Winograd Convolution in Pytorch

Has Winograd convolution been implemented in Pytorch? Is there a way that I can replace Conv2d with its Winograd implementation?
I want to compare the run times of Winograd and simple convolution.

PyTorch dispatches to multiple backends like CuDNN or MKLDNN which in turn use Winograd depending (mainly) on the kernel size.
PyTorch’s design is to abstract the implementation details and does not support targeted invocation of the Winograd variants.
The usual way to do such benchmarking would be to implement the functions you want to benchmark in a C++ extension.

Best regards


1 Like