• State Resolved
  • Locked Locked
  • Replies 7 replies
  • Answers 1 answer
  • Subscribers 97 subscribers
  • Views 689 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.

Original question:

TDA2EXEVM: How can I use TIDL's tensorflow framework？

TDA2EXEVM: The ssdJacintoNetV2 is ported to the EVM board to display black screen

Focus_lee
Expert 1055 points

Part Number: TDA2EXEVM

Hi ,

I use the ssdJacintoNetV2 which isTi provided Object detect net in caffe-jacinto do a binary task of Vehicle recognition。

1、And after training，it can work on the PC,but when I import it into bin to run on the evm, it show black screen，The transformation process is as follows.  But i import the demo trained model of voc0712 is OK.

2、Regardless of whether I use this network for binary or multi-category training tasks, the loss value remains around 4.4,show as blow. why?

  • 0
    TI__Genius 17580 points
    Hi,
    >>> it can work on the PC,but when I import it into bin to run on the evm, it show black screen
    Can you please describe the steps how are you running on EVM? Also share the config file used to run these bin files on EVM?

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Expert 1055 points

    Fullscreen 8688.tidl_import_JDetNet_desay.txt Download 
    # Default - 0
randParams         = 0 

# 0: Caffe, 1: TensorFlow, Default - 0
modelType          = 0 

# 0: Fixed quantization By tarininng Framework, 1: Dyanamic quantization by TIDL, Default - 1
quantizationStyle  = 1 

# quantRoundAdd/100 will be added while rounding to integer, Default - 50
quantRoundAdd      = 25

numParamBits       = 8
# 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType      = 0 

inputNetFile       = "deploy.prototxt"
inputParamsFile    = "ti-custom-cfg1_ssdJacintoNetV2_iter_4000.caffemodel"
outputNetFile      = "tidl_net_jdetNet_ssd_caronly.bin"
outputParamsFile   = "tidl_param_jdetNet_ssd_caronly.bin"

rawSampleInData = 1
preProcType   = 4
sampleInData = "trace_dump_0_768x320.y"
tidlStatsTool = "eve_test_dl_algo.out.exe"
layersGroupId = 0	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	2	0
conv2dKernelType = 0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1
    step1、modify the deploy.prototxt in the end filed " keep_top_k: 20   confidence_threshold: 0.15 "

    step2、I use the cammand .\tidl_model_import.out.exe tidl_import_JDetNet_desay.txt transform the caffemodel to bin

    step3、Modify  the file "TIDL_SSD_CFG.TXT" netFileName=tidl_net_jdetNet_ssd_caronly.bin paramFileName=tidl_param_jdetNet_ssd_caronly.bin

  • 0 in reply to Praveen Eppa1
    Expert 1055 points
    I am also try to train the voc dateset into 3classes, It can run on the EVM after tramform caffemodel to bin .is this ssdJacintoNetV2 net can't do one class task?
  • 0 in reply to Focus_lee
    TI__Genius 17580 points
    Hi,

    Did you try the steps in this below thread?
    e2e.ti.com/.../689617

    Also, are you able to see the detection's in the OD use case with pre-built bin files ?

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    TI__Genius 17580 points
    Hi,

    Are you able to try the suggested in the above thread ?

    Thanks,
    Praveen
  • 0 in reply to Praveen Eppa1
    Expert 1055 points

    ya,

    thank you Praveen , I have solve this problem.

    The reason is that some of my dataset label are a little too small, As a result, the training accuracy is not high, the loss value is always around 4.5, and the sparse training process is also very bad.

    Can you share some  training  experience to me about the ssdJacintoNetV2. My recognition rate is not very good now.

  • 0 in reply to Focus_lee
    TI__Genius 17580 points

    Hi,

    Glad that you could solve your problem.

    Please create a new thread for this question on training accuracy.

    Thanks,

    Praveen