Index
Problem types for neuroevolution
baseneproblem
¶
gymne
¶
This namespace contains the GymNE class.
GymNE (NEProblem)
¶
Representation of a NEProblem where the goal is to maximize
the total reward obtained in a gym environment.
Source code in evotorch/neuroevolution/gymne.py
class GymNE(NEProblem):
"""
Representation of a NEProblem where the goal is to maximize
the total reward obtained in a `gym` environment.
"""
def __init__(
self,
env: Optional[Union[str, Callable]] = None,
network: Optional[Union[str, nn.Module, Callable[[], nn.Module]]] = None,
*,
env_name: Optional[Union[str, Callable]] = None,
network_args: Optional[dict] = None,
env_config: Optional[Mapping] = None,
observation_normalization: bool = False,
num_episodes: int = 1,
episode_length: Optional[int] = None,
decrease_rewards_by: Optional[float] = None,
alive_bonus_schedule: Optional[tuple] = None,
action_noise_stdev: Optional[float] = None,
num_actors: Optional[Union[int, str]] = "max",
actor_config: Optional[dict] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
initial_bounds: Optional[BoundsPairLike] = (-0.00001, 0.00001),
):
"""
`__init__(...)`: Initialize the GymNE.
Args:
env: The gym environment to solve. Expected as a Callable
(maybe a function returning a gym.Env, or maybe a gym.Env
subclass), or as a string referring to a gym environment
ID (e.g. "Ant-v4", "Humanoid-v4", etc.).
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network policy whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
Note that this network can be a recurrent network.
When the network's `forward(...)` method can optionally accept
an additional positional argument for the hidden state of the
network and returns an additional value for its next state,
then the policy is treated as a recurrent one.
When the network is given as a callable object (e.g.
a subclass of `nn.Module` or a function) and this callable
object is decorated via `evotorch.decorators.pass_info`,
the following keyword arguments will be passed:
(i) `obs_length` (the length of the observation vector),
(ii) `act_length` (the length of the action vector),
(iii) `obs_shape` (the shape tuple of the observation space),
(iv) `act_shape` (the shape tuple of the action space),
(v) `obs_space` (the Box object specifying the observation
space, and
(vi) `act_space` (the Box object specifying the action
space). Note that `act_space` will always be given as a
`gym.spaces.Box` instance, even when the actual gym
environment has a discrete action space. This because `GymNE`
always expects the neural network to return a tensor of
floating-point numbers.
env_name: Deprecated alias for the keyword argument `env`.
It is recommended to use the argument `env` instead.
network_args: Optionally a dict-like object, storing keyword
arguments to be passed to the network while instantiating it.
env_config: Keyword arguments to pass to `gym.make(...)` while
creating the `gym` environment.
observation_normalization: Whether or not to do online observation
normalization.
num_episodes: Number of episodes over which a single solution will
be evaluated.
episode_length: Maximum amount of simulator interactions allowed
in a single episode. If left as None, whether or not an episode
is terminated is determined only by the `gym` environment
itself.
decrease_rewards_by: Some gym env.s are defined in such a way that
the agent gets a constant reward for each timestep
it survives. This constant reward can also be called
"survival bonus". Such a rewarding scheme can lead the
evolution to local optima where the agent does nothing
but does not die either, just to collect the survival
bonuses. To prevent this, it can be desired to
remove the survival bonuses from each reward obtained.
If this is the case with the problem at hand,
the user can set the argument `decrease_rewards_by`
to a positive float number, and that number will
be subtracted from each reward.
alive_bonus_schedule: Use this to add a customized amount of
alive bonus.
If left as None (which is the default), additional alive
bonus will not be added.
If given as a tuple `(t, b)`, an alive bonus `b` will be
added onto all the rewards beyond the timestep `t`.
If given as a tuple `(t0, t1, b)`, a partial (linearly
increasing towards `b`) alive bonus will be added onto
all the rewards between the timesteps `t0` and `t1`,
and a full alive bonus (which equals to `b`) will be added
onto all the rewards beyond the timestep `t1`.
action_noise_stdev: If given as a real number `s`, then, for
each generated action, Gaussian noise with standard
deviation `s` will be sampled, and then this sampled noise
will be added onto the action.
If action noise is not desired, then this argument can be
left as None.
num_actors: Number of actors to create for parallelized
evaluation of the solutions.
