Custom TQDMProgressBar changes not reflected

[oseymour]

Bug description

I wrote a custom TQDMProgressBar class with some changes. When I run train.fit() in JupyterLab the default progress bar is still used, however.

What version are you seeing the problem on?

v2.4

How to reproduce the bug

from lightning.pytorch.callbacks import TQDMProgressBar

class CustomProgBar(TQDMProgressBar):
    def __init__(self, ncols: int = 100):
        super().__init__(leave=True)
        self.ncols = ncols
    
    def init_sanity_tqdm(self):
        bar = super().init_sanity_tqdm()
        bar.ncols = self.ncols
        return bar

    def init_train_tqdm(self):
        bar = super().init_train_tqdm()
        bar.ncols = self.ncols
        return bar

    def init_validation_tqdm(self):
        bar = super().init_validation_tqdm()
        bar.ncols = self.ncols
        return bar


trainer = L.Trainer(accelerator="cpu", max_epochs=5, callbacks=[CustomProgBar(),], log_every_n_steps=1)

# `model` and `data` are LightningModule and LightningDataModule instances, respectively.
# I can include the code for this if you think it's needed for debugging this.
trainer.fit(model, datamodule=data)

Error messages and logs

Printout without callbacks argument passed:

 | Name  | Type     | Params | Mode
------------------------------------------
0 | model | UNet     | 3.0 M  | eval
1 | loss  | DiceLoss | 0      | eval
------------------------------------------
3.0 M     Trainable params
0         Non-trainable params
3.0 M     Total params
11.893    Total estimated model params size (MB)
0         Modules in train mode
112       Modules in eval mode
Sanity Checking: |                                                                               | 0/? [00:00<…
Training: |                                                                                      | 0/? [00:00<…

Printout with just the CustomProgBar callaback:

| Name  | Type     | Params | Mode
------------------------------------------
0 | model | UNet     | 3.0 M  | eval
1 | loss  | DiceLoss | 0      | eval
------------------------------------------
3.0 M     Trainable params
0         Non-trainable params
3.0 M     Total params
11.893    Total estimated model params size (MB)
0         Modules in train mode
112       Modules in eval mode
Sanity Checking: |                                                                               | 0/? [00:00<…
Training: |                                                                                      | 0/? [00:00<…

Environment

Current environment

#- PyTorch Lightning Version (e.g., 2.4.0): 2.4.0
#- PyTorch Version (e.g., 2.4): 2.4.0
#- Python version (e.g., 3.12): 3.12.3
#- OS (e.g., Linux): Windows 10
#- CUDA/cuDNN version: n/a
#- GPU models and configuration: none, CPU only
#- How you installed Lightning(`conda`, `pip`, source): pip
#- TQDM version: 4.66.6

More info

No response

cc @lantiga

[SkafteNicki]

@oseymour thanks for reporting this issue. The reason you are not seeing that effect of changing the n_cols parameter in your custom TQDMProgressBar is that the dynamic_ncols parameter is set to True in the parent class

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

Line 205 in a08b64e

dynamic_ncols=True,

which

If set, constantly alters ncols and nrows to the environment (allowing for window resizes) [default: False].
https://tqdm.github.io/docs/tqdm/
the solution here is to simply change this in your custom class:

from lightning.pytorch.callbacks import TQDMProgressBar

class CustomProgBar(TQDMProgressBar):
    def __init__(self, ncols: int = 100):
        super().__init__(leave=True)
        self.ncols = ncols
    
    def init_sanity_tqdm(self):
        bar = super().init_sanity_tqdm()
        bar.ncols = self.ncols
        bar.dynamic_ncols = self.use_dynamic_ncols
        return bar

    def init_train_tqdm(self):
        bar = super().init_train_tqdm()
        bar.ncols = self.ncols
        bar.dynamic_ncols = self.use_dynamic_ncols
        return bar

    def init_validation_tqdm(self):
        bar = super().init_validation_tqdm()
        bar.ncols = self.ncols
        bar.dynamic_ncols = self.use_dynamic_ncols
        return bar

Closing issue, but feel free to ping and reopen if necessary.