SK-TDA4VM: How to run the Edgei-yolov5 trained network on TDA4

Hi,

May I know is there any updates?

Thanks and Best Regards,

Cherry

Hello Cherry,

Apologies for the delayed response.

If you are following this github https://github.com/TexasInstruments/edgeai-yolov5, then the prototxt files can be found in the compressed files here: https://github.com/TexasInstruments/edgeai-yolov5/tree/master/pretrained_models/modelartifacts/8bits

Regards,

Takuma

Hi Takuma,

Thanks for your response.

The customer has checked the file you provided and did not get the answer, customer would like to know how this prototxt file was generated and how the numbers were calculated.

Thanks and Best Regards,

Cherry

Hi Takuma,

May I know is there any updates about the question above? Thanks.

Best,

Cherry

Hi Cherry,

I have forwarded this question to the developer of edgeai-yolov5

He should be replying soon.

Regards,

Manu

While we are waiting for Debapriya Maji to respond, you may look here for example prototxt files for yolov5:

https://github.com/TexasInstruments/edgeai-yolov5/tree/master/pretrained_models/models

Specifically, is this the model that you are looking for?

https://github.com/TexasInstruments/edgeai-yolov5/tree/master/pretrained_models/models/yolov5s6_640_ti_lite/weights

Note: the .link files are text files. If you open it, you can get the url for the actual file and you can download it using wget.

Regards,

Hi Cherry,

 Will you please share the onnx model that you have exported and the error you are getting while using the shared prototxt for import.

Regards, Debapriya

Hi dear all,

Thanks for your support and I have updated to the customer, if there is any new feedback I will reach you ASAP.

Thanks again!

Best Regards,

Cherry

Hi,

The updates are as follows:

When using edgeai-tidl-tools to convert the semantic split network onnx file into bin, run the command python onnxrt_ep.py -c. The generated .bin file performs very poorly on TDA4. Warning during conversion: Warning: [TIDL_E_DATAFLOW_INFO_NULL] ti_cnnperfsim.out fails to allocate memory in MSMC. Please look into perfsim log. This model can only be used on PC emulation, it will get fault on target.

There is also a something that looks abnormal: Substitute string tidl_net_ not Found

Could you help check what's happening?

Thanks and Best Regards,

Cherry

Hi Cherry,

It will be great if the ONNX model and the corresponding prototxt can be shared.

Regards, Debapriya

Hi Debapriya,

http://software-dl.ti.com/jacinto7/esd/modelzoo/gplv3/08_00_00_05/edgeai-yolov5/pretrained_models/models/yolov5s6_384_ti_lite/weights/yolov5s6_384_ti_lite_32p8_51p2.onnx

http://software-dl.ti.com/jacinto7/esd/modelzoo/gplv3/08_00_00_05/edgeai-yolov5/pretrained_models/models/yolov5s6_384_ti_lite/weights/yolov5s6_384_ti_lite_metaarch.prototxt

Please check if these works.

Thanks and Best Regards,

Cherry

Hi Debapriya,

May I know is there any feedback regarding the info above?

Thanks and Best Regards,

Cherry

Hi Cherry,

I tried this model and it is working fine. These models and prototxt were shared by TI and was validated. Now, based on the pytorch and ONNX version, the prototxt may have to be changed.

Hence, if you can share the exact model and prototxt that will help.. 

Regards, Debapriya

HI Cherry,

 The first error comes during calibration when the model run an epoch of floating point inference. That should be fine. It will not come after the calibration error.

  The second warning can also be ignored. 

  I think the issue is in defining the correct prototxt. If the prototxt is not defined correctly, some layers will go to ARM and result in high latency.

Regards, Debapriya

Hi Debapriya,

Thanks for your kind support again!

The customer would like to know how to ensure the date of prototxt, is there any related documents?

Thanks and Best Regards,

Cherry