GAE with multiple graphs

[jangkj09]

Hi,

I'm trying to understand how to train a GAE (or VGAE) with multiple graphs. I see bits and pieces in the discussions and issues (#458, #3769, #3046, #1176, etc.) but nothing conclusive. Is there a tutorial or implementation that utilizes multiple graphs to train a GAE?

If not, do I need to simply do a RandomLinkSpit on each graph in a batch during the training loop as done in here (#458)?

I'm also not sure "integrate batch support in the decoder to fix (using the to_dense_batch method)" (#458 (comment)) means.

Could some guidance be provided? Unless there is already an implementation.

[rusty1s]

We sadly don't have an example for this yet, I am sorry. The process to do this would be

1. You call RandomLinkSplit on each data object before training.
2. You call batched_negative_sampling within your training loop to generate negative examples.

Everything else stays mostly the same as in our link_pred.py example. Happy to help you further. We can also work on this together in a small PR. Let me know if you are interested.

[jangkj09]

Sure, I would like to figure this out. It would be an opportunity to understand pyg more in depth as well.
Just one quick question, would I call RandomLinkSplit on each data object before I create the dataloader? Or after?

I tried it after and during a training loop that examines the batches, but it seemed to have some weird behavior where some of the data examples lose their edges. I'm not sure if its my dataset or if its something else. I will look into this myself.

Either way, please let me know how to proceed with the PR.

Thanks!

[rusty1s]

The split needs to be fixed in advance, so what you can do is:

dataset = MyDataset(..., transform=RandomLinkSplit(...))
train_dataset, val_dataset, test_dataset = zip(*[data for data in dataset])