[ENH] Add tqdm Progress Bar for model.predict()

[jgyasu]

Closes #818

This PR adds a tqdm progress bar to visualise the progress during the running of model.predict() method. The progress bar is put around the loop in _compute_likelihood_params() method where distributional_components are getting processed.

• Passes black
• Passes pylint

Thanks :)

[jgyasu]

Should I list tqdm in pyproject.toml?

[tomicapretto]

Hi again @jgyasu and thanks for all the contributions!

I'm not sure it's easy to resolve #818. While what you're doing effectively adds a progress bar, it adds it at the distributional component level (i.e. parameters of a likelihood for which we have at least one predictor). That is too high level. The user will see one or two "steps" and that's it.

The real difficulty is in what happens in the .predict() method (see here

bambi/bambi/model_components.py

Line 128 in 6b66691

def predict(

). There are several steps that can take time (predict_common, predict_group_specific, etc.)

Another pain point, and I think it's the one that usually takes the longest, is what happens when users do model.predict(... kind="response"). After computing the draws for the response parameters (e.g. mu and sigma in gaussian models), it uses pymc.draw() which usually takes some time, especially when we pass large arrays (that happens here

bambi/bambi/families/family.py

Line 161 in 6b66691

output_array = pm.draw(response_dist.dist(**kwargs))

)

It's not that I want to ruin the party, but I think this one is not possible to be resolved until we change how we get predictions, which would be through the proper usage of pymc.sample_posterior_predictive(). But for that to happen, we first need to modify other internal things.