the gradient of sinkhorn layer

[Dark-98]

Sorry, I don't find backward of sinkhorn layer. So how do you use the gradient of this layer. I really need your help.

[pemami4911]

Hi,

The Sinkhorn layer (

sinkhorn-policy-gradient.pytorch/spg/layers.py

Line 10 in e029640

class Sinkhorn(Module):

) is differentiable, and PyTorch will compute its gradient automatically whenever .backward() is called on the SPG loss (e.g.,

sinkhorn-policy-gradient.pytorch/train_spg.py

Line 395 in e029640

actor_loss.backward()

)

Hope this helps!