def weighted_bce_loss(input,
target,
ignore_index=None,
reduction='mean',
from_logits=True):
n = target.size()[1:].numel()
n = torch.full((len(target), ),
n,
dtype=target.dtype,
device=target.device)
dim = dims(target)[1:]
if ignore_index is not None:
weight = (target != ignore_index).float()
n = torch.sum(weight, dim=dim)
target = target.masked_fill(target == ignore_index, 0)
else:
weight = 1
n_pos = torch.sum(target, dim=dim)
neg_weight = n_pos / n
neg_weight = unsqueeze(neg_weight, dim)
pos_weight = 1 - neg_weight
weight *= target * pos_weight + (1 - target) * neg_weight
if from_logits:
return F.binary_cross_entropy_with_logits(input,
target,
weight,
reduction=reduction)
else:
return F.binary_cross_entropy(input,
target,
weight,
reduction=reduction)
def f1_loss(pred, target, eps=1e-8, average='micro'):
assert pred.shape == target.shape
if average == 'samples':
dim = dims(pred)[1:]
tp = torch.sum(pred * target, dim=dim)
fp = torch.sum((1 - pred) * target, dim=dim)
fn = torch.sum(pred * (1 - target), dim=dim)
elif average == 'micro':
tp = torch.sum(pred * target)
fp = torch.sum((1 - pred) * target)
fn = torch.sum(pred * (1 - target))
else:
raise ValueError(
"`average` must be one of [`samples`, 'micro'], got `%s`" %
average)
p = tp / (tp + fp + eps)
r = tp / (tp + fn + eps)
f1 = 2 * p * r / (p + r + eps)
return 1 - torch.mean(f1)
def dice_loss(pred, target):
dim = dims(pred)
numerator = 2 * torch.sum(pred * target, dim=dim)
denominator = torch.sum(pred + target, dim=dim)
losses = 1 - (numerator + 1) / (denominator + 1)
return torch.mean(losses)
