Vendor efficientnet-pytorch

[adamjstewart]

Supersedes #1018

EfficientNet-PyTorch doesn't have wheels on PyPI, and is no longer maintained. Let's vendor it and maintain it ourselves. This will make SMP easier to install.

This PR doesn't deprecate the non-timm EfficientNet implementation, but we can also do that if we want to. Let's wait until we have download statistics.

Will do the same thing with pretrainedmodels once this is merged.

[codecov]

Codecov Report

Attention: Patch coverage is 71.47766% with 83 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...mentation_models_pytorch/encoders/_efficientnet.py	71.37%	83 Missing ⚠️

Files with missing lines	Coverage Δ
...gmentation_models_pytorch/encoders/efficientnet.py	94.33% <100.00%> (-0.11%)	⬇️
...mentation_models_pytorch/encoders/_efficientnet.py	71.37% <71.37%> (ø)

[qubvel]

Thanks for moving it and removing dependency! What do you think re removing unnecessary code? I mean not all functions/methods are required, such as from_pretrained

[qubvel]

Thanks!

[qubvel]

Thanks, a few things to remove and we can merge it

[qubvel]

Thanks!

[adamjstewart]

pretrainedmodels is up next. This one is a bit more complicated because we have multiple models. Should we have a segmentation_models_pytorch/encoders/_pretrainedmodels/ directory containing each model, or multiple segmentation_models_pytorch/encoders/_foo.py files for each model?

[qubvel]

I prefer segmentation_models_pytorch/encoders/_foo.py, but feel free to choose the one you think is better.