Overhaul 2.1 Remove Dependencies / Add Full Timm Support

isaaccorley commented 5 months ago

This PR seeks to do a few things:

remove the dependency on pretrainedmodels and efficientnet-pytorch as they are no longer maintained.
remove the mock and torchvision dependency (torchvision isn't used anywhere anyway)
Replace any lost model support by fully using timm (torchseg.encoders.TimmEncoder)
add support for timm ViT encoders (torchseg.encoders.TimmViTEncoder`)
Test all supported and unsupported timm encoders (they are defined here torchseg.encoders.supported). This includes ConvNext and Swin pretrained backbones
Update tests to be more thorough
Remove unecessary forward/backward calls in tests (this speeds up the tests significantly)

There's some other misc cleanup that is done:

remove Activation class. We instead let the user choose the head activation when creating a model. Still defaults to nn.Identity()
Allow for passing additional encoder_params to timm.create_model(**kwargs) in case users want to further customize the backbone

JulienMaille commented 4 months ago

Looking forward to test this update! Have you checked if you can use timm convenext_nano encoder with this PR?

isaaccorley commented 4 months ago

Looking forward to test this update!

Have you checked if you can use timm convenext_nano encoder with this PR?

Yes this PR allows the use of convnext as the encoder!

isaaccorley commented 4 months ago

@notprime Can you give this a review and maybe give some thoughts?

ogencoglu commented 4 months ago

Looking forward to this update...

isaaccorley commented 4 months ago

@JulienMaille @ogencoglu @notprime These features are now merged and you can install them using

pip install --pre torchseg or pip install 'torchseg==0.0.1a2'

isaaccorley / torchseg