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.

  • Resolved

TDA2EXEVM: TIDL - tidlstasTool result matching

Part Number: TDA2EXEVM

Hello,

I am using TIDL ,and try the imagenet classification model inside the caffe-jacinto-models.

And i use the TIDL model import tool and get the result.

the input  rawsample data is a data inside the folder : ..\..\test\testvecs\input\preproc_0_224x224.y  which is a aircraft

But the output of the TIDLstatsTool is not the same as the the one i run the same models on the caffe-jacinto(the output of the caffe jacinto is class # 403)

but the output of the statsTool is class #704 (I open stats_tool_out.bin to get this result)

here is my TIDL model import tool config_files .

-------------------------------------------------------------------------------------

# 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 = 50

numParamBits = 8

#inNumChannels = 3

Cnov2dKernelType = 0

inputNetFile          = "..\..\test\testvecs\config\caffe_jacinto_models\trained\image_classification\imagenet_jacintonet11v2\sparse\deploy.prototxt"
inputParamsFile   = "..\..\test\testvecs\config\caffe_jacinto_models\trained\image_classification\imagenet_jacintonet11v2\sparse\imagenet_jacintonet_iter_160000.caffemodel"
outputNetFile        = "NET.BIN"
outputParamsFile = "PRM.BIN"

rawSampleInData = 1
#sampleInData = "..\..\test\testvecs\input\000100_1024x512_bgr.y"
sampleInData = "..\..\test\testvecs\input\preproc_0_224x224.y"
tidlStatsTool = "..\quantStatsTool\eve_test_dl_algo.out.exe"

----------------------------------------------------------------------------------------------------------------------------

also i test the segmentation model but the result seems to be wrong either.

Did i miss something? 

or what should i do to get the same result??

Thank you in advance.

  • Hello ,
    I use TIDLSRC.01.00.00.00.
    Does anyone know this problem or any suggestion?
  • Hello,
    Does anyone know this problem or any suggestion?
  • In reply to Eric Lai:

    Hi,

    I checked the TIDL model import tool config file that you used, looks like you missed setting "inElementType      = 0",  can you please check now by setting this.

    After setting "inElementType" to 0, new config file is

    -----------------------------------------------------------------------------

    # 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 = 50

    numParamBits = 8

    #inNumChannels = 3

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

    conv2dKernelType = 0

    inputNetFile = "..\..\test\testvecs\config\caffe_jacinto_models\trained\image_classification\imagenet_jacintonet11v2\sparse\deploy.prototxt"
    inputParamsFile = "..\..\test\testvecs\config\caffe_jacinto_models\trained\image_classification\imagenet_jacintonet11v2\sparse\imagenet_jacintonet_iter_160000.caffemodel"
    outputNetFile = "NET.BIN"
    outputParamsFile = "PRM.BIN"

    rawSampleInData = 1
    #sampleInData = "..\..\test\testvecs\input\000100_1024x512_bgr.y"
    sampleInData = "..\..\test\testvecs\input\preproc_0_224x224.y"
    tidlStatsTool = "..\quantStatsTool\eve_test_dl_algo.out.exe"

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------

    Sorry for the delay, we had long weekend here.

    Thanks,

    Praveen

  • In reply to Praveen Eppa:

    Hello Praveen,

    Thank you for your response.

    The result is reasonable now.

    Few more questions please,

    In the TIDL import configuration parameters:

    1. What is the meaning of dynamic quantization?

    2. What is the preProcType 0 to 3 stands for?

    3.In the TIDL import log , I want to know the exactly meaning of :

         1.Max PASS and values  

         2.The sec value of OUT Q

         3.The sec and third values of #MMACs

         4.The values of sparsity.    I know the definition of sparsity,but there are two values confusing me.

    Thank you!

    Regards,

    Eric

  • In reply to Eric Lai:

    Find the answers below:

    1.What is the meaning of dynamic quantization?

    Kumar: The Q formats/Scaling factor used for activation are selected dynamically at run-time.

    2. What is the preProcType 0 to 3 stands for?
    Kumar: This test tench specific for reading input image and resizing before giving to TIDL library. This can be ignored RAW image (RGB) is used

    3.In the TIDL import log , I want to know the exactly meaning of :
    Kumar : All the logs are for trace purpose, Some of them are not relevant for the final user.

    1.Max PASS and values
    Kumar : This Shall be ignored during import, this have meaning only during inference

    2.The sec value of OUT Q
    Kumar : Only first value is relevant Q of the output data. Second is debug info, can be ignored

    3.The sec and third values of #MMACs
    Kumar : Only first value is relevant. Second and third are debug info considering sparsity, can be ignored

    4.The values of sparsity. I know the definition of sparsity,but there are two values confusing me.
    Kumar : First value is sparsity considering actual processing computations. Second is sparsity considering available number of zero weights. These are different because of alignments during processing.

    Rgerads,
    Kumar.D

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.