garage.torch.policies¶

PyTorch Policies.

class CategoricalCNNPolicy(env, kernel_sizes, hidden_channels, strides=1, hidden_sizes=(32, 32), hidden_nonlinearity=torch.tanh, hidden_w_init=nn.init.xavier_uniform_, hidden_b_init=nn.init.zeros_, paddings=0, padding_mode='zeros', max_pool=False, pool_shape=None, pool_stride=1, output_nonlinearity=None, output_w_init=nn.init.xavier_uniform_, output_b_init=nn.init.zeros_, layer_normalization=False, name='CategoricalCNNPolicy')

CategoricalCNNPolicy.

A policy that contains a CNN and a MLP to make prediction based on a categorical distribution.

It only works with akro.Discrete action space.

Parameters
• env (garage.envs) – Environment.

• kernel_sizes (tuple[int]) – Dimension of the conv filters. For example, (3, 5) means there are two convolutional layers. The filter for first layer is of dimension (3 x 3) and the second one is of dimension (5 x 5).

• strides (tuple[int]) – The stride of the sliding window. For example, (1, 2) means there are two convolutional layers. The stride of the filter for first layer is 1 and that of the second layer is 2.

• hidden_channels (tuple[int]) – Number of output channels for CNN. For example, (3, 32) means there are two convolutional layers. The filter for the first conv layer outputs 3 channels

• hidden_sizes (list[int]) – Output dimension of dense layer(s) for the MLP for mean. For example, (32, 32) means the MLP consists of two hidden layers, each with 32 hidden units.

• hidden_nonlinearity (callable) – Activation function for intermediate dense layer(s). It should return a torch.Tensor. Set it to None to maintain a linear activation.

• hidden_w_init (callable) – Initializer function for the weight of intermediate dense layer(s). The function should return a torch.Tensor.

• hidden_b_init (callable) – Initializer function for the bias of intermediate dense layer(s). The function should return a torch.Tensor.

• paddings (tuple[int]) – Zero-padding added to both sides of the input

• padding_mode (str) – The type of padding algorithm to use, either ‘SAME’ or ‘VALID’.

• max_pool (bool) – Bool for using max-pooling or not.

• pool_shape (tuple[int]) – Dimension of the pooling layer(s). For example, (2, 2) means that all the pooling layers have shape (2, 2).

• pool_stride (tuple[int]) – The strides of the pooling layer(s). For example, (2, 2) means that all the pooling layers have strides (2, 2).

• output_nonlinearity (callable) – Activation function for output dense layer. It should return a torch.Tensor. Set it to None to maintain a linear activation.

• output_w_init (callable) – Initializer function for the weight of output dense layer(s). The function should return a torch.Tensor.

• output_b_init (callable) – Initializer function for the bias of output dense layer(s). The function should return a torch.Tensor.

• layer_normalization (bool) – Bool for using layer normalization or not.

• name (str) – Name of policy.

forward(self, observations)

Compute the action distributions from the observations.

Parameters

observations (torch.Tensor) – Batch of observations on default torch device.

Returns

Batch distribution of actions. dict[str, torch.Tensor]: Additional agent_info, as torch Tensors.

Do not need to be detached, and can be on any device.

Return type

torch.distributions.Distribution

get_action(self, observation)

Get a single action given an observation.

Parameters

observation (np.ndarray) – Observation from the environment. Shape is $$env_spec.observation_space$$.

Returns

• np.ndarray: Predicted action. Shape is

$$env_spec.action_space$$.

• dict:
• np.ndarray[float]: Mean of the distribution

• np.ndarray[float]: Standard deviation of logarithmic

values of the distribution.

Return type

tuple

get_actions(self, observations)

Get actions given observations.

Parameters

observations (np.ndarray) – Observations from the environment. Shape is $$batch_dim \bullet env_spec.observation_space$$.

