• State Resolved
  • Locked Locked
  • Replies 11 replies
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 894 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.

BEAGLE-3P-BBONE-AI: AM5729

kavya m
Prodigy 155 points

Part Number: BEAGLE-3P-BBONE-AI
Other Parts Discussed in Thread: AM5729

Hi, Can anyone tell me how to implement the ssd_multibox or any tidl files on the host device rather than on eve/DSP accelerators? Since I am trying to understand the elapsed time between host execution and accelerator execution.

  • 0
    TI__Mastermind 40600 points

    Hi Kavya, TIDL examples use model binaries imported for our TIDL HWA (in case of AM57x EVE/DSP). One option could be to use a generic model, say TFLite Mobilenetv1 and build it, run it for ARM only. And them import it for TIDL and run it.

    This could be tricky, but possible.. In other words I am afraid we don't have a demo ready which you can switch from running the model in ARM to complete offload to TIDL. But we have done some benchmarking that shows improvement. I will get in touch with a colleague to see if we have something that can be shared.

    thank you,

    Paula

  • 0 in reply to Paula Carrillo
    TI__Mastermind 40600 points

    Hi Kavya, for your reference our datasheet has some performance numbers. Which might be helpful for you.

    7266.TIDeepLearningLibrary_DataSheet.pdf

    thank you,

    Paula

  • 0 in reply to Paula Carrillo
    Prodigy 155 points

    Thank You, Paula.

    Please do get back to share relevant details, a demo would be great. Since I am just starting out with this technology, having a demo always works..!

  • 0 in reply to Paula Carrillo
    Prodigy 155 points

    Thank you so much for the datasheet, hope to find out more materials for building the models on host AM5729 to understanding its performance efficiency against the efficiency of the accelerator.

  • 0 in reply to kavya m
    TI__Mastermind 40600 points

    Hi Kavya, please see some additional TIDL performance data here:

    Hope this helps,

    Paula

  • 0 in reply to Paula Carrillo
    Prodigy 155 points

    Hi Paula, I went through the link and it helped. Can I know how to get started with deploying custom models using TIDL. Basically, I would like to know how to get started with TIDL and implement my own models using it on BBAI.

    Correct me if I am wrong!

    Thank You.

  • 0 in reply to kavya m
    TI__Mastermind 40600 points

    Hi Kavya, you could import your own models (if all the layers are supported by TIDL) by using our TIDL import tools. After a successful import, you will have two *.bin one for the network model and one for the params. You can check our OOB models to see how we use those models in our Object Detection, or Segmentation demo for example.. so you can replace our *.bin models for yours as a starting point for your development.

    Another option is to use NEO-AI-DLR/TVM. We recently implemented heterogenous subgraph support with NEO/TVM and rolled out the service as part of AWS SageMaker. This option gives the advantage that if there is any unsupported layer by TIDL, your model will still run on AM57x. Supported layers will run as subgraphs in TIDL, unsupported layers run in ARM. Performance could be reduced depending on the model, but we increase support of number of models.

    Let me know, for your type of models and application, which approach you think you would like to take and I can share you more details.

    thank you,

    Paula

  • 0 in reply to Paula Carrillo
    Prodigy 155 points

    Thank you for this information. I have my own Alex net object detection model in python, so i am searching for the conversion tool by tidl , or how these model from pc is imported using tidl import tools.

  • 0 in reply to kavya m
    TI__Mastermind 40600 points

    Hi Kavya, some information in section 3.15.1.4.7. Import Process. TIDL import tool is part of PSDKL filesystem and there are some configuration files examples inside /usr/share/ti/tidl/utils/test/testvecs/config/import that you can use as a reference.

    Thank you,

    Paula

  • 0 in reply to Paula Carrillo
    Prodigy 155 points

     Hi, how to obtain the PSDKL filesystem, for this import process.

    What is the starting point to start using the import tool and network translation tool for my custom development.

    I am now using Beaglebone ai , cloud 9 ide for execution, if i need to use TIDL what is the development process. Please share the necessary files

  • 0 in reply to kavya m
    TI__Mastermind 40600 points

    Hi Kavya, you can donwload Linux PSDK from:

    https://software-dl.ti.com/processor-sdk-linux/esd/AM57X/latest/index_FDS.html

    In this link you will also find some other links to User Guide and how to create an SD card

    thank you,

    Paula