Smooth L1 loss shape

Greetings,

Could anyone let me know the shape of inputs of smooth_l1_loss?
I know there are 2 inputs - prediction output and target (gt).
Furthermore, what’s gonna be changed if I want to implement ‘class-wise smooth l1 loss’?

Thank you

The prediction output and the target must have the same shape: https://pytorch.org/docs/master/nn.html?highlight=smoothl1#torch.nn.SmoothL1Loss

What do you mean by “class-wise smooth l1 loss”?

Oh, actually, I found that in https://github.com/yjxiong/action-detection
It is a code about detection and authors called a loss function they used as ‘class-wise smooth l1 loss’.
I searched that keyword on Google but I couldn’t find any information at all.
Honestly, I doubt the effect of their formulation…