remove empty windows during predict

[jmoralez]

No description provided.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[elephaint]

#1059 is a generalization of this PR, right? I.e. the default behaviour of #1059 could be to remove empty windows during all steps of the pipeline - train/val/predict; and have the hyperparameter of that PR serve as a setting for the amount of bad quality windows to consider. For example, we could have the data_availability_threshold as a Union[None, float], with the following behavour:

• None means all windows are considered - empty and 'sparse';
• A float value between 0 and 1 indicates the threshold for the amount of zeros in the windows to drop that window.

Wdyt?

[jmoralez]

Seems like it, but that's more focused on the train stage. We can use that PR but having the threshold as an argument to the create windows method or similar so that we can always provide 0 during the predict step, because predicting windows full of zeros is just wasting compute.

The ideal scenario in my opinion would be not even creating those windows at all, but that requires a refactor on the dataset which impacts all models and thus may be harder.