• State Resolved
  • Locked Locked
  • Replies 3 replies
  • Subscribers 98 subscribers
  • Views 350 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: TIDL Importer for quantized modles

Tsurumoto
Genius 15254 points
Part Number: TDA4VM

Hi

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_model_import.html

current release doesn't support quantize models.

Q1: Do you have a plan to support the quontized models in the future?

Q2: When is it released?


Thanks and Best regards,

HaTa

  • 0
    TI__Guru 56820 points

    Hi Hata,

        Currently we only have support for Quantized aware trained (QAT) models from Jacinto AI Devkit (https://github.com/TexasInstruments/jacinto-ai-devkit). We will have support for other frameworks like TFLite in future but right now I don't have exact timelines about when it will be available.

    Regards,

    Anshu

  • 0 in reply to Anshu Jain
    TI__Genius 15254 points

    Hi Anshu

    I understand from your comment that importing Float model after QuantizedAwareTraining is supported.

    It seems that Jacinto AI Devkit's Training Repository is currently only available on a Pytorch base as shown below, and Tensorflow based training repository seems to be coming later.

    I need to use Tensorflow due to Google-derived technology patents such as BatchNormization, but when will the Tensorflow-based Training Repository be available?

    Thanks and Best regards,

    HaTa

  • +1 in reply to Tsurumoto
    TI__Guru 56820 points

    Hi HaTa,

          We have tried using QAT from TensorFlow model Optimization toolkit in the current form. We have not found good results with imagenet training. There is no example script provided for Imagenet in the toolkit. MNIST example is provided but it didn’t yield good result for imagenet training for network like MobileNets. Last release was in 3Q-2020 after that there is no further release. However their repo is quite active so we are hopeful of using it once they make next release. We can only comment on timelines ones the next release is available.

    Regards,
    Anshu