Multi-Dimensional Metric

[dolevshapira]

Hi,
I wanted to create a batch-wise confusion matrix and sum it over using a callback, but I figured that it is not possible to create a multi-dimensional value passed as a metric after getting a little more familiar with the source code.
I think it is a very useful feature, so I wondered how to implement it.
As someone who's understanding of Keras internals is very shallow, I thought to ask for guidelines (Unless someone else would like to do it themselves)
would it be possible to implement this feature using return score_array instead of this line in _weighted_masked_objective and changing the basic metrics to do K.mean before returning?