def prediction(model, gt, noisy, white_level):
_, pred = model(noisy.view(1, -1, 512, 512), noisy[:, :8, ...],
white_level)
pred = pred.clamp(0.0, 1.0)
pred = sRGBGamma(pred)
pred = pred.cpu()
psnr = calculate_psnr(pred, gt)
ssim = calculate_ssim(pred, gt)
# psnr.append(psnr_t)
# ssim.append(ssim_t)
# print(pred.size())
return pred, psnr, ssim
def eval(args):
color = args.color
print('Eval Process......')
burst_length = args.burst_length
# print(args.checkpoint)
checkpoint_dir = "checkpoints/" + args.checkpoint
if not os.path.exists(checkpoint_dir) or len(
os.listdir(checkpoint_dir)) == 0:
print('There is no any checkpoint file in path:{}'.format(
checkpoint_dir))
# the path for saving eval images
eval_dir = "eval_img"
if not os.path.exists(eval_dir):
os.mkdir(eval_dir)
# dataset and dataloader
data_set = SingleLoader_DGF(noise_dir=args.noise_dir,
gt_dir=args.gt_dir,
image_size=args.image_size,
burst_length=burst_length)
data_loader = DataLoader(data_set,
batch_size=1,
shuffle=False,
num_workers=args.num_workers)
# model here
if args.model_type == "attKPN":
model = Att_KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attWKPN":
model = Att_Weight_KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
elif args.model_type == "KPN":
model = KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
else:
print(" Model type not valid")
return
if args.cuda:
model = model.cuda()
if args.mGPU:
model = nn.DataParallel(model)
# load trained model
ckpt = load_checkpoint(checkpoint_dir,
cuda=args.cuda,
best_or_latest=args.load_type)
state_dict = ckpt['state_dict']
if not args.mGPU:
new_state_dict = OrderedDict()
if not args.cuda:
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
else:
model.load_state_dict(ckpt['state_dict'])
print('The model has been loaded from epoch {}, n_iter {}.'.format(
ckpt['epoch'], ckpt['global_iter']))
# switch the eval mode
model.eval()
# data_loader = iter(data_loader)
trans = transforms.ToPILImage()
with torch.no_grad():
psnr = 0.0
ssim = 0.0
torch.manual_seed(0)
for i, (image_noise_hr, image_noise_lr,
image_gt_hr) in enumerate(data_loader):
if i < 100:
# data = next(data_loader)
if args.cuda:
burst_noise = image_noise_lr.cuda()
gt = image_gt_hr.cuda()
else:
burst_noise = image_noise_lr
gt = image_gt_hr
if color:
b, N, c, h, w = image_noise_lr.size()
feedData = image_noise_lr.view(b, -1, h, w)
else:
feedData = image_noise_lr
pred_i, pred = model(feedData, burst_noise[:, 0:burst_length,
...],
image_noise_hr)
psnr_t = calculate_psnr(pred, gt)
ssim_t = calculate_ssim(pred, gt)
print("PSNR : ", str(psnr_t), " : SSIM : ", str(ssim_t))
pred = torch.clamp(pred, 0.0, 1.0)
if args.cuda:
pred = pred.cpu()
gt = gt.cpu()
burst_noise = burst_noise.cpu()
if args.save_img:
trans(burst_noise[0, 0, ...].squeeze()).save(os.path.join(
eval_dir, '{}_noisy.png'.format(i)),
quality=100)
trans(pred.squeeze()).save(os.path.join(
eval_dir, '{}_pred_{:.2f}dB.png'.format(i, psnr_t)),
quality=100)
trans(gt.squeeze()).save(os.path.join(
eval_dir, '{}_gt.png'.format(i)),
quality=100)
else:
break
def eval(config, args):
train_config = config['training']
arch_config = config['architecture']
use_cache = train_config['use_cache']
print('Eval Process......')
checkpoint_dir = train_config['checkpoint_dir']
if not os.path.exists(checkpoint_dir) or len(
os.listdir(checkpoint_dir)) == 0:
print('There is no any checkpoint file in path:{}'.format(
checkpoint_dir))
# the path for saving eval images
eval_dir = train_config['eval_dir']
if not os.path.exists(eval_dir):
os.mkdir(eval_dir)
files = os.listdir(eval_dir)
for f in files:
os.remove(os.path.join(eval_dir, f))
# dataset and dataloader
data_set = TrainDataSet(train_config['dataset_configs'],
img_format='.bmp',
degamma=True,
color=False,
blind=arch_config['blind_est'],
train=False)
data_loader = DataLoader(data_set,
batch_size=1,
shuffle=False,
num_workers=args.num_workers)
dataset_config = read_config(train_config['dataset_configs'],
_configspec_path())['dataset_configs']
# model here
model = KPN(color=False,
burst_length=dataset_config['burst_length'],
blind_est=arch_config['blind_est'],
kernel_size=list(map(int, arch_config['kernel_size'].split())),
sep_conv=arch_config['sep_conv'],
channel_att=arch_config['channel_att'],
spatial_att=arch_config['spatial_att'],
upMode=arch_config['upMode'],
core_bias=arch_config['core_bias'])
if args.cuda:
model = model.cuda()
if args.mGPU:
model = nn.DataParallel(model)
