Improve model implementation + weight loading from inswapper_128.onnx + inference code

[blend-er]

StyleTransferModel_128.py Changes

• Rewrote AdaIN using InstanceNorm2d
• Simplified decoder path

weight_transfer.ipynb

• Maps weights directly from inswapper_128.onnx to StyleTransferModel
  todo: figure out why swap.py produces samples that slightly differ from inswapper.

StyleTransferModel:

Inswapper:

[mikecokina]

@blend-er , have you tried your version of the architecture with the checkpoints provided in this project? For example, using the reswapper-1019500.pth checkpoint? I converted the weights of INSwapper using the code you shared in the Jupyter notebook and successfully executed it for INSwapper. Unfortunately, all results with the RESwapper checkpoints turn out as completely gray images.

[blend-er]

Hi @mikecokina, I have added conversion from old reswapper checkpoints in the notebook.

[mikecokina]

@blend-er , can confirm it works perfectly now. Thank you for your quick response!