Part Number: TDA2
Tool/software: Linux
I am using TDA2 evm SDK_VISION_03_04_00_00 , i run a cnn in sample_app\src\rtos\alg_plugins\framecopy\frameCopyAlgoLocalDma.c
my cnn have a big model 15MB about i have to read it by a15 and send buf to C66 , i read the model file in vision_sdk\sample_app\src\hlos\usecases\nullSrc_display\chains_nullSrcDisplay.c and send the address of model buffer to c66. in frameCopyAlgoLocalDma.c\Alg_FrameCopyCreate i get the address of model buffer .
in frameCopyAlgoLocalDma.c\Alg_FrameCopyProcess Everything went smoothly. but i have a problem at last.
when i run the program ， some times the out put of the CNN is always right ,some times front 3-5 frames is wrong but little difference, some times front 300-500 frames is wrong but little differenc
So it's uncertain how many frames the results are incorrect, sometimes more, sometimes less, but after running 500 frames, they're basically correct. Later, I found that when the CNN output was incorrect, it was because the obtained model was partially incorrect. So the results are incorrect, but the gap is not big. So I think it's risky to get the CNN model in this way. It takes a while for C66 to get the correct CNN model.
What causes this problem? Is there a good way to solve it?
BESTWISHES
Shuai!
