jcy132 / PatchGCL

Official Pytorch implementation of "Patch-wise Graph Contrastive Learning for Image Translation" (AAAI 2024)
MIT License
8 stars 1 forks source link

Parallel Training does not work. #4

Open koraberisha opened 1 month ago

koraberisha commented 1 month ago

So, ive tried multiple times to get parallel training to work, but to no avail, many modifications needed to the top_k_graph function, and the GNN_Loss class, which makes things a bit trickier, especially given only batch size 1 works!

Just wondering if this project is still in development or i should give solving these ago myself. Thanks.

Things tried:

DataParallel DataDistributedParrallel

Reason:

GPU Utilization is always inhibited, my hunch is onloading/offloading from CUDA too often in the patchGCL model class. Specifically the nonzero_graph function, leveraging scipy&numpy is questionable but looking to hear more about it before i make any judgements!

Hope i can help somehow, will provide more info at request.