-
Notifications
You must be signed in to change notification settings - Fork 40
/
train.py
executable file
·291 lines (259 loc) · 12.4 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
import argparse
import shutil
import importlib
import logging
import numpy as np
import torch
from torch import nn
from torch.utils.data import DataLoader
from datasets import process
from datasets.load_dataset import GaussianPossionNoiseImages, GaussianNoiseImages, PixelShift
from model.common import cal_model_parm_nums
from TorchTools.ArgsTools.base_args import BaseArgs
from TorchTools.LossTools.metrics import PSNR, AverageMeter
from TorchTools.model_util import load_pretrained_models
from torch.utils.tensorboard import SummaryWriter
def main():
logging.info('===> Creating dataloader...')
if 'pixelshift' in args.dataset:
train_set = PixelShift(args.train_list, 'train',
args.patch_size,
args.downsampler, args.scale,
args.in_type, args.mid_type, args.out_type)
val_set = PixelShift(args.val_list, 'val',
args.val_patch_size,
args.downsampler, args.scale,
args.in_type, args.mid_type, args.out_type)
else:
if "p" in args.noise_model: # if gaussian-possion is used
train_set = GaussianPossionNoiseImages(args.train_list, 'train',
args.patch_size,
args.downsampler, args.scale,
args.in_type, args.mid_type, args.out_type)
val_set = GaussianPossionNoiseImages(args.val_list, 'val',
args.val_patch_size,
args.downsampler, args.scale,
args.in_type, args.mid_type, args.out_type)
else: # otherwise, use gaussian noise only
train_set = GaussianNoiseImages(args.train_list, 'train',
args.patch_size,
args.downsampler, args.scale,
args.in_type, args.mid_type, args.out_type,
args.sigma
)
val_set = GaussianNoiseImages(args.val_list, 'val',
args.val_patch_size,
args.downsampler, args.scale,
args.in_type, args.mid_type, args.out_type,
args.sigma
)
train_loader = DataLoader(dataset=train_set, num_workers=4, batch_size=args.batch_size,
shuffle=True, pin_memory=True)
val_loader = DataLoader(dataset=val_set, num_workers=4, batch_size=args.batch_size,
shuffle=False, pin_memory=True)
# =================
logging.info('===> Loading the network ...')
module = importlib.import_module("model.{}".format(args.model))
model = module.Net(**vars(args))
if args.n_gpus > 1:
model = nn.DataParallel(model)
model = model.to(args.device)
logging.info(model)
model_size = cal_model_parm_nums(model)
logging.info('Number of params: %.4f M' % (model_size / (1e6)))
# =================
logging.info('===> Loading the Optimizers ...')
criterion = torch.nn.L1Loss().to(args.device)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
if args.scheduler == 'step':
scheduler = torch.optim.lr_scheduler.StepLR(
optimizer, step_size=1, gamma=args.gamma)
elif args.scheduler == 'cos':
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, args.max_epochs, args.lr * 1e-3)
else:
raise NotImplementedError(
f'{args.scheduler} is not supported. support step or cos lr scheduler. ')
best_psnr, start_epoch = 0., 0
if args.pretrain is not None:
best_psnr, start_epoch = load_pretrained_models(
model, args.pretrain, optimizer, scheduler)
# =================
# train + val
logging.info('---------- Start training -------------\n')
iters = len(train_loader)
last_loss = np.inf
for epoch in range(start_epoch, args.max_epochs):
# train
losses = AverageMeter()
mid_losses = AverageMeter()
main_losses = AverageMeter()
model.train()
for idx, data in enumerate(train_loader):
# train, data convert
if 'noisy' in args.in_type:
img = torch.cat(
(data[args.in_type], data['variance']), dim=1).to(args.device)
else:
