I'm a little out of sync with machine learning but this looks really awesome =) I do hope as well, that we move forward with a unified array interface in JuliaGPU!
Am Samstag, 28. Februar 2015 16:19:18 UTC+1 schrieb Deniz Yuret: > > KUnet.jl <https://github.com/denizyuret/KUnet.jl> (beginning deep > learning with 500 lines of Julia > <http://www.denizyuret.com/2015/02/beginning-deep-learning-with-500-lines.html> > ) is out with its alpha release. Only the basic functionality is in > place (i.e. backprop with relu, softmax, sgd, momentum, nesterov, adagrad, > dropout, l1-l2 regularization etc.) but the GPU functionality is in, its > speed is competitive with Caffe <http://caffe.berkeleyvision.org/>, and I > think convolutional and recurrent nets can be added without too much > effort. I wrote a blog post > <http://www.denizyuret.com/2015/02/beginning-deep-learning-with-500-lines.html> > about > the code structure and there is some basic documentation > <https://github.com/denizyuret/KUnet.jl/blob/master/README.md>. You can > send me suggestions for improvement (both in coding style and new > functionality) using comments > <http://www.blogger.com/comment.g?blogID=8540876&postID=328231440874481473> > to > the blog post > <http://www.denizyuret.com/2015/02/beginning-deep-learning-with-500-lines.html>, > > or using issues <https://github.com/denizyuret/KUnet.jl/issues> or pull > requests <https://help.github.com/articles/fork-a-repo/> on GitHub > <https://github.com/denizyuret/KUnet.jl>. > > I tried to make the code (cpu/gpu) generic and high level. Getting the > same code working on the GPU and the CPU in Julia proved to be a bit > challenging and showed that both a more standard treatment of CPU and GPU > arrays, and a standard syntax for in-place operations would be welcome > additions to the language. I'd like to thank Tim Holy (CUDArt), Nick > Henderson (CUBLAS), and Simon Byrne (InplaceOps) for their generous help. > > best, > deniz >
