abhuse / polyloss-pytorch Goto Github PK

logit is [4,19,1024,1024] and label is [4,1024,1024]
Thank you very much

I'm using torch==1.11.0+cu113 and the example for Poly1 Cross-Entropy Loss throws an error RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn.

I just have to change line 7 to logits = torch.rand([batch_size, num_classes], requires_grad=True) for it to work.

Hi,
Thanks a lot for providing a PyTorch implementation of PolyLoss.

I have a binary classification problem in which the class weights are as below:
class 0: 86.2% of 462 data points
class 1: 13.8% of 462 data points

When I use your loss with the same default parameters, all my validation predictions turn out to be 0, which is the majority class. Could you please show me how I should exactly set my criterion weight?

criterion = Poly1CrossEntropyLoss(num_classes=num_classes, reduction='mean')

Hello, my tensor size for the label is (batch, channel, H, W), how do I apply polyloss-focal

Could you please tell me how to change poly Loss if I use it in the object detection？
When I directly added poly loss to object detection, there was a error:
RuntimeError: Class values must be non-negatives
thanks in advance !

Hi.
First of all, thanks for your great repo.

When I'm passing weights like this:

Poly1FocalLoss(num_classes=num_classes, 
                      reduction='mean',
                      weight=torch.tensor([0.1, 0.9]))

I get this error:
RuntimeError: The size of tensor a (256) must match the size of tensor b (2) at non-singleton dimension 3

B, C, H, W = 32, 2, 256, 256
Label is not one_hoted.