LitProgressBar ignores TQDM_MINITERS

[wouterzwerink]

Bug description

tqdm supports setting the environment variable TQDM_MINITERS to reduce the frequency of progress bar updates.
Since lightning trainer defaults to a TQDM progress bar, I expected this to work for lightning. However, pytorch lightnings trainer seems to ignore this.

What version are you seeing the problem on?

v2.1

How to reproduce the bug

export TQDM_MINITERS=5
python your_script.py


where script can be anything using a `pytorch_lightning.Trainer` with `progress_bar_enabled=True`

Error messages and logs

Environment

Current environment

#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow): Trainer
#- PyTorch Lightning Version (e.g., 1.5.0): 2.1.3
#- Lightning App Version (e.g., 0.5.2):
#- PyTorch Version (e.g., 2.0): 2.0.1
#- Python version (e.g., 3.9): 3.9
#- OS (e.g., Linux): Linux
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source): poetry
#- Running environment of LightningApp (e.g. local, cloud):

More info

No response

cc @awaelchli

[awaelchli]

@wouterzwerink Could you point us to the documentation for this? There is nothing like that in the source code of tqdm: https://github.com/search?q=repo:tqdm/tqdm+TQDM_MINITERS&type=code
and
https://github.com/search?q=repo:tqdm/tqdm+MINITERS&type=code

[wouterzwerink]

@awaelchli yes unfortunately it's documented quite poorly!
They use this wrapper to overwrite their function defaults with environment variables
https://github.com/tqdm/tqdm/blob/4c956c20b83be4312460fc0c4812eeb3fef5e7df/tqdm/utils.py#L34

Applied here to TQDM:
https://github.com/tqdm/tqdm/blob/4c956c20b83be4312460fc0c4812eeb3fef5e7df/tqdm/std.py#L953

[awaelchli]

Thanks for the pointer! My blind guess is that maybe it doesn't work because we have our own small wrapper:

pytorch-lightning/src/lightning/pytorch/callbacks/progress/tqdm_progress.py

Line 39 in ee9f17e

class Tqdm(_tqdm):

(but it should, since we call super().__init__?). Might be worth investigating what's going on there. Apart from that, I don't see why it wouldn't work because we're using tqdm quite plainly.

[wouterzwerink]

I tried just adding

if __name__ == "__main__":
    import time

    for _ in Tqdm(range(100)):
        time.sleep(1)

to the bottom of this file: pytorch-lightning/src/lightning/pytorch/callbacks/progress/tqdm_progress.py

Calling it with and without TQDM_MINITERS set works as expected. So this Tqdm wrapper class is not causing any issues

[awaelchli]

Ah ok, I understand what miniters does now. miniters=x is supposed to update only ever x ticks. But then it makes sense that it doesn't work with Lightning because our progress bar implementation directly updates the bar (there is no iterable over it), because we need explicit and dynamic control over when the update happens:

pytorch-lightning/src/lightning/pytorch/callbacks/progress/tqdm_progress.py

Lines 448 to 451 in 1a59097

	def _update_n(bar: _tqdm, value: int) -> None:
	if not bar.disable:
	bar.n = value
	bar.refresh()

But I think there is probably an easy way to incorporate it by just grabbing the default refresh rate from the env variable here:

pytorch-lightning/src/lightning/pytorch/callbacks/progress/tqdm_progress.py

Lines 429 to 434 in 1a59097

	def _resolve_refresh_rate(refresh_rate: int) -> int:
	if os.getenv("COLAB_GPU") and refresh_rate == 1:
	# smaller refresh rate on colab causes crashes, choose a higher value
	rank_zero_debug("Using a higher refresh rate on Colab. Setting it to `20`")
	return 20
	return refresh_rate

Would you be interested to contribute this change?