Update on Roadmap

Just an update on where Keras is going in the short term.

Here are the goals for the next release:

• More unit test coverage. @phreeza and @tleeuwenburg have been making significant progress on this front.
• Caffe model importing. @pranv has a PR for this, hopefully we can merge it asap.
• Support for automatic architecture/hyperparam search. I am working on it.
• Stateful recurrent layers for efficient sequence generation. See: https://github.com/fchollet/keras/issues/98
• TimeDistributed layer to make it possible to include any layer type as part of a recurrent architecture.

Let me know if I’m forgetting anything important.

Regarding visualization tools: these will be developed as a standalone app separate from Keras, in order to be usable with other deep learning frameworks as well. The project is still essentially at the ideation phase. Hulaos is a first concrete step.

[scott-gray]

I’d hold out another week or so for the fully refactored neon release. The documentation should be fairly complete.

I think I’m covering most everything that theano can do in my backend. The new conv kernels use a dimshuffle on the filter to speed things up. I still need to generalize that code for full transpose(axes=()) support, but that won’t be too hard.

I also want to fully support any dimension of tensor in reduction and broadcast operations (ew already works) but that might take me a bit of time to get to. Right now those things just work with 1 or 2.

At the moment, I’m just fleshing out all my conv kernel tiling sizes so performance will be good for a wide range of params. Trying to make it as hard as possible for the FFT guys to overtake me… but I suppose it might be inevitable. Then again, smaller parallel layer configs (ala googlenet) seem to be the fashion now. It’s hard to see FFT catching up there. There’s not enough dimensionality to hide your latencies and the filters are all small.

In another overlooked area, my gemm kernels are killing cublas on RNN networks. The new neon will have lots of examples of those with benchmarks.

We’re also playing with nvrtc to see if we can’t speed up pycuda compilation times some, though that overhead is already pretty minimal. You generally have a few seconds to wait the very first time you run something, but after that everything runs instantly. Autodif works at runtime with no overhead (after being cached).

[scott-gray]

It’s pretty high on my list of priorities to finish an embeddings lookup table kernel for bprop. This allows you to sum the gradient over unique words. I should be able to avoid atomics and do it deterministicly. Hope to have that sometime this week. I believe that’s the only major thing we’re missing for RNNs.