perceptron weight averaging is not standard

My understanding is that for the averaged perceptron, the weights are averaged across each example in each iteration, not just each iteration as in the current implementation (and the perl one here from @sagae: http://people.ict.usc.edu/~sagae/parser/simpledep.pl).

A nice reference with a clever efficiency trick:
http://www.ciml.info/dl/v0_8/ciml-v0_8-ch03.pdf