I have read the pytorch tutorial and have some toy examples. But I am still very confused how to know if the code can be autograded or not. I just know that each variable (especially those involved in the computation graph) in the custom loss function should be in Varaible type. While if the loss computed by tensor.sum (or .dot) is float type, then we return a float type, can the custom loss function still be autograded?
I also give a toy example:
input is a torch variable of size BatchxnclassesxHxW representing probabilities for each class
target is a a also tensor, with batchx1xHxW
eps = 0.000001
assert set(list(uniques))<=set([0,1]), “target must only contain zeros and ones”
probs = F.softmax(input) #maybe it is not necessary target = target_one_hot.contiguous().view(-1,1).squeeze(1) result = probs.contiguous().view(-1,1).squeeze(1) intersect = torch.dot(result, target) target_sum = torch.sum(target) result_sum = torch.sum(result) union = result_sum + target_sum + (2*eps) IoU = intersect / union dice_total = 1 - 2*IoU return dice_total