NotImplementedError

[onlyonewater]

Hi, pyg team, I am a freshman for pgy package, when I run the code, I get the error like this:
File "/mnt/workspace/duhao/.conda/envs/diffdock_pp/lib/python3.10/site-packages/torch/nn/parallel/parallel_apply.py", line 61, in _worker output = module(*input, **kwargs) File "/mnt/workspace/duhao/.conda/envs/diffdock_pp/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl return forward_call(*input, **kwargs) File "/mnt/workspace/duhao/.conda/envs/diffdock_pp/lib/python3.10/site-packages/torch_geometric/nn/data_parallel.py", line 69, in forward inputs = self.scatter(data_list, self.device_ids) File "/mnt/workspace/duhao/.conda/envs/diffdock_pp/lib/python3.10/site-packages/torch_geometric/nn/data_parallel.py", line 77, in scatter count = torch.tensor([data.num_nodes for data in data_list]) File "/mnt/workspace/duhao/.conda/envs/diffdock_pp/lib/python3.10/site-packages/torch_geometric/data/feature_store.py", line 515, in __iter__ raise NotImplementedError
I do not why, since I use the data_parallel and the batch size is 4 with 4 GPUs

[rusty1s]

It looks like you are using FeatureStore in combination with nn.DataParallel. This is currently a mis-use, as nn.DataParallel cannot deal with single graphs (it expects a dataset of graphs instead). We have collected a few examples on how to use NeighborLoader on multi-GPU here.

[rusty1s]

Can you show me an example script of yours?

[onlyonewater]

it seems that I want to use four GPUs in DataParallel with model =DataParallel(model, device_ids=[4,5,6,7]), but in the beginning code, I use torch.cuda.set_device(4), so there exists a mismatch between the true GPU num (1) with my expected GPU numbers (4)? I think this may be the reason? I am not sure

[onlyonewater]

well, the code in https://github.com/ketatam/DiffDock-PP may be a little complicated.

[onlyonewater]

and now I confirm that using single GPU with DataParallel is OK, I use dataloader class in single GPU situation

[rusty1s]

Sorry for late reply. Is there any chance you can provide a minimal example here to reproduce? How does the input of the model look like? Note also that we deprecated DataParallel a while ago, so I strongly suggest to switch to DistributedDataParallel instead.