• State Resolved
  • Locked Locked
  • Replies 2 replies
  • Subscribers 97 subscribers
  • Views 458 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.

Original question:

TDA4VM: TIDL Importer for quantized modles

TDA4VM: Failure when trying to compile a quantized model

Tsurumoto
Genius 15254 points
Part Number: TDA4VM

Hi

When tried to compile a quantized model, it's finished with failure.

According to the description at

    https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_02_00_06/exports/docs/tidl_j7_01_04_00_08/ti_dl/docs/user_guide_html/md_tidl_osr_tflrt_tidl.html

a quontized model is supported at "TFLiteRuntime + TIDL HeterogeneousExecution".

The model I use is "Efficientnet-Lite4" which is lined at the following TIDL manual.

　　https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_02_00_06/exports/docs/tidl_j7_01_04_00_08/ti_dl/docs/user_guide_html/md_tidl_models_info.html

When I tried to compile quontized tiflite which is linked to above by using "tflrt_delegate.py", it's finished with the following error.


Q1: Is a quontized model really supported?

Q2: If it's not supported yet then please let me know when it's released.

Q3: If it's already supported then please let me know why above compile is finished with error.

Thanks and Best regards,

HaTa

  • +1
    TI__Genius 9065 points

    Hi HaTa,

    Quantized models are supported for tflite runtime.

    In the currently release version, uint8 support exists, but int8 support needs to be added.

    I took a look at the site for efficientnet_lite4 models, and looks like the model under uint8 also has tensor type int8,so you may not be able to run that as well.

    The above snapshot has the fp32 version validated. You can give a try with fp32 for this model, that should work. But in general,uint8 tensor type models should work, int8 won't.

    Since this bug has also come after code freeze for SDK 7.3, we will include support for int8 in the next release (SDK 7.4).

    Internal JIRA reference:

    https://jira.itg.ti.com/browse/TIDL-1341

    Regards,

    Anand

  • 0 in reply to Anand Pathak
    TI__Genius 9065 points

    Hi HaTa,

    Just to clarify, next release will be SDK 8.0 not SDK 7.4 as I mentioned above, sorry for the typo.

    Regards,

    Anand