
b,c,h,w = img_list[0].shape
batch_list = []
for img in img_list:
clip(img)
tmp_batch = isp(img[0,:,:,:], img_name[0], data_config["file_list"], alpha[0])
for i in range(b-1):
tmp_batch = np.concatenate((tmp_batch, isp(img[i+1,:,:,:], img_name[i+1], data_config["file_list"], alpha[i+1])), axis=1)
batch_list.append(tmp_batch)
new_img_array = np.concatenate(batch_list, axis=2) * 255
new_img = Image.fromarray(np.transpose(new_img_array, [1,2,0]).astype('uint8'), 'RGB')
new_img.save(image_path, quality=95)
def clip(img):
img[img>1] = 1
img[img<0] = 0
def get_variance_map(input_raw, shot_noise, read_noise, mul=None):
if not type(mul) == type(None):
shot_noise = shot_noise * mul
read_noise = read_noise * mul * mul
b, c, h, w = input_raw.size()
read_noise = torch.unsqueeze(read_noise, 2)
read_noise = torch.unsqueeze(read_noise, 3)
read_noise = read_noise.repeat(1,1,h,w)
shot_noise = torch.unsqueeze(shot_noise, 2)
shot_noise = torch.unsqueeze(shot_noise, 3)
shot_noise = shot_noise.repeat(1,1,h,w)
variance = torch.add(input_raw * shot_noise, read_noise)
n = tdist.Normal(loc=torch.zeros_like(variance), scale=torch.sqrt(variance))
noise = n.sample()
var_map = input_raw + noise
return var_map
def train(output_directory, epochs, learning_rate1, learning_rate2, learning_rate3, aperture,
iters_per_checkpoint, batch_size, epoch_size, loss_type1, loss_type2, loss_type3,
net_type, net_type_ap, seed, checkpoint_path1, checkpoint_path2, checkpoint_path3, residual_learning1,
residual_learning2, parallel, variance_map, isp_save, multi_stage=None, multi_stage2=None):
# set manual seed
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
# build exposure module
model_exposure = ExposureNet().cuda()
# build noise model
if net_type == "u_net":
model_noise = UNet(**network_config).cuda()
else:
print("unsupported network type")
return 0
# build aperture model
if aperture:
if net_type_ap == "unet_att_style":
model_aperture = AttenUnet_style(**network_config2).cuda()
else:
print("unsupported network type")
return 0
if parallel:
model_exposure = nn.DataParallel(model_exposure)
model_noise = nn.DataParallel(model_noise)
if aperture:
model_aperture = nn.DataParallel(model_aperture)
optimizer_1 = torch.optim.Adam(model_exposure.parameters(), lr=learning_rate1)
optimizer_2 = torch.optim.Adam(model_noise.parameters(), lr=learning_rate2)
scheduler_2 = MultiStepLR(optimizer_2, milestones=[20, 40], gamma=0.1)
if aperture:
optimizer_3 = torch.optim.Adam(model_aperture.parameters(), lr=learning_rate3)
scheduler_3 = MultiStepLR(optimizer_3, milestones=[20, 40], gamma=0.1)
# Load checkpoint if one exists
iteration = 0
if checkpoint_path1 != "":
model_exposure, optimizer_1, iteration = load_checkpoint(checkpoint_path1, model_exposure, optimizer_1)
if checkpoint_path2 != "":
model_noise, optimizer_2, iteration = load_checkpoint(checkpoint_path2, model_noise,
optimizer_2)
if checkpoint_path3 !="":
model_aperture, optimizer_3, iteration = load_checkpoint(checkpoint_path3, model_aperture,
optimizer_3)
iteration += 1
# build dataset
trainset = autoTrainSetRaw2jpgProcessed(**data_config)
epoch_size = min(len(trainset), epoch_size)
train_sampler = torch.utils.data.RandomSampler(trainset, True, epoch_size)
train_loader = DataLoader(trainset, num_workers=5, shuffle=False,
sampler=train_sampler,
batch_size=batch_size,
pin_memory=False,
drop_last=True)
# Get shared output_directory ready
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
os.chmod(output_directory, 0o775)
print("output directory", output_directory)
model_noise.train()
model_exposure.train()
if aperture:
model_aperture.train()
epoch_offset = max(0, int(iteration / len(train_loader)))
# ================ MAIN TRAINNIG LOOP! ===================
for epoch in range(epoch_offset, epochs):
print("Epoch: {}".format(epoch))
for i, batch in enumerate(train_loader):
model_exposure.zero_grad()
model_noise.zero_grad()
if aperture:
model_aperture.zero_grad()
exp_params, ap_params, noise_params, input_raw, input_jpg, output_raw, mask, \
input_shot_noise, input_read_noise, output_shot_noise, output_read_noise, \
img_name = batch
if aperture:
ap_params = torch.autograd.Variable(ap_params.cuda())
exp_params = torch.autograd.Variable(exp_params.cuda())
noise_params = torch.autograd.Variable(noise_params.cuda())
input_shot_noise = torch.autograd.Variable(input_shot_noise.cuda())
input_read_noise = torch.autograd.Variable(input_read_noise.cuda())
output_shot_noise = torch.autograd.Variable(output_shot_noise.cuda())
