TDA2EXEVM: how to match network results between caffe-jacinto(float) and TIDL(int8)

ZhangQiang

Part Number: TDA2EXEVM

I tried to match the results layer by layer between caffe-jacinto and TIDL. But the results are quite different.

how to use quantization params to get float point weights, bias, inData and outData? the TIDL layer's output is int8 or uint8?

int32_t weightsQ;
int32_t biasQ;
int32_t inDataQ;
int32_t outDataQ;
int32_t interDataQ;

over 7 years ago

0 Praveen Eppa over 7 years ago

Expert 1420 points

Hi,

Similar question is already answered in below e2e thread, please refer to this

Thanks,

Praveen

0 ZhangQiang over 7 years ago in reply to Praveen Eppa

Intellectual 415 points

it's different question.
i imported one model, and can NOT match ConvolutionLayer output, so i want to match layer's weights and output.
i do NOT use caffe-jacinto's quantization.

0 Praveen Eppa over 7 years ago in reply to ZhangQiang

Expert 1420 points

Hi,

Please refer to function "TIDL_QuantizeUnsignedMax" in tidl_import_common.cpp file, which converts float values to int values. You may need to do reverse of this to get float point values from int.

For getting dataQ for inData and outData, please refer to below thread,
e2e.ti.com/.../2370081

Thanks,
Praveen

0 ZhangQiang over 7 years ago in reply to Praveen Eppa

Intellectual 415 points

I got it, thank you very much!

Processors

Processors forum

TDA2EXEVM: how to match network results between caffe-jacinto(float) and TIDL(int8)