SK-AM62A-LP: [edgeai-modelmaker] which shell/python script produces outputs folder?

Thanks Qutaiba,

I uploaded the same image into SK-AM62A-LP and do inference, I found out that their performances are different. Inference results of the compiled model that is simulated on the host machine are significantly better than inference results produced by SK-AM62A-LP.

My question is:

From my understanding, the performance should be similar if not the same as both scenarios show the result of the inference on the SAME generated artifacts. The only difference should be on the inference speed. 
But now that on the host machine (rtx3080 in my case) is better than that on SK-AM62A-LP. Why?

Is the inference performance related to the hardware limitation?

Hi YL,

Yes, the results of the inference on the simulation on PC should be similar to the inference on target. I suggest checking the settings for the post process for the mode:

• Is the detection_threshold in the model config the same on pc and on target?
• Check the label_offset in the param.yaml anc config.yaml.

Best regards,

Qutaiba

Hi Quitaiba, 

Thank you for your quick reply.

I had compared the detection threshold of both platforms (host pc and SK-AM62A-LP). Both threshold is 0.1. Below is the evidence.

Above shows the config.yaml in host pc. The detection_threshold under postprocess is 0.1.

Above shows the object_detection.yaml. viz_threshold is 0.1.


Is there something I missed out? I still don't understand why the inference results in host pc and SK-AM62A-LP are different despite the extremely similar (if not same) settings and configs.

Hi YL,

I suggest following the debug steps presented in this document which targets exactly what you are observing: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_osr_debug.md#steps-to-debug-error-scenarios-for-targetevmdevice-execution

This document is part of the tidl-tools which is more comprehensive than model maker when in comes to model compilation and debugging. 

Best regards,

Qutaiba

Hi Quitaiba,

So you mean when it comes to do model compilation alone, you suggest using edgeai-tidl-tools rather than edgeai-benchmark?

Also, is it possible to compile my model using edgeai-benchmark alone without going through the edgeai-modelmaker pipeline? I ask because edgeai-modelmaker includes model training, which can be quite time-consuming.

Regards,
Matt