I set two states, the eos token (zeros). And the excess of the max length (-1 s). And add a punishment on last state on the reward.

In order to include the mini_reverse_model. And changes provided by Antonio and Rene

I set two states, the eos token (zeros). And the excess of the max length (-1 s). And add a punishment on last state on the reward.

I modified the reward function using different parameters of the sacrebleu. And from a prior test, seems to be improved the learning.

Add the option to use attention as the state. Is already learning something, but in a very slow way....

Included the complete q-learn sequence. But the Qnet, still is not working(not improving scores). I write a TODO list on readme.

Step back with the implementation. Coded another idea, from extraction of state, a, r, state_, a basic idea of the reward(to be improved). Deletd the get_states...

Code a class QManager, to init model, data, and dqn parameters, all in DQN_loop.py, and added the option dqn_train, in model add the option to extract the atention vectors. To check fit as states.

On prediction.py, use the validate_on_data to create the funtion _states_data, wich return the states representation on the same way as the translation funtion