List number of batches for each local rank

[obrookes]

Hi everyone,

Is there an existing mechanism to print the number of batches i.e. len(dataloader) being processed by each rank in a multi-gpu setting?

Thanks :)

[akihironitta]

trainer.estimated_stepping_batches may be what you're looking for.
https://pytorch-lightning.readthedocs.io/en/1.6.4/api/pytorch_lightning.trainer.trainer.Trainer.html#pytorch_lightning.trainer.trainer.Trainer.estimated_stepping_batches

[obrookes]

Thanks for this @akihironitta!

[rohitgr7]

in a multi-GPU setting with DDP all the ranks processes same number of batches.

[rohitgr7]

trainer.estimated_stepping_batches gives you total optimization steps. for batches you can just do len(self.train_dataloader)

[obrookes]

Thanks for your reply @rohitgr7, it's much appreciated 👍. As I understand it, samples are replicated to ensure each rank processes the same number of batches - is there a mechanism to display / control which classes are up-sampled? I'm working on a long-tailed recognition problem where this could potentially have a large impact.