Angle Based loss versus cosine inverse loss

[rohun-tripathi]

Thanks for your great work!

Have you experimented with arccos based normal loss. Does the performance vary when using arccos instead of (1 - normalized inner product)?

The code I am referring to is -

prod = ( grad_fake[:,:,None,:] @ grad_real[:,:,:,None] ).squeeze(-1).squeeze(-1)
fake_norm = torch.sqrt( torch.sum( grad_fake2, dim=-1 ) )
real_norm = torch.sqrt( torch.sum( grad_real2, dim=-1 ) )