Returns

• np.ndarray: Predicted actions.

$$batch_dim \bullet env_spec.action_space$$.

• dict:
• np.ndarray[float]: Mean of the distribution.

• np.ndarray[float]: Standard deviation of logarithmic

values of the distribution.

Return type

tuple

get_param_values(self)

Get the parameters to the policy.

This method is included to ensure consistency with TF policies.

Returns

The parameters (in the form of the state dictionary).

Return type

dict

set_param_values(self, state_dict)

Set the parameters to the policy.

This method is included to ensure consistency with TF policies.

Parameters

state_dict (dict) – State dictionary.

property name(self)

Name of policy.

Returns

Name of policy

Return type

str

reset(self, do_resets=None)

Reset the policy.

This is effective only to recurrent policies.

do_resets is an array of boolean indicating which internal states to be reset. The length of do_resets should be equal to the length of inputs, i.e. batch size.

Parameters

do_resets (numpy.ndarray) – Bool array indicating which states to be reset.

property env_spec(self)

Policy environment specification.

Returns

Environment specification.

Return type

garage.EnvSpec

property observation_space(self)

Observation space.

Returns

The observation space of the environment.

Return type

akro.Space

property action_space(self)

Action space.

Returns

The action space of the environment.

Return type

akro.Space

class ContextConditionedPolicy(latent_dim, context_encoder, policy, use_information_bottleneck, use_next_obs)

Bases: torch.nn.Module

A policy that outputs actions based on observation and latent context.

In PEARL, policies are conditioned on current state and a latent context (adaptation data) variable Z. This inference network estimates the posterior probability of z given past transitions. It uses context information stored in the encoder to infer the probabilistic value of z and samples from a policy conditioned on z.

Parameters
• latent_dim (int) – Latent context variable dimension.

• context_encoder (garage.torch.embeddings.ContextEncoder) – Recurrent or permutation-invariant context encoder.

• policy (garage.torch.policies.Policy) – Policy used to train the network.

• use_information_bottleneck (bool) – True if latent context is not deterministic; false otherwise.

• use_next_obs (bool) – True if next observation is used in context for distinguishing tasks; false otherwise.

Reset $$q(z \| c)$$ to the prior and sample a new z from the prior.

Parameters

sample_from_belief(self)

Sample z using distributions from current means and variances.

update_context(self, timestep)

Append single transition to the current context.

Parameters

timestep (garage._dtypes.TimeStep) – Timestep containing transition information to be added to context.

infer_posterior(self, context)

Compute $$q(z \| c)$$ as a function of input context and sample new z.

Parameters

context (torch.Tensor) – Context values, with shape $$(X, N, C)$$. X is the number of tasks. N is batch size. C is the combined size of observation, action, reward, and next observation if next observation is used in context. Otherwise, C is the combined size of observation, action, and reward.

forward(self, obs, context)

Given observations and context, get actions and probs from policy.

Parameters
• obs (torch.Tensor) –

Observation values, with shape $$(X, N, O)$$. X is the number of tasks. N is batch size. O

is the size of the flattened observation space.

• context (torch.Tensor) – Context values, with shape $$(X, N, C)$$. X is the number of tasks. N is batch size. C is the combined size of observation, action, reward, and next observation if next observation is used in context. Otherwise, C is the combined size of observation, action, and reward.

Returns

• torch.Tensor: Predicted action values.
• np.ndarray: Mean of distribution.

• np.ndarray: Log std of distribution.

• torch.Tensor: Log likelihood of distribution.

• torch.Tensor: Sampled values from distribution before

applying tanh transformation.

torch.Tensor: z values, with shape $$(N, L)$$. N is batch size.

L is the latent dimension.

Return type

tuple

get_action(self, obs)

Sample action from the policy, conditioned on the task embedding.

Parameters

obs (torch.Tensor) – Observation values, with shape $$(1, O)$$. O is the size of the flattened observation space.

