def obs_to_tensor(
self, observation: Union[np.ndarray, Dict[str, np.ndarray]]
) -> Tuple[th.Tensor, bool]:
"""
Convert an input observation to a PyTorch tensor that can be fed to a model.
Includes sugar-coating to handle different observations (e.g. normalizing images).
:param observation: the input observation
:return: The observation as PyTorch tensor
and whether the observation is vectorized or not
"""
vectorized_env = False
if isinstance(observation, dict):
# need to copy the dict as the dict in VecFrameStack will become a torch tensor
observation = copy.deepcopy(observation)
for key, obs in observation.items():
obs_space = self.observation_space.spaces[key]
if is_image_space(obs_space):
obs_ = maybe_transpose(obs, obs_space)
else:
obs_ = np.array(obs)
vectorized_env = vectorized_env or is_vectorized_observation(
obs_, obs_space)
# Add batch dimension if needed
if key != "scene_graph":
observation[key] = obs_.reshape(
(-1, ) + self.observation_space[key].shape)
else:
observation[key] = obs_
elif is_image_space(self.observation_space):
# Handle the different cases for images
# as PyTorch use channel first format
observation = maybe_transpose(observation, self.observation_space)
else:
observation = np.array(observation)
if not isinstance(observation, dict):
# Dict obs need to be handled separately
vectorized_env = is_vectorized_observation(observation,
self.observation_space)
# Add batch dimension if needed
observation = observation.reshape((-1, ) +
self.observation_space.shape)
observation = obs_as_tensor(observation, self.device)
return observation, vectorized_env
def predict(self,
observation: np.ndarray,
state: Optional[np.ndarray] = None,
mask: Optional[np.ndarray] = None,
deterministic: bool = False,
use_behav=None) -> Tuple[np.ndarray, Optional[np.ndarray]]:
"""
Overrides the base_class predict function to include epsilon-greedy exploration.
:param observation: the input observation
:param state: The last states (can be None, used in recurrent policies)
:param mask: The last masks (can be None, used in recurrent policies)
:param deterministic: Whether or not to return deterministic actions.
:return: the model's action and the next state
(used in recurrent policies)
"""
if not deterministic and np.random.rand() < self.exploration_rate:
if is_vectorized_observation(
maybe_transpose(observation, self.observation_space),
self.observation_space):
n_batch = observation.shape[0]
action = np.array(
[self.action_space.sample() for _ in range(n_batch)])
else:
action = np.array(self.action_space.sample())
else:
action, state = self.policy.predict(observation, state, mask,
deterministic)
return action, state
def predict(self, observation: np.ndarray,
state: Optional[np.ndarray] = None,
mask: Optional[np.ndarray] = None,
deterministic: bool = False) -> Tuple[np.ndarray, Optional[np.ndarray]]:
"""
Overrides the base_class predict function to include epsilon-greedy exploration.
:param observation: (np.ndarray) the input observation
:param state: (Optional[np.ndarray]) The last states (can be None, used in recurrent policies)
:param mask: (Optional[np.ndarray]) The last masks (can be None, used in recurrent policies)
:param deterministic: (bool) Whether or not to return deterministic actions.
:return: (Tuple[np.ndarray, Optional[np.ndarray]]) the model's action and the next state
(used in recurrent policies), 'is_random_action'(0 or 1) indicates if the action taken was random or not
"""
if not deterministic and np.random.rand() < self.exploration_rate:
# choose random action
n_batch = observation.shape[0]
# action = np.array([self.action_space.sample() for _ in range(n_batch)])
action = np.array([self.action_space.sample()])
is_random_action = 1
vectorized_env = is_vectorized_observation(observation, self.policy.observation_space)
if not vectorized_env:
action = action[0]
# print("is random action",action)
else:
action, state = self.policy.predict(observation, state, mask, deterministic)
is_random_action = 0
# print("is nonrandom action", action, observation.shape[0])
# if type(action) is not np.array:
# action = np.array([action])
return action, state, is_random_action
def predict(
self,
observation: np.ndarray,
state: Optional[np.ndarray] = None,
mask: Optional[np.ndarray] = None,
deterministic: bool = False,
) -> Tuple[np.ndarray, Optional[np.ndarray]]:
"""
Overriden to create proper Octree batch.
Get the policy action and state from an observation (and optional state).
