Unable to load weights for pytorch checkpoint file

[ElnuDev]

My main OS is NixOS, but I was having some issues getting the Python dependencies packaged, so I wanted to give mokuro a shot on Windows 10 first.

I have Python 3.9.13 installed (I checked and apparently PyTorch supports 3.10 now despite what the README says; I had the same issue in 3.10), and I have gone through the setup instructions. I have an Nvidia RTX 2060 SUPER, so I installed PyTorch with CUDA 11.7 per the instructions on PyTorch's website with pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117.

When I entered the directory with my scan image files and ran mokuro ., I got this error:

OSError: Unable to load weights from pytorch checkpoint file for 'C:\Users\elnu/.cache\huggingface\hub\models--kha-white--manga-ocr-base\snapshots\aa6573bd10b0d446cbf622e29c3e084914df9741\pytorch_model.bin' at 'C:\Users\elnu/.cache\huggingface\hub\models--kha-white--manga-ocr-base\snapshots\aa6573bd10b0d446cbf622e29c3e084914df9741\pytorch_model.bin'. If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True.
2023-04-18 13:13:24.868 | INFO     | mokuro.run:run:51 - Processed successfully: 0/1

Full output here.

Any ideas what's gone wrong? Have I missed something? Thanks in advance.

Update: Just wanted to add that I was reading through Xelieu's guide and they mentioned that there sometimes is a problem where torch.cuda.is_available() may return False. I checked and for me this was the case, and even after reinstalling torch and verifying that it started detecting CUDA as available, I still ran into the same issue.

[kha-white]

Seems like something wrong with the model cache, you might try to delete the cache C:\Users\elnu/.cache\huggingface\hub\models--kha-white--manga-ocr-base and try again