Open John1231983 opened 4 years ago
Thanks for interesting repo. I am investigating your method on resnet101 and found that number of flop of your conv bigger than traditional conv three time with same setting. Any wrong for my observation
There is no extra computation nor extra memory usage introduced in Rectified Convolution.
Only the boundaries are rectified. Please see the technical report https://hangzhang.org/RectifiedConv/RectifiedConv.pdf
Thanks for interesting repo. I am investigating your method on resnet101 and found that number of flop of your conv bigger than traditional conv three time with same setting. Any wrong for my observation