• State Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 948 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.

PROCESSOR-SDK-TDAX: Deployment of the TI import tool result

safwan ismael
Intellectual 300 points

Part Number: PROCESSOR-SDK-TDAX

Hello,

after using the impot tool, one obtains two bin files: one for parameters (say parameters.bin) and one for the model (say model.bin).

For which processor the instructions in model.bin are meant? How to deploy this file on TDA? Does it contain instructions which are directly executable on the TDA after linking model.bin to the proper frameworks? Doesn't it contain calls to functions in the TIDL? In that case, how the link is done?

Thanks a lot in advance.

Best,

Safwan

  • 0
    TI__Genius 17580 points
    Hi Safwan,

    These two bin flies are not executables or does not contains any functions of TIDL.
    1. Model.bin is having the model(layers) imported from framework
    2. Parameters.bin is having the kernel coefficients for each layer

    Please refer to Vision SDK use case to how to use these bin files.

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    TI__Genius 13395 points
    Hi Safwan,
    refer VisionSDK_UserGuide_TI_DeepLearning.pdf to prepare SD card and run TIDL demos

    regards, Shiju
  • 0 in reply to Shiju Sivasankaran
    Intellectual 300 points
    Hello,

    thank you for the answers of both of you.
    I want to understand the behaviour of TIDL at RUNTIME but to tell the truth, it is not so clear from the document you mentioned.
    So far, it seems to me that the role of the TIDL at runtime is to carry out a sequence of convolutions, activations, pooling, etc (guided by the information saved in the output of the import_tool). Right?
    A user may also write code to read the information in the output of the import_tool and carry out these operations (if the format of the output is disclosed). My question is: What does TIDL do differently at RUNTIME in these elementary operations compared to an implementation done by a user?

    Best,
    Safwan
  • 0 in reply to safwan ismael
    TI__Mastermind 22145 points
    Hi Safwan,
    The TIDL accepts the information (CNN net structure) generated by the import tool and processes these layers one by one (convolutions, activations, pooling). TIDL has optimized implementation for all these layers (~18 types of layes) on EVE and DSP core. The user may not need to write code on EVE or DSP to deptoy his CNN network.

    Please refer TIDL user guide for mopre information importing a CNN network. Refer the VSDK usecase or TIDL tetsbench to pass imported information to TIDL libary.

    Please refer below white paper on TIDL to get a high-level understanding about the software eco system from TI for deep learning on TDA2x
    www.ti.com/.../spry314.pdf


    Hope this is helpful.