nnabla-ext-cuda
nnabla-ext-cuda copied to clipboard
Fix F.clip_grad_by_norm in the case of small norm
The operation of clip_by_norm must take the maximum value between the L2 norm and the clipping value.
See https://github.com/sony/nnabla/pull/572 .
cc. @TE-AkioHayakawa
I've added some changes.
- the unused kernel function has been removed
- the backward function is corrected
@TE-AkioHayakawa