Frederik Künstner

Not obvious from the logs without the code. It's weird that the error appears after the call on `CrossEntropyLoss()`. This should be the start of a new backward pass and...

Hi @Ludvins, From what I understand, your setup looks like this. You have a function `f ∶ Rᵃ x Rᵇ → Rᶜ` with inputs `x ∈ Rᵃ, 𝜃 ∈ Rᵇ`....

Got it, assuming the other difference is that `c = 1`. If you only want to use this matrix by doing matrix-vector products, ie computing `v.T @ J.T @ J...

If I got your setup right, that will still be difficult without implementing the batchnorm operation. There is no option to disable the extension on the batchnorm parameters only, because...