:param observation: the input observation
:param state: The last states (can be None, used in recurrent policies)
:param mask: The last masks (can be None, used in recurrent policies)
:param deterministic: Whether or not to return deterministic actions.
:return: the model's action and the next state
(used in recurrent policies)
"""
if isinstance(observation, dict):
observation = ObsDictWrapper.convert_dict(observation)
else:
observation = np.array(observation)
vectorized_env = is_vectorized_observation(observation,
self.observation_space)
if self._debug_write_octree:
ocnn.write_octree(th.from_numpy(observation[-1]), 'octree.octree')
# Make batch out of tensor (consisting of n-stacked octrees)
octree_batch = preprocess_stacked_octree_batch(
observation,
self.device,
separate_batches=self._separate_networks_for_stacks)
with th.no_grad():
actions = self._predict(octree_batch, deterministic=deterministic)
# Convert to numpy
actions = actions.cpu().numpy()
if isinstance(self.action_space, gym.spaces.Box):
if self.squash_output:
# Rescale to proper domain when using squashing
actions = self.unscale_action(actions)
else:
# Actions could be on arbitrary scale, so clip the actions to avoid
# out of bound error (e.g. if sampling from a Gaussian distribution)
actions = np.clip(actions, self.action_space.low,
self.action_space.high)
if not vectorized_env:
if state is not None:
raise ValueError(
"Error: The environment must be vectorized when using recurrent policies."
)
actions = actions[0]
return actions, state
def predict(
self,
observation: np.ndarray,
state: Optional[np.ndarray] = None,
mask: Optional[np.ndarray] = None,
deterministic: bool = False,
) -> Tuple[np.ndarray, Optional[np.ndarray]]:
"""
Get the policy action and state from an observation (and optional state).
Includes sugar-coating to handle different observations (e.g. normalizing images).
:param observation: the input observation
:param state: The last states (can be None, used in recurrent policies)
:param mask: The last masks (can be None, used in recurrent policies)
:param deterministic: Whether or not to return deterministic actions.
:return: the model's action and the next state
(used in recurrent policies)
"""
# TODO (GH/1): add support for RNN policies
# if state is None:
# state = self.initial_state
# if mask is None:
# mask = [False for _ in range(self.n_envs)]
if isinstance(observation, dict):
observation = ObsDictWrapper.convert_dict(observation)
else:
observation = np.array(observation)
# Handle the different cases for images
# as PyTorch use channel first format
observation = maybe_transpose(observation, self.observation_space)
vectorized_env = is_vectorized_observation(observation, self.observation_space)
observation = observation.reshape((-1,) + self.observation_space.shape)
observation = th.as_tensor(observation).to(self.device)
with th.no_grad():
actions = self._predict(observation, deterministic=deterministic)
# Convert to numpy
actions = actions.cpu().numpy()
if isinstance(self.action_space, gym.spaces.Box):
if self.squash_output:
# Rescale to proper domain when using squashing
actions = self.unscale_action(actions)
else:
# Actions could be on arbitrary scale, so clip the actions to avoid
# out of bound error (e.g. if sampling from a Gaussian distribution)
actions = np.clip(actions, self.action_space.low, self.action_space.high)
if not vectorized_env:
if state is not None:
raise ValueError("Error: The environment must be vectorized when using recurrent policies.")
actions = actions[0]
return actions, state
def preprocesses_obs_for_model(
self,
observation: Union[np.ndarray, Dict[str, np.ndarray]]
) -> Tuple[Union[np.ndarray, Dict[str, np.ndarray]], bool]:
"""
Preporcesses obs both for prediction and evaluate action
:param observation
:return: Observation as PyTorch tensor
:return:
"""
vectorized_env = False
if isinstance(observation, dict):
# need to copy the dict as the dict in VecFrameStack will become a torch tensor
observation = copy.deepcopy(observation)
for key, obs in observation.items():
obs_space = self.observation_space.spaces[key]
if is_image_space(obs_space):
obs_ = maybe_transpose(obs, obs_space)
else:
obs_ = np.array(obs)
vectorized_env = vectorized_env or is_vectorized_observation(obs_, obs_space)
# Add batch dimension if needed
observation[key] = obs_.reshape((-1,) + self.observation_space[key].shape)
elif is_image_space(self.observation_space):
# Handle the different cases for images
# as PyTorch use channel first format
observation = maybe_transpose(observation, self.observation_space)
else:
observation = np.array(observation)
if not isinstance(observation, dict):
# Dict obs need to be handled separately
vectorized_env = is_vectorized_observation(observation, self.observation_space)
# Add batch dimension if needed
observation = observation.reshape((-1,) + self.observation_space.shape)
return observation, vectorized_env
def predict(
self,
observation: Union[np.ndarray, Dict[str, np.ndarray]],
state: Optional[np.ndarray] = None,
mask: Optional[np.ndarray] = None,
deterministic: bool = False,
) -> Tuple[np.ndarray, Optional[np.ndarray]]:
"""
Get the policy action and state from an observation (and optional state).
