• State
  • Locked Locked
  • Replies 3 replies
  • Subscribers 98 subscribers
  • Views 341 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA4VM: An error occurred during Segmentation demo on TDA4VM.

Jaemin Lee
Prodigy 90 points
Part Number: TDA4VM

Hi.

Error when running ./run_app_tidl_seg.sh on TDA4VM.

I've also tried running ./vision_app_init.sh but I get the same error.

The error looks like the image below.


The segmentation model is a custom model, and it has been confirmed that both compilation and inference work well using tidl on the PC.

Compilation results and inference results are included in the compressed file below.

segmentation.zip

I don't know what the cause is. It compiles well and the model graph looks good, so I think there is no problem with the model itself.

The basic segmentation model works well on TDA4VM.

How can I solve this problem?

Thank you.

SDK version: 8.4.2

  • 0
    TI__Guru 56820 points

    Hi,

        Can you enable more traces by setting debugTraceLevel = 3 in inference configuration file and share the log?

    Regards,
    Anshu

  • 0 in reply to Anshu Jain
    Prodigy 90 points

    Thank you for your answer.

    I found the cause of the error.

    The cause of the error is group convolution with kernel_size=(1,3).

    So, I have one question.

    How can use group convolution with kernel_size=(1,3) on TDA4VM?

    I made a binary file of my model to compile but, my model is not working on TDA4VM.

    Additionally, if I use group convolution with kernel_size=(3,3), my model is working on TDA4VM.

    Thank you.

  • 0 in reply to Jaemin Lee
    TI__Guru 56820 points

    Hi,

       I am assuming here you meant depth wise separable convolution, if thats the case then 1x3 convolution is not supported for depth wise convolution. 


    Regards,

    Anshu