Disable Progress bars for Train/Validation/Predict separately

[CompRhys]

🚀 Feature

Currently we can turn off progress bars by setting enable_progress_bar=False but this is an all or nothing solution.

Motivation

I am using PL trainer to make predictions in a small customer facing prototype, the progress bars make the output look very different to the previous solution we are proposing a drop in solution for. We can disable all the progress bars but it's useful to have them for monitoring training even if we do not want them when calling predict.

Pitch

Additional kwargs to disable progress bars for Trainer.predict calls

cc @justusschock @kaushikb11 @awaelchli @Borda @rohitgr7

[carmocca]

You could create a new Trainer instance for prediction with the progress bar disabled.

[justusschock]

@carmocca for such cases should we create a function like Trainer.new() that basically creates an identical copy of the trainer and takes in some kwargs to overwrite only specific values?

This function would basically gather all the initial kwargs and update them with the given kwargs

[rohitgr7]

This function would basically gather all the initial kwargs and update them with the given kwargs

just wondering how it will work in case the kwargs states are updated during trainer.some_call. For eg. logger, callbacks, etc.

[justusschock]

@rohitgr7 there are two possible options for this:
either we say that this is expected (so we don't care about this) as for example when you call .test on the new trainer, you'd want the checkpointing callback to already have the correct checkpoints for best model etc. from the preceeding training

The second thing would be to reset the original state (could be done for example by dumping an initial checkpoint, which we then load again)

[carmocca]

@justusschock We discussed adding such function in the past. I think it makes sense.

Do you mind writing a proposal in a separate issue about it? Closing this as I don't want to sidetrack the original question.