Description
User Flow
I am training a model, I want to see the current progress and how much is left of the work. If possible I would also like to see some key metrics attached to the end of the progress bar that is being updated with each round of new results. Something like:
https://devblog.pytorchlightning.ai/super-charged-progress-bars-with-rich-lightning-669653d6ab97
Current state
AIR is showing a trial table (the same UI as if a single trial Tune) that shows "RUNNING" state and some key metrics with cut-offs every few seconds. There are also other not so relevant information like node_ip etc. The issue is that this table is not tailored to the train case.
Proposal
Add an option for user to instead use progress bar. This will automatically disable the Tune type of logging.
API
Provide an in-hourse Callback implementation:
pbc = ProgressBarCallback(num) # number of times that `report` is expected to be called by the training function.
Trainer = TorchTrainer(train_loop_per_worker=..., run_config=RunConfig(callbacks=[pbc]))