Shengchao Liu

Slower than TF?

3

I'm testing the speed-up of ResNet on TF and PyTorch. In TF, typically it can converge within 80k steps, which is 80k batches, and when we set batch-size=128, that should...

In the paper, it says `the lth P-GNN layer first samples k random anchor-sets...` I noticed that you called two sampling functions in the [`preselect_anchor` function](https://github.com/JiaxuanYou/P-GNN/blob/master/utils.py#L210-L222). One is for each...

Hi, I'm wondering which function did you use to generate the visualization on MUTAG? (Figure 4 and 5 in the paper) Or if authors have any follow-up on [this](https://github.com/RexYing/gnn-model-explainer/blob/master/explainer_main.py#L254-L256)? Some...

Sort the one-hot-encoding for reproducible results

3

To #58, the results are not reproducible. It's caused by the one-hot encoding for the label while loading the data. We can get reproducible results after fixing the class order.

Hi, thank you for sharing the code. I'm wondering if you have provided the pretrained checkpoints somewhere.

Hi there, Thank you for providing the codes. I got this exception when running `shapenet_airplane_test.sh` in the docker image: ``` RuntimeError: CUDA kernel failed : 209 ``` Any help would...

Hi there, Thank you again for providing this nice benchmark tool! Previously, I posted a confirmation issue (#254) on the data splits for SchNet training. Now I want to double-check...