One can also set this as "max", which means that
an actor will be created on each available CPU.
When the parallelization is enabled each actor will have its
own instance of the `gym` environment.
In the case of `GymNE`, the default value for this argument
is "max", which means there will be full parallelization,
utilizing all the available CPUs.
actor_config: A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass `actor_config=dict(num_gpus=1)`.
Can also be given as None (which is the default),
if no such options are to be passed.
num_subbatches: If `num_subbatches` is None (assuming that
`subbatch_size` is also None), then, when evaluating a
population, the population will be split into n pieces, `n`
being the number of actors, and each actor will evaluate
its assigned piece. If `num_subbatches` is an integer `m`,
then the population will be split into `m` pieces,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
subbatch_size: If `subbatch_size` is None (assuming that
`num_subbatches` is also None), then, when evaluating a
population, the population will be split into `n` pieces, `n`
being the number of actors, and each actor will evaluate its
assigned piece. If `subbatch_size` is an integer `m`,
then the population will be split into pieces of size `m`,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
When there can be significant difference across the solutions
in terms of computational requirements, specifying a
`subbatch_size` can be beneficial, because, while one
actor is busy with a subbatch containing computationally
challenging solutions, other actors can accept more
tasks and save time.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
initial_bounds: Specifies an interval from which the values of the
initial policy parameters will be drawn.
"""
# Store various environment information
if (env is not None) and (env_name is None):
self._env_maker = env
elif (env is None) and (env_name is not None):
self._env_maker = env_name
elif (env is not None) and (env_name is not None):
raise ValueError(
f"Received values for both `env` ({repr(env)}) and `env_name` ({repr(env_name)})."
f" Please specify the environment to solve via only one of these arguments, not both."
)
else:
raise ValueError("Environment name is missing. Please specify it via the argument `env`.")
# Make sure that the network argument is not missing.
if network is None:
raise ValueError(
"Received None via the argument `network`."
"Please provide the network as a string, or as a `Callable`, or as a `torch.nn.Module` instance."
)
# Store various environment information
self._env_config = {} if env_config is None else deepcopy(dict(env_config))
self._decrease_rewards_by = 0.0 if decrease_rewards_by is None else float(decrease_rewards_by)
self._alive_bonus_schedule = alive_bonus_schedule
self._action_noise_stdev = None if action_noise_stdev is None else float(action_noise_stdev)
self._observation_normalization = bool(observation_normalization)
self._num_episodes = int(num_episodes)
self._episode_length = None if episode_length is None else int(episode_length)
self._info_keys = dict(cumulative_reward="avg", interaction_count="sum")
self._env: Optional[gym.Env] = None
self._obs_stats: Optional[RunningStat] = None
self._collected_stats: Optional[RunningStat] = None
# Create a temporary environment to read its dimensions
tmp_env = _make_env(self._env_maker, **(self._env_config))
# Store the temporary environment's dimensions
self._obs_length = len(tmp_env.observation_space.low)
if isinstance(tmp_env.action_space, gym.spaces.Discrete):
self._act_length = tmp_env.action_space.n
self._box_act_space = gym.spaces.Box(low=float("-inf"), high=float("inf"), shape=(self._act_length,))
else:
self._act_length = len(tmp_env.action_space.low)
self._box_act_space = tmp_env.action_space
self._act_space = tmp_env.action_space
self._obs_space = tmp_env.observation_space
self._obs_shape = tmp_env.observation_space.low.shape
# Validate the space types of the environment
ensure_space_types(tmp_env)
if self._observation_normalization:
self._obs_stats = RunningStat()
self._collected_stats = RunningStat()
else:
self._obs_stats = None
self._collected_stats = None
self._interaction_count: int = 0
self._episode_count: int = 0
super().__init__(
objective_sense="max", # RL is maximization
network=network, # Using the policy as the network
network_args=network_args,
initial_bounds=initial_bounds,
num_actors=num_actors,
actor_config=actor_config,
subbatch_size=subbatch_size,
device="cpu",
)
self.after_eval_hook.append(self._extra_status)
@property
def _network_constants(self) -> dict:
