update LRCN in models.py

added batch norm and L2 regularization, as well as glorot initialization, for much more robust training (at least in my local project)