def define_criterion(self, name):
if name.lower() == 'bce+dice':
self.criterion = Loss.BCE_Dice()
elif name.lower() == 'dice':
self.criterion = Loss.DiceLoss()
elif name.lower() == 'bce':
self.criterion = nn.BCEWithLogitsLoss()
elif name.lower() == 'robustfocal':
self.criterion = Loss.RobustFocalLoss2d()
elif name.lower() == 'lovasz-hinge' or name.lower() == 'lovasz':
self.criterion = Loss.Lovasz_Hinge(per_image=True)
elif name.lower() == 'bce+lovasz':
self.criterion = Loss.BCE_Lovasz(per_image=True)
else:
raise NotImplementedError(
'Loss {} is not implemented'.format(name))
def __init__(self, lr=0.005, fold=None, val_mode='max', criterion_name='lovasz', comment=''):
super(SegmentationNetwork, self).__init__()
self.lr = lr
self.fold = fold
self.scheduler = None
self.best_model_path = None
self.epoch = 0
self.val_mode = val_mode
if criterion_name == 'lovasz':
self.criterion = Loss.Lovasz_Hinge(per_image=True)
if self.val_mode == 'max':
self.best_metric = -np.inf
elif self.val_mode == 'min':
self.best_metric = np.inf
self.comment = comment
self.train_log = dict(loss=[], iou=[], mAP=[])
self.val_log = dict(loss=[], iou=[], mAP=[])
self.create_save_folder()
def define_criterion(self, name):
if name.lower() == 'bce+dice':
self.criterion = Loss.BCE_Dice()
elif name.lower() == 'dice':
self.criterion = Loss.DiceLoss()
elif name.lower() == 'bce':
# self.criterion = nn.BCEWithLogitsLoss()
# self.criterion = nn.CrossEntropyLoss(size_average=False, reduction='sum')
self.criterion = nn.CrossEntropyLoss()
# self.criterion = Loss.CE_SOFT()
elif name.lower() == 'robustfocal':
self.criterion = Loss.RobustFocalLoss2d()
elif name.lower() == 'lovasz-hinge' or name.lower() == 'lovasz':
self.criterion = Loss.Lovasz_Hinge(per_image=True)
elif name.lower() == 'bce+lovasz':
self.criterion = Loss.BCE_Lovasz(per_image=True)
else:
raise NotImplementedError(
'Loss {} is not implemented'.format(name))
def define_criterion(self, name):
if name.lower == 'lovasz':
self.criterion = Loss.Lovasz_Hinge(per_image=True)
