TDA4VM: MNIST model - Wrong inference with OpenVX

Padmasree N

Part Number: TDA4VM

Hello Team,

In reference to the related question (e2e.ti.com/.../tda4vm-mnist-model-import-error-tidl), I was able to succesfully import and infer the mnist model after upgrading my SDK version to 8.

Now, when I try to infer using TIOVX framework, I observe the inference results are not matching. The class labels are same for all numbers.

Please find the inference results in TIDL.

Kindly check if its right and advice me where I am going wrong.

Also, Please do connect to me regarding the TI representative support for my organisation

Thanks and Regards,

Padmasree N.

over 2 years ago

0 kumar.desappan over 2 years ago

TI__Mastermind 22145 points

Can you please share the application that you are using for "TIOVX framework based test".

You can also try TFlite/Onnxrt for running this model by referring to below

https://github.com/TexasInstruments/edgeai-tidl-tools

Ntoe : Please checkout to right SDK version this repo, based the SDK that you are using in EVM

0 Padmasree N over 2 years ago in reply to kumar.desappan

Intellectual 470 points

Hello Kumar,

Thanks for your reply!

I used TIDL import and infer tool to validate my network. I followed the below links for the same.

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_02_00_05/exports/docs/tidl_j721e_08_02_00_11/ti_dl/docs/user_guide_html/md_tidl_model_import.html

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_02_00_05/exports/docs/tidl_j721e_08_02_00_11/ti_dl/docs/user_guide_html/md_tidl_sample_test.html

I have few questions regarding TIDL-OSRT.

1) TIDL- RT has OpenVX framework which takes care of hardware acceleration. How does the same happen with TIDL-OSRT ?

2) Do you recommend TIDL-OSRT as a final complete embedded product to be used in production ?

Kindly resolve my queries.

Regards,

Padmasree N.

0 kumar.desappan over 2 years ago in reply to Padmasree N

TI__Mastermind 22145 points

TIDL-OSRT. internally uses TIDL- RT for all the supported layers acceleration and uses open standard APIs for ease of use. Also supports running un supported layers in ARM.

Refer table in the below page for more details

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_05_00_11/exports/docs/tidl_j721e_08_05_00_16/ti_dl/docs/user_guide_html/index.html

0 Padmasree N over 2 years ago in reply to kumar.desappan

Intellectual 470 points

Hello Kumar,

Thanks for your reply! I have referred the docs you shared. But still I am not clear about my query.

Please do answer my (2) query above

Regards,

Padmasree N.

+1 kumar.desappan over 2 years ago in reply to Padmasree N

TI__Mastermind 22145 points

Yes, We recommend users to use OSRT for a better user experience and a richer coverage of neural network.

If your application safety requirement can NOT be met with this, then we would recommend to use TIDL-OSRT python interface foe model compilation and validation. Then use the compiled artifacts with TIDL-RT for final deployment on device.

0 Padmasree N over 2 years ago in reply to kumar.desappan

Intellectual 470 points

Hello Kumar,

Thanks for your reply!

For model compilation and validation, I have used TIDL import and infer tool to validate my network. I followed the below links for the same.

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_02_00_05/exports/docs/tidl_j721e_08_02_00_11/ti_dl/docs/user_guide_html/md_tidl_model_import.html

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_02_00_05/exports/docs/tidl_j721e_08_02_00_11/ti_dl/docs/user_guide_html/md_tidl_sample_test.html

Kindly let me know if this is correct or I need to use only TIDL-OSRT for the same?

Regards,

Padmasree N.

0 kumar.desappan over 2 years ago in reply to Padmasree N

TI__Mastermind 22145 points

It is not mandatory to use TIDL-OSRT , but debugging function issue would be easer with TIDL-OSRT

Processors

Processors forum

TDA4VM: MNIST model - Wrong inference with OpenVX