TDA4VM: Porting Unsupported OD model (yolov8) to TDA4VM

Hi Ramaseshan,

I do apologize for the delay in getting to your question. Is this a blocking issue for you currently?

If so, can you attach the compilation logs as a text file so I can see the errors more clearly? 

Best,

Asha

Hello Asha Bhandarkar,

I am facing some issues with the compilation process.

As mentioned earlier, I have modified the YOLOv8 architecture to use the "ReLU" activation function instead of "Sigmoid". Additionally, I have configured the max-pool layers to have a kernel shape of 3.

Please find the attached logs for the errors I encountered in compilation. As you could see accuracy is 0 while making inference using compiled artifacts.

Kindly review the logs and assist with the errors.

run_import(compilation).txt

run_inference(after compilation).txt

Hi Ramaseshan,

I'm seeing errors in your compilation log so I do not expect you to have a successful inference.

Looking at the first error:

Error: Layer 123, /model.22/Slice_1:/model.22/Slice_1_output_0 is missing inputs in the network and cannot be topologically sorted
Input 0: /model.22/dfl/Reshape_1_output_0, dataId=308

This is something we will need to debug further. If you are running into many compilation issues, I would recommend using edgeai-tidl-tools with the python OSRT approach for importing which will allow you to better debug the model as you adjust to to be supported in the TIDL framework. Further details are here. I believe this will also give you more flexibility in terms of the compilation options - such as delegating user specified layers to the ARM core. 

If it is possible to share the model and model artifacts, that will be helpful so we can look at it further on our end. 

Best,

Asha

Hi Asha,

Thanks for your suggestions. I will try compiling the model using TIDL tools and get back to you. Meanwhile, could you please review the model artifacts generated using benchmark, which I have attached here?

yolov8_model_artifacts.zip

Best regards,
Ramaseshan

Hi Ramaseshan, 

Thank you for providing the model artifacts. Taking a glance, the output svg and ONNX model don't seem to line up (the SVG is showing multiple outputs). I would be interested in debugging further with the edgeai_tidl_tools path and seeing if we gain more information that way. I would recommend attaching the same output files as earlier and also providing the compilation with the debug_level = 2 (for more detailed compilation logs)

Best,

Asha

Hi Asha,

I tried compiling the model using TIDL tools, but the calibration stopped after the 4th iteration with an error message. For your information, I set the debug level to 2. I have attached the TIDL tools compilation logs. Please check them.

tidl-compilation-logs.txt

Best,  
Ramaseshan

Hi Ramaseshan,

Asha is on vacation for the rest of this week, and will get back to you sometime early-mid next week.

regards

Suman

Hi Suman,

Thanks for letting me know. I'll follow up with Asha next week.

Best regards,  
Ramaseshan

Hi Asha,

Hope you are back, kindly look into my previous message in the same forum in which i have shared the error faced in TIDL tools.

Best regards,
Ramaseshan