return {
"obs_length": self._obs_length,
"act_length": self._act_length,
"obs_space": self._obs_space,
"act_space": self._box_act_space,
"obs_shape": self._obs_space.shape,
"act_shape": self._box_act_space.shape,
}
@property
def _str_network_constants(self) -> dict:
return {
"obs_space": self._obs_space.shape,
"act_space": self._box_act_space.shape,
}
def _instantiate_new_env(self, **kwargs) -> gym.Env:
env_config = {**kwargs, **(self._env_config)}
env = _make_env(self._env_maker, **env_config)
if self._alive_bonus_schedule is not None:
env = AliveBonusScheduleWrapper(env, self._alive_bonus_schedule)
return env
def _get_env(self) -> gym.Env:
if self._env is None:
self._env = self._instantiate_new_env()
return self._env
def _normalize_observation(self, observation: Iterable, *, update_stats: bool = True) -> Iterable:
observation = np.asarray(observation, dtype="float32")
if self.observation_normalization:
if update_stats:
self._obs_stats.update(observation)
self._collected_stats.update(observation)
return self._obs_stats.normalize(observation)
else:
return observation
def _use_policy(self, observation: Iterable, policy: nn.Module) -> Iterable:
with torch.no_grad():
result = policy(torch.as_tensor(observation, dtype=torch.float32, device="cpu")).numpy()
if self._action_noise_stdev is not None:
result = (
result
+ self.make_gaussian(len(result), center=0.0, stdev=self._action_noise_stdev, device="cpu").numpy()
)
env = self._get_env()
if isinstance(env.action_space, gym.spaces.Discrete):
result = np.argmax(result)
elif isinstance(env.action_space, gym.spaces.Box):
result = np.clip(result, env.action_space.low, env.action_space.high)
return result
def _prepare(self) -> None:
super()._prepare()
self._get_env()
@property
def network_device(self) -> Device:
"""The device on which the problem should place data e.g. the network
In the case of GymNE, supported Gym environments return numpy arrays on CPU which are converted to Tensors
Therefore, it is almost always optimal to place the network on CPU
"""
return torch.device("cpu")
def _rollout(
self,
*,
policy: nn.Module,
update_stats: bool = True,
visualize: bool = False,
decrease_rewards_by: Optional[float] = None,
) -> dict:
"""Peform a rollout of a network"""
if decrease_rewards_by is None:
decrease_rewards_by = self._decrease_rewards_by
else:
decrease_rewards_by = float(decrease_rewards_by)
policy = ensure_stateful(policy)
policy.reset()
if visualize and (not _gym_older_than_0_26):
# Beginning with gym 0.26, we need to specify the render mode when instantiating the environment.
env = self._instantiate_new_env(render_mode="human")
else:
env = self._get_env()
observation = self._normalize_observation(reset_env(env), update_stats=update_stats)
if visualize and _gym_older_than_0_26:
env.render()
t = 0
cumulative_reward = 0.0
while True:
observation, raw_reward, done, info = take_step_in_env(env, self._use_policy(observation, policy))
reward = raw_reward - decrease_rewards_by
t += 1
if update_stats:
self._interaction_count += 1
if visualize:
env.render()
observation = self._normalize_observation(observation, update_stats=update_stats)
cumulative_reward += reward
if done or ((self._episode_length is not None) and (t >= self._episode_length)):
if update_stats:
self._episode_count += 1
final_info = dict(cumulative_reward=cumulative_reward, interaction_count=t)
for k in self._info_keys:
if k not in final_info:
final_info[k] = info[k]
return final_info
@property
def _nonserialized_attribs(self) -> List[str]:
return super()._nonserialized_attribs + ["_env"]
def run(
self,
policy: Union[nn.Module, Iterable],
*,
update_stats: bool = False,
visualize: bool = False,
num_episodes: Optional[int] = None,
decrease_rewards_by: Optional[float] = None,
) -> dict:
"""
Evaluate the policy on the gym environment.
Args:
policy: The policy to be evaluated. This can be a torch module
or a sequence of real numbers representing the parameters
of a policy network.
update_stats: Whether or not to update the observation
normalization data while running the policy. If observation
normalization is not enabled, then this argument will be
ignored.
visualize: Whether or not to render the environment while running
the policy.
num_episodes: Over how many episodes will the policy be evaluated.
Expected as None (which is the default), or as an integer.
If given as None, then the `num_episodes` value that was given
while initializing this GymNE will be used.
decrease_rewards_by: How much each reward value should be
decreased. If left as None, the `decrease_rewards_by` value
value that was given while initializing this GymNE will be
used.
