• State TI Thinks Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 2 answers
  • Subscribers 54 subscribers
  • Views 333 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.

CCS/PROCESSOR-SDK-DRA8X-TDA4X: WARNING: PReLU Layer fc256_relu's

shuai wang32
Genius 4770 points

Part Number: PROCESSOR-SDK-DRA8X-TDA4X

Tool/software: Code Composer Studio

HI:

            I   am   using    psdk_rtos_auto_j7_06_02_00_21   SDK ,         for  TIDL       i     have   try      some    example   model  in  the     TIDL Importer    document.

           but    i      Import    my   CNN    have   some   problem.    log  lis  like   follow!

IPLayer,=128
WARNING: PReLU Layer pre_lu6's bias cannot be found(or not match) in coef file, Random bias will be generated! Only for evaluation usage! Results are all random!
IPLayer,=256
WARNING: PReLU Layer pre_lu7's bias cannot be found(or not match) in coef file, Random bias will be generated! Only for evaluation usage! Results are all random! 

       We use Caffe framework and tidl to run the same instance model, and compare the output results of each layer. It is found that the error of some layers in front of the network is relatively small, and the error is expanding after the warning prelu layer.

           I have given my example model, my tidl, in the attachment_ Import configuration file, data generated by Caffe and tidl dump respectively, and comparison chart of results.

Thanks!

Shuai

  • 0 in reply to shuai wang32
    Genius 4770 points

    Hi :

              model_flatten    is   the   model   for   test

               test_bgr.raw     is   image   for   test

              tidl_import_test.txt   is       Import configuration file.

              import_log       is     my    Import log .

              pc_result  is  dump    result   of     caffe 

             trace-import_model_flatten_8bit     is  dump    result   of     TIDL.

            compear     is  result  compear .

    Thanks

    Shuai

       

              

  • 0 in reply to shuai wang32
    TI__Genius 17580 points

    Hi Shuai,

    Sorry for the delay.

    >> WARNING: PReLU Layer pre_lu6's bias cannot be found(or not match) in coef file, Random bias will be generated! Only for evaluation usage!

    This warning is because there are flatten and Inner product layers before this PReLU layer and those are making the tensor to 1D data, and this 1D data can not be handled by PReLU layer so import tool can not support importing this model.

    Thanks,

    Praveen 

  • 0 in reply to Praveen Eppa1
    Genius 4770 points

    Hi :Praveen

            so    how   can     i      slove   this   problem? 

            Is it feasible for me to connect the pre-u layer with the custom layer?

           Thanks!

    Shuai

  • 0 in reply to shuai wang32
    TI__Mastermind 22145 points

    You can user custom layer API to mark all the Prelu layer with 1D input as unsupported by TIDL and use custom layer implementation for the same.

    if it is only one layer at the end, you implement it as post processing layer also outside TIDL