Recently, I have been comparing some CNNs in terms of time, #Multiple Add (MAC) operations, #parameters and model size. I saw some similar SO questions ( here and here ), and in the latter they suggest using the Netscope CNN Analyzer . This tool allows me to calculate most of the things I need by simply entering the definition of my Caffe network.
However, the number of re-adding operations of some architectures that I have seen in documents and on the Internet does not match what Netscope displays, while other architectures do. I always compare FLOP or MAC with the MACC column in netscope, but there is a ~ 10x factor that I forget at some point (see table below for more details).
Architecture ---- MAC (paper/internet) ---- macc column in netscope VGG 16 ~15.5G ~157G GoogLeNet ~1.55G ~16G
Link to GoogLeNet macro number and VGG16 macc number in Netscope.
Could anyone who used this tool point me to the mistake I am making while reading the Netscope output?
deep-learning caffe flops conv-neural-network
rafaspadilha
source share