Ok, Windows 10 wasn't that hard. I wish there were GPU instructions - it's straightforward and you don't need Cygwin or AWS if you have a recent Nvidia GPU.
Well, it wouldn't have been hard if I'd had a complete set of instructions. 7 hours later, I got it to finish the cat/dog dataset with GPU acceleration. Anaconda install was straightforward, and conda worked ok for adding dependencies. The instructions on their website for theano and karas worked ok. There were 2 different bugs that broke it - it uses hugely long path names, which break in windows, and I had to downgrade keras to 1.2.2 or it wouldn't work.
I had to guess it wanted Python 2.7, was correct.
Then, naturally, you have to create an account to be allowed to download the neural net CUDA stuff, and I had to guess that MSVC version 2013 was not too new but not too old that it would work. (MSVC has abysmal compatibility forward and back, you basically need the exact correct version of the compiler) And I had to add it to the path or it wouldn't work.
Then I find out the data set download is long gone, and I had to make up the data set. I found out that Windows 10 has some horrific optimization and even with an SSD and a machine with 32 gb of RAM, it took a long time to create two folders with dogs in one and cats in the other. Maybe I should have used a batch file or written a script for it.
After all that, I'm getting ~300 second runtimes. Nothing spectacular. I've noticed that my GPU seems to average only about 60% load when processing the dataset, so maybe there's settings that can be tweaked.
NOTE : NO Cygwin. NO Bash for Windows. Both are counterproductive and useless! (I wasted time installing Cygwin, then deleted it, and Bash for windows is just going to put a bunch of native linux stuff on your machine that will NOT work with the GPU acceleration so it's a waste of time! About 3 of the 7 hours are wasted because both were suggested above...)