# load trained model
ckpt = load_checkpoint(checkpoint_dir, args.checkpoint)
model.load_state_dict(ckpt['state_dict'])
print('The model has been loaded from epoch {}, n_iter {}.'.format(
ckpt['epoch'], ckpt['global_iter']))
# switch the eval mode
model.eval()
# data_loader = iter(data_loader)
burst_length = dataset_config['burst_length']
data_length = burst_length if arch_config['blind_est'] else burst_length + 1
patch_size = dataset_config['patch_size']
trans = transforms.ToPILImage()
with torch.no_grad():
psnr = 0.0
ssim = 0.0
for i, (burst_noise, gt, white_level) in enumerate(data_loader):
if i < 100:
# data = next(data_loader)
if args.cuda:
burst_noise = burst_noise.cuda()
gt = gt.cuda()
white_level = white_level.cuda()
pred_i, pred = model(burst_noise,
burst_noise[:, 0:burst_length,
...], white_level)
pred_i = sRGBGamma(pred_i)
pred = sRGBGamma(pred)
gt = sRGBGamma(gt)
burst_noise = sRGBGamma(burst_noise / white_level)
psnr_t = calculate_psnr(pred.unsqueeze(1), gt.unsqueeze(1))
ssim_t = calculate_ssim(pred.unsqueeze(1), gt.unsqueeze(1))
psnr_noisy = calculate_psnr(
burst_noise[:, 0, ...].unsqueeze(1), gt.unsqueeze(1))
psnr += psnr_t
ssim += ssim_t
pred = torch.clamp(pred, 0.0, 1.0)
if args.cuda:
pred = pred.cpu()
gt = gt.cpu()
burst_noise = burst_noise.cpu()
trans(burst_noise[0, 0, ...].squeeze()).save(os.path.join(
eval_dir, '{}_noisy_{:.2f}dB.png'.format(i, psnr_noisy)),
quality=100)
trans(pred.squeeze()).save(os.path.join(
eval_dir, '{}_pred_{:.2f}dB.png'.format(i, psnr_t)),
quality=100)
trans(gt.squeeze()).save(os.path.join(eval_dir,
'{}_gt.png'.format(i)),
quality=100)
print('{}-th image is OK, with PSNR: {:.2f}dB, SSIM: {:.4f}'.
format(i, psnr_t, ssim_t))
else:
break
print('All images are OK, average PSNR: {:.2f}dB, SSIM: {:.4f}'.format(
psnr / 100, ssim / 100))
def train(config, num_workers, num_threads, cuda, restart_train, mGPU):
# torch.set_num_threads(num_threads)
train_config = config['training']
arch_config = config['architecture']
batch_size = train_config['batch_size']
lr = train_config['learning_rate']
weight_decay = train_config['weight_decay']
decay_step = train_config['decay_steps']
lr_decay = train_config['lr_decay']
n_epoch = train_config['num_epochs']
use_cache = train_config['use_cache']
print('Configs:', config)
# checkpoint path
checkpoint_dir = train_config['checkpoint_dir']
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
# logs path
logs_dir = train_config['logs_dir']
if not os.path.exists(logs_dir):
os.makedirs(logs_dir)
shutil.rmtree(logs_dir)
log_writer = SummaryWriter(logs_dir)
# dataset and dataloader
data_set = TrainDataSet(train_config['dataset_configs'],
img_format='.bmp',
degamma=True,
color=False,
blind=arch_config['blind_est'])
data_loader = DataLoader(data_set,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
dataset_config = read_config(train_config['dataset_configs'],
_configspec_path())['dataset_configs']
# model here
model = KPN(color=False,
burst_length=dataset_config['burst_length'],
blind_est=arch_config['blind_est'],
kernel_size=list(map(int, arch_config['kernel_size'].split())),
sep_conv=arch_config['sep_conv'],
channel_att=arch_config['channel_att'],
spatial_att=arch_config['spatial_att'],
upMode=arch_config['upMode'],
core_bias=arch_config['core_bias'])
if cuda:
model = model.cuda()
if mGPU:
model = nn.DataParallel(model)
model.train()
# loss function here
loss_func = LossFunc(coeff_basic=1.0,
coeff_anneal=1.0,
gradient_L1=True,
alpha=arch_config['alpha'],
beta=arch_config['beta'])
# Optimizer here
if train_config['optimizer'] == 'adam':
optimizer = optim.Adam(model.parameters(), lr=lr)
elif train_config['optimizer'] == 'sgd':
optimizer = optim.SGD(model.parameters(),
lr=lr,
momentum=0.9,
weight_decay=weight_decay)
else:
raise ValueError(
"Optimizer must be 'sgd' or 'adam', but received {}.".format(
train_config['optimizer']))
optimizer.zero_grad()
# learning rate scheduler here
scheduler = lr_scheduler.StepLR(optimizer, step_size=10, gamma=lr_decay)
average_loss = MovingAverage(train_config['save_freq'])
if not restart_train:
try:
checkpoint = load_checkpoint(checkpoint_dir, 'best')
start_epoch = checkpoint['epoch']
global_step = checkpoint['global_iter']
best_loss = checkpoint['best_loss']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['lr_scheduler'])
print('=> loaded checkpoint (epoch {}, global_step {})'.format(
start_epoch, global_step))
except:
start_epoch = 0
global_step = 0
best_loss = np.inf
print('=> no checkpoint file to be loaded.')
