hive.agents.qnets.qnet_heads module

class hive.agents.qnets.qnet_heads.DQNNetwork(base_network, hidden_dim, out_dim, linear_fn=None)

Bases: Module

Implements the standard DQN value computation. Transforms output from base_network with output dimension hidden_dim to dimension out_dim, which should be equal to the number of actions.

Parameters

• base_network (torch.nn.Module) – Backbone network that computes the representations that are used to compute action values.

• hidden_dim (int) – Dimension of the output of the network.

• out_dim (int) – Output dimension of the DQN. Should be equal to the number of actions that you are computing values for.

• linear_fn (torch.nn.Module) – Function that will create the torch.nn.Module that will take the output of network and produce the final action values. If None, a torch.nn.Linear layer will be used.

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class hive.agents.qnets.qnet_heads.DuelingNetwork(base_network, hidden_dim, out_dim, linear_fn=None, atoms=1)

Bases: Module

Computes action values using Dueling Networks (https://arxiv.org/abs/1511.06581). In dueling, we have two heads—one for estimating advantage function and one for estimating value function.

Parameters

• base_network (torch.nn.Module) – Backbone network that computes the representations that are shared by the two estimators.

• hidden_dim (int) – Dimension of the output of the base_network.

• out_dim (int) – Output dimension of the Dueling DQN. Should be equal to the number of actions that you are computing values for.

• linear_fn (torch.nn.Module) – Function that will create the torch.nn.Module that will take the output of network and produce the final action values. If None, a torch.nn.Linear layer will be used.

• atoms (int) – Multiplier for the dimension of the output. For standard dueling networks, this should be 1. Used by DistributionalNetwork.

init_networks()

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class hive.agents.qnets.qnet_heads.DistributionalNetwork(base_network, out_dim, vmin=0, vmax=200, atoms=51)

Bases: Module

Computes a categorical distribution over values for each action (https://arxiv.org/abs/1707.06887).

Parameters

• base_network (torch.nn.Module) – Backbone network that computes the representations that are used to compute the value distribution.

• out_dim (int) – Output dimension of the Distributional DQN. Should be equal to the number of actions that you are computing values for.

• vmin (float) – The minimum of the support of the categorical value distribution.

• vmax (float) – The maximum of the support of the categorical value distribution.

• atoms (int) – Number of atoms discretizing the support range of the categorical value distribution.

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

dist(x)

Computes a categorical distribution over values for each action.

training: bool