How to change properties of learner's progress bars

fastai
1

I have defined a fastai tabular_learner in a .py script. When I run the script locally, when the script reaches learn.fit_one_cycle(2, wd = 2) the progress bar appears as plain text (not an html widget). If I understand correctly, this corresponds to the ConsoleMasterBar and/or ConsoleProgressBar from fastprogress. This is good and as expected.

When I run the same script in a google colab notebook (running from a cell with contents !python filename.py), I get output <IPython.core.display.HTML object> where I’d expect to see some form of progress bar. I would like to see the progress bar. Any format is okay. Note that this behavior results from running the model from a script. When run directly from a cell, the html progress bar is displayed as expected.

I suspect that forcing the progress bar to follow console behavior in colab will result in a usable progress bar, so that’s been my main angle on this problem so far.

I have tried the following:

  1. placing the following either in the cell from which the script is executed, or within the script just after defining the learner, or both:
from fastprogress.fastprogress import force_console_behavior
master_bar, progress_bar = force_console_behavior()

The results were the same; <IPython.core.display.HTML object>

  1. Following the discussion here, I tried the same but with fastai.basic_train.master_bar, fastai.basic_train.progress_bar. This returned error module 'fastai' has no attribute 'basic_train', and as far as I can tell. the basic_train versions of the bars were only from fastai1.

  2. I made sure ipywidgets was installed, following some suggestions.

  3. I’ve spent a lot of time digging through the fastai documentation and source code for references to progress_bar or fastprogress or master_bar. I’ve gathered that there’s probably something I can do with the progress callback here. And that the mbar and pbar attributes are probably relevant. But I’m hitting a wall trying to figure out how to make those bars follow console behavior, which I suspect/hope will output correctly.

Lastly, in the meantime, I’ve just used:

with learn.no_bar():
    learn.fit_one_cycle(2, wd = 2)

which is preferable to repeated lines of <IPython.core.display.HTML object>.

Any suggestions would be appreciated!