SK-AM62A-LP: EDGE AI Arc Fault Model

Hi Sopheap,

As you mentioned, the Arc Fault Detection model is available on the C2000 Devices. Theoretically, C7x on AM62A is more than capable of implementing this application. AM62A AI solutions currently focuses on vision/audio applications. That said, we would like to understand learn more about your use case. Would you please, share more details about your use case? You can also communicate with your local TI representative. 

Best regards,

Qutaiba

Hello Qutaiba,

I am aware of its capability to implement arc fault detection model on c7x. Currently, I do not have C2000 devices available in my places, only c7x on am62A7. When I was looking on EDGE AI model composer online, I followed the tutorial and guidances provided by TI DEV. When it comes to select device for training in EDGE AI model composer, I could not find any related model (AM6xA). My question is that even if I train the model on EDGE AI using devices (F28P55x, or F28P65x), is it possible to convert it to be used on AM6xA after exported the model on CCS using above link code?

Hi Sopheap,

Sopheap Key said:

My question is that even if I train the model on EDGE AI using devices (F28P55x, or F28P65x), is it possible to convert it to be used on AM6xA after exported the model on CCS using above link code?

This is currently not possible. The output artifacts for the F28Pxxx are not compatible with the SW for AM62A and can not be converted. 

That said, we would like to hear more from you about your use case to try to find alternative solutions and also consider in future planning for our tools.

Best regards,

Qutaiba

Hello Qutaiba,

Sorry, for the late reply. I was busy on another task and did not check the response here.

It would be nice if there is an example regarding image classification or object detection that I can follow or implement in CCS. I found an arc fault model project that can be used and implemented using CCS.

Regards,

Key

Hi Sopheap,

No worries, there is no Edge AI example on AM62A using CCS. It might be possible, but the vision Edge AI applications are a bit complicated to run/mange with CCS. We provide several examples which you can run directly on AM62A Linux SDK. See the documentation for the Edge AI SDK: https://software-dl.ti.com/processor-sdk-linux/esd/AM62AX/10_01_00/exports/edgeai-docs/common/sample_apps.html. This link provides three methods to run demos such as gStreamer and TIOVX. The rest of the documentation provides more details bout configuring the demos and importing models. 

In order for a model to work on AM62A and use the AI accelerator (C7x-MMA), it has to be compiled using the TIDL tool: https://github.com/TexasInstruments/edgeai-tidl-tools. The documentation for this repo provides more details. 

If the above do not answer your question, please provide more details about the model you found so that I can provide more informed suggestions. 

Best regards,

Qutaiba

Hello Qutaiba,

Is it not neccesary to train model on EDGE AI? Arc fault or motor fault detection is not available for AM62Ax devices on Edge AI Studio. From what I understand, I can just train model on python using tensorflow/keras and export model as tflite file. Then, it is needed to be compiled using TIDL tool in above link using tidl_model_import for C7x-MMA. Am I correct?

Regards,
Key

Hi Qutaiba,

Is it necessary to compile on linux? 

Regards,

Key

Hello Qutaiba,

Currently, I am using Ubuntu 18.04 LTS on WSL. I follow the tutorial provided in https://github.com/TexasInstruments/edgeai-tidl-tools. I successfully git clone edge-ai TIDL tool into my linux environment. Could you please let me know on git checkout? Is it to fetch edgeai TIDL tool version?

Regards,
Key

Hello Qutaiba,

I was looking at the installation guide edgeai sdk linux from https://software-dl.ti.com/processor-sdk-linux/esd/AM62AX/09_02_00/exports/docs/devices/AM62AX/linux/Overview/Download_and_Install_the_SDK.html. It mentioned that it should be installed on Ubuntu 18.04 as it was tested and verified on that version. Another thing, how to know which version of SDK match TIDL tool version? Currently, I installed version 10_00_00_08 on linux system. 

As I am working on signal processing, any kind of signal processing problem is my interest and arc fault detection is also one of them. 

Regards, 

Key

Hello Qutaiba,

I successfully installed and compiled the example using TIDL tool in above link. I have a question regarding the script to convert tflite model to obtain artifacts in this link https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/tfl/tflrt_delegate.py. I trained a simple sequential model on tensorflow and use tflite model converted by tensorflow lite converter. I copied the file to model folder on edgeai_tidl_tools as shown below.

Then, I compiled the tflite model using python3 tflrt_delegate.py to obtain the artifact file. However, there is no artifact file of my trained model in artifact folder. Could you please enlighten me on this point?

Regards,

Key

Hello Qutaiba,

Thank you for guiding me through. I already provided a model config in above file and also run the tflite_delegate file.
I will try to look into it again and create a new thread.

Regards,
Key