OK, so I have to admit that this is actually not the actual replication of the Google DeepMind paper. This is only a toy LSTM network written in Python with Theano, related to some sort of trajectory prediction.
ContNet2.py and DiscNet2.py are directly modified from DynNet.py. All these codes are pretty messy and therefore not for serious business.