Using torch.distributed.launch, not horovod

[kjw9899]

What a great work! There are some questions related to your work.

Can I implemente your code by leveraging torch.distributed.launch without using horovod?

And I wonder if the performance in the paper comes out even when stage 3 is run without running stage 1 and stage 2.

Finally, I can't download the posetrack dataset at the moment, so I wonder if the posetrack dataset has a lot of impact on performance.

Thank you!