else:
start_epoch = 0
global_step = 0
best_loss = np.inf
if os.path.exists(checkpoint_dir):
pass
# files = os.listdir(checkpoint_dir)
# for f in files:
# os.remove(os.path.join(checkpoint_dir, f))
else:
os.mkdir(checkpoint_dir)
print('=> training')
burst_length = dataset_config['burst_length']
data_length = burst_length if arch_config['blind_est'] else burst_length + 1
patch_size = dataset_config['patch_size']
for epoch in range(start_epoch, n_epoch):
epoch_start_time = time.time()
# decay the learning rate
lr_cur = [param['lr'] for param in optimizer.param_groups]
if lr_cur[0] > 5e-6:
scheduler.step()
else:
for param in optimizer.param_groups:
param['lr'] = 5e-6
print(
'=' * 20,
'lr={}'.format([param['lr'] for param in optimizer.param_groups]),
'=' * 20)
t1 = time.time()
for step, (burst_noise, gt, white_level) in enumerate(data_loader):
if cuda:
burst_noise = burst_noise.cuda()
gt = gt.cuda()
# print('white_level', white_level, white_level.size())
#
pred_i, pred = model(burst_noise, burst_noise[:, 0:burst_length,
...], white_level)
#
loss_basic, loss_anneal = loss_func(sRGBGamma(pred_i),
sRGBGamma(pred), sRGBGamma(gt),
global_step)
loss = loss_basic + loss_anneal
# backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
# update the average loss
average_loss.update(loss)
# calculate PSNR
psnr = calculate_psnr(pred.unsqueeze(1), gt.unsqueeze(1))
ssim = calculate_ssim(pred.unsqueeze(1), gt.unsqueeze(1))
# add scalars to tensorboardX
log_writer.add_scalar('loss_basic', loss_basic, global_step)
log_writer.add_scalar('loss_anneal', loss_anneal, global_step)
log_writer.add_scalar('loss_total', loss, global_step)
log_writer.add_scalar('psnr', psnr, global_step)
log_writer.add_scalar('ssim', ssim, global_step)
# print
print(
'{:-4d}\t| epoch {:2d}\t| step {:4d}\t| loss_basic: {:.4f}\t| loss_anneal: {:.4f}\t|'
' loss: {:.4f}\t| PSNR: {:.2f}dB\t| SSIM: {:.4f}\t| time:{:.2f} seconds.'
.format(global_step, epoch, step, loss_basic, loss_anneal,
loss, psnr, ssim,
time.time() - t1))
t1 = time.time()
# global_step
global_step += 1
if global_step % train_config['save_freq'] == 0:
if average_loss.get_value() < best_loss:
is_best = True
best_loss = average_loss.get_value()
else:
is_best = False
save_dict = {
'epoch': epoch,
'global_iter': global_step,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
'lr_scheduler': scheduler.state_dict()
}
save_checkpoint(save_dict,
is_best,
checkpoint_dir,
global_step,
max_keep=train_config['ckpt_to_keep'])
print('Epoch {} is finished, time elapsed {:.2f} seconds.'.format(
epoch,
time.time() - epoch_start_time))
def test_multi(args):
color = True
burst_length = args.burst_length
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if args.model_type == "attKPN":
model = Att_KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attKPN_Wave":
model = Att_KPN_Wavelet_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attWKPN":
model = Att_Weight_KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "KPN":
model = KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
else:
print(" Model type not valid")
return
checkpoint_dir = "checkpoints/" + args.checkpoint
if not os.path.exists(checkpoint_dir) or len(
os.listdir(checkpoint_dir)) == 0:
print('There is no any checkpoint file in path:{}'.format(
checkpoint_dir))
# load trained model
ckpt = load_checkpoint(checkpoint_dir,
cuda=device == 'cuda',
best_or_latest=args.load_type)
state_dict = ckpt['state_dict']
# if not args.cuda:
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
# else:
# model.load_state_dict(ckpt['state_dict'])
model.to(device)
print('The model has been loaded from epoch {}, n_iter {}.'.format(
ckpt['epoch'], ckpt['global_iter']))
# switch the eval mode
model.eval()
# model= save_dict['state_dict']
trans = transforms.ToPILImage()
torch.manual_seed(0)
all_noisy_imgs = scipy.io.loadmat(
args.noise_dir)['ValidationNoisyBlocksSrgb']
all_clean_imgs = scipy.io.loadmat(args.gt)['ValidationGtBlocksSrgb']
i_imgs, i_blocks, _, _, _ = all_noisy_imgs.shape
psnrs = []
ssims = []
for i_img in range(i_imgs):
for i_block in range(i_blocks):
image_noise = transforms.ToTensor()(Image.fromarray(
all_noisy_imgs[i_img][i_block]))
image_noise = transforms.ToTensor()(Image.fromarray(
all_noisy_imgs[i_img][i_block]))
image_noise, image_noise_hr = load_data(image_noise, burst_length)
image_noise_hr = image_noise_hr.to(device)
