PROCESSOR-SDK-AM62A: How to generate pt file under mmyolo repo

Hi Ryuuichi,

The final purpose of the steps in https://github.com/TexasInstruments/edgeai-tensorlab/issues/7 are to export an .onnx file to be ported to the TIDL (more step by step details are included in: https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1457711/faq-am62a7-q1-deploy-mmyolo-on-am62a-yolov8-nano-as-example).

Currently, the .pth file is serving the purpose above and for that reason there is no intent to use .pt instead. 

May I ask why the .pt is preferred in your use case?

Best regards,

Qutaiba

Hello,

Thank you for your reply.

The final purpose of the steps in https://github.com/TexasInstruments/edgeai-tensorlab/issues/7 are to export an .onnx file to be ported to the TIDL (more step by step details are included in: https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1457711/faq-am62a7-q1-deploy-mmyolo-on-am62a-yolov8-nano-as-example).
=> Yes, I understood about above.

>May I ask why the .pt is preferred in your use case?
We would like to make benchmark by using other platform with same model.
As you said, TI platform need pth, however other platform which we will use require pt file. So I asked this.

Best Regards,

Hi Ryuuichi,

Thank you for providing the details. My understanding is that the .pth and .pt can be interchangeably used in modern pytorch version.

At this time, I can only provide an experimental suggestions:

• Experiment with manually changing the extension for the generated file.
• If above did not work, I would suggest to look up the code where https://github.com/open-mmlab/mmyolo use to save the model and add another line like torch.save(model, path/name.pt). 

Best regards,

Qutaiba

Hello,

I'm trying to perform your idea 2.

• If above did not work, I would suggest to look up the code where https://github.com/open-mmlab/mmyolo use to save the model and add another line like torch.save(model, path/name.pt). 

At first, I'm checking which code will perform "torch.save".
I understood that following file perform save file (process_checkpoint(in_file, out_file)).
mmyolo/.dev_scripts/gather_models.py at main · open-mmlab/mmyolo · GitHub

However, it seems that "out_file" come from "training result".
I'm not sure which file perform training itself. It seems that "pth" character is given under training session.
So, I think that I need to change character under training session.

But I could not find which file perform this. Do you know which file perform training ?

Best Regards,

Hi,

I am not entirely sure. I would suggest to start tracking with tools/train.py but I don't have specific idea. I will post back here if I figured a way to save .pt file. 

Best regards,

Qutaiba