Sporadic dug pit several, did not fill the soil, is too much of the outstanding, gossip less say, or more records summed up. Today's theme is around convolution and accelerating
Remember to read lecun their group of an article, is the FFT accelerated convolution. According to convolution theorem, the convolution on the time domain can be transformed into the Fourier transform of the spatial domain.
LeCun's article is to accelerate by turning convolution into Fourier transform. As seen from the experiment, the speedup is about twice times higher. This part of the code is open source, but it does not seem to merge into Caffe, perhaps because of the limited speed of acceleration, but also to consume space. Of course, I think the main problem is to accelerate the ratio, because of the acceleration process, because of its principle, when the convolution core is small, there is no acceleration, when the nucleus is 3 or 5 o'clock, the speed of some slower or equivalent, and in the CNN convolution of the core is relatively small, the acceleration effect is small, and based on the image processing itself, Convolutional nuclei are generally not too big, googlenet with 7x7 is already exploding days. On the other hand, for Caffe to achieve multi-GPU card acceleration or multi-machine acceleration, it is real deal acceleration, no matter how big your convolution core, you can accelerate.
Cond......
Convolution,fft, speed up.