Facundo Sapienza
Facundo Sapienza
The tests to generate these results are included in #274
Maybe... I am playing closer to the implementation of discrete and continuous this days. It is important to mention that continuous will not give you the same as finite differences,...
I thing I found a reason for the discrepancy: the continuos VJP I wrote is derived with an extra integration by parts that is not really required. This is the...
I am not sure more cases will be covered with what we have in `SIA2D`: a diffusivity where we parametrized a neural network NN that takes as an input H...
@JordiBolibar @albangossard this is still work in progress and I am sure there is a lot of place for improvement, but I made some changes in the code to handle...
@JordiBolibar @albangossard please take a look at the latest changes in the code. Any feedback is welcome! Please still don't merge into main, although I don't think we will merge...
@JordiBolibar would you please take a quick look at then Enzyme code? I didn't make major changes to the existing code with Enzyme, but this still breaks. Happy to debug...
I changed the API to be now based on multiple distpatch. The target object and all the code looks much clean now and I imagine this will serve better in...
Hi @albangossard ! Thank you for the comment. I fixed this in the last commit, where I restored the VJPs test to work again. A model of \theta needs to...
@albangossard I change the inversion tests to see if now they pass. Also notice I added new tests for the pure D inversion, so now we have three different inversions...