Change regression task to timm support

[nilsleh]

This PR changes the RegressionTask to support Timm models instead of torchvision.models for greater model variety and consistency, because Timm is used in the other tasks as well.

• new keyword argument num_outputs which allows the regression task to have multiple regression outputs, either if there are multiple regression targets or some form of probabilistic neural networks
• other keyword arguments will be the same as ClassificationTask after Docs/trainers for Classification and Segmentation Task #852 is merged with the exception of the loss argument because regression only has mse loss at the moment

[calebrob6]

Pulling out discussion from the comments, I think it makes most sense to call the deep learning model, "model", in the trainer code. In semantic segmentation tasks this will be a "unet", "deeplabv3+", or whatever else smp accepts, as that is the "model" that is being used, elsewhere this will be any timm architecture. In the segmentation trainer, it makes sense to call the encoder, "encoder".

I.e. what Adam said above but "encoder" instead of "backbone" (it makes sense to use "backbone" for BYOL and other SSL trainers).

[nilsleh]

Pulling out discussion from the comments, I think it makes most sense to call the deep learning model, "model", in the trainer code. In semantic segmentation tasks this will be a "unet", "deeplabv3+", or whatever else smp accepts, as that is the "model" that is being used, elsewhere this will be any timm architecture. In the segmentation trainer, it makes sense to call the encoder, "encoder".

I.e. what Adam said above but "encoder" instead of "backbone" (it makes sense to use "backbone" for BYOL and other SSL trainers).

Let me know if this is what you had in mind and I understood you correctly. I will then open another PR to change the segmentation task names.

[calebrob6]

Yep, exactly that!

[adamjstewart]

One minor change and I think this is ready to go.

[adamjstewart]

Would love to get one more review from someone before merging

[nilsleh]

I.e. what Adam said above but "encoder" instead of "backbone" (it makes sense to use "backbone" for BYOL and other SSL trainers).

@adamjstewart looking back through the comments, I interpret the following:

• Segmentation task: segmentation_model -> model and encoder_name -> encoder
• Regression task: regression_model -> model
• Classifcation task: classification_model -> model
• BYOL task: encoder_name -> encoder (although Caleb seems to suggest encoder_name -> backbone?)
• Object Detection: detection_model -> model and backbone -> ?

[adamjstewart]

Hmm, I would personally use "model" for the high-level model in all tasks and "encoder" iff that model has the ability to replace some kind of encoder/backbone with a pre-trained one. I guess I don't see a big distinction between SL encoders and SSL backbones, but I'll let @calebrob6 clarify.

[calebrob6]

Encoder or backbone is fine for BYOL

[nilsleh]

Okay so,

• Segmentation task: segmentation_model -> model and encoder_name -> encoder
• Regression task: regression_model -> model
• Classifcation task: classification_model -> model
• BYOL task: encoder_name -> encoder
• Object Detection: detection_model -> model

is the final verdict?

[calebrob6]

LGTM!