# begin = time.time()
image_noise = image_noise.to(device)
# print(image_noise_batch.size())
# burst_size = image_noise.size()[1]
# print(burst_noise.size())
# print(image_noise_hr.size())
if color:
b, N, c, h, w = image_noise.size()
feedData = image_noise.view(b, -1, h, w)
else:
feedData = image_noise
# print(feedData.size())
pred_i, pred = model(feedData, image_noise[:, 0:burst_length, ...],
image_noise_hr)
del pred_i
pred = pred.detach().cpu()
# print("Time : ", time.time()-begin)
gt = transforms.ToTensor()(Image.fromarray(
all_clean_imgs[i_img][i_block]))
gt = gt.unsqueeze(0)
# print(pred_i.size())
# print(pred[0].size())
psnr_t = calculate_psnr(pred, gt)
ssim_t = calculate_ssim(pred, gt)
psnrs.append(psnr_t)
ssims.append(ssim_t)
print(i_img, " ", i_block, " UP : PSNR : ", str(psnr_t),
" : SSIM : ", str(ssim_t))
if args.save_img != '':
if not os.path.exists(args.save_img):
os.makedirs(args.save_img)
plt.figure(figsize=(15, 15))
plt.imshow(np.array(trans(pred[0])))
plt.title("denoise KPN DGF " + args.model_type, fontsize=25)
image_name = str(i_img) + "_" + str(i_block)
plt.axis("off")
plt.suptitle(image_name + " UP : PSNR : " + str(psnr_t) +
" : SSIM : " + str(ssim_t),
fontsize=25)
plt.savefig(os.path.join(
args.save_img,
image_name + "_" + args.checkpoint + '.png'),
pad_inches=0)
"""
if args.save_img:
# print(np.array(trans(mf8[0])))
plt.figure(figsize=(30, 9))
plt.subplot(1,3,1)
plt.imshow(np.array(trans(pred[0])))
plt.title("denoise DGF "+args.model_type, fontsize=26)
plt.subplot(1,3,2)
plt.imshow(np.array(trans(gt[0])))
plt.title("gt ", fontsize=26)
plt.subplot(1,3,3)
plt.imshow(np.array(trans(image_noise_hr[0])))
plt.title("noise ", fontsize=26)
plt.axis("off")
plt.suptitle(str(i)+" UP : PSNR : "+ str(psnr_t)+" : SSIM : "+ str(ssim_t), fontsize=26)
plt.savefig("checkpoints/22_DGF_" + args.checkpoint+str(i)+'.png',pad_inches=0)
"""
print(" AVG : PSNR : " + str(np.mean(psnrs)) + " : SSIM : " +
str(np.mean(ssims)))
def train(num_workers, cuda, restart_train, mGPU):
# torch.set_num_threads(num_threads)
color = True
batch_size = args.batch_size
lr = 2e-4
lr_decay = 0.89125093813
n_epoch = args.epoch
# num_workers = 8
save_freq = args.save_every
loss_freq = args.loss_every
lr_step_size = 100
burst_length = args.burst_length
# checkpoint path
checkpoint_dir = "checkpoints/" + args.checkpoint
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
# logs path
logs_dir = "checkpoints/logs/" + args.checkpoint
if not os.path.exists(logs_dir):
os.makedirs(logs_dir)
shutil.rmtree(logs_dir)
log_writer = SummaryWriter(logs_dir)
# dataset and dataloader
data_set = SingleLoader_DGF(noise_dir=args.noise_dir,gt_dir=args.gt_dir,image_size=args.image_size,burst_length=burst_length)
data_loader = DataLoader(
data_set,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers
)
# model here
if args.model_type == "attKPN":
model = Att_KPN_noise_DGF(
color=color,
burst_length=burst_length,
blind_est=False,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False
)
elif args.model_type == "attWKPN":
model = Att_Weight_KPN_noise_DGF(
color=color,
burst_length=burst_length,
blind_est=False,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False
)
elif args.model_type == 'KPN':
model = KPN_noise_DGF(
color=color,
burst_length=burst_length,
blind_est=False,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False
)
else:
print(" Model type not valid")
return
if cuda:
model = model.cuda()
if mGPU:
model = nn.DataParallel(model)
model.train()
# loss function here
loss_func = LossBasic()
if args.wavelet_loss:
print("Use wavelet loss")
loss_func2 = WaveletLoss()
# Optimizer here
optimizer = optim.Adam(
model.parameters(),
lr=lr
)
optimizer.zero_grad()
# learning rate scheduler here
scheduler = lr_scheduler.StepLR(optimizer, step_size=lr_step_size, gamma=lr_decay)
average_loss = MovingAverage(save_freq)
if not restart_train:
try:
checkpoint = load_checkpoint(checkpoint_dir,cuda , best_or_latest=args.load_type)
start_epoch = checkpoint['epoch']
global_step = checkpoint['global_iter']
best_loss = checkpoint['best_loss']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['lr_scheduler'])
print('=> loaded checkpoint (epoch {}, global_step {})'.format(start_epoch, global_step))
except:
start_epoch = 0
global_step = 0
best_loss = np.inf
print('=> no checkpoint file to be loaded.')
