Batch Training and DataLoader in PyG

[khaled-rahman]

Hi, I was trying to train a model in PyTorch Lightning for large graphs.
To create a batch, I used NeighborSampler following pl examples.
To create the batch graph, is it possible to take all neighbors instead of size parameters?
More specifically, I would like to create the batch on nodes, e.g., if node_ids = [0, 1, 2, 3, 4, ..., 49], and batch_size=4, then I would like to take node ids [0, 1, 2, 3] in a batch, [4, 5, 6, 7] in the second batch and so on. Also, the edge_index for the first batch should contain all the edges of a graph that are incoming/outgoing from the nodes [0, 1, 2, 3], i.e., Data(x = [0, 1, 2, 3], edge_index = [[0, 1, 2, 2, 3,. ...], [2, 3, 5, 6, 0, ...]])

Is it possible to create such batches using NeighborSample or NeighborLoader?
It would be helpful to get some pointers on this.

[wsad1]

Setting sizes=-1 will get you all neighbors of a node in NeighborSampler as explained here.

Now NeighborSampler is depreciated so its better to use NeighborLoader. In NeighborLoader although its not mentioned in the docs, and I am not entirely sure, it looks like you can get all neighbors by setting num_neigbhors=[-1], @rusty1s can confirm this.

[khaled-rahman]

Thank you so much!

[rusty1s]

Added the missing piece of information to the doc, thanks @wsad1 :)