Progress bar in Jupyter adds newlines for each validation step

[tcwalther]

🐛 Bug

Pytorch lightning prints newlines when showing the validation progress bar. See attached screenshot.

This may be a bug with nested progress bars in tqdm. I'm using the recent tqdm 4.45.0; I also had the issue with tqdm 4.43.0.

To Reproduce

Steps to reproduce the behavior:

In a Jupyter notebook, run trainer.fit(model) on any model that has a validation step.

Code sample

Expected behavior

No newlines should be added.

Environment

cuda:
        GPU:
                Tesla K80
        available:           True
        version:             10.1
packages:
        numpy:               1.17.4
        pyTorch_debug:       False
        pyTorch_version:     1.4.0
        pytorch-lightning:   0.7.1
        tensorboard:         2.2.0
        tqdm:                4.45.0
system:
        OS:                  Linux
        architecture:
                64bit

        processor:           
        python:              3.7.3
        version:             #1 SMP Debian 4.9.168-1+deb9u4 (2019-07-19)

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!

[williamFalcon]

upgrade to master. i think it’s fixed there

[williamFalcon]

happy to reopen if it isn’t

[tcwalther]

Thanks for your swift reply. I'm afraid it's still an issue after upgrading.

[tcwalther]

What happens in Jupyter is that there's a bunch of empty divs left over. I expanded one for illustration, too.

[williamFalcon]

maybe submit a PR?

Looks like this issue might have some solutions.
tqdm/tqdm#375

[tcwalther]

Turns out this is a bug in tqdm. I will file an issue there.
Easiest way to reproduce:

from tqdm.auto import tqdm
from time import sleep

for _ in tqdm(range(5)):
    for _ in tqdm(range(5), leave=False):
        sleep(0.01)

Feel free to close this issue. And thanks for replying so swiftly, it really motivated to dig into the issue myself :-).

[tcwalther]

I had a look at that issue you linked, but it didn't seem fully relevant. I've filed a new issue:

tqdm/tqdm#935

[tcwalther]

Looks like this has been an issue between jupyterlab and ipywidgets for quite some time (jupyter-widgets/ipywidgets#1845), and seems to be a main motivator why the fastai people wrote their own progress bar library:

AnswerDotAI/fastprogress#6

Would you consider switching to fastprogress, too? I understand the library is a lot less common than tqdm, but it may be more suitable for this project.

[Borda]

let's open a new issue with a discussion about fastprogress,
see https://github.com/fastai/fastprogress

[tcwalther]

@Borda will do so now.