Includes sugar-coating to handle different observations (e.g. normalizing images).
:param observation: the input observation
:param state: The last states (can be None, used in recurrent policies)
:param mask: The last masks (can be None, used in recurrent policies)
:param deterministic: Whether or not to return deterministic actions.
:return: the model's action and the next state
(used in recurrent policies)
"""
# TODO (GH/1): add support for RNN policies
# if state is None:
# state = self.initial_state
# if mask is None:
# mask = [False for _ in range(self.n_envs)]
# Switch to eval mode (this affects batch norm / dropout)
self.eval()
vectorized_env = False
if isinstance(observation, dict):
# need to copy the dict as the dict in VecFrameStack will become a torch tensor
observation = copy.deepcopy(observation)
for key, obs in observation.items():
obs_space = self.observation_space.spaces[key]
if is_image_space(obs_space):
obs_ = maybe_transpose(obs, obs_space)
else:
obs_ = np.array(obs)
vectorized_env = vectorized_env or is_vectorized_observation(obs_, obs_space)
# Add batch dimension if needed
observation[key] = obs_.reshape((-1,) + self.observation_space[key].shape)
elif is_image_space(self.observation_space):
# Handle the different cases for images
# as PyTorch use channel first format
observation = maybe_transpose(observation, self.observation_space)
else:
observation = np.array(observation)
if not isinstance(observation, dict):
# Dict obs need to be handled separately
vectorized_env = is_vectorized_observation(observation, self.observation_space)
# Add batch dimension if needed
observation = observation.reshape((-1,) + self.observation_space.shape)
observation = obs_as_tensor(observation, self.device)
with th.no_grad():
actions = self._predict(observation, deterministic=deterministic)
# Convert to numpy
actions = actions.cpu().numpy()
if isinstance(self.action_space, gym.spaces.Box):
if self.squash_output:
# Rescale to proper domain when using squashing
actions = self.unscale_action(actions)
else:
# Actions could be on arbitrary scale, so clip the actions to avoid
# out of bound error (e.g. if sampling from a Gaussian distribution)
actions = np.clip(actions, self.action_space.low, self.action_space.high)
if not vectorized_env:
if state is not None:
raise ValueError("Error: The environment must be vectorized when using recurrent policies.")
actions = actions[0]
return actions, state
def test_is_vectorized_observation():
# with pytest.raises("ValueError"):
# pass
# All vectorized
box_space = spaces.Box(-1, 1, shape=(2, ))
box_obs = np.ones((1, ) + box_space.shape)
assert is_vectorized_observation(box_obs, box_space)
discrete_space = spaces.Discrete(2)
discrete_obs = np.ones((3, ), dtype=np.int8)
assert is_vectorized_observation(discrete_obs, discrete_space)
multidiscrete_space = spaces.MultiDiscrete([2, 3])
multidiscrete_obs = np.ones((1, 2), dtype=np.int8)
assert is_vectorized_observation(multidiscrete_obs, multidiscrete_space)
multibinary_space = spaces.MultiBinary(3)
multibinary_obs = np.ones((1, 3), dtype=np.int8)
assert is_vectorized_observation(multibinary_obs, multibinary_space)
dict_space = spaces.Dict({"box": box_space, "discrete": discrete_space})
dict_obs = {"box": box_obs, "discrete": discrete_obs}
assert is_vectorized_observation(dict_obs, dict_space)
# All not vectorized
box_obs = np.ones(box_space.shape)
assert not is_vectorized_observation(box_obs, box_space)
discrete_obs = np.ones((), dtype=np.int8)
assert not is_vectorized_observation(discrete_obs, discrete_space)
multidiscrete_obs = np.ones((2, ), dtype=np.int8)
assert not is_vectorized_observation(multidiscrete_obs,
multidiscrete_space)
multibinary_obs = np.ones((3, ), dtype=np.int8)
assert not is_vectorized_observation(multibinary_obs, multibinary_space)
