)
# NETWORK SETTINGS
# Data
loaders = myutils.get_module(args.net_dir, 'loaders')
dataloaders, dataset_sizes = loaders.get_loaders(
dfs,
mean,
std,
size=params.size,
batch_size=params.batch_size,
num_workers=params.num_workers)
# Net
net = myutils.get_network(args.net_dir, params.network)
optimizer = myutils.get_optimizer(params.optimizer, net,
params.learning_rate,
params.momentum, params.weight_decay)
# Schedulers
#vgg
if params.network.startswith('vgg'):
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
mode='max')
#efficientnetb
if params.network.startswith('efficientnetb'):
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=2.4,
gamma=0.97)
#resnet
if params.network.startswith('resnet'):
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
fname = os.path.join(args.data_dir, f'val{fold}.csv')
val = pd.read_csv(fname)
dfs['train'] = train
dfs['val'] = val
# NETWORK SETTINGS
# Data
loaders = myutils.get_loaders(args.net_dir, 'loaders')
dataloaders, dataset_sizes = loaders.get_loaders(
dfs,
size=params.size,
batch_size=params.batch_size,
num_workers=params.num_workers)
# Net
net = myutils.get_network(args.net_dir, params.network)
optimizer = myutils.get_optimizer(params.optimizer, net,
params.learning_rate)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=7,
gamma=0.1)
# Loss function
weight = myutils.get_weight(train)
criterion = myutils.get_loss_fn(args.net_dir, params.network, weight)
logging_process.info(
f'Model: {args.model_dir}\tFile: train{fold}.csv\tWeight: {weight}'
)
# Train
print(f'Fold {fold}')
print('-' * 10)
logging_process.info(
f'Model: {args.model_dir}\tFold: {fold}, training has started for {params.num_epochs} epochs'