For the purposes of performance and multi-GPU training, CNN has been studying cuda-convnet2 for a while.
Search, online incredibly a decent research Cuda-convnet2 code articles are not found, it seems that the holiday has been busy.
Caffe author Jiayanqing also expressed his admiration for Convnet2 author Alex in a number of occasions, showing the gap between the two CNN implementations.
Caffe more in line with popular tastes, while Convnet2 is in line with the pursuit of GPU enthusiasts.
Convnet2 code style is not as organized as Caffe.
Caffe Nature is single-threaded or CPU-thinking. Convnet2 is multi-threaded and belongs to the GPU thinking mode.
Caffe relies heavily on library functions (Glob, GFlags, Leveldb, Lmdb, Mkl/blas ...). ), and Convnet2 almost all of them.
The caffe parameter is set more freely, and the Convnet2 is constrained by many parameters for performance reasons.
Caffe approach the software, while the Convnet2 is close to the hardware.
Caffe for lazy people, convnet2 for geeks.
Comparison of Cuda-convnet2 and Caffe