gnn_benchmark_dataset raw data format

[devnkong]

https://github.com/rusty1s/pytorch_geometric/blob/e425622d6efc6832b15e9fe577710a7119d76cef/torch_geometric/datasets/gnn_benchmark_dataset.py#L53-L54

Hello Matthias,

I wanna know how I can generate the raw data used above. From the original paper their raw data is loaded here, which seems quite diffrerent from yours. Thanks!

Best Kezhi

[rusty1s]

The raw data is pre-processed by me, so that it can be accessed without using pickle. Note that pickle has trouble loading the data without having access to the classes and file structure of the GNNBenchmark repo). Other-wise, the data is equivalent.

[devnkong]

I see, is it possible for you to share the preprocessing script? I wanna create a dataset using GNN Benchmark‘s method but do wanna code under the PyG scheme, thx! Could possibly move faster with your help!

[rusty1s]

This is the one I used for loading the SBM datasets (others are similar):

name = 'PATTERN'

dataset = SBMsDataset(f'SBM_{name}')

num_train = len(dataset.train)
num_val = len(dataset.val)

data_list = []
for G, y in chain(dataset.train, dataset.val, dataset.test):
    y = y.to(torch.long)
    x = torch.nn.functional.one_hot(G.ndata['feat']).to(torch.float)

    row, col = G.edges()
    edge_index = torch.stack([row, col], dim=0).to(torch.long)
    data = {'x': x, 'y': y, 'edge_index': edge_index}
    data_list.append(data)

train_data_list = data_list[:num_train]
val_data_list = data_list[num_train:num_train + num_val]
test_data_list = data_list[num_train + num_val:]

[devnkong]

Thank you!