Write sqeeze step -> ensemble step

[kordc]

No description provided.

[kordc]

I'd focus there on the 2 steps he tells us about:

• ensembles
• leave it training -> but this one may be relevant.

So maybe instead of the squeeze, we should enable ensemble DL training?

[kordc]

We agreed on making just ensemble step

[kordc]

Deep learning ensembles' training is a huge topic. I found a nice package Ensembles-PyTorch. They provide a wide selection of already implemented ensemble types. I'll play around with this library using our ArtModule

[SebChw]

Nice, if we won't need to write it by ourselves it would be great.

[kordc]

It supports only classification and regression. Even multi-label classification is not supported.

I think doing this step a task-agnostic is simply not possible. I see three options:

• Do separate steps for classification, regression, multi-label classification etc.
• Do one step, that will create a very simple ensemble basing on voting, especially using torch.mean(). This could be calculated by for at least regression, and both types of classification. For other it's simply not that simple
• Do a template step as we do with DataAnalisys

[trebacz626]

I think weighted mean might be good enough

[trebacz626]

user can specify weights

[mmaecki]

Same as @trebacz626 suggests. Any more sophisticated ensembles can be done by the user.

[SebChw]

Actually, most tasks are classification or regression. If we can utilize this library quickly/easily lets use it, even though it doesn't support everything. If user wants something more sophisticated they must implement it by themselves.