Summary

This repository contains attempts to adapt the PyTorch autotagger model in TensorflowJS + Node environment. You can see more details in the paper on Habr.com.

Converting a weight file from PyTorch to Tensorflow

• Exporting *.pth file to ONNX (torch.onnx.export(learn.eval().to('cuda'), torch.randn(1, 3, 224, 224).to('cuda'), "danbooru.onnx"))
• Using onnx-tf: $ onnx-tf convert -i danbooru.onnx -o danbooru
• Next stage is converting *.pb to a package for browsers, but I failed it. :( BTW *.pb model is good for my Electron pet project so I don't care about that.

Requirements

• Node 16
• Python 3.9
• fastbook 0.0.26
• scipy 1.8.1