hsjang0/dpm-snc

dpm-snc's Issues

Multi-GPU Processing

Hi, I appreciate your hard work and it helps me really a lot.

I'm currently running the code on fully supervised setting with CORA dataset and GCN-based DPM in my server, which has 2 RTX 3090 NVIDIA GPU. I find that if there are more than two available GPUs, it automatically utilize parallelism. (Here)

However, I find error when using more than two GPUs, especially in adjacency list part. Adjacency list has a shape of (2, # number of edges) and when utilizing GPU parallelism, it splits adjacency list into two tensors where each has shape of (1, # number of edges), which is a not intended purpose and results in an error.

If I only use one GPU even I have more than that, code works fine. However, I want to utilize more than one GPU for fast training. Is there any possible way to solve this problem?

Issue in loss function

It's very interesting to see your work.

Is your loss only consider a simplified objective function?

DPM-SNC/DPM-GSP-semi-supervised/method_series/gaussian_ddpm_losses.py

Line 91 in 46ea3d0

torch.sum(((pred_y[train_mask]-noise[train_mask])**2), dim = -1)

I think that a variable is an epsilon, not a \hat{\mu} because of :

DPM-SNC/DPM-GSP-semi-supervised/method_series/gaussian_ddpm_losses.py

Line 91 in 46ea3d0

torch.sum(((pred_y[train_mask]-noise[train_mask])**2), dim = -1)