else:
start_epoch = 0
global_step = 0
best_loss = np.inf
if os.path.exists(checkpoint_dir):
pass
# files = os.listdir(checkpoint_dir)
# for f in files:
# os.remove(os.path.join(checkpoint_dir, f))
else:
os.mkdir(checkpoint_dir)
print('=> training')
for epoch in range(start_epoch, n_epoch):
epoch_start_time = time.time()
# decay the learning rate
# print('='*20, 'lr={}'.format([param['lr'] for param in optimizer.param_groups]), '='*20)
t1 = time.time()
for step, (image_noise_hr,image_noise_lr, image_gt_hr, _) in enumerate(data_loader):
# print(burst_noise.size())
# print(gt.size())
if cuda:
burst_noise = image_noise_lr.cuda()
gt = image_gt_hr.cuda()
image_noise_hr = image_noise_hr.cuda()
noise_gt = (image_noise_hr-image_gt_hr).cuda()
else:
burst_noise = image_noise_lr
gt = image_gt_hr
noise_gt = image_noise_hr - image_gt_hr
#
_, pred,noise = model(burst_noise,image_noise_hr)
# print(pred.size())
#
loss_basic = loss_func(pred, gt)
loss_noise = loss_func(noise,noise_gt)
loss = loss_basic + loss_noise
if args.wavelet_loss:
loss_wave = loss_func2(pred,gt)
loss_wave_noise = loss_func2(noise,noise_gt)
# print(loss_wave)
loss = loss_basic + loss_wave + loss_noise + loss_wave_noise
# backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
# update the average loss
average_loss.update(loss)
# global_step
if not color:
pred = pred.unsqueeze(1)
gt = gt.unsqueeze(1)
if global_step %loss_freq ==0:
# calculate PSNR
print("burst_noise : ",burst_noise.size())
print("gt : ",gt.size())
psnr = calculate_psnr(pred, gt)
ssim = calculate_ssim(pred, gt)
# add scalars to tensorboardX
log_writer.add_scalar('loss_basic', loss_basic, global_step)
log_writer.add_scalar('loss_total', loss, global_step)
log_writer.add_scalar('psnr', psnr, global_step)
log_writer.add_scalar('ssim', ssim, global_step)
# print
print('{:-4d}\t| epoch {:2d}\t| step {:4d}\t| loss_basic: {:.4f}\t|'
' loss: {:.4f}\t| PSNR: {:.2f}dB\t| SSIM: {:.4f}\t| time:{:.2f} seconds.'
.format(global_step, epoch, step, loss_basic, loss, psnr, ssim, time.time()-t1))
t1 = time.time()
if global_step % save_freq == 0:
if average_loss.get_value() < best_loss:
is_best = True
best_loss = average_loss.get_value()
else:
is_best = False
save_dict = {
'epoch': epoch,
'global_iter': global_step,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
'lr_scheduler': scheduler.state_dict()
}
save_checkpoint(
save_dict, is_best, checkpoint_dir, global_step, max_keep=10
)
global_step += 1
print('Epoch {} is finished, time elapsed {:.2f} seconds.'.format(epoch, time.time()-epoch_start_time))
lr_cur = [param['lr'] for param in optimizer.param_groups]
if lr_cur[0] > 5e-6:
scheduler.step()
else:
for param in optimizer.param_groups:
param['lr'] = 5e-6
def test_multi(args):
color = True
burst_length = 8
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if args.model_type == "attKPN":
model = Att_KPN(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attWKPN":
model = Att_Weight_KPN(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "KPN":
model = KPN(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
else:
print(" Model type not valid")
return
# model2 = KPN(
# color=color,
# burst_length=burst_length,
# blind_est=True,
# kernel_size=[5],
# sep_conv=False,
# channel_att=False,
# spatial_att=False,
# upMode="bilinear",
# core_bias=False
# )
checkpoint_dir = "checkpoints/" + args.checkpoint
if not os.path.exists(checkpoint_dir) or len(
os.listdir(checkpoint_dir)) == 0:
print('There is no any checkpoint file in path:{}'.format(
checkpoint_dir))
# load trained model
ckpt = load_checkpoint(checkpoint_dir,
cuda=device == 'cuda',
best_or_latest=args.load_type)
state_dict = ckpt['state_dict']
# if not args.cuda:
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
# else:
# model.load_state_dict(ckpt['state_dict'])
#############################################
# checkpoint_dir = "checkpoints/" + "kpn"
# if not os.path.exists(checkpoint_dir) or len(os.listdir(checkpoint_dir)) == 0:
# print('There is no any checkpoint file in path:{}'.format(checkpoint_dir))
# # load trained model
# ckpt = load_checkpoint(checkpoint_dir,cuda=device=='cuda')
# state_dict = ckpt['state_dict']
# new_state_dict = OrderedDict()
# if not args.cuda:
# for k, v in state_dict.items():
# name = k[7:] # remove `module.`
# new_state_dict[name] = v
# # model.load_state_dict(ckpt['state_dict'])
# model2.load_state_dict(new_state_dict)
###########################################
print('The model has been loaded from epoch {}, n_iter {}.'.format(
ckpt['epoch'], ckpt['global_iter']))
# switch the eval mode
model.to(device)
model.eval()
# model2.eval()
# model= save_dict['state_dict']
trans = transforms.ToPILImage()
torch.manual_seed(0)
noisy_path = sorted(glob.glob(args.noise_dir + "/*.png"))
clean_path = [i.replace("noisy", "clean") for i in noisy_path]
for i in range(len(noisy_path)):
image_noise = load_data(noisy_path[i], burst_length)
begin = time.time()
image_noise_batch = image_noise.to(device)
# print(image_noise.size())
# print(image_noise_batch.size())
burst_noise = image_noise_batch.to(device)
if color:
b, N, c, h, w = burst_noise.size()
feedData = burst_noise.view(b, -1, h, w)
else:
feedData = burst_noise
# print(feedData.size())
pred_i, pred = model(feedData, burst_noise[:, 0:burst_length, ...])
