v0.3 -- Second DCNN

Trained net after 10 epochs. This network is a larger DCNN, trained on 20,000 19x19 games from KGS. Should provide a half-decent game already, assuming I can get the web-app code to work.

v0.2 -- First nets

Just getting started here! Here's checkpoints for 4 & 5 epochs of training on 5,000 7-dan+ ranked games on KGS. Final validation accuracy was 29.43%. Sounds rubbish, except the chance of randomly predicting the right move on a 361-point Go board is like 0.3%, so it's not too bad really. Still a long way to go of course!

This was trained using SGD on a small DCNN with a simple one-plane input encoder. Will try again with the 11-plane encoder + Adagrad for comparison's sake.