Minerva: A New Deep Learning Framework
DNN framework comparison
李沐: 我的理解是 caffe/cxxnet是适合需要用dnn来跑自己应用的人,准备好数据,使用或者稍微改下已有的网络配置就能跑。torch/minerva/theano 适合折腾dnn的人,可以很方便写个脚本跑起来。(purine还来得及看)。从实现技术上每个库都有自己的特点,这可以写长文了
yangqing: cudaconv偏hacky,caffe相对比较注重稳定,minerva不很熟悉,cxxnet和purine偏重lightweight + cutting edge,DIGITS是个wrapper所以和其他不是很一样。话说我超喜欢cxxnet的template和purine的operator,我自己的code也在用… 又话说caffe现在的interface的确不太能忍 :P
Alex Smola: See this short video in youtube Points:
- Never use Matlab Libraries for dnn
- Never start from scratch(unless you want to learn how it is done)
- Torch, theano is good but slow, caffe is kind of good, Minerva should be one of the bests now.
