def accumulate(self, learn):
for i in range(len(learn.pred)):
pred = learn.pred[i]["masks"]
targ = learn.yb[i]["masks"].squeeze(1)
# In case not predicted mask. Set all pixels to background
if pred.shape[0] > 0:
pred = pred[0]
else:
pred = torch.zeros(targ.shape).to(targ.device)
pred, targ = flatten_check(pred, targ)
self.inter += (pred * targ).float().sum().item()
self.union += (pred + targ).float().sum().item()
def medae(inp, targ):
"Mean absolute error between `inp` and `targ`."
inp, targ = flatten_check(inp, targ)
e = torch.abs(inp - targ)
return torch.median(e).item()
def mpe(y_pred, target) -> Tensor:
y_pred, target = flatten_check(y_pred, target)
loss = (y_pred - target) / (target + 1e-8)
return loss.mean()
def smape(y_pred: Tensor, target: Tensor) -> Tensor:
y_pred, target = flatten_check(y_pred, target)
loss = 2 * (y_pred - target).abs() / (y_pred.abs() + target.abs() + 1e-8)
return loss.mean()
def huber(inp, targ):
"""Huber error between `inp` and `targ`."""
return F.smooth_l1_loss(*flatten_check(inp, targ))
def accumulate(self, learn):
pred, targ = flatten_check(learn.pred, learn.y)
self.inter += (pred * targ).float().sum().item()
self.union += (pred + targ).float().sum().item()