Closed afternone closed 2 years ago
Computing gradient for the whole graph could take longer time. It is suggested to compute less than 64 nodes for gradient for each updates. A proper node number to update would be 32. I will release a new feature for node sampling to act like batch size in classical deep learning model. It will update gradient for some nodes in a graph randomly.
This line takes more than 200s on CPU, but the forward pass takes less than 1s on the same CPU.