Returns

Output action value, with shape $$(1, A)$$.

A is the size of the flattened action space.

dict:
• np.ndarray[float]: Mean of the distribution.

• np.ndarray[float]: Standard deviation of logarithmic values

of the distribution.

Return type

torch.Tensor

compute_kl_div(self)

Compute $$KL(q(z|c) \| p(z))$$.

Returns

$$KL(q(z|c) \| p(z))$$.

Return type

float

property networks(self)

Return context_encoder and policy.

Returns

Encoder and policy networks.

Return type

list

property context(self)

Return context.

Returns

Context values, with shape $$(X, N, C)$$.

X is the number of tasks. N is batch size. C is the combined size of observation, action, reward, and next observation if next observation is used in context. Otherwise, C is the combined size of observation, action, and reward.

Return type

torch.Tensor

class DeterministicMLPPolicy(env_spec, name='DeterministicMLPPolicy', **kwargs)

Implements a deterministic policy network.

The policy network selects action based on the state of the environment. It uses a PyTorch neural network module to fit the function of pi(s).

forward(self, observations)

Compute actions from the observations.

Parameters

observations (torch.Tensor) – Batch of observations on default torch device.

Returns

Batch of actions.

Return type

torch.Tensor

get_action(self, observation)

Get a single action given an observation.

Parameters

observation (np.ndarray) – Observation from the environment.

Returns

• np.ndarray: Predicted action.

• dict:
• np.ndarray[float]: Mean of the distribution

• np.ndarray[float]: Log of standard deviation of the

distribution

Return type

tuple

get_actions(self, observations)

Get actions given observations.

Parameters

observations (np.ndarray) – Observations from the environment.

Returns

• np.ndarray: Predicted actions.

• dict:
• np.ndarray[float]: Mean of the distribution

• np.ndarray[float]: Log of standard deviation of the

distribution

Return type

tuple

get_param_values(self)

Get the parameters to the policy.

This method is included to ensure consistency with TF policies.

Returns

The parameters (in the form of the state dictionary).

Return type

dict

set_param_values(self, state_dict)

Set the parameters to the policy.

This method is included to ensure consistency with TF policies.

Parameters

state_dict (dict) – State dictionary.

property name(self)

Name of policy.

Returns

Name of policy

Return type

str

reset(self, do_resets=None)

Reset the policy.

This is effective only to recurrent policies.

do_resets is an array of boolean indicating which internal states to be reset. The length of do_resets should be equal to the length of inputs, i.e. batch size.

Parameters

do_resets (numpy.ndarray) – Bool array indicating which states to be reset.

property env_spec(self)

Policy environment specification.

Returns

Environment specification.

Return type

garage.EnvSpec

property observation_space(self)

Observation space.

Returns

The observation space of the environment.

Return type

akro.Space

property action_space(self)

Action space.

Returns

The action space of the environment.

Return type

akro.Space

class DiscreteQFArgmaxPolicy(qf, env_spec, name='DiscreteQFArgmaxPolicy')

Policy that derives its actions from a learned Q function.

The action returned is the one that yields the highest Q value for a given state, as determined by the supplied Q function.

Parameters
• qf (object) – Q network.

• env_spec (EnvSpec) – Environment specification.

• name (str) – Name of this policy.

forward(self, observations)

Get actions corresponding to a batch of observations.

Parameters

observations (torch.Tensor) – Batch of observations of shape $$(N, O)$$. Observations should be flattened even if they are images as the underlying Q network handles unflattening.

Returns

Batch of actions of shape $$(N, A)$$

Return type

torch.Tensor

get_action(self, observation)

Get a single action given an observation.

Parameters

observation (np.ndarray) – Observation with shape $$(O, )$$.

Returns

Predicted action with shape $$(A, )$$. dict: Empty since this policy does not produce a distribution.

Return type

torch.Tensor

get_actions(self, observations)

Get actions given observations.

