Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Datasets are downloading only on node 0, rank 0 in Ignite examples.

See original GitHub issue

🐛 Bug description

Datasets are downloading only on node 0, rank 0 in Ignite examples. idist.get_rank() should be idist.get_local_rank()

https://github.com/pytorch/ignite/blob/13a42b90589112e67bcc6e1780e7c826a5a34199/examples/contrib/cifar10/main.py#L204-L214

https://github.com/pytorch/ignite/blob/13a42b90589112e67bcc6e1780e7c826a5a34199/examples/contrib/cifar10_qat/main.py#L189-L199

https://github.com/pytorch/ignite/blob/13a42b90589112e67bcc6e1780e7c826a5a34199/examples/contrib/transformers/main.py#L207-L219

cc @vfdev-5 for confirmation

Issue Analytics

State:
Created 2 years ago
Reactions:1
Comments:9 (6 by maintainers)

Top GitHub Comments

1reaction

sdesroziscommented, Apr 10, 2021

I think that is correct to download dataset only on rank 0.

Usually, if you have multiple nodes, dataset is downloaded on a shared directory on the file system. We don’t want to have multiple datasets on memory.

0reactions

sdesroziscommented, Apr 11, 2021

I’m not sure. Cifar10 is a small dataset, I think that what is currently done is good.