ML on the Mind

I made fun of PyBrain a while back.  Found myself using it again, still dissatisfied but reluctantly putting stuff together.  If you’ve got a simple ML problem and can’t be arsed to hash together a backprop learner, it does serve it’s purpose.  I’m still peeved, however, because it’s slow as balls.  I’m aware that backprop isn’t the fastest learning algorithm, but I would have expected a little bit of a speed bump.  I wonder if they’re accelerating with numpy, of if they’re doing everything sequentially.

I have a 256 input visible layer, a 512 unit hidden layer, a 128 unit layer, and a 256 unit output layer.  With 1000 training examples, each epoch is taking 20-30 minutes.  :\