output_read_noise = torch.autograd.Variable(output_read_noise.cuda())
mask = mask.cuda()
input_raw = torch.autograd.Variable(input_raw.cuda())
output_raw = torch.autograd.Variable(output_raw.cuda())
# simple exposure correction
output_exp, exp_params_m = model_exposure([exp_params, input_raw])
# noise correction
# Estimate variance map
if variance_map:
# input variance map
variance_input = get_variance_map(output_exp, input_shot_noise, input_read_noise, exp_params_m)
# output variance map (As long as the output iso is known, it can be estimated)
input_cat = torch.cat([output_exp, variance_input], 1)
if net_type == "u_net":
output_ns = model_noise(input_cat)
else:
if net_type == "u_net":
output_ns = model_noise(output_exp)
#aperture module
if aperture:
output_ap = model_aperture([ap_params, output_ns+output_exp])
# define exposure loss
if loss_type1 == "l1":
loss_f1 = torch.nn.L1Loss()
elif loss_type1 == "l2":
loss_f1 = torch.nn.MSELoss()
loss1 = loss_f1(output_exp*mask, output_raw*mask)
# define noise loss
if loss_type2 == "l1":
loss_f2 = torch.nn.L1Loss()
elif loss_type2 == "l2":
loss_f2 = torch.nn.MSELoss()
if residual_learning1:
loss2 = loss_f2(output_ns*mask, (output_raw-output_exp)*mask)
else:
loss2 = loss_f2(output_ns*mask, output_raw*mask)
if aperture:
if loss_type3 == "l1":
loss_f3 = torch.nn.L1Loss()
elif loss_type3 == "l2":
loss_f3 = torch.nn.MSELoss()
if residual_learning2:
loss3 = loss_f3((output_exp+output_ns+output_ap)*mask, output_raw*mask)
else:
loss3 = loss_f3(output_ap*mask, output_raw*mask)
if not multi_stage:
loss1.backward(retain_graph=True)
optimizer_1.step()
loss2.backward(retain_graph=True)
optimizer_2.step()
if aperture:
loss3.backward(retain_graph=True)
optimizer_3.step()
else:
if not aperture:
if epoch < multi_stage:
loss1.backward(retain_graph=True)
optimizer_1.step()
else:
loss2.backward(retain_graph=True)
optimizer_2.step()
else:
if epoch < multi_stage:
loss1.backward(retain_graph=True)
optimizer_1.step()
elif epoch >= multi_stage and epoch < multi_stage2:
loss2.backward(retain_graph=True)
optimizer_2.step()
else:
loss3.backward(retain_graph=True)
optimizer_3.step()
if aperture:
print("epochs{} iters{}:\t{:.9f}\t{:.9f}\t{:.9f}".format(epoch, iteration, loss1, loss2, loss3))
else:
print("epochs{} iters{}:\t{:.9f}\t{:.9f}".format(epoch, iteration, loss1, loss2))
if (iteration % iters_per_checkpoint == 0):
checkpoint_path1 = "{}/exp_{}".format(
output_directory, iteration)
checkpoint_path2 = "{}/unet_{}".format(
output_directory, iteration)
checkpoint_path3 = "{}/unet_att_{}".format(
output_directory, iteration)
image_path = "{}/img_{}.jpg".format(
output_directory, iteration)
# save checkpoints
save_checkpoint(model_exposure, "exposure", optimizer_1, learning_rate1, iteration,
checkpoint_path1, parallel)
save_checkpoint(model_noise, net_type, optimizer_2, learning_rate2, iteration,
checkpoint_path2, parallel)
save_checkpoint(model_aperture, net_type_ap, optimizer_3, learning_rate3, iteration,
checkpoint_path3, parallel)
# save testing images
if residual_learning1:
if isp_save:
if aperture:
if residual_learning2:
save_training_images_raw([input_raw.cpu().data.numpy(),
output_exp.cpu().data.numpy(),
(output_ns+output_exp).cpu().data.numpy(),
(output_ap+output_ns+output_exp).cpu().data.numpy(),
output_raw.cpu().data.numpy()], image_path, img_name, exp_params_m.cpu().data.numpy())
else:
save_training_images_raw([input_raw.cpu().data.numpy(),
output_exp.cpu().data.numpy(),
(output_ns+output_exp).cpu().data.numpy(),
output_ap.cpu().data.numpy(),
output_raw.cpu().data.numpy()], image_path, img_name, exp_params_m.cpu().data.numpy())
else:
save_training_images_raw([input_raw.cpu().data.numpy(),
output_exp.cpu().data.numpy(),
(output_ns+output_exp).cpu().data.numpy(),
output_raw.cpu().data.numpy()], image_path, img_name, exp_params_m.cpu().data.numpy())
iteration += 1
if epoch > multi_stage2:
scheduler_3.step()
elif epoch > multi_stage:
scheduler_2.step()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str,
help='JSON file for configuration')
args = parser.parse_args()
global config_path
config_path = args.config
with open(config_path) as f:
data = f.read()
config = json.loads(data)
train_config = config["train_config"]
global data_config
data_config = config["data_config"]
global network_config
network_config = config["network_config"]
global network_config2
network_config2 = config["network_config2"]
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
train(**train_config)