TDA4VM: deeplabV2 segmentation model API inference result is different from the PC (Emulation)

? ?

Part Number: TDA4VM

Tool/software:

We use the comand "./out/tidl_model_import.out" to import and get the int16 TI model. Then I use the "./PC_dsp_test_dl_algo.out" to infer the model in PC (Emulation) and use the API code to infer the model in TDA4VM board . But ,I get different result in the 500 inference images , the inference in the TDA4VM board get the lower miou ( 0.97 ) and the inference in PC (Emulation) get higher iou (0.98) . I compared the layers of the single image , I find there is difference in every layer between the PC (Emulation) and in TDA4VM board . What is wrong with the result ? Any one can help me ?

over 1 year ago

0 Pratik Kedar over 1 year ago

TI__Mastermind 24041 points

Hi,

Please check out our trouble shooting guide here : https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_osr_debug.md

0 Asha Bhandarkar over 1 year ago

TI__Genius 10170 points

Hi,

I apologize for the delay in getting a full fledged response back to you.

From my understanding across all your inference images you are seeing a consistent mismatch between host emulation and target.

Can you give more information on what SDK you seeing this behavior on? It looks like you have gone through the process of generating the corresponding traces with writeTraceLevel set when quantization is int16

Can you report back with what exact layer you are first seeing the deviation as well as attach your model and related configuration scripts for compilation and inference?

Best,

Asha

Processors

Processors forum

TDA4VM: deeplabV2 segmentation model API inference result is different from the PC (Emulation)