img = data[args.in_type].to(args.device)
gt = data[args.out_type].to(args.device)
batch_size = gt.size(0)
output = model(img)
# forward+backward+optimization
# if main_loss > 4 * last_loss:
# continue
if args.mid_type is not None:
main_loss = criterion(output[-1], gt)
mid_losses = []
for i, mid_type in enumerate(args.mid_type):
mid_gt = data[mid_type].to(args.device)
mid_loss = criterion(output[i], mid_gt)
mid_losses.append(args.mid_lambda[i] * mid_loss)
loss = main_loss + sum(mid_losses)
else:
main_loss = criterion(output, gt)
loss = main_loss
# zero parameters
optimizer.zero_grad()
loss.backward()
if args.grad_norm_clip > 0:
torch.nn.utils.clip_grad_norm_(
model.parameters(), args.grad_norm_clip, norm_type=2)
optimizer.step()
losses.update(loss.item(), batch_size)
main_losses.update(main_loss.item(), batch_size)
if idx % args.print_freq == 0:
print_str = f'Epoch: [{epoch}]/[{args.max_epochs}] Iter:[{idx}]/[{iters}]\t Loss: {loss.item():.4f} \t main_loss: {main_loss.item():.4f}\t'
if args.mid_type is not None:
for i, mid_type in enumerate(args.mid_type):
print_str += f'{mid_type} loss: {mid_losses[i].item():.4f}\t'
logging.info(print_str)
scheduler.step()
writer.add_scalar('train_loss', losses.avg, epoch)
writer.add_scalar('main_loss', main_losses.avg, epoch)
writer.add_scalar('lr', scheduler.get_last_lr()[-1], epoch)
# ================
# val
if epoch % args.eval_freq == 0 or epoch == args.max_epochs - 1:
args.epoch = epoch
cur_psnr = val(val_loader, model, args.vis_eval)
is_best = (cur_psnr > best_psnr)
best_psnr = max(cur_psnr, best_psnr)
model_cpu = {k: v.cpu() for k, v in model.state_dict().items()}
save_checkpoint({
'epoch': epoch,
'state_dict': model_cpu,
'best_psnr': best_psnr,
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict()
}, is_best, args=args)
writer.add_scalar('eval_psnr', cur_psnr, epoch)
logging.info('Saving the final model.')
# wandb
if args.use_wandb:
args.Wandb.add_file(
f'{args.ckpt_dir}/{args.jobname}_checkpoint_best.pth')
args.Wandb.add_file(
f'{args.ckpt_dir}/{args.jobname}_checkpoint_latest.pth')
def val(val_loader, model, vis_eval=False):
psnrs = AverageMeter()
model.eval()
with torch.no_grad():
for i, data in enumerate(val_loader):
if 'noisy' in args.in_type:
img = torch.cat(
(data[args.in_type], data['variance']), dim=1).to(args.device)
else:
img = data[args.in_type].to(args.device)
gt = data[args.out_type].to(args.device)
output = model(img)
# psnr
if args.output_mid:
output = output[-1]
mse = (output.clamp(0, 1) - gt).pow(2).mean()
psnr = PSNR(mse)
psnrs.update(psnr, gt.size(0))
if i == 0 and vis_eval and (args.epoch % args.img_freq == 0):
batch_size = img.shape[0]
n_img = min(5, batch_size)
n_stride = batch_size // n_img # show 5 imgs only
if 'rgb' in args.in_type and 'linrgb' not in args.out_type:
rgb_out = output[::n_stride]
rgb_gt = gt[::n_stride]
else:
red_g = data['metadata']['red_gain'][::n_stride].to(
args.device)
blue_g = data['metadata']['blue_gain'][::n_stride].to(
args.device)
ccm = data['metadata']['ccm'][::n_stride].to(args.device)
if 'raw' in args.out_type:
rgb_out = process.raw2srgb(
output[::n_stride], red_g, blue_g, ccm)
rgb_gt = process.raw2srgb(
gt[::n_stride], red_g, blue_g, ccm)
elif 'linrgb' in args.out_type:
rgb_out = process.rgb2srgb(
output[::n_stride], red_g, blue_g, ccm)
rgb_gt = process.rgb2srgb(
gt[::n_stride], red_g, blue_g, ccm)
B, C, H, W = rgb_out.shape
writer.add_images('eval_result',
torch.stack((rgb_gt, rgb_out), dim=1).contiguous().view(-1, C, H, W), args.epoch)
logging.info('\nEpoch: [{}]/[{}] \t''TEST PSNR: {psnrs.avg: .4f})\n'.
