saurabhdash
Abnormal computation time when carrying out experiment?
I tested with my real data which has ~300 nodes and 8000 edges. It takes more than 1 day to perform the training. (Also tested data with 40 nodes and takes 10s every epoch. ) While the paper shows a computation time 1000ms for 1k nodes. Any probable reason from configuration? I used python 3.8.5, cuda 11.1, pytorch1.8.1 and other dependencies same as the requirements.txt. I used one Quadro P5000 graphic. On the other hand, are there any computation time result with larger data provided for reference? Thanks a lot.
The Custom Loss involves a for loop in python which is extremely inefficient. Since the adjacency matrix is small, you can use the default PyTorch backprop for the loss.
