Summary:
- Delete SingleQValue and AllActionQValues; returns tensor directly
- CategoricalDQN is now taking a ModelBase which returns 3-D tensor (batch_size, num_actions, num_atoms)
- FullyConnectedDQN's constructor has one more argument, num_atoms, which default to None. When it's not None, returns 3-D tensor like above
- DuelingQNetwork takes 3 sub networks: shared_network, value_network, advantage_network. The parametric version is split out to ParametricDuelingQNetwork for better readability.
- Introduces a subclass of Sequential conforming to ModelBase API. `input_prototype()` returns the value from the first module.
Reviewed By: czxttkl
Differential Revision: D21336534
fbshipit-source-id: 629c688b0de4754effb30cd61f4a6da3cca021ab