There was a bug I missed in the EXACT v0.25 client which will cause it to crash from over allocating memory. Please abort any WUs running with this version ASAP so they don't eat up all your machines memory. I really apologize, it ran correctly for my local tests but I had missed a call to convert one of the input filenames to it's proper location in the client.

That being said, v0.26 should be running fine. I made quite a few improvements to the code, so now it should require about 1/4th the memory as the v0.24 client (so the CIFAR work units will now take less memory that the MNIST work units from v0.24 and before). It's also running ~2x faster. That and I upgraded the back propagation algorithm to use batch normalization which also means we'll be getting better results from it. Pretty excited to see how it runs.

More...