ActNorm is incorrectly implemented

[mjack3]

A tensor with shape (1) is always returned as the log det Jacobean.

[fdraxler]

Thanks for raising this issue! Can you maybe produce a minimal code example reproducing the issue so that we can have a closer look? Thanks!

[mjack3]

You can reproduce the error by ingesting a simple tensor of shape batch x channnel x height x width.

For example:

x = torch.rand(32,128,28,28)

And then ingest it to a INN using the sequinn API

[ju-w]

I fail to see the problem.

def build():
    nodes = [     Ff.InputNode(3,32,32) ]
    nodes.append( Ff.Node(nodes[-1], Fm.ActNorm) )
    nodes.append( Ff.OutputNode(nodes[-1]))
    inn = Ff.GraphINN(nodes)
    return inn
inn = build()
x, jac_det = inn(torch.rand(190,3,32,32))
print(jac_det.shape)

As expected

torch.Size([190])

Of course it is 190 times the same number, as jac_det is only determined by the scaling inside ActNorm.

[mjack3]

You are right. I have checked it in the main and 0.2 versions. Both work fine. Sorry for the false alarm. This issue can be closed :)

((I wonder if this repository is going to be more active, it is not common to see the team responding))