if cfg.loss == "ce":
criterion = nn.CrossEntropyLoss(reduction='mean')
elif cfg.loss == "sce":
criterion = SymmetricCrossEntropyLoss(alpha=cfg.loss_cfg['sce']['alpha'],
beta=cfg.loss_cfg['sce']['beta'],
num_classes=cfg.n_class)
# criterion4 = NormalizedSymmetricCrossEntropyLoss(alpha=cfg.alpha, beta=cfg.beta, num_classes=cfg.n_class)
elif cfg.loss == "focal":
criterion = FocalLoss(gamma=cfg.loss_cfg['focal']['gamma'])
elif cfg.loss == "ce-dice":
criterion = nn.CrossEntropyLoss(reduction='mean')
elif cfg.loss == 'decouple-v1':
criterion = DecoupledSegLoss_v1(alpha=cfg.loss_cfg['sce']['alpha'],
beta=cfg.loss_cfg['sce']['beta'],
num_classes=cfg.n_class)
elif cfg.loss == 'decouple-v2':
criterion = DecoupledSegLoss_v2(alpha=cfg.loss_cfg['sce']['alpha'],
beta=cfg.loss_cfg['sce']['beta'],
gamma=cfg.loss_cfg['focal']['gamma'],
num_classes=cfg.n_class)
runner.train(dataset_train,
dataset_val,
criterion,
get_optimizer,
Trainer,
Evaluator,
collate,
dataset_test=dataset_test,
tester_func=SlideInference)
transformer_val = TransformerVal()
dataset_val = OralDataset(valset_cfg["img_dir"],
valset_cfg["mask_dir"],
valset_cfg["meta_file"],
label=valset_cfg["label"],
transform=transformer_val)
if cfg.loss == "ce":
criterion = nn.CrossEntropyLoss(reduction='mean')
elif cfg.loss == "sce":
criterion = SymmetricCrossEntropyLoss(alpha=cfg.loss_cfg['sce']['alpha'],
beta=cfg.loss_cfg['sce']['beta'],
num_classes=cfg.n_class)
# criterion4 = NormalizedSymmetricCrossEntropyLoss(alpha=cfg.alpha, beta=cfg.beta, num_classes=cfg.n_class)
elif cfg.loss == "focal":
criterion = FocalLoss(gamma=cfg.loss_cfg['focal']['gamma'])
elif cfg.loss == "ce-dice":
criterion = nn.CrossEntropyLoss(reduction='mean')
elif cfg.loss == 'decouple-v1':
criterion = DecoupledSegLoss_v1(alpha=cfg.loss_cfg['sce']['alpha'],
beta=cfg.loss_cfg['sce']['beta'],
num_classes=cfg.n_class)
elif cfg.loss == 'decouple-v2':
criterion = DecoupledSegLoss_v2(alpha=cfg.loss_cfg['sce']['alpha'],
beta=cfg.loss_cfg['sce']['beta'],
gamma=cfg.loss_cfg['focal']['gamma'],
num_classes=cfg.n_class)
runner.train(dataset_train, dataset_val, criterion, get_optimizer, Trainer,
Evaluator, collate)