dict_obs = {"box": box_obs, "discrete": discrete_obs}
assert not is_vectorized_observation(dict_obs, dict_space)
# A mix of vectorized and non-vectorized things
with pytest.raises(ValueError):
discrete_obs = np.ones((1, ), dtype=np.int8)
dict_obs = {"box": box_obs, "discrete": discrete_obs}
is_vectorized_observation(dict_obs, dict_space)
# Vectorized with the wrong shape
with pytest.raises(ValueError):
discrete_obs = np.ones((1, ), dtype=np.int8)
box_obs = np.ones((1, 2) + box_space.shape)
dict_obs = {"box": box_obs, "discrete": discrete_obs}
is_vectorized_observation(dict_obs, dict_space)
# Weird shape: error
with pytest.raises(ValueError):
discrete_obs = np.ones((1, ) + box_space.shape, dtype=np.int8)
is_vectorized_observation(discrete_obs, discrete_space)
# wrong shape
with pytest.raises(ValueError):
multidiscrete_obs = np.ones((2, 1), dtype=np.int8)
is_vectorized_observation(multidiscrete_obs, multidiscrete_space)
# wrong shape
with pytest.raises(ValueError):
multibinary_obs = np.ones((2, 1), dtype=np.int8)
is_vectorized_observation(multidiscrete_obs, multibinary_space)
# Almost good shape: one dimension too much for Discrete obs
with pytest.raises(ValueError):
box_obs = np.ones((1, ) + box_space.shape)
discrete_obs = np.ones((1, 1), dtype=np.int8)
dict_obs = {"box": box_obs, "discrete": discrete_obs}
is_vectorized_observation(dict_obs, dict_space)
def predict(
self,
observation: np.ndarray,
state: Optional[np.ndarray] = None,
mask: Optional[np.ndarray] = None,
deterministic: bool = False
) -> Tuple[np.ndarray, Optional[np.ndarray]]:
"""
Get the policy action and state from an observation (and optional state).
Includes sugar-coating to handle different observations (e.g. normalizing images).
:param observation: (np.ndarray) the input observation
:param state: (Optional[np.ndarray]) The last states (can be None, used in recurrent policies)
:param mask: (Optional[np.ndarray]) The last masks (can be None, used in recurrent policies)
:param deterministic: (bool) Whether or not to return deterministic actions.
:return: (Tuple[np.ndarray, Optional[np.ndarray]]) the model's action and the next state
(used in recurrent policies)
"""
# if state is None:
# state = self.initial_state
# if mask is None:
# mask = [False for _ in range(self.n_envs)]
observation = np.array(observation)
# Handle the different cases for images
# as PyTorch use channel first format
if is_image_space(self.observation_space):
if (observation.shape == self.observation_space.shape
or observation.shape[1:] == self.observation_space.shape):
pass
else:
# Try to re-order the channels
transpose_obs = VecTransposeImage.transpose_image(observation)
if (transpose_obs.shape == self.observation_space.shape
or transpose_obs.shape[1:]
== self.observation_space.shape):
observation = transpose_obs
vectorized_env = is_vectorized_observation(observation,
self.observation_space)
observation = observation.reshape((-1, ) +
self.observation_space.shape)
observation = th.as_tensor(observation).to(self.device)
with th.no_grad():
actions = self._predict(observation, deterministic=deterministic)
# Convert to numpy
actions = actions.cpu().numpy()
# Rescale to proper domain when using squashing
if isinstance(self.action_space,
gym.spaces.Box) and self.squash_output:
actions = self.unscale_action(actions)
clipped_actions = actions
# Clip the actions to avoid out of bound error when using gaussian distribution
if isinstance(self.action_space,
gym.spaces.Box) and not self.squash_output:
clipped_actions = np.clip(actions, self.action_space.low,
self.action_space.high)
if not vectorized_env:
if state is not None:
raise ValueError(
"Error: The environment must be vectorized when using recurrent policies."
)
clipped_actions = clipped_actions[0]
return clipped_actions, state