del pred_i
# pred_i2, pred2 = model2(feedData, burst_noise[:, 0:burst_length, ...])
# print("Time : ", time.time()-begin)
pred = pred.detach().cpu()
gt = transforms.ToTensor()(Image.open(clean_path[i]).convert('RGB'))
# print(pred_i.size())
# print(pred.size())
# print(gt.size())
gt = gt.unsqueeze(0)
_, _, h_hr, w_hr = gt.size()
_, _, h_lr, w_lr = pred.size()
gt_down = F.interpolate(gt, (h_lr, w_lr),
mode='bilinear',
align_corners=True)
pred_up = F.interpolate(pred, (h_hr, w_hr),
mode='bilinear',
align_corners=True)
# print("After interpolate")
# print(pred_up.size())
# print(gt_down.size())
psnr_t_up = calculate_psnr(pred_up, gt)
ssim_t_up = calculate_ssim(pred_up, gt)
psnr_t_down = calculate_psnr(pred, gt_down)
ssim_t_down = calculate_ssim(pred, gt_down)
print(i, " UP : PSNR : ", str(psnr_t_up), " : SSIM : ",
str(ssim_t_up), " : DOWN : PSNR : ", str(psnr_t_down),
" : SSIM : ", str(ssim_t_down))
if args.save_img != '':
if not os.path.exists(args.save_img):
os.makedirs(args.save_img)
plt.figure(figsize=(15, 15))
plt.imshow(np.array(trans(pred_up[0])))
plt.title("denoise KPN split " + args.model_type, fontsize=25)
image_name = noisy_path[i].split("/")[-1].split(".")[0]
plt.axis("off")
plt.suptitle(image_name + " UP : PSNR : " + str(psnr_t_up) +
" : SSIM : " + str(ssim_t_up),
fontsize=25)
plt.savefig(os.path.join(
args.save_img, image_name + "_" + args.checkpoint + '.png'),
pad_inches=0)
# print(np.array(trans(mf8[0])))
"""
def eval(args):
color = True
print('Eval Process......')
burst_length = 8
checkpoint_dir = "checkpoints/" + args.checkpoint
if not os.path.exists(checkpoint_dir) or len(
os.listdir(checkpoint_dir)) == 0:
print('There is no any checkpoint file in path:{}'.format(
checkpoint_dir))
# the path for saving eval images
eval_dir = "eval_img"
if not os.path.exists(eval_dir):
os.mkdir(eval_dir)
files = os.listdir(eval_dir)
for f in files:
os.remove(os.path.join(eval_dir, f))
# dataset and dataloader
data_set = MultiLoader(noise_dir=args.noise_dir,
gt_dir=args.gt_dir,
image_size=args.image_size)
data_loader = DataLoader(data_set,
batch_size=1,
shuffle=False,
num_workers=args.num_workers)
# model here
if args.model_type == "attKPN":
model = Att_KPN_noise(color=color,
burst_length=burst_length,
blind_est=False,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attWKPN":
model = Att_Weight_KPN_noise(color=color,
burst_length=burst_length,
blind_est=False,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "KPN":
model = KPN_noise(color=color,
burst_length=burst_length,
blind_est=False,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
else:
print(" Model type not valid")
return
if args.cuda:
model = model.cuda()
if args.mGPU:
model = nn.DataParallel(model)
# load trained model
ckpt = load_checkpoint(checkpoint_dir, cuda=args.cuda)
model.load_state_dict(ckpt['state_dict'])
print('The model has been loaded from epoch {}, n_iter {}.'.format(
ckpt['epoch'], ckpt['global_iter']))
# switch the eval mode
model.eval()
# data_loader = iter(data_loader)
trans = transforms.ToPILImage()
with torch.no_grad():
psnr = 0.0
ssim = 0.0
for i, (burst_noise, gt) in enumerate(data_loader):
if i < 100:
# data = next(data_loader)
if args.cuda:
burst_noise = burst_noise.cuda()
gt = gt.cuda()
pred_i, pred = model(burst_noise)
if not color:
psnr_t = calculate_psnr(pred.unsqueeze(1), gt.unsqueeze(1))
ssim_t = calculate_ssim(pred.unsqueeze(1), gt.unsqueeze(1))
psnr_noisy = calculate_psnr(
burst_noise[:, 0, ...].unsqueeze(1), gt.unsqueeze(1))
else:
psnr_t = calculate_psnr(pred, gt)
ssim_t = calculate_ssim(pred, gt)
psnr_noisy = calculate_psnr(burst_noise[:, 0, ...], gt)
psnr += psnr_t
ssim += ssim_t
pred = torch.clamp(pred, 0.0, 1.0)
if args.cuda:
pred = pred.cpu()
gt = gt.cpu()
burst_noise = burst_noise.cpu()
trans(burst_noise[0, 0, ...].squeeze()).save(os.path.join(
eval_dir, '{}_noisy_{:.2f}dB.png'.format(i, psnr_noisy)),
quality=100)
trans(pred.squeeze()).save(os.path.join(
eval_dir, '{}_pred_{:.2f}dB.png'.format(i, psnr_t)),
quality=100)
trans(gt.squeeze()).save(os.path.join(eval_dir,
'{}_gt.png'.format(i)),
quality=100)
print('{}-th image is OK, with PSNR: {:.2f}dB, SSIM: {:.4f}'.
