• State Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 2 answers
  • Subscribers 97 subscribers
  • Views 765 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.

Does TI have Deep-Learning tools : Pruning, Quantization, Compression ?

Peter Li79
Expert 8940 points
Other Parts Discussed in Thread: TDA2

Hi, expert:

My customer is asking whether TI has the tools as following? 

  • Pruning
  • Quantization (both post-training quantization and quantization aware training)
  • Compression

I can find some slides about Quantization in the TIDL training slides. it seems the Quantization  is part of TIDL Import tool.

Please kindly share your comments.

Many thanks.

  • 0
    TI__Genius 11506 points

    Hi,

    We are planning to publish tools for quantization - both post training calibration and quantization aware training. There are also several other training examples that we are including. It is expected to be made available at the following location very soon - hopefully in just a few days:

    https://git.ti.com/jacinto-ai-devkit/

    In each repository at that link, click on the "about" tab to read the documentation for that repository.

    Also, here is a collection of the e2e queries related to jacinto-ai-devkit: https://e2e.ti.com/support/j721e/f/1026/tags/jacinto_2D00_ai_2D00_devkit

    When you post a question, you can include: jacinto-ai-devkit in the tags as you create the query, for a faster response.

    Best regards,

    Manu.

  • 0 in reply to Manu Mathew
    TI__Expert 8940 points

    Hi, Manu:

    Many thanks for the information regarding quantization tool.

    How about Pruning and Compression?

    Thanks.

  • 0 in reply to Peter Li79
    TI__Genius 11506 points

    Hi Peter,

    This devkit can be used to train models for Jacinto7 TDA4x. Jacinto7 TDA4x does not benefit from sparsity. The MAC throughput is quite high anyway. So there is no point in applying pruning of weight coefficients like we used to do for Jacinto6 TDA2/TDA3. 

    We need to add pruning/weight sparsity only if we want to enable training of sparse models for Jacinto6 TDA2/TDA3 through this devkit. We do not want to rule out adding that support - but it is not in the current plan. At present, you can use caffe-jacinto for that.

  • 0 in reply to Manu Mathew
    TI__Genius 11506 points

    To explain further, TIDL import tool does a simple calibration. It works reasonably well for common networks, but not so well for networks have Depthwise convolution layers. These is also a plan to address these short comings and implement advanced calibration in TIDL import tool for Jacinto7, but that will take a few months to be released. 

    You can use jacinto-ai-devkit if:

    - you want to use advanced calibration before it is incorporated in TIDL import tool.

    - you want to use quantization aware training (which may be slightly better than advanced calibration).

    I hope it is clear. Let me know if you have further questions.

  • 0 in reply to Manu Mathew
    TI__Genius 11506 points

    Hi, 

    We have make pytorch-jacinto-ai-devkit available. Please start by reading here:

    https://github.com/TexasInstruments/jacinto-ai-devkit

    It has tools to help with Calibration for Quantization and Quantization Aware Training.

    Best regards,

    Manu