def compute_actions(self, observation, add_noise=False, update=True):
observation = self.preprocessor.transform(observation)
observation = self.observation_filter(observation[None], update=update)
action = self.sess.run(self.sampler,
feed_dict={self.inputs: observation})
action = unbatch_actions(action)
if add_noise and isinstance(self.action_space, gym.spaces.Box):
action += np.random.randn(*action.shape) * self.action_noise_std
return action
def compute_actions(self, observation, add_noise=False, update=True):
observation = self.preprocessor.transform(observation)
observation = self.observation_filter(observation[None], update=update)
# `actions` is a list of (component) batches.
actions = self.sess.run(self.sampler,
feed_dict={self.inputs: observation})
if add_noise:
actions = tree.map_structure(self._add_noise, actions,
self.action_space_struct)
# Convert `flat_actions` to a list of lists of action components
# (list of single actions).
actions = unbatch_actions(actions)
return actions
def _compute_actions(policy, obs_batch, add_noise=False, update=True):
observation = policy.preprocessor.transform(obs_batch)
observation = policy.observation_filter(
observation[None], update=update)
observation = convert_to_torch_tensor(observation)
dist_inputs, _ = policy.model({
SampleBatch.CUR_OBS: observation
}, [], None)
dist = policy.dist_class(dist_inputs, policy.model)
action = dist.sample().detach().numpy()
action = unbatch_actions(action)
if add_noise and isinstance(policy.action_space, gym.spaces.Box):
action += np.random.randn(*action.shape) * policy.action_noise_std
return action