I'm dying to know how to implement the int8 winograd in conv3x3s1

[BUG1989]

I find the mkl-dnn has been supported the int8 winograd in convolution.
Is there any paper to introduce how to quantization winograd.

Thanks a lot!

[vpirogov]

@BUG1989, there's a section in documentation describing quantization approach Intel MKL-DNN supports. The same approach is applied to Winograd convolution algorithm.

[vpirogov]

Summoning @tprimak or @nshustrov to comment on the nuances of the implementation.

[nshustrov]

@BUG1989, the implementation of int8 winograd convolution in MLK-DNN does not require special quantization. The approach to quantization of input data and weights is the same with quantization used for direct int8 convolution with unsigned int8 input and signed int8 weights. All needed data adjustment happens in the weights reorder and transformation of input and output data.

[BUG1989]

I see. Thank you very much for your help.And the quantize of weights before or after winograd transformation ?

[vpirogov]

Quantization is independent on the Winograd algorithm itself. We apply the same quantization scheme and the same parameters for all convolution implementations, including direct, GEMM and Winograd.

[BUG1989]

Thank you very much for your advice. I have realized it in other opensource project.I have implemented the int8 winograd F(2,3) in arm platform and it has the same accuracy as original int8 conv3x3s1 : )
ncnn pr
P.S. : ncnn is a high-performance neural network inference framework optimized for the mobile platform

[xiaoweiChen]

@BUG1989 can show the opensource project name or address, which you find int8 winograd in it?

[BUG1989]

@xiaoweiChen
https://github.com/Tencent/ncnn/blob/master/src/layer/x86/convolution_3x3_int8.h

The detail codes,please see the two functions:

• conv3x3s1_winograd23_transform_kernel_int8_sse
• conv3x3s1_winograd23_int8_sse

[xiaoweiChen]

@BUG1989 Thanks for your sharing. :)