Where the jacobian is being computed?

[mjack3]

Hello, this repository is very interesting. I am studying for a PhD in normalizing flow, and when reviewing your code, I could not see where the Jacobian matrix is calculated. Is it a repository for building flow-based networks?

Regards

[silvandeleemput]

Hi, thanks for your interest in MemCNN. The repository is mainly focused on rendering invertible operations memory-efficient and isn't designed in particular for flow-based networks. However, since flow-based networks rely on invertible mappings it should be feasible to get some utility out of the library.

The library doesn't contain an explicit method for computing a Jacobian matrix. PyTorch does have such methods in functorch: https://pytorch.org/functorch/stable/notebooks/jacobians_hessians.html