Parameters

observations (np.ndarray) – Batch of observations, should have shape $$(N, O)$$.

Returns

Predicted actions. Tensor has shape $$(N, A)$$. dict: Empty since this policy does not produce a distribution.

Return type

torch.Tensor

get_param_values(self)

Get the parameters to the policy.

This method is included to ensure consistency with TF policies.

Returns

The parameters (in the form of the state dictionary).

Return type

dict

set_param_values(self, state_dict)

Set the parameters to the policy.

This method is included to ensure consistency with TF policies.

Parameters

state_dict (dict) – State dictionary.

property name(self)

Name of policy.

Returns

Name of policy

Return type

str

reset(self, do_resets=None)

Reset the policy.

This is effective only to recurrent policies.

do_resets is an array of boolean indicating which internal states to be reset. The length of do_resets should be equal to the length of inputs, i.e. batch size.

Parameters

do_resets (numpy.ndarray) – Bool array indicating which states to be reset.

property env_spec(self)

Policy environment specification.

Returns

Environment specification.

Return type

garage.EnvSpec

property observation_space(self)

Observation space.

Returns

The observation space of the environment.

Return type

akro.Space

property action_space(self)

Action space.

Returns

The action space of the environment.

Return type

akro.Space

class GaussianMLPPolicy(env_spec, hidden_sizes=(32, 32), hidden_nonlinearity=torch.tanh, hidden_w_init=nn.init.xavier_uniform_, hidden_b_init=nn.init.zeros_, output_nonlinearity=None, output_w_init=nn.init.xavier_uniform_, output_b_init=nn.init.zeros_, learn_std=True, init_std=1.0, min_std=1e-06, max_std=None, std_parameterization='exp', layer_normalization=False, name='GaussianMLPPolicy')

MLP whose outputs are fed into a Normal distribution..

A policy that contains a MLP to make prediction based on a gaussian distribution.

Parameters
• env_spec (EnvSpec) – Environment specification.

• hidden_sizes (list[int]) – Output dimension of dense layer(s) for the MLP for mean. For example, (32, 32) means the MLP consists of two hidden layers, each with 32 hidden units.

• hidden_nonlinearity (callable) – Activation function for intermediate dense layer(s). It should return a torch.Tensor. Set it to None to maintain a linear activation.

• hidden_w_init (callable) – Initializer function for the weight of intermediate dense layer(s). The function should return a torch.Tensor.

• hidden_b_init (callable) – Initializer function for the bias of intermediate dense layer(s). The function should return a torch.Tensor.

• output_nonlinearity (callable) – Activation function for output dense layer. It should return a torch.Tensor. Set it to None to maintain a linear activation.

• output_w_init (callable) – Initializer function for the weight of output dense layer(s). The function should return a torch.Tensor.

• output_b_init (callable) – Initializer function for the bias of output dense layer(s). The function should return a torch.Tensor.

• learn_std (bool) – Is std trainable.

• init_std (float) – Initial value for std. (plain value - not log or exponentiated).

• min_std (float) – Minimum value for std.

• max_std (float) – Maximum value for std.

• std_parameterization (str) –

How the std should be parametrized. There are two options: - exp: the logarithm of the std will be stored, and applied a

exponential transformation

• softplus: the std will be computed as log(1+exp(x))

• layer_normalization (bool) – Bool for using layer normalization or not.

• name (str) – Name of policy.

forward(self, observations)

Compute the action distributions from the observations.

Parameters

observations (torch.Tensor) – Batch of observations on default torch device.

Returns

Batch distribution of actions. dict[str, torch.Tensor]: Additional agent_info, as torch Tensors

Return type

torch.distributions.Distribution

get_action(self, observation)

Get a single action given an observation.

Parameters

observation (np.ndarray) – Observation from the environment. Shape is $$env_spec.observation_space$$.

Returns

• np.ndarray: Predicted action. Shape is

$$env_spec.action_space$$.

