TDA4VM: Frame rate test issue

Cherry Zhou

Part Number: TDA4VM
Other Parts Discussed in Thread: MATHLIB

Hi Team,

1) The ONNX model of shufflenet_v2 reported an error "The ONNX operator ReduceMean is not suported now" when converting bin file.

2) The ONNX model of mobilenet_v1 reported the following error when converting bin file:

3) The ONNX model of Yolop reported "ONNX operator Div is not suported now ；ONNX operator ScatterND is not suported now" when converting bin file.

4) The ONNX model of deeplabV3_resnet50 reported the following error when converting bin file:

5) Does 8-bit fixed-point and 16-bit fixed-point have a significant impact on model frame rate testing when the model is converted?

6) Is the maths supported in the DSP?

Could you help check this case? Thanks.

Best Regards,

Cherry

over 1 year ago

0 Doredla Sudheer Kumar over 1 year ago

TI__Guru 64870 points

Hi,

Due to a holiday, Our expert is on vacation kindly expect a 1~2 day delay in responses.

Best Regards,
Sudheer

0 Pratik Kedar over 1 year ago in reply to Doredla Sudheer Kumar

TI__Mastermind 24041 points

Cherry Zhou said:
1) The ONNX model of shufflenet_v2 reported an error "The ONNX operator ReduceMean is not suported now" when converting bin file.

Could you try out the model compilation on latest SDK, per my knowledge this operator is internally implemented as global mean.

Cherry Zhou said:
2) The ONNX model of mobilenet_v1 reported the following error when converting bin file:

We support the mobileNetV1 as part of TIDL, you can refer our model zoo repo example here : https://github.com/TexasInstruments/edgeai-modelzoo/tree/r9.0/modelartifacts/TDA4VM/8bits

Cherry Zhou said:
The ONNX model of Yolop reported "ONNX operator Div is not suported now ；ONNX operator ScatterND is not suported now" when converting bin file.

ScatterND operator should be supported as part of our latest release, please try the same on latest sdk.

Cherry Zhou said:
4) The ONNX model of deeplabV3_resnet50 reported the following error when converting bin file:

DeepLabV3 is supported as part of tidl as well, we recommend to refer out model zoo for the same. https://github.com/TexasInstruments/edgeai-modelzoo/tree/r9.0/modelartifacts/TDA4VM/8bits

Cherry Zhou said:
5) Does 8-bit fixed-point and 16-bit fixed-point have a significant impact on model frame rate testing when the model is converted?

Lets understand how to quantization specific to speed vs accuracy tradeoffs.

In layman terms, higher the quantization bits higher the accuracy and lower the performance wrt lower quantization bits (Speed of processing) on the other hand lower the quantization bits lower will be accuracy (wrt higher quantized models) but better the speed/performance.

Cherry Zhou said:
6) Is the maths supported in the DSP?

Not sure what is the ask here, please try to elaborate the question/contest in details.

On side not check out this MathLIb repo https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/09_00_01_01/exports/docs/psdk_rtos/docs/user_guide/sdk_components_j721e.html#c7x-mathlib

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/09_00_01_01/exports/docs/psdk_rtos/docs/user_guide/sdk_components_j721e.html#c7x-mathlib

0 Cherry Zhou over 1 year ago in reply to Pratik Kedar

TI__Mastermind 22235 points

Hi Pratick,

Thank you for the support.

After upgrading sdk09_00_01_01, the following error is reported when executing the command make Linux_FS_install_SD in the log Copy executable files to SD card:

It was found that there was no out file in the new downloaded SDK (/ti-processor-SDK-RTOS-j721e-EVM-09_00_01_01/c7x-MMA-tidl /arm-tidl/tfl _delegate directory). Why's this and how to get the related contents?

Thanks and regards,

Cherry

0 Pratik Kedar over 1 year ago in reply to Cherry Zhou

TI__Mastermind 24041 points

Did you built the SDK before running above experiments ?

0 Cherry Zhou over 1 year ago in reply to Pratik Kedar

TI__Mastermind 22235 points

Hi Pratik,

Pratik Kedar said:
Did you built the SDK before running above experiments ?

Yes, it was built and the model(bin.) has been converted successfully. The customer is wondering that this file exists in the previous 8.0 SDK release after downloading and extracting, but in the latest version, was it generated when building the SDK?

Also, manually upgraded glibc using the SDK built with Ubuntu18.04. Is it possible to build with Ubuntu18.04?

Thanks and regards,

Cherry

0 Pratik Kedar over 1 year ago in reply to Cherry Zhou

TI__Mastermind 24041 points

Hi Cherry,

Greeting of the day !

Recommending you to please Read the documentations.

The Files that you are enquiring about will be generated as part of SDK BUILD.

Could you please follow the documentation on how to build sdk here : https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/09_00_01_01/exports/docs/vision_apps/docs/user_guide/BUILD_INSTRUCTIONS.html

Cherry Zhou said:
Also, manually upgraded glibc using the SDK built with Ubuntu18.04. Is it possible to build with Ubuntu18.04?

What is the need here ?? and may I know why you want to build the sdk on 18.04 OS despite of documented as 22.04 ?

What is thought process behind changing GLIBS/GNUS ??

Processors

Processors forum

TDA4VM: Frame rate test issue