• State Resolved
  • Locked Locked
  • Replies 13 replies
  • Answers 4 answers
  • Subscribers 99 subscribers
  • Views 849 views
  • Users 0 members are here
Support feedback
Options
Options
Similar topics

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.

RTOS/TDA2EVM5777: Caffe-Jacinto-Models Pre-trained OD model Accurate

zane low
Prodigy 70 points

Part Number: TDA2EVM5777

Tool/software: TI-RTOS

HI,

I using Caffe-Jacinto-Models Pre-trained model and testing SDK03_06 TIDL import tools.

Question:

1. Did i use the import tool correct ?
2. The Import config param can be modify ? Which Docs maybe describe about these param ?
3. After import the bin files size look difference with TIDL_files in SDK apps/tools, caffemodel src is not same ?
4. Only JDetNet for OD support currently?

Src:

Caffe-Jacinto-Models: version 0.17 and Copy deploy.prototxt from 0.16

Import Config: /cfs-file/__key/communityserver-discussions-components-files/791/5430.tidl_5F00_import_5F00_JDetNet_5F00_voc0712.zip

Deploy File: /cfs-file/__key/communityserver-discussions-components-files/791/6011.deploy.zip

Input video: https://www.youtube.com/watch?v=MNn9qKG2UFI

Bin file i generate

Bin file found in SDK

After Generate *.bin, put file to SD card and use TIDL usecase (step F / 3):

1. keep_top_k: 200 / confidence_threshold: 0.01 - Video section all black and not show

2. keep_top_k: 20 / confidence_threshold: 0.15 - Video show but have many false boxes

(Use *.bin in SDK TIDL_files only miss boxes but no false boxes)

Reference:

https://e2e.ti.com/support/processors/f/791/t/689617#pi320966=1

  • 0
    TI__Genius 17580 points
    Hi,

    1. No, please use import config file from latest release (the config file in the reference e2e post was from an older release at time of that post created).

    2. Yes, you can refer to TIDL user guide (TIDeepLearningLibrary_UserGuide.pdf) in the TIDL release

    3. This expected as the config file is not the latest one, If you use latest config these sizes should match.

    4. Yes.

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Prodigy 70 points

    Hi Praveen,

    Where can i get the latest release config ?
    I use download link from reference e2e post.

    Thanks for reply.

  • 0 in reply to zane low
    TI__Genius 17580 points
    From the latest TIDL release, which is part of VSDK release, please make sure it should be TIDL 01.01.03.00.

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Prodigy 70 points

    Hi Praveen,

    i tried tidl_import_JDetNet to generate bin files, but the sizes also 3047KB.

    but i consider these two inputs:

    inputNetFile = "..\..\test\testvecs\config\caffe_jacinto_models\trained\image_detection\jdetNet_768x320\deploy.prototxt
    inputParamsFile = "..\..\test\testvecs\config\caffe_jacinto_models\trained\image_detection\jdetNet_768x320\ti-jdetNet_768x320.caffemodel"

    Is that means these files in Caffe-Jacinto-Models 0.16 / object detection / 768x320 / sparse? because i make change input files from Caffe-Jacinto-Models.

  • 0 in reply to zane low
    TI__Genius 17580 points
    Can you please share import config file used to get 3047 KB size

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Prodigy 70 points

    Hi Praveen,

    Caffemodel using Github repo 0.16 sparse.

    Zip file provide:

    1. two import config (both get 3047KB bin file)
    2. output bin files
    3. import tools bat file

    Thanks for help.

    /cfs-file/__key/communityserver-discussions-components-files/791/TIDL_5F00_JDetNet.zip

  • 0 in reply to zane low
    TI__Genius 17580 points
    Looks like you did not change below values in the Prototxt for the import output sahred. Please try with below values and with latest import config,
    keep_top_k: 20
    confidence_threshold: 0.15

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Prodigy 70 points

    Hi Praveen,

    Change the values still output bin files with 3047KB.

    I try two case:

    1. Use Default Values ( keep_top_k:200 / confidence_threshold: 0.01 ): TIDL OD Usecase show screen but no video seems fps too low
    2. Change Suggest Value ( keep_top_k:20 / confidence_threshold: 0.15 ): TIDL OD Usecase show screen and video but got many false boxes and not same with VSDK apps/tools provided bin files.

    So i suppose the caffe model source or config maybe not same?
    The OD accurate and Category for these two model ?

    Thanks.

  • 0 in reply to zane low
    TI__Genius 17580 points
    Hi ,

    The change in sizes of output bin files could because of the pre-built bin files had only 4 classes, but this model in the git-hub had 21 classes, so build VSDK with below one line change for this new model to get proper output in VSDK for this git hub model as explained in the first post of the reference link (e2e.ti.com/.../689617.

    File : vision_sdk\apps\src\rtos\alg_plugins\objectdetection\objectDrawLink_algPlugin.c
    Line : tempoutPutList->objDesc[tempoutPutList->numObjects].objType = (label - 1)%3;

    So, please try with above change by referring to the reference e2e link.

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Prodigy 70 points

    Hi Praveen,

    I have notice that classes increase to 21 and check the code in objectDrawLink_algPlugin.c.

    The code is set to ((label - 1)%3) not the comment as below in the reference said (label - 1).

    Anyway, i change ((label - 1)%3) to (label - 1)%3 and try build image again. then see what will be difference.

    Thanks.

  • 0 in reply to zane low
    TI__Genius 17580 points
    Okay, please let us know the results after you try them.

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Prodigy 70 points

    Hi Praveen,

    it's look better a lit bit (eg. big false boxes not stuck in video for long time), but using trained model at github still display false boxes more than the version from VSDK provided.

    Github 0.16&0.17 trained model seems worse than VSDK Version (PROCESSOR_SDK_VISION_03_06_00_00\vision_sdk\apps\tools\TIDL_files\*.Bin) .

    For Import Tools tutorials maybe ok, but result look not fine (accuracy and weird boxes).

    So the Src caffemodel and deploy.prototxt not the same i guess, because i tried difference config file (default / reference / etc.) and the output param_Bin file size not match.
    If you have suggest config, maybe can share and i double check the trained model.

    By the way, Files and Path from github place in, need to do some modify by user that using the latest SDK. But the Problem can be trace with Import Tools Log Files.

    Thanks.

  • 0 in reply to zane low
    TI__Genius 17580 points

    Hi,

    Thanks for your feedback. Yes, it is expected that bin size in VSDK will not match with github generated bin files (as both are trained with different classes) and also user to need to update import config file to use git hub models. 

    Regards,

    Praveen