yolov5 port

Jérôme Thomas

Part Number: TDA4VM

Hi,

I'm exploring the port of yolov5 object detection network on TDA4.

My starting point is (https://github.com/TexasInstruments/edgeai-yolov5). I trained this network on a pedestrian dataset.

My first question is there a preferred runtime between onnx, tflite and TVM to move it to the TDA4?

Nevertheless, I tried the onnx runtime and then moved it to https://dev.ti.com/edgeai/ . I then modified the Custom model onnx runtime jupiter notebook in order to try it. I get the following results:

Detection are ok but one can see that half of the inference time was spent on the cpu. I guess there are some layers that are not supported by the mma? if so, is there a way to determine which ones in order to replace them with supported layers if possible? Would it be different if I use the tflite runtime?

Thanks in advance,

Cheers,

Jerome

over 3 years ago

0 Anand Pathak over 3 years ago

TI__Genius 9065 points

Hi,

You can continue using ONNX runtime to port this model to TDA4.

For compilation of this object detection model, you need to specify some additional compilation options to get better performance, specifically "object_detection:meta_layers_names_list" and "object_detection:meta_arch_type".

You can refer to the below scripts for example code to specify these options as part of the "compile_options" in the custom notebook: https://github.com/TexasInstruments/edgeai-tidl-tools/tree/master/examples/osrt_python/ort

The corresponding prototxt file required for above options should be present at the following link: https://github.com/TexasInstruments/edgeai-yolov5/tree/master/pretrained_models/modelartifacts/8bits. Please refer to the folder here corresponding to the model you are using.

Let me know if you have any additional questions or face any issues.

Regards,

Anand

Processors

Processors forum

yolov5 port