GitLab

The moving average object was shared between training and evaluation. As
a result, the cost during evaluation was averaged with the training
cost. This affects only objective, not precision.

Uses more NumPy now :)