Oceanusity
commented
2 years ago Hello, I think the optimization step uses gradient descent (GD). In this way, it collects the gradients for each batch and updates the model weights when all the graph inputs are included.
As shown here, it uses the code loss.backward() for each batch. However, we only update the model weights with optimizer.step() when the gradients from all the graphs are collected. In this case, we can save the GPU memory to train PGExplainer over datasets with thousands of graphs.
Hi, I read pseudo code of paper and it seems the loss calculation and backpropagation should be done after getting prediction probabilities for all graph inputs (graph explanation) or nodes (node explanation), though current code calculates loss and backpropagates it per sample