Why PyTorch does not use cudnnPoolingForward/Backward?
It is only used in Caffe2.
Is there any problem when you switch from cuda to cudnn?
(It should be performance improvement generally when you switch from cuda to cudnn)
Why PyTorch does not use cudnnPoolingForward/Backward?
It is only used in Caffe2.
Is there any problem when you switch from cuda to cudnn?
(It should be performance improvement generally when you switch from cuda to cudnn)