About reproduction issue

[SteveTanggithub]

could u plz provide the every single hyperparameters setting in training and models? We are not able to reproduct the results in the paper. Thank you very much!

[xlzhou01]

I have the same problem with you. I found one error in the author code. I don't know if it's right. Specifically, it is in the train function in the main.py. Each for loop is equivalent to taking the first batch_size sample from the scrambled training set, so I think it ends at the end of the for loop, and it may not have completely traversed the training set.

for pos in pbar:
    selected_idx = np.random.permutation(len(train_graphs))[:args.batch_size]
    batch_graph = [train_graphs[idx] for idx in selected_idx]
    output = model(batch_graph)
    labels = torch.LongTensor([graph.label for graph in batch_graph]).to(device)
    loss = criterion(output, labels)

    # backprop
    if optimizer is not None:
        optimizer.zero_grad()  
        loss.backward()  
        optimizer.step()_