how could I stack the graph with different number of node and edge_index in to a mini-batch?

[luciaL]

I have thousands of graphs but with the different number of nodes and edge_index. During training, it will take a long time to stack along one dimension with Dataloader. So I'd like to try the mini-batch method. But the DenseDataLoader and cat_dim methods need the number of nodes is all the same.
class MyData(Data): def __cat_dim__(self, key, item): if key == 'foo': return None else: return super().__cat_dim__(key, item)

How could I accelerate the training? I think the first one is to batching along a new dimension without stacking. The two above methods not gonna work with the different number of nodes.
The second method is using SparseTensor, but this works with batch_size =1.

[rusty1s]

If you have a dataset with graphs of varying size, I suggest you use torch_geometric.data.DataLoader. If you notice slow batching, I suggest you try to increase the num_workers argument.

[luciaL]

Thank you so much for your reply.