Closed dfalbel closed 2 years ago
@rdinnager Could you check if this works in your setup?
I'm testing locally now. To get the libraries to be built on CI, you need to make a commit with #build-libs
somewhere in the commit message.
By the way, thanks for working on this! I was wondering whether the changes from the lantern-types
branch would mean I would need to rewrite some parts of styleganr
. I didn't expect you to do it for me, but it is much appreciated! I have kind of put this project on hold though until https://github.com/rdinnager/styleganr/issues/1 gets resolved on the torch
side (please let me know if there is anything I can do to help with this). I tried your suggestion of writing C++ autograd functions, but so far haven't been able to get it to work. I think the original stylegan
uses a lot of the features of the Python autograd API, which makes it difficult to properly translate to C++ (like calling an autograd function object within an autograd function to get doubly differentiable functions, which it seems you can't do in C++? or at least I can't figure out how). At this point I'm happy to wait for a fix to the R autograd API (but I know this is a tricky problem, so please don't feel pressure, I am in no hurry).
Thanks @rdinnager ! I think I solved the last issues.
Ok! It makes sense that writing the C++ version of the autograd stuff is quite complicated. It's high-priority to me to fix the leak in torch. I did a small investigation while in lantern-types and I think I found the culprit, I'll try to fix it soon.
Regarding the testing failures, this will land into the main branch really soon.
Thanks @dfalbel! Wonderful! I've just switched back to using feature/install-headers
in DESCRIPTION
since I noticed you merged feature/lantern-types
into that, so I could test the build, but good to hear it will be on the main branch soon!
This is using the new lantern types to simplify a little object handling. Still a work in progress.
Missing:
TensorTensorInt
correctly.