• dict:
• np.ndarray[float]: Mean of the distribution

• np.ndarray[float]: Standard deviation of logarithmic

values of the distribution.

Return type

tuple

get_actions(self, observations)

Get actions given observations.

Parameters

observations (np.ndarray) – Observations from the environment. Shape is $$batch_dim \bullet env_spec.observation_space$$.

Returns

• np.ndarray: Predicted actions.

$$batch_dim \bullet env_spec.action_space$$.

• dict:
• np.ndarray[float]: Mean of the distribution.

• np.ndarray[float]: Standard deviation of logarithmic

values of the distribution.

Return type

tuple

get_param_values(self)

Get the parameters to the policy.

This method is included to ensure consistency with TF policies.

Returns

The parameters (in the form of the state dictionary).

Return type

dict

set_param_values(self, state_dict)

Set the parameters to the policy.

This method is included to ensure consistency with TF policies.

Parameters

state_dict (dict) – State dictionary.

property name(self)

Name of policy.

Returns

Name of policy

Return type

str

reset(self, do_resets=None)

Reset the policy.

This is effective only to recurrent policies.

do_resets is an array of boolean indicating which internal states to be reset. The length of do_resets should be equal to the length of inputs, i.e. batch size.

Parameters

do_resets (numpy.ndarray) – Bool array indicating which states to be reset.

property env_spec(self)

Policy environment specification.

Returns

Environment specification.

Return type

garage.EnvSpec

property observation_space(self)

Observation space.

Returns

The observation space of the environment.

Return type

akro.Space

property action_space(self)

Action space.

Returns

The action space of the environment.

Return type

akro.Space

class Policy(env_spec, name)

Bases: torch.nn.Module, garage.np.policies.Policy, abc.ABC

Policy base class.

Parameters
• env_spec (EnvSpec) – Environment specification.

• name (str) – Name of policy.

abstract get_action(self, observation)

Get action sampled from the policy.

Parameters

observation (np.ndarray) – Observation from the environment.

Returns

Action and extra agent

info.

Return type

Tuple[np.ndarray, dict[str,np.ndarray]]

abstract get_actions(self, observations)

Get actions given observations.

Parameters

observations (np.ndarray) – Observations from the environment.

Returns

Actions and extra agent

infos.

Return type

Tuple[np.ndarray, dict[str,np.ndarray]]

get_param_values(self)

Get the parameters to the policy.

This method is included to ensure consistency with TF policies.

Returns

The parameters (in the form of the state dictionary).

Return type

dict

set_param_values(self, state_dict)

Set the parameters to the policy.

This method is included to ensure consistency with TF policies.

Parameters

state_dict (dict) – State dictionary.

property name(self)

Name of policy.

Returns

Name of policy

Return type

str

reset(self, do_resets=None)

Reset the policy.

This is effective only to recurrent policies.

do_resets is an array of boolean indicating which internal states to be reset. The length of do_resets should be equal to the length of inputs, i.e. batch size.

Parameters

do_resets (numpy.ndarray) – Bool array indicating which states to be reset.

property env_spec(self)

Policy environment specification.

Returns

Environment specification.

Return type

garage.EnvSpec

property observation_space(self)

Observation space.

Returns

The observation space of the environment.

Return type

akro.Space

property action_space(self)

Action space.

Returns

The action space of the environment.

Return type

akro.Space

class TanhGaussianMLPPolicy(env_spec, hidden_sizes=(32, 32), hidden_nonlinearity=nn.ReLU, hidden_w_init=nn.init.xavier_uniform_, hidden_b_init=nn.init.zeros_, output_nonlinearity=None, output_w_init=nn.init.xavier_uniform_, output_b_init=nn.init.zeros_, init_std=1.0, min_std=np.exp(- 20.0), max_std=np.exp(2.0), std_parameterization='exp', layer_normalization=False)

Multiheaded MLP whose outputs are fed into a TanhNormal distribution.