format(i, psnr_t, ssim_t))
else:
break
gt = gt.cuda()
white_level = white_level.cuda()
gt = gt / white_level
gt = gt.clamp(0.0, 1.0)
gt = sRGBGamma(gt)
if 'kpn_5x5' in val_dict:
_, kpn_5x5_pred = kpn_5x5(noisy.view(1, -1, 512, 512),
noisy[:, :8, ...], white_level)
kpn_5x5_pred = kpn_5x5_pred.clamp(0.0, 1.0)
kpn_5x5_pred = sRGBGamma(kpn_5x5_pred)
kpn_5x5_pred = kpn_5x5_pred.cpu()
psnr_t = calculate_psnr(kpn_5x5_pred, gt)
ssim_t = calculate_ssim(kpn_5x5_pred, gt)
psnr_kpn_5x5.append(psnr_t)
ssim_kpn_5x5.append(ssim_t)
trans_rgb(kpn_5x5_pred.squeeze()).save(
'./eval_images/{}_kpn_5x5_{:.2f}dB_{:.4f}.png'.format(
index, psnr_t, ssim_t),
quality=100)
del kpn_5x5_pred
if 'kpn_7x7' in val_dict:
_, kpn_7x7_pred = kpn_7x7(noisy.view(1, -1, 512, 512),
noisy[:, :8, ...], white_level)
kpn_7x7_pred = kpn_7x7_pred.clamp(0.0, 1.0)
def test_multi(args):
color = True
burst_length = args.burst_length
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if args.model_type == "attKPN":
model = Att_KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attKPN_Wave":
model = Att_KPN_Wavelet_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "attWKPN":
model = Att_Weight_KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=True,
spatial_att=True,
upMode="bilinear",
core_bias=False)
elif args.model_type == "KPN":
model = KPN_DGF(color=color,
burst_length=burst_length,
blind_est=True,
kernel_size=[5],
sep_conv=False,
channel_att=False,
spatial_att=False,
upMode="bilinear",
core_bias=False)
else:
print(" Model type not valid")
return
checkpoint_dir = "checkpoints/" + args.checkpoint
if not os.path.exists(checkpoint_dir) or len(
os.listdir(checkpoint_dir)) == 0:
print('There is no any checkpoint file in path:{}'.format(
checkpoint_dir))
# load trained model
ckpt = load_checkpoint(checkpoint_dir,
cuda=device == 'cuda',
best_or_latest=args.load_type)
state_dict = ckpt['state_dict']
# if not args.cuda:
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
model.load_state_dict(new_state_dict)
# else:
# model.load_state_dict(ckpt['state_dict'])
model.to(device)
print('The model has been loaded from epoch {}, n_iter {}.'.format(
ckpt['epoch'], ckpt['global_iter']))
# switch the eval mode
model.eval()
# model= save_dict['state_dict']
trans = transforms.ToPILImage()
torch.manual_seed(0)
noisy_path = sorted(glob.glob(args.noise_dir + "/*.png"))
clean_path = [i.replace("noisy", "clean") for i in noisy_path]
upscale_factor = int(math.sqrt(burst_length))
for i in range(len(noisy_path)):
image_noise, image_noise_hr = load_data(noisy_path[i], burst_length)
image_noise_hr = image_noise_hr.to(device)
# begin = time.time()
image_noise_batch = image_noise.to(device)
# print(image_noise_batch.size())
burst_size = image_noise_batch.size()[1]
burst_noise = image_noise_batch.to(device)
# print(burst_noise.size())
# print(image_noise_hr.size())
if color:
b, N, c, h, w = burst_noise.size()
feedData = burst_noise.view(b, -1, h, w)
else:
feedData = burst_noise
# print(feedData.size())
pred_i, pred = model(feedData, burst_noise[:, 0:burst_length, ...],
image_noise_hr)
# del pred_i
pred_i = pred_i.detach().cpu()
print(pred_i.size())
pred_full = pixel_shuffle(pred_i, upscale_factor)
pred_full = pred_full
print(pred_full.size())
pred = pred.detach().cpu()
# print("Time : ", time.time()-begin)
gt = transforms.ToTensor()(Image.open(clean_path[i]).convert('RGB'))
gt = gt.unsqueeze(0)
# print(pred_i.size())
# print(pred[0].size())
psnr_t = calculate_psnr(pred, gt)
ssim_t = calculate_ssim(pred, gt)
print(i, " pixel_shuffle UP : PSNR : ",
str(calculate_psnr(pred_full, gt)), " : SSIM : ",
str(calculate_ssim(pred_full, gt)))
print(i, " UP : PSNR : ", str(psnr_t), " : SSIM : ", str(ssim_t))
if args.save_img != '':
if not os.path.exists(args.save_img):
os.makedirs(args.save_img)
plt.figure(figsize=(15, 15))
plt.imshow(np.array(trans(pred[0])))
plt.title("denoise KPN DGF " + args.model_type, fontsize=25)
image_name = noisy_path[i].split("/")[-1].split(".")[0]
plt.axis("off")
plt.suptitle(image_name + " UP : PSNR : " + str(psnr_t) +
" : SSIM : " + str(ssim_t),
fontsize=25)
plt.savefig(os.path.join(
args.save_img, image_name + "_" + args.checkpoint + '.png'),
pad_inches=0)
"""
def train(config, restart_training, num_workers, num_threads):
torch.set_num_threads(num_threads)
print("Using {} CPU threads".format(torch.get_num_threads()))
# TODO: de-hardcode this one.
N_CHANNEL = 3
train_config = config["training"]
batch_size = train_config["batch_size"]
lr = train_config["learning_rate"]
w_decay = train_config["weight_decay"]
step_size = train_config["decay_steps"]
gamma = train_config["lr_decay"]
betas = (train_config["beta1"], train_config["beta2"])
n_epochs = train_config["num_epochs"]
dataset_configs = train_config["dataset_configs"]
use_cache = train_config["use_cache"]
print("Configs:", config)
# create dir for model
checkpoint_dir = train_config["checkpoint_dir"]
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
logger = Logger(train_config["logs_dir"])
use_gpu = torch.cuda.is_available()
num_gpu = list(range(torch.cuda.device_count()))
print("Using On the fly TRAIN datasets")
train_data = OnTheFlyDataset(train_config["dataset_configs"],
im_size=(train_config["image_width"],
train_config["image_height"]),
use_cache=use_cache)
train_loader = DataLoader(train_data,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
model = get_model(config["architecture"])
l1_loss = nn.SmoothL1Loss()
if use_gpu:
ts = time.time()
model = model.cuda()
model = nn.DataParallel(model, device_ids=num_gpu)
print("Finish cuda loading, time elapsed {}".format(time.time() - ts))
# for sanity check
all_parameters = [
p for n, p in model.named_parameters() if p.requires_grad
]
if train_config["optimizer"] == "adam":
print("Using Adam.")
