jinheonbaek / gmt Goto Github PK

when i run the code. The torch version used is 1.4.0 the conda version is 10.1 but when I use torch_sparse version 0.5.1 the conda version is 10.0 and the torch_sparse version 0.4.4 not has SparseTensor, I want to know the right version of torch_sparse. I look forward to your early reply.

Hi! This is a fantastic work! However, can I ask if the reported performance is based on 100 times of training? In Section C.2 GRAPH CLASSIFICATION, you mentioned:

We then report the average performances on the validation and test sets, by performing overall experiments 10 times with different seeds.

Does this means your reported performance is based on 10folds*10seeds, which means 100 trained models in total? Or 10folds with different seeds?

Thanks for the codes and this is a interesting work.

I have a question about the implementation about GMPool_I.

In GMPool_I, the seed matrix S \in R^{1 \times k}, where the k is the number of nodes after GMPool_G.
But, In module GMPool_I, whatever S is, the attentional matrix A is a all-1 Matrix \in R^{1 \times k} because of the col-wise softmax

tensor([[[1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.]],

        ...,

        [[1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.]]], device='cuda:1',
       grad_fn=<SoftmaxBackward0>)

It seems that the output of GMPool_I is equivalent to sum(WH+b) with FFN and shortcut . It looks like more reasonable to use a row-wise softmax for GMPool_I. Is there some wrong with my understanding ?

Looking forward your reply.

Hi, this is a fantastic work. I'm wondering if you can provide the code for Efficiency section in you paper, e.g., experiments on the Erdos-Renyi graphs (I've check your repo and HaarPool's, and didn't find code about efficiency part). Furthermore, could you please provide the raw data of Figire 3 and Figire 4? We want to follow your work and settings.

Best wishes!

Hi, I'm very interested in this method.

There is a question about code in trainers. Is the dataset for validation the same as the test dataset?

    train_idxes = torch.as_tensor(np.loadtxt('./datasets/%s/10fold_idx/train_idx-%d.txt' % (self.args.data, fold_number),
                                            dtype=np.int32), dtype=torch.long)
    val_idxes = torch.as_tensor(np.loadtxt('./datasets/%s/10fold_idx/test_idx-%d.txt' % (self.args.data, val_fold_number),
                                            dtype=np.int32), dtype=torch.long)     
    test_idxes = torch.as_tensor(np.loadtxt('./datasets/%s/10fold_idx/test_idx-%d.txt' % (self.args.data, fold_number),
                                            dtype=np.int32), dtype=torch.long)

i look forward to your reply.