Returns:
A dictionary containing the score and the timestep count.
"""
if not isinstance(policy, nn.Module):
policy = self.make_net(policy)
if num_episodes is None:
num_episodes = self._num_episodes
try:
policy.eval()
episode_results = [
self._rollout(
policy=policy,
update_stats=update_stats,
visualize=visualize,
decrease_rewards_by=decrease_rewards_by,
)
for _ in range(num_episodes)
]
results = _accumulate_all_across_dicts(episode_results, self._info_keys)
return results
finally:
policy.train()
def visualize(
self,
policy: Union[nn.Module, Iterable],
*,
update_stats: bool = False,
num_episodes: Optional[int] = 1,
decrease_rewards_by: Optional[float] = None,
) -> dict:
"""
Evaluate the policy and render its actions in the environment.
Args:
policy: The policy to be evaluated. This can be a torch module
or a sequence of real numbers representing the parameters
of a policy network.
update_stats: Whether or not to update the observation
normalization data while running the policy. If observation
normalization is not enabled, then this argument will be
ignored.
num_episodes: Over how many episodes will the policy be evaluated.
Expected as None (which is the default), or as an integer.
If given as None, then the `num_episodes` value that was given
while initializing this GymNE will be used.
decrease_rewards_by: How much each reward value should be
decreased. If left as None, the `decrease_rewards_by` value
value that was given while initializing this GymNE will be
used.
Returns:
A dictionary containing the score and the timestep count.
"""
return self.run(
policy=policy,
update_stats=update_stats,
visualize=True,
num_episodes=num_episodes,
decrease_rewards_by=decrease_rewards_by,
)
def _ensure_obsnorm(self):
if not self.observation_normalization:
raise ValueError("This feature can only be used when observation_normalization=True.")
def get_observation_stats(self) -> RunningStat:
"""Get the observation stats"""
self._ensure_obsnorm()
return self._obs_stats
def _make_sync_data_for_actors(self) -> Any:
if self.observation_normalization:
return dict(obs_stats=self.get_observation_stats())
else:
return None
def set_observation_stats(self, rs: RunningStat):
"""Set the observation stats"""
self._ensure_obsnorm()
self._obs_stats.reset()
self._obs_stats.update(rs)
def _use_sync_data_from_main(self, received: dict):
for k, v in received.items():
if k == "obs_stats":
self.set_observation_stats(v)
def pop_observation_stats(self) -> RunningStat:
"""Get and clear the collected observation stats"""
self._ensure_obsnorm()
result = self._collected_stats
self._collected_stats = RunningStat()
return result
def _make_sync_data_for_main(self) -> Any:
result = dict(episode_count=self.episode_count, interaction_count=self.interaction_count)
if self.observation_normalization:
result["obs_stats_delta"] = self.pop_observation_stats()
return result
def update_observation_stats(self, rs: RunningStat):
"""Update the observation stats via another RunningStat instance"""
self._ensure_obsnorm()
self._obs_stats.update(rs)
def _use_sync_data_from_actors(self, received: list):
total_episode_count = 0
total_interaction_count = 0
for data in received:
data: dict
total_episode_count += data["episode_count"]
total_interaction_count += data["interaction_count"]
if self.observation_normalization:
self.update_observation_stats(data["obs_stats_delta"])
self.set_episode_count(total_episode_count)
self.set_interaction_count(total_interaction_count)
def _make_pickle_data_for_main(self) -> dict:
# For when the main Problem object (the non-remote one) gets pickled,
# this function returns the counters of this remote Problem instance,
# to be sent to the main one.
return dict(interaction_count=self.interaction_count, episode_count=self.episode_count)
def _use_pickle_data_from_main(self, state: dict):
# For when a newly unpickled Problem object gets (re)parallelized,
# this function restores the inner states specific to this remote
# worker. In the case of GymNE, those inner states are episode
# and interaction counters.
for k, v in state.items():
if k == "episode_count":
self.set_episode_count(v)
elif k == "interaction_count":
self.set_interaction_count(v)
else:
raise ValueError(f"When restoring the inner state of a remote worker, unrecognized state key: {k}")
def _extra_status(self, batch: SolutionBatch):
return dict(total_interaction_count=self.interaction_count, total_episode_count=self.episode_count)
@property
def observation_normalization(self) -> bool:
"""
Get whether or not observation normalization is enabled.