optimizer = optim.Adam([
{
'params': all_parameters
},
],
lr=lr,
betas=betas,
weight_decay=w_decay,
amsgrad=True)
elif train_config["optimizer"] == "sgd":
print("Using SGD.")
optimizer = optim.SGD([
{
'params': all_parameters
},
],
lr=lr,
momentum=betas[0],
weight_decay=w_decay)
else:
raise ValueError(
"Optimizer must be 'sgd' or 'adam', received '{}'".format(
train_config["optimizer"]))
scheduler = lr_scheduler.StepLR(optimizer,
step_size=step_size,
gamma=gamma)
n_global_iter = 0
average_loss = MovingAverage(train_config["n_loss_average"])
best_loss = np.inf
checkpoint_loaded = False
if not restart_training:
try:
checkpoint = load_checkpoint(checkpoint_dir, 'best')
start_epoch = checkpoint['epoch']
n_global_iter = checkpoint['global_iter']
best_loss = checkpoint['best_loss']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
checkpoint_loaded = True
print("=> loaded checkpoint (epoch {})".format(
checkpoint['epoch']))
except:
start_epoch = 0
n_global_iter = 0
best_loss = np.inf
print("=> load checkpoint failed, training from scratch")
else:
start_epoch = 0
print("=> training from scratch")
for epoch in range(start_epoch, n_epochs):
scheduler.step()
ts = time.time()
t4 = None
t_generate_data = []
t_train_disc = []
t_train_gen = []
t_vis = []
t_save = []
for iter, batch in enumerate(train_loader):
if t4 is not None:
# collect information and print out average time.
t0_old = t0
t0 = time.time()
if t4 is not None:
t_generate_data.append(t0 - t4)
t_train_disc.append(t1 - t0_old)
t_train_gen.append(t2 - t1)
t_vis.append(t3 - t2)
t_save.append(t4 - t3)
N_report = 100
N_print = 1000
if (iter % N_report) == 0:
t_generate_data = np.mean(t_generate_data)
t_train_disc = np.mean(t_train_disc)
t_train_gen = np.mean(t_train_gen)
t_vis = np.mean(t_vis)
t_save = np.mean(t_save)
t_total = t_generate_data + t_train_disc + t_train_gen + t_vis + t_save
if (iter % N_print) == 0:
print("t_generate_data: {:0.4g} s ({:0.4g}%)".format(
t_generate_data, t_generate_data / t_total * 100))
print("t_train_disc: {:0.4g} s ({:0.4g}%)".format(
t_train_disc, t_train_disc / t_total * 100))
print("t_train_gen: {:0.4g} s ({:0.4g}%)".format(
t_train_gen, t_train_gen / t_total * 100))
print("t_vis: {:0.4g} s ({:0.4g}%)".format(
t_vis, t_vis / t_total * 100))
print("t_save: {:0.4g} s ({:0.4g}%)".format(
t_save, t_save / t_total * 100))
logger.scalar_summary('Steps per sec', 1.0 / t_total,
n_global_iter)
t_generate_data = []
t_train_disc = []
t_train_gen = []
t_vis = []
t_save = []
should_vis = ((n_global_iter + 1) % train_config["vis_freq"]) == 0
if use_gpu:
degraded_img = batch['degraded_img'].cuda()
target_img = batch['original_img'].cuda()
else:
degraded_img = batch['degraded_img']
target_img = batch['original_img']
t1 = time.time()
optimizer.zero_grad()
# Run the input through the model.
output_img = model(degraded_img)
loss = l1_loss(output_img, target_img)
loss.backward()
optimizer.step()
logger.scalar_summary('Loss', loss.data[0], n_global_iter)
psnr = calculate_psnr(output_img, target_img)
logger.scalar_summary('Train PSNR', psnr, n_global_iter)
average_loss.update(loss.data[0])
t2 = time.time()
if iter % 10 == 0:
print("epoch{}, iter{}, loss: {}" \
.format(epoch, iter, loss.data[0]))
n_global_iter += 1
if should_vis:
exp = batch['vis_exposure'] if 'vis_exposure' in batch else None
img = create_vis(degraded_img[:, :3, ...], target_img,
output_img, exp)
logger.image_summary("Train Images", img, n_global_iter)
t3 = time.time()
if (n_global_iter % train_config["save_freq"]) == 0:
if average_loss.get_value() < best_loss:
is_best = True
best_loss = average_loss.get_value()
else:
is_best = False
save_dict = {
'epoch': epoch,
'global_iter': n_global_iter,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
}
save_checkpoint(save_dict, is_best, checkpoint_dir,
n_global_iter)
t4 = time.time()
print("Finish epoch {}, time elapsed {}" \
.format(epoch, time.time() - ts))
