fastest way to get off the ground with convnets

I've been building a neural network library as a learning tool for myself (so I'm not expecting to beat state-of-the-art systems but still want reasonable performance). I've gotten quite far with vanilla feed-forward nets but am now trying to implements conv-nets and am finding it quite hard to find decent resources.

Since I actually want to use this system for something interesting like image classification I'm looking for a GPU based backend and I'd like to reduce the dependencies I need. Currently I'm looking at incorporating cudarray but am having problems getting it to compile so I'd also like to consider alternatives. My main requirements are something reasonably fast (GPU based on nvidia GPU), with a python interface or simple enough that I could write a cython wrapper, and low level enough that it just does convolutions and maybe pooling (I don't want to call another NN library).

I've also found a general dirth of good CNN resources so would appreciate any tutorials links, etc... on implementing a CNN.

thanks!