Hi,
I am using SDK ti-processor-sdk-rtos-j721e-evm-07_01_00_11, when I am converting CenterNet model from onnx to ti format, test bench is saving results correctly
But when I use same model (Model is converted in floating point) in modified app of ap_tidl_od for centernet model.
This modified app is working okay with int8 or int16 quantized models (even though some deviation in PC emulator and Evm results)
- When I try to save intermediate outputs from model I realize so many nan values are generated, saveing script from file ap_tidl_module.c is as below
for(k = 0; k < tensor_sizes[2]; k++)
{
float *pOut = (float *)data_ptr + (tensor_sizes[0] * tensor_sizes[1] * k) + (ioBufDesc->outPadT[0] * tensor_sizes[0]) + ioBufDesc->outPadL[0];
for(i = 0; i < ioBufDesc->outHeight[tensor_id]; i++)
{
fwrite(pOut, 1, ioBufDesc->outWidth[tensor_id], fp);
for(int i =0; i <ioBufDesc->outWidth[tensor_id]; i++){
fprintf(fp_1, "%f\n", pOut[i]);
//printf("Number is in float %.6f \n", pOut[i]);
}
//fwrite("\n", sizeof(char), 1, fp_1);
pOut += tensor_sizes[0];
}
}
Can you please let me know why nan values are being generated?
what is wrong here?
Thank you
Mudasar
