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.

TDA4VM: An post-processing issue based on Yolov5 model

Part Number: TDA4VM


Hi Team,

We use our own model trained on the basis of YoloV5. The output feature dimension of this model is different from the default dimension supported in the current TIDL library. When using TIDL's tidl_model_import.out tool for conversion, I need to modify the value of the numInCodeElements variable in the TIDL_findValidLocation_cn() function to get the correct result. But when running the converted model on the TDA4VM development board, I found that the program would call the countM = pKerPrivArgsL->execute((void *)pKerPrivArgsL, NULL, NULL, NULL, NULL) function. We initially detected that this step would execute the function pointed to by the pFxnTIDL_odFindValidLocAndScore_ixX_oxX_exec function pointer. We have not found the definition of the function so far. Could you please provide the specific implementation of this function? If there is no specific implementation, could you please specify its differences from the function implementation on the PC side (that is, the TIDL_findValidLocation_cn() function)? I would appreciate your help.

Kind regards,

Katherine

  • Hi Katherine,

    When you run the inference on PC then it runs by default in reference mode where _cn() functions are called. Is this flow functional for your model?

    You can set flowCtrl=1 in inference cfg file to run the reference flow on h/w otherwise DSP optimized code will run. I will recommend to make sure that you get the expected functionality on PC first. 

    On your specific question --You can search for the function TIDL_odFindValidLocAndScore_exec_c7x() in whole project, you should be able to find it in the file <tidl_dir>/ti_dl/algo/src/tidsp/c7x/tidl_odFindValidLocAndScore_ixX_c7x.c.

    regards

    Deepak Poddar