Rich progress bar doesn't play nicely with other prints

[ruro]

🐛 Bug

Some internal printing mechanisms don't play nicely with RichProgressBar.
This can result in the progress bar getting interleaved into the printed data.

For example, this affects the pretty table of metrics that is printed after Trainer.testing your model:

Testing ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1/1 0:00:00 • 0:00:00 0.00it/s  ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│         test_loss         │    -4.4067792892456055    │
└───────────────────────────┴───────────────────────────┘

A similar issue can also come up, when using RichModelSummary.summarize.

To Reproduce

Take the BoringModel sample code and add

callbacks=[pl.callbacks.RichProgressBar()]

to the Trainer.

Expected behavior

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│         test_loss         │    -4.4067792892456055    │
└───────────────────────────┴───────────────────────────┘
Testing ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1/1 0:00:00 • 0:00:00 0.00it/s

Environment

* CUDA:
	- GPU:
		- NVIDIA GeForce GTX 1050
	- available:         True
	- version:           11.3
* Packages:
	- numpy:             1.22.3
	- pyTorch_debug:     False
	- pyTorch_version:   1.11.0+cu113
	- pytorch-lightning: 1.6.0
	- tqdm:              4.63.2
* System:
	- OS:                Linux
	- architecture:
		- 64bit
		- ELF
	- processor:         
	- python:            3.10.4
	- version:           #1 SMP PREEMPT Mon Mar 28 09:16:36 UTC 2022

Fix

Instead of creating a new Console and printing to it, RichModelSummary.summarize and EvaluationLoop._print_results should check if trainer has a progress_bar_callback of type RichProgressBar and if so, it should use trainer.progress_bar_callback.progress.console.

Edit: Actually a better solution might be using rich.get_console() instead of creating new instances of rich.console.Console() everywhere and if you need to change the kwargs (because user passed console_kwargs in RichProgressBar) use rich.reconfigure(**console_kwargs).

cc @kaushikb11 @rohitgr7

[ruro]

Also, you might want to initialize the logging using the Rich handler, if it is available, otherwise all the warnings and pl.utilities.rank_zero.rank_zero_* print functions also break the progress bars.

Like this (in top level __init__.py):

if not _root_logger.hasHandlers():
    from pytorch_lightning.utilities.imports import _RICH_AVAILABLE
    if _RICH_AVAILABLE:
        from rich.logging import RichHandler
        _logger.addHandler(RichHandler(
            omit_repeated_times=False,
            show_level=False,
            show_path=False,
            show_time=False,
        ))
    else:
        _logger.addHandler(logging.StreamHandler())
    _logger.propagate = False

[carmocca]

These are both great suggestions. Would you like to work on the fixes yourself?

You can open 2 PRs. One for each of your comments

[ruro]

Sorry, I am kind of busy right now. I am trying to submit issues as I run into them, but I don't currently have the time to properly fix them. If this issue is still not fixed in a couple of weeks, I might revisit it and submit the PRs, but for now I'll have to pass.