Thanks for your exiting work. I have a small question. In the implementation of cuda, there seems the winograd is fixed at 4x4, does it means we can only use cylinder convolution with kernel size 3x3. Is it possible to use other kernel size, such as 5x5, 7x7, etc?
https://github.com/mcrl/CyCNN/blob/master/cycnn-extension/cycnn_cuda.cu#L813
Thanks for your exiting work. I have a small question. In the implementation of cuda, there seems the winograd is fixed at 4x4, does it means we can only use cylinder convolution with kernel size 3x3. Is it possible to use other kernel size, such as 5x5, 7x7, etc?
https://github.com/mcrl/CyCNN/blob/master/cycnn-extension/cycnn_cuda.cu#L813