def __init__(self, cfg):
super(Solver, self).__init__(cfg)
self.init_epoch = self.cfg['schedule']
net_name = self.cfg['algorithm'].lower()
lib = importlib.import_module('model.' + net_name)
net = lib.Net
self.model = net(num_channels=self.cfg['data']['n_colors'],
base_filter=64,
args=self.cfg)
self.optimizer = maek_optimizer(self.cfg['schedule']['optimizer'], cfg,
self.model.parameters())
self.loss = make_loss(self.cfg['schedule']['loss'])
self.log_name = self.cfg['algorithm'] + '_' + str(
self.cfg['data']['upsacle']) + '_' + str(self.timestamp)
# save log
self.writer = SummaryWriter('log/' + str(self.log_name))
save_net_config(self.log_name, self.model)
save_yml(cfg, os.path.join('log/' + str(self.log_name), 'config.yml'))
save_config(
self.log_name,
'Train dataset has {} images and {} batches.'.format(
len(self.train_dataset), len(self.train_loader)))
save_config(
self.log_name, 'Val dataset has {} images and {} batches.'.format(
len(self.val_dataset), len(self.val_loader)))
save_config(
self.log_name, 'Model parameters: ' +
str(sum(param.numel() for param in self.model.parameters())))
def __init__(self, cfg):
super(Solver, self).__init__(cfg)
self.init_epoch = self.cfg['schedule']
net_name = self.cfg['algorithm'].lower()
lib = importlib.import_module('model.' + net_name)
net = lib.Net
if self.cfg['algorithm'] == 'VDSR' or self.cfg['algorithm'] == 'SRCNN':
train_dataset = DatasetFromHdf5("data/train.h5")
self.train_loader = DataLoader(train_dataset,
cfg['data']['batch_size'],
shuffle=False,
num_workers=1)
val_dataset = DatasetFromHdf5("data/test.h5")
self.val_loader = DataLoader(val_dataset,
cfg['data']['batch_size'],
shuffle=False,
num_workers=1)
self.model = net(args=self.cfg)
self.optimizer = maek_optimizer(self.cfg['schedule']['optimizer'], cfg,
self.model.parameters())
self.milestones = list(
map(lambda x: int(x), self.cfg['schedule']['decay'].split('-')))
self.scheduler = lr_scheduler.MultiStepLR(
self.optimizer,
self.milestones,
gamma=self.cfg['schedule']['gamma'],
last_epoch=-1)
self.loss = make_loss(self.cfg['schedule']['loss'])
self.log_name = self.cfg['algorithm'] + '_' + str(
self.cfg['data']['upsacle']) + '_' + str(self.timestamp)
# save log
self.writer = SummaryWriter('log/' + str(self.log_name))
# if not self.cfg['algorithm'] == 'SMSR':
# summary(self.model, (3, self.cfg['data']['patch_size'], self.cfg['data']['patch_size']), device='cpu')
save_net_config(self.log_name, self.model)
save_net_py(self.log_name, net_name)
save_yml(cfg, os.path.join('log/' + str(self.log_name), 'config.yml'))
save_config(
self.log_name,
'Train dataset has {} images and {} batches.'.format(
len(self.train_dataset), len(self.train_loader)))
save_config(
self.log_name, 'Val dataset has {} images and {} batches.'.format(
len(self.val_dataset), len(self.val_loader)))
save_config(
self.log_name, 'Model parameters: ' +
str(sum(param.numel() for param in self.model.parameters())))
def __init__(self, cfg):
super(Solver, self).__init__(cfg)
self.init_epoch = self.cfg['schedule']
net_name = self.cfg['algorithm'].lower()
lib = importlib.import_module('model.' + net_name)
net = lib.Net
if self.cfg['algorithm'] == 'VDSR' or self.cfg['algorithm'] == 'SRCNN':
train_dataset = DatasetFromHdf5("data/train.h5")
self.train_loader = DataLoader(train_dataset,
cfg['data']['batch_size'],
shuffle=False,
num_workers=1)
val_dataset = DatasetFromHdf5("data/test.h5")
self.val_loader = DataLoader(val_dataset,
cfg['data']['batch_size'],
shuffle=False,
num_workers=1)
self.model = net(num_channels=self.cfg['data']['n_colors'],
base_filter=64,
scale_factor=self.cfg['data']['upsacle'],
args=self.cfg)
self.optimizer = maek_optimizer(self.cfg['schedule']['optimizer'], cfg,
self.model.parameters())
self.loss = make_loss(self.cfg['schedule']['loss'])
self.log_name = self.cfg['algorithm'] + '_' + str(
self.cfg['data']['upsacle']) + '_' + str(self.timestamp)
# save log
self.writer = SummaryWriter('log/' + str(self.log_name))
summary(self.model, (3, self.cfg['data']['patch_size'],
self.cfg['data']['patch_size']),
device='cpu')
save_net_config(self.log_name, self.model)
save_net_py(self.log_name, net_name)
save_yml(cfg, os.path.join('log/' + str(self.log_name), 'config.yml'))
save_config(
self.log_name,
'Train dataset has {} images and {} batches.'.format(
len(self.train_dataset), len(self.train_loader)))
save_config(
self.log_name, 'Val dataset has {} images and {} batches.'.format(
len(self.val_dataset), len(self.val_loader)))
save_config(
self.log_name, 'Model parameters: ' +
str(sum(param.numel() for param in self.model.parameters())))
def __init__(self, cfg, name):
super(Solver, self).__init__(cfg)
self.init_epoch = self.cfg['schedule']
net_name = self.cfg['algorithm'].lower()
lib = importlib.import_module('model.' + net_name)
net = lib.Net
self.model = net(num_channels=self.cfg['data']['n_colors'],
base_filter=64,
scale_factor=self.cfg['data']['upsacle'],
args=self.cfg)
self.train_dataset = get_data(
self.cfg,
str(self.cfg['train_dataset']) + '/' + str(name) + '.png',
str(self.cfg['train_dataset']) + '/' + str(name) + '.png',
self.cfg['data']['upsacle'])
self.train_loader = DataLoader(self.train_dataset,
self.cfg['data']['batch_size'],
shuffle=False,
num_workers=self.num_workers)
for iteration, batch in enumerate(self.train_loader, 1):
lr, hr, bic, hr_ref, bic_ref, file_name = Variable(
batch[0]), Variable(batch[1]), Variable(batch[2]), Variable(
batch[4]), Variable(batch[5]), (batch[6])
self.hr_ref = hr_ref
self.lr = lr
self.file_name = file_name
self.noise_init = get_noise(
32, 'noise',
(self.cfg['data']['patch_size'] * self.cfg['data']['upsacle'],
self.cfg['data']['patch_size'] * self.cfg['data']['upsacle']))
self.noise = self.noise_init.detach().clone()
self.optimizer = maek_optimizer(self.cfg['schedule']['optimizer'], cfg,
self.model.parameters())
self.loss = CycleLoss(scale=1 / 4, loss_type='MSE')
self.log_name = self.cfg['algorithm'] + '_' + str(
self.cfg['data']['upsacle']) + '_' + str(self.timestamp)
# save log
self.writer = SummaryWriter('log/' + str(self.log_name))
save_net_config(self.log_name, self.model)
save_yml(cfg, os.path.join('log/' + str(self.log_name), 'config.yml'))