A policy that contains a MLP to make prediction based on a gaussian distribution with a tanh transformation.

Parameters
• env_spec (EnvSpec) – Environment specification.

• hidden_sizes (list[int]) – Output dimension of dense layer(s) for the MLP for mean. For example, (32, 32) means the MLP consists of two hidden layers, each with 32 hidden units.

• hidden_nonlinearity (callable) – Activation function for intermediate dense layer(s). It should return a torch.Tensor. Set it to None to maintain a linear activation.

• hidden_w_init (callable) – Initializer function for the weight of intermediate dense layer(s). The function should return a torch.Tensor.

• hidden_b_init (callable) – Initializer function for the bias of intermediate dense layer(s). The function should return a torch.Tensor.

• output_nonlinearity (callable) – Activation function for output dense layer. It should return a torch.Tensor. Set it to None to maintain a linear activation.

• output_w_init (callable) – Initializer function for the weight of output dense layer(s). The function should return a torch.Tensor.

• output_b_init (callable) – Initializer function for the bias of output dense layer(s). The function should return a torch.Tensor.

• init_std (float) – Initial value for std. (plain value - not log or exponentiated).

• min_std (float) – If not None, the std is at least the value of min_std, to avoid numerical issues (plain value - not log or exponentiated).

• max_std (float) – If not None, the std is at most the value of max_std, to avoid numerical issues (plain value - not log or exponentiated).

• std_parameterization (str) –

How the std should be parametrized. There are two options: - exp: the logarithm of the std will be stored, and applied a

exponential transformation

• softplus: the std will be computed as log(1+exp(x))

• layer_normalization (bool) – Bool for using layer normalization or not.

forward(self, observations)

Compute the action distributions from the observations.

Parameters

observations (torch.Tensor) – Batch of observations on default torch device.

Returns

Batch distribution of actions. dict[str, torch.Tensor]: Additional agent_info, as torch Tensors

Return type

torch.distributions.Distribution

get_action(self, observation)

Get a single action given an observation.

Parameters

observation (np.ndarray) – Observation from the environment. Shape is $$env_spec.observation_space$$.

Returns

• np.ndarray: Predicted action. Shape is

$$env_spec.action_space$$.

• dict:
• np.ndarray[float]: Mean of the distribution

• np.ndarray[float]: Standard deviation of logarithmic

values of the distribution.

Return type

tuple

get_actions(self, observations)

Get actions given observations.

Parameters

observations (np.ndarray) – Observations from the environment. Shape is $$batch_dim \bullet env_spec.observation_space$$.

Returns

• np.ndarray: Predicted actions.

$$batch_dim \bullet env_spec.action_space$$.

• dict:
• np.ndarray[float]: Mean of the distribution.

• np.ndarray[float]: Standard deviation of logarithmic

values of the distribution.

Return type

tuple

get_param_values(self)

Get the parameters to the policy.

This method is included to ensure consistency with TF policies.

Returns

The parameters (in the form of the state dictionary).

Return type

dict

set_param_values(self, state_dict)

Set the parameters to the policy.

This method is included to ensure consistency with TF policies.

Parameters

state_dict (dict) – State dictionary.

property name(self)

Name of policy.

Returns

Name of policy

Return type

str

reset(self, do_resets=None)

Reset the policy.

This is effective only to recurrent policies.

do_resets is an array of boolean indicating which internal states to be reset. The length of do_resets should be equal to the length of inputs, i.e. batch size.

Parameters

do_resets (numpy.ndarray) – Bool array indicating which states to be reset.

property env_spec(self)

Policy environment specification.

Returns

Environment specification.

Return type

garage.EnvSpec

property observation_space(self)

Observation space.

Returns

The observation space of the environment.

Return type

akro.Space

property action_space(self)

Action space.

Returns

The action space of the environment.

Return type

akro.Space