"""
return self._observation_normalization
def set_episode_count(self, n: int):
"""
Set the episode count manually.
"""
self._episode_count = int(n)
def set_interaction_count(self, n: int):
"""
Set the interaction count manually.
"""
self._interaction_count = int(n)
@property
def interaction_count(self) -> int:
"""
Get the total number of simulator interactions made.
"""
return self._interaction_count
@property
def episode_count(self) -> int:
"""
Get the total number of episodes completed.
"""
return self._episode_count
def _get_local_episode_count(self) -> int:
return self.episode_count
def _get_local_interaction_count(self) -> int:
return self.interaction_count
def _evaluate_network(self, policy: nn.Module) -> Union[float, torch.Tensor]:
result = self.run(
policy,
update_stats=True,
visualize=False,
num_episodes=self._num_episodes,
decrease_rewards_by=self._decrease_rewards_by,
)
return result["cumulative_reward"]
def to_policy(self, x: Iterable, *, clip_actions: bool = True) -> nn.Module:
"""
Convert the given parameter vector to a policy as a PyTorch module.
If the problem is configured to have observation normalization,
the PyTorch module also contains an additional normalization layer.
Args:
x: An sequence of real numbers, containing the parameters
of a policy. Can be a PyTorch tensor, a numpy array,
or a SolutionVector.
clip_actions: Whether or not to add an action clipping layer so
that the generated actions will always be within an
acceptable range for the environment.
Returns:
The policy expressed by the parameters.
"""
policy = self.make_net(x)
if self.observation_normalization and (self._obs_stats.count > 0):
policy = ObsNormWrapperModule(policy, self._obs_stats)
if clip_actions and isinstance(self._get_env().action_space, gym.spaces.Box):
policy = ActClipWrapperModule(policy, self._get_env().action_space)
return policy
def save_solution(self, solution: Iterable, fname: Union[str, Path]):
"""
Save the solution into a pickle file.
Among the saved data within the pickle file are the solution
(as a PyTorch tensor), the policy (as a `torch.nn.Module` instance),
and observation stats (if any).
Args:
solution: The solution to be saved. This can be a PyTorch tensor,
a `Solution` instance, or any `Iterable`.
fname: The file name of the pickle file to be created.
"""
# Convert the solution to a PyTorch tensor on the cpu.
if isinstance(solution, torch.Tensor):
solution = solution.to("cpu")
elif isinstance(solution, Solution):
solution = solution.values.clone().to("cpu")
else:
solution = torch.as_tensor(solution, dtype=torch.float32, device="cpu")
if isinstance(solution, ReadOnlyTensor):
solution = solution.as_subclass(torch.Tensor)
policy = self.to_policy(solution).to("cpu")
# Store the solution and the policy.
result = {
"solution": solution,
"policy": policy,
}
# If available, store the observation stats.
if self.observation_normalization and (self._obs_stats is not None):
result["obs_mean"] = torch.as_tensor(self._obs_stats.mean)
result["obs_stdev"] = torch.as_tensor(self._obs_stats.stdev)
result["obs_sum"] = torch.as_tensor(self._obs_stats.sum)
result["obs_sum_of_squares"] = torch.as_tensor(self._obs_stats.sum_of_squares)
# Some additional data.
result["interaction_count"] = self.interaction_count
result["episode_count"] = self.episode_count
result["time"] = datetime.now()
# If the environment is specified via a string ID, then store that ID.
if isinstance(self._env_maker, str):
result["env"] = self._env_maker
# Save the dictionary which stores the data.
with open(fname, "wb") as f:
pickle.dump(result, f)
def get_env(self) -> gym.Env:
"""
Get the gym environment stored by this GymNE instance
"""
return self._get_env()
episode_count: int
property
readonly
¶
Get the total number of episodes completed.
interaction_count: int
property
readonly
¶
Get the total number of simulator interactions made.
network_device: Union[str, torch.device]
property
readonly
¶
The device on which the problem should place data e.g. the network In the case of GymNE, supported Gym environments return numpy arrays on CPU which are converted to Tensors Therefore, it is almost always optimal to place the network on CPU
observation_normalization: bool
property
readonly
¶
Get whether or not observation normalization is enabled.
