search

LMZimmer / Auto-PyTorch_refactor

Apache License 2.0
0 stars 2 forks source link

New networks #51

Open bastiscode opened 3 years ago

bastiscode commented 3 years ago

PR includes ShapedResnet backbone for tabular data and Resnet backbone for image data. More coming soon.

LMZimmer commented 3 years ago

I have a general question about the backbone part. Do you think it is possible to share code between image.py, tabular.py and time_series.py - for instance, putting conv/dense layers into a layers.py, or even have a conv net that can be used by all of them?

bastiscode commented 3 years ago

I also thought about that. I think for stuff like attention layers which could be possibly added later a layers.py makes sense. If one wants to include more NAS-like capabilities, I think you could implement searchable dense and conv layers or other module blocks that can be used in backbones across tasks. Then you could also add more general backbones like NASImageBackbone for example.

But as of now I think all the implemented networks have a rather specific architecture, and sharing code between them is not something I would do.

LMZimmer commented 3 years ago

Okay, then I will add that to the todo/future work 👍