• State TI Thinks Resolved
  • Locked Locked
  • Replies 8 replies
  • Answers 4 answers
  • Subscribers 99 subscribers
  • Views 739 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.

TDA2xx_EVM_BIOS : Error in importing tensorflow pretrained model

Pratik Brid
Intellectual 750 points
Other Parts Discussed in Thread: TDA2

Hi

I have installed PROCESSOR_SDK_VISION_03_07_00_00 which i have configured for tda2xx_evm_bios_all.
I have copied the REL.TIDL.01.01.03.00 folder to my windows pc and using the tidl import tool tidl_model_import.out.exe.
when i give the pretrained tensorflow model frozen_graph file x.pb to import tool tidl_model_import.out.exe
the command executed is tidl_model_import.out.exe ..\..\test\testvecs\config\import\tidl_import_x.txt
then it gives me this message
"
TF Model File : ..\..\test\testvecs\config\tensorflow_models\x\x.pb
Op Type Cast is Not suported will be By passed
Op Type Shape is Not suported will be By passed
Could not find the requested input Data : mul !!
"
I was referring the text file in the path ..\..\test\testvecs\config\import\*.txt file There are paramters mentioned as
quantizationStyle
quantRoundAdd
numParamBits
preProcType
conv2dKernelType
I referred the documentation TIDeepLearningLibrary_UserGuide.pdf where the explanation of each parameters is given but how can we determine that for my particular model which paramters need to set and to what value. 
Request you for your inputs.
Regards
Pratik
  • 0
    Intellectual 750 points

    Hello ,

    Any update on this ??

    Regards

    Pratik Brid

  • 0 in reply to Pratik Brid
    TI__Genius 17580 points

    Hi Pratik,

    Sorry for the delay.

    The error is because of your models has unsupported layers, please refer to "TIDeepLearningLibrary_DataSheet.pdf" for supported layers.

    Refer to section "3.6.2 Importing the Model and Parameters" for the explanation about the parameters in the import.txt file.

    Thanks,

    Praveen 

  • 0 in reply to Praveen Eppa1
    Prodigy 100 points

    Hi Praveen,

    For the above two errors on the unsupported layers,

    Op Type Cast is Not suported will be By passed
    Op Type Shape is Not suported will be By passed

    is there any alternative available which can be used in model re-training so that the layers would be supported with respect to TIDL?

    I am trying to import face-detection which was trained in tensorflow

  • 0 in reply to Sudhanshu Mishra
    TI__Genius 17580 points

    Hi Sudhanshu,

    One way to use only those layers which are supported in TIDL in your network, please refer to TIDL datasheet for supported layers.

    Other way would be split network up to TIDL supported layers and run that using TIDL and run remaining network outside the TIDL.

    Thanks,

    Praveen

  • 0 in reply to Praveen Eppa1
    Prodigy 100 points

    Hi Praveen,

    Thank you for the reply

    I had referred to TIDL datasheet and understood about the unsupported layers. I am a beginner in DL, so can you suggest how am I supposed to split my layers/networks in model file(x.pb). The model I am using is pre-trained model on Face-detection(Facebox).

    Also, with respect to running supported layers on TIDL and remaining outside the TIDL, by outside do you mean importing the layers on EVE cores or is it something else?

    Regards

    Sudhanshu

  • 0 in reply to Sudhanshu Mishra
    TI__Genius 17580 points

    Hi Sudhanshu,

    You need to freeze and optimize the graph with all the layers from input to the last supported layer and run this on EVE core and remaining on the DSP core, please refer tp FAQ 24 in the TIDL user guide for more details.

    Thanks,

    Praveen

  • 0 in reply to Praveen Eppa1
    Prodigy 100 points

    Hi Praveen,

    Sorry the late reply and thanks for your above response.

    I am using PROCESSOR_SDK_VISION_03_07_00_00 which had TIDL version 01.01.03.00... So, in the UserGuide, there is no FAQ 24(Kindly mention what version u are referring to, so that latest TIDL import tool can be referred)

    However, in the above mentioned TIDL user-guide's FAQ 22, it mentions about how to run the splitted networks on EVE and DSP respectively..

    I saw that DSP take input of bin files.. I am still not able to understand how are we going to create binaries for the unsupported layers? SInce TIDL's import tool will create bin file only for the supported layers.

    Thanks

  • 0 in reply to Sudhanshu Mishra
    TI__Genius 17580 points

    Hi 

    I am referring to latest release of PROCESSOR_SDK_VISION_03_08_00_00 which had TIDL version 01.02.00.00... which had this FAQ24.Please refer to that.

    If you have unsupported layers in the middle of the network then that model can't be run on TDA2 (you may have to re-train your model with supported layers), but if you have unsupported layers at the end then those can be offloaded to DSP.

    Thanks,

    Praveen