Question regarding PyTorch / .safetensors weights for fine-tuning

#2
by johnlockejrr - opened

Hi team, thanks for releasing Indra OCR—the CPU throughput and performance look fantastic!

Right now, the repository only includes .onnx weights, which makes downstream fine-tuning difficult. Do you have any plans to publish the native PyTorch checkpoints or *.safetensors weights for the models so they can be fine-tuned? Would be so cool to have the OCR model, for example, to work on other languages (Hebrew for example, but not only).

Thanks!

Disregard, I converted the weights to pytorch. A friend already finetuned the OCR model.

Very good model! In training same fast as lightning.

image

image

For personal use.

And the detector.

image

@johnlockejrr Could you share the model and the training code?

@johnlockejrr Could you share the model and the training code?

Check my github

And the detector.

image

@johnlockejrr Could you share the model and the training code?

I can't. Datalab should do that if they intend. Sorry.

Sign up or log in to comment