Multi-gpu training with custom dataset in pytorch-geometric

1

Hello,

I created a custom dataset list exactly using your example at the end of this link: (https://pytorch-geometric.readthedocs.io/en/latest/notes/create_dataset.html)

It looks like this:
data_list = [Data(…), …, Data(…)]

I then try to use DatalistLoader instead of Dataloader to perform multi-gpu training but they dont seem to be compatible. What is the way to do it based on the data I provide?

2 Likes
2

Hi,
I have an related issue and I didn’t find a solution.
I have a custom dataset which inherits from torch_geometric Data.
In the single-GPU version I use torch_geometric DataLoader with follow_batch argument - because the collate of my dataset is not trivial.

I want to convert my code to multi-GPU - so I tried to yse torch_geometric DataParallel with DataListLoader, but this loader doesn’t support follow_batch.
What can I do?

Thanks.