I understand that using some hacks I could train xnor net, but what about efficient inference? Does pytorch support fair bitwise xnor-conv operations?
No bitwise convolutions are not supported.
But the ones in the paper have the input as floats and the weights as binary values?
Most element-wise bitwise operations are implemented for
ByteTensors such as and, or, nor, xor… You can do
dir(torch.ByteTensor()) to see what is there, most like
This is not what I want
Can I do Conv2d with torch.ByteTensor or torch.CharTensor?
No only float convolutions are supported.