__init__(self, env=None, network=None, *, env_name=None, network_args=None, env_config=None, observation_normalization=False, num_episodes=1, episode_length=None, decrease_rewards_by=None, alive_bonus_schedule=None, action_noise_stdev=None, num_actors='max', actor_config=None, num_subbatches=None, subbatch_size=None, initial_bounds=(-1e-05, 1e-05))
special
¶
__init__(...): Initialize the GymNE.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env |
Union[str, Callable] |
The gym environment to solve. Expected as a Callable (maybe a function returning a gym.Env, or maybe a gym.Env subclass), or as a string referring to a gym environment ID (e.g. "Ant-v4", "Humanoid-v4", etc.). |
None |
network |
Union[str, torch.nn.modules.module.Module, Callable[[], torch.nn.modules.module.Module]] |
A network structure string, or a Callable (which can be
a class inheriting from |
None |
env_name |
Union[str, Callable] |
Deprecated alias for the keyword argument |
None |
network_args |
Optional[dict] |
Optionally a dict-like object, storing keyword arguments to be passed to the network while instantiating it. |
None |
env_config |
Optional[collections.abc.Mapping] |
Keyword arguments to pass to |
None |
observation_normalization |
bool |
Whether or not to do online observation normalization. |
False |
num_episodes |
int |
Number of episodes over which a single solution will be evaluated. |
1 |
episode_length |
Optional[int] |
Maximum amount of simulator interactions allowed
in a single episode. If left as None, whether or not an episode
is terminated is determined only by the |
None |
decrease_rewards_by |
Optional[float] |
Some gym env.s are defined in such a way that
the agent gets a constant reward for each timestep
it survives. This constant reward can also be called
"survival bonus". Such a rewarding scheme can lead the
evolution to local optima where the agent does nothing
but does not die either, just to collect the survival
bonuses. To prevent this, it can be desired to
remove the survival bonuses from each reward obtained.
If this is the case with the problem at hand,
the user can set the argument |
None |
alive_bonus_schedule |
Optional[tuple] |
Use this to add a customized amount of
alive bonus.
If left as None (which is the default), additional alive
bonus will not be added.
If given as a tuple |
None |
action_noise_stdev |
Optional[float] |
If given as a real number |
None |
num_actors |
Union[int, str] |
Number of actors to create for parallelized
evaluation of the solutions.
One can also set this as "max", which means that
an actor will be created on each available CPU.
When the parallelization is enabled each actor will have its
own instance of the |
'max' |
actor_config |
Optional[dict] |
A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass |
None |
num_subbatches |
Optional[int] |
If |
None |
subbatch_size |
Optional[int] |
If |
None |
initial_bounds |
Union[Iterable[Union[float, Iterable[float], torch.Tensor]], evotorch.core.BoundsPair] |
Specifies an interval from which the values of the initial policy parameters will be drawn. |
(-1e-05, 1e-05) |
Source code in evotorch/neuroevolution/gymne.py
def __init__(
self,
env: Optional[Union[str, Callable]] = None,
network: Optional[Union[str, nn.Module, Callable[[], nn.Module]]] = None,
*,
env_name: Optional[Union[str, Callable]] = None,
network_args: Optional[dict] = None,
env_config: Optional[Mapping] = None,
observation_normalization: bool = False,
num_episodes: int = 1,
episode_length: Optional[int] = None,
decrease_rewards_by: Optional[float] = None,
alive_bonus_schedule: Optional[tuple] = None,
action_noise_stdev: Optional[float] = None,
num_actors: Optional[Union[int, str]] = "max",
actor_config: Optional[dict] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
initial_bounds: Optional[BoundsPairLike] = (-0.00001, 0.00001),
):
"""
`__init__(...)`: Initialize the GymNE.
Args:
env: The gym environment to solve. Expected as a Callable
(maybe a function returning a gym.Env, or maybe a gym.Env
subclass), or as a string referring to a gym environment
ID (e.g. "Ant-v4", "Humanoid-v4", etc.).
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network policy whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
Note that this network can be a recurrent network.
When the network's `forward(...)` method can optionally accept
an additional positional argument for the hidden state of the
network and returns an additional value for its next state,
then the policy is treated as a recurrent one.
