Pytorch1.3 Quantization error in Resnet50

When applying quantization supported by pytorch 1.3 to Resnet 50, the following error occurs.

RuntimeError: Didn’t find engine for operation quantized::linear_prepack NoQEngine (operator () at …\aten\src\ATen\native\quantized\cpu\qlinear_prepack.cpp:202)
(no backtrace available)

Does anyone know how

Could you try set USE_FBGEMM=1 before running python install? If it doesn’t work, please file a Github issue.

I installed pytorch 1.3 using anconda,
and there seems to be no environment for python setup. I think it need additional library(FBGMM) settings, do you know how?

I think FBGEMM should be turned for machines/compilers that support it by default, can you print torch.backends.quantized.supported_engines to see if fbgemm is there?

Got the same error. torch.backends.quantized.supported_engines is showing None
I am running on windows. How to enable it on windows?

cc @hx89 @dskhudia do you know how to do this?

We haven’t yet tested it on windows.

@dskhudia this issue is resolved?
If yes, can you tell?


Hi Mohit,

I responded to the question here in detail:

It still needs some work from our side to make it work on windows.