format(args.epoch, args.max_epochs, psnrs=psnrs))
return psnrs.avg
def save_checkpoint(state, is_best, args):
filename = '{}/{}_checkpoint_latest.pth'.format(
args.ckpt_dir, args.jobname)
torch.save(state, filename)
if is_best:
shutil.copyfile(
filename, '{}/{}_checkpoint_best.pth'.format(args.ckpt_dir, args.jobname))
if args.save_freq > 0 and args.epoch % args.save_freq == 0:
shutil.copyfile(
filename, '{}/{}_checkpoint_epoch{}.pth'.format(args.ckpt_dir, args.jobname, args.epoch))
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='PyTorch implementation of ISP-Net')
baseargs = BaseArgs(parser)
baseargs._print_args()
args=baseargs.args
writer = SummaryWriter(log_dir=args.exp_dir)
main()
# below is the code for debug:
# from datasets import process
# def vis_rgb(x):
# import matplotlib.pyplot as plt
# plt.imshow(x)
# plt.show()
# vis_rgb(gt[0].permute(1,2,0).cpu())
# gt_srgb = process.rgb2srgb(gt, data['red_gain'].to(args.device), data['blue_gain'].to(args.device), data['cam2rgb'].to(args.device))
# vis_rgb(gt_srgb[0].permute(1,2,0).cpu())
# noisy_rgb = process.raw2srgb(data['input'].to(args.device), data['red_gain'].to(args.device), data['blue_gain'].to(args.device), data['cam2rgb'].to(args.device))
# vis_rgb(noisy_rgb[0].permute(1,2,0).cpu())
# clean_rgb = process.raw2srgb(data['mid_gt'].to(args.device), data['red_gain'].to(args.device), data['blue_gain'].to(args.device), data['cam2rgb'].to(args.device))
# vis_rgb(clean_rgb[0].permute(1,2,0).cpu())
# Debug the RawDeno:
# def vis_rgb(x):
# import matplotlib.pyplot as plt
# plt.imshow(x)
# plt.show()
# red_g = data['metadata']['red_gain'][0:1].to(args.device)
# blue_g = data['metadata']['blue_gain'][0:1].to(args.device)
# ccm = data['metadata']['cam2rgb'][0:1].to(args.device)
# rgb_in = process.raw2srgb(img[0:1, :4], red_g, blue_g, ccm)
# rgb_gt = process.raw2srgb(gt[0:1], red_g, blue_g, ccm)
# vis_rgb(rgb_in[0].permute(1,2,0).detach().cpu().numpy())
# vis_rgb(rgb_gt[0].permute(1,2,0).detach().cpu().numpy())
# if debug:
# red_g = torch.tensor(2.7385).to('cuda')
# blue_g = torch.tensor(1.3687).to('cuda')
# ccm = torch.tensor([[[1.7365, -0.5612, -0.1753], [-0.1531, 1.5584, -0.4053], [0.0199, -0.4041, 1.3842]]],
# device='cuda')
# rgb_out = process.rgb2srgb(output, red_g, blue_g, ccm)
# # debug:
# red_g, blue_g, ccm = data['metadata']['red_gain'].to(args.device), \
# data['metadata']['blue_gain'].to(args.device), \
# data['metadata']['ccm'].to(args.device)
# rgb_gt = process.rgb2srgb(gt, red_g, blue_g, ccm)
# out_gt = process.rgb2srgb(output, red_g, blue_g, ccm)