When the network is given as a callable object (e.g.
a subclass of `nn.Module` or a function) and this callable
object is decorated via `evotorch.decorators.pass_info`,
the following keyword arguments will be passed:
(i) `obs_length` (the length of the observation vector),
(ii) `act_length` (the length of the action vector),
(iii) `obs_shape` (the shape tuple of the observation space),
(iv) `act_shape` (the shape tuple of the action space),
(v) `obs_space` (the Box object specifying the observation
space, and
(vi) `act_space` (the Box object specifying the action
space). Note that `act_space` will always be given as a
`gym.spaces.Box` instance, even when the actual gym
environment has a discrete action space. This because `GymNE`
always expects the neural network to return a tensor of
floating-point numbers.
env_name: Deprecated alias for the keyword argument `env`.
It is recommended to use the argument `env` instead.
network_args: Optionally a dict-like object, storing keyword
arguments to be passed to the network while instantiating it.
env_config: Keyword arguments to pass to `gym.make(...)` while
creating the `gym` environment.
observation_normalization: Whether or not to do online observation
normalization.
num_episodes: Number of episodes over which a single solution will
be evaluated.
episode_length: Maximum amount of simulator interactions allowed
in a single episode. If left as None, whether or not an episode
is terminated is determined only by the `gym` environment
itself.
decrease_rewards_by: Some gym env.s are defined in such a way that
the agent gets a constant reward for each timestep
it survives. This constant reward can also be called
"survival bonus". Such a rewarding scheme can lead the
evolution to local optima where the agent does nothing
but does not die either, just to collect the survival
bonuses. To prevent this, it can be desired to
remove the survival bonuses from each reward obtained.
If this is the case with the problem at hand,
the user can set the argument `decrease_rewards_by`
to a positive float number, and that number will
be subtracted from each reward.
alive_bonus_schedule: Use this to add a customized amount of
alive bonus.
If left as None (which is the default), additional alive
bonus will not be added.
If given as a tuple `(t, b)`, an alive bonus `b` will be
added onto all the rewards beyond the timestep `t`.
If given as a tuple `(t0, t1, b)`, a partial (linearly
increasing towards `b`) alive bonus will be added onto
all the rewards between the timesteps `t0` and `t1`,
and a full alive bonus (which equals to `b`) will be added
onto all the rewards beyond the timestep `t1`.
action_noise_stdev: If given as a real number `s`, then, for
each generated action, Gaussian noise with standard
deviation `s` will be sampled, and then this sampled noise
will be added onto the action.
If action noise is not desired, then this argument can be
left as None.
num_actors: Number of actors to create for parallelized
evaluation of the solutions.
One can also set this as "max", which means that
an actor will be created on each available CPU.
When the parallelization is enabled each actor will have its
own instance of the `gym` environment.
In the case of `GymNE`, the default value for this argument
is "max", which means there will be full parallelization,
utilizing all the available CPUs.
actor_config: A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass `actor_config=dict(num_gpus=1)`.
Can also be given as None (which is the default),
if no such options are to be passed.
num_subbatches: If `num_subbatches` is None (assuming that
`subbatch_size` is also None), then, when evaluating a
population, the population will be split into n pieces, `n`
being the number of actors, and each actor will evaluate
its assigned piece. If `num_subbatches` is an integer `m`,
then the population will be split into `m` pieces,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
subbatch_size: If `subbatch_size` is None (assuming that
`num_subbatches` is also None), then, when evaluating a
population, the population will be split into `n` pieces, `n`
being the number of actors, and each actor will evaluate its
assigned piece. If `subbatch_size` is an integer `m`,
then the population will be split into pieces of size `m`,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
When there can be significant difference across the solutions
in terms of computational requirements, specifying a
`subbatch_size` can be beneficial, because, while one
actor is busy with a subbatch containing computationally
challenging solutions, other actors can accept more
tasks and save time.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
initial_bounds: Specifies an interval from which the values of the
initial policy parameters will be drawn.
"""
# Store various environment information
if (env is not None) and (env_name is None):
self._env_maker = env
elif (env is None) and (env_name is not None):
self._env_maker = env_name
elif (env is not None) and (env_name is not None):
raise ValueError(
f"Received values for both `env` ({repr(env)}) and `env_name` ({repr(env_name)})."
f" Please specify the environment to solve via only one of these arguments, not both."
)
else:
raise ValueError("Environment name is missing. Please specify it via the argument `env`.")
# Make sure that the network argument is not missing.
if network is None:
raