wandb loses "global_step" in PyTorch event_writer

There doesn’t seem to be a correspondence between global_step and step in my graphs. I’m using wandb through eventwriter API in PyTorch. Debugging a bit, it seems

step gets converted to global_step here https://github.com/wandb/client/blob/6417dd926abe76dbb7c56e7017d2ee7d1c918eb5/wandb/tensorboard/__init__.py#L205

Then wandb.log gets called with this dict, it doesn’t see step so it assigns one automatically

Custom global step is useful to compare data efficiency consistency across runs – using “forward calls” as x-axis means your curves look 2x better when doubling batch size or number of workers. Using global data counter for steps gives easier to interpret curves @vanpelt

[vanpelt]

Yeah, we look for any metrics that are monotonically increasing and make them available as x-axis options. MLFlow, Tensorboard, and others allow you to go back in time during training and we don’t. By putting the step in metrics we’re punting the missing feature by allowing users to log historic steps at the cost of using them as an x-axis (but still being able to export via the python ap). The big rewrite will enable this behaviour. We’re seeing a bunch of users syncing tensorboard, so we should atleast default to global_step as the x-axis in this case.

[yaroslavvb]

Thanks for the info, closing it as duplicate of https://github.com/wandb/client/issues/613