Yes, if you are going with exposed/wrapped Java APIs there is all kinds of stuff. In particular, if you are going that route, probably clj-ml for weka and dl4clj for deeplearning4j would be on the short list

I was just looking into this. @joelkuiper as about Keras specifically … so I’d say the closest thing I found is Cortex. Yes, there are java bindings to tensorflow, mxnet, etc. but they tend to be at a lower ‘computation graph’ level than Keras is. And the core.matrix, neandertal options are great for matrix math (on and off the GPU) but are at an even lower level still.

I’d love for there to be a Keras style api/library/dsl for Clojure on top of the java tensorflow bindings. That would rock.

BTW the new CUDA backend for neanderthal is coming up nicely: https://twitter.com/draganrocks/status/862063626001285120

it is almost at the theoretical limit for that GPU thanks to nvidia's assembly-level hardware-specific optimizations (transparent to the user)

compared to the open-source OpenCL engine, it runs in 141 ms vs 220 ms for that specific size of matrices a, b, and c...