Running bulk predictions locally using 100% CPU

[ishiland]

Is there a working example or suggested approach to running predictions locally using 100% CPU? I am only able to achieve 40-50% cpu usage when running predictions across several images. My script is looping over an array of input image paths and executing the prediction one at a time like this:
predictor.predict([in_image_path], output_predict_path)

The pattern I am using is similar as in this issue.

[AdeelH]

Just to be clear, you are not using a GPU at all, right? Also, I'm assuming these are relatively large GeoTIFFs and that we are talking about the utilization during prediction on a single such image.

I don't know if CPU utilization scales the same way as GPU utilization, but you might try increasing the batch size.

[ishiland]

Just to be clear, you are not using a GPU at all, right? Also, I'm assuming these are relatively large GeoTIFFs and that we are talking about the utilization during prediction on a single such image.

Yes this is correct. I have tried multiprocessing my for-loop but run into pickling issues. I will attempt to increase the batch size, thanks.

[AdeelH]

I wonder if running two instances of the same script in parallel might work. You might also want to try increasing num_workers.

Also, see also these tutorials for making predictions in a way that is more flexible than using the Predictor class: one, two.