Weighted binary cross entropy loss: is it deprecated?

Deep Learning
1

I’m trying to implement a Unet architecture for binary semantic segmentation which has PyTorch’s weighted BCELoss as loss function:

learn = unet_learner(dls, 
                     resnet34, 
                     loss_func=BCELoss(weight=weights),
                     opt_func=ranger,
                     self_attention=True,
                     act_cls=Mish,
).to_fp32()

but I get this error:

ValueError: Using a target size (torch.Size([32, 256, 256])) that is different to the input size (torch.Size([32, 2, 256, 256])) is deprecated. Please ensure they have the same size.

How can I get rid of the dimension in excess?
Does the Fastai library have a specific binary weighted loss?

2

Hello,

BCELoss requires the model to predict only one value per pixel, denoting the probability of it belonging to the positive category, whereas your neural net is generating two channels. Thus, the solution would be to either force the U-Net to output a single plane by passing n_out=1 or replace the loss function with CrossEntropyLoss, which expects as many channels as there are classes and also accepts weights.

Please let me know if you run into any other issues.

1 Like