The ocr_predictor utilizes only 1 CPU core.

[sergei-voronkov]

I have a set of images from real documents.
While I try to recognize text with default ocr_predictor setting + torch, it takes about 10 seconds per page.
I notice the CPU load it too low in the same time. It seems the model uses only 1 CPU core for the recognition.
Is there a way to utilize all my core to speed up the recognition.

My configuration:
Windows 10
Python 3.12.4
Pytorch
No GPU

Thanks!

[felixdittrich92]

Hi @sergei-voronkov 👋,

Looks like an issue with the PyTorch installation on your machine, this does not come from docTR itself.
Could you try to force it with torch.set_num_threads(num_cores) ?

Otherwsie you can also try: https://github.com/felixdittrich92/OnnxTR which is optimized for inference, especially on CPU / and or limited hardware.

Best regards,
Felix

[sergei-voronkov]

The conda-base torch fixes the problem! Thanks a lot!

[felixdittrich92]

Nice 😊

[sergei-voronkov]

Could you explain the reason? Wrong torch version?

[felixdittrich92]

I think anything was broken by building the wheel with pip on your windows machine, because conda comes with precompiled binaries where no wheel is required. :)

[sergei-voronkov]

Thanks a lot!