Pytorch_geometric: nn.DataParallel

[bolinpu]

Hi,
I just would like to confirm If could use the "nn.DataParallel" and how?

[rusty1s]

I do not think that is possible because we do not have a dedicated batch dimensionality :(
This is however a really important request and I leave this issue open till I come up with a way to tackle it.

[bolinpu]

we have data.batch. Hope you could solve it ASAP. Many thanks

[bolinpu]

Hi rusty1s,
Could you please tell me when you are going to finish this?

[rusty1s]

Will look at it in more detail in the upcoming week.

[rusty1s]

Hey,
I implemented a first version of torch_geometric.nn.DataParallel. See here for the doc and here for an example. The basic idea is borrowed from PyTorch. However, as we allow variable sized input, I created a new dataloader (torch_geometric.data.DataListLoader) which outputs a list of data objects, and the DataParallel module converts these data objects to a batch per device. Let me know what you think.

[bolinpu]

@ @rusty1s I will have a try ASAP. Many thanks

[bolinpu]

@rusty1s I got a error like the following:

RuntimeError: Expected tensor for argument #1 'input' to have the same device as tensor for argument #2 'weight'; but device 1 does not equal 2 (while checking arguments for cudnn_batch_norm)

[rusty1s]

Can you show me a minimal example of your code?

[rusty1s]

Closed in favour of #137.