• State Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 1 answer
  • Subscribers 99 subscribers
  • Views 468 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.

TDA2E: Issues running caffe trained SSD model on TDA2 development board

WEI CHIH CHERN
Intellectual 320 points
Part Number: TDA2E
Other Parts Discussed in Thread: TDA2

Hello,

I've used scripts 'train_image_object_detection.sh' to train SSD for detection.  However, some trained models failed to run on TDA2 development board after conversion from tidl_model_import.out, the issues are:

1. It's does run for few frames, but then the output screen/detection process froze.

2. I only changed hyper-parameters listed in the 'train_image_object_detection.sh', but no matter how I changed ssd_size, resize_width/height, and ds_fac, the converted models still showed same error behavior.

I've attached a converted caffe SSD (0420_3.7z) & a .caffemodel & deploy.prototxt & lable_map.prototxt (caffeModel.7z), would you please test it and perhaps offer any suggestion that might solve the issue?

Thank you,

Wei Chih

https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/791/0420_5F00_3.7z

https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/791/6724.caffeModel.7z

  • 0
    TI__Genius 17580 points

    Hi Wei Chih,

    Did you get the default use case working at your end ?

    Also, which Vision SDK version you are using ?

    Thanks,

    Praveen

  • 0 in reply to Praveen Eppa1
    Intellectual 390 points

    Hi Praveen:

    This is wxchen who is colleague with Wei-Chin,

    i was trying to inspect image by this transferred model by default usecase in TIDL OD(RTOS, TDA2EVM, 4-eves, Processor VisionSDK: 3.07),

    here is the manifest on SDK.Processor_SDK_Vision_manifest.html

    but the output was same as Wei-Chin's saying.

    by the way, 

    i was trying to use pre-built binary(3.08), and load the network file to board, then i got the following error:

    [IPU1-0]  TIDL Usecases
[IPU1-0]  ---------------
[IPU1-0]  1: TIDL File I/O Usecase
[IPU1-0]  2: Semantic Segmentation Usecase
[IPU1-0]  3: TIDL OD Usecase
[IPU1-0]
[IPU1-0]  x: Exit
[IPU1-0]
[IPU1-0]  Enter Choice:
[IPU1-0]
[IPU1-0]     17.395199 s:
[IPU1-0]     17.421734 s:  TIDL Configuration parameters
[IPU1-0]     17.421826 s:  -----------------------------
[IPU1-0]     17.421887 s:  inputWidth         = 768
[IPU1-0]     17.421978 s:  inputHeight        = 320
[IPU1-0]     17.422039 s:  inputFile          = inData_OD
[IPU1-0]     17.422100 s:  inputFile          = inHeader_OD
[IPU1-0]     17.422161 s:  netFileName        = tidl_net_jdetNet_ssd.bin
[IPU1-0]     17.422222 s:  paramFileName      = tidl_param_jdetNet_ssd.bin
[IPU1-0]     17.422283 s:  inputfps           = 8
[IPU1-0]     17.422344 s:  threshold          = 0.200000
[IPU1-0]     17.422436 s:  -----------------------------
[IPU1-0]     17.517049 s:  Assertion @ Line: 103 in /adasuser/surya/vsdk_3_8_6/vision_sdk/apps/src/rtos/usecases/common/chains_common_tidl.c: readSize == sizeof(sTIDL_Network_t) : failed !!!
[IPU1-0]     17.517873 s:  Assertion @ Line: 103 in /adasuser/surya/vsdk_3_8_6/vision_sdk/apps/src/rtos/usecases/common/chains_common_tidl.c: readSize == sizeof(sTIDL_Network_t) : failed !!!

    wxchen

  • 0 in reply to Praveen Eppa1
    Intellectual 320 points

    Hi Praveen,

    It turned out, it was the conversion mistake made by one of my coworker.  It does run on the board as expected now.  

    #update:

    It turns out only the newly cloned caffe-jacinto-model/scripts/train_image_object_detection.sh trained SSD models worked; every parameter left unchanged, and trained with my custom dataset

    Once changing parameters like ds_fac and ssd_size, the trained model will only run the first few frames on the board then froze.  I will verify the change of ds_fac and ssd_size does make the trained models un-runable on the board tomorrow at work with the newly cloned caffe-jacinto-models again, then report back.

    Thanks a lot,

    Wei Chih

  • 0 in reply to WEI CHIH CHERN
    Intellectual 320 points

    Hi Praveen,

    After cloning new caffe-jacinto-models, and use 'train_image_object_detection.sh' for training, I discovered that the models that used batch normalization for training cannot be ran on TDA2 dev-borad.

    Thank you,

    Wei Chih

  • 0 in reply to WEI CHIH CHERN
    TI__Genius 17580 points

    Hi Wei Chih,

    Thanks for the update.

    Regards,

    Praveen