CUDA-based neural networks in Python

I have spent the last couple of weeks coding on two projects: CUDArray is a CUDA-based subset of NumPy and deeppy is a neural network framework built on top of CUDArray.

While there is still a lot to be done on both projects, I think they have reached a point where they start to look promising. Highlights include:

• Since CUDArray imitates NumPy, its user interface is well-known to NumPy-programmers.
• CUDArray has CPU fall-back when no GPU is available.
• CUDArray relies on cuBLAS, cuRAND and cuDNN for fast array operations.
• deeppy has a pythonic user interface (that is, no YAML configuration files).
• In terms of speed, CUDArray/deeppy is very competitive with other popular neural network frameworks. At least, that's my experience so far.

Be sure to check out some deeppy examples:

Multi-layer perceptron with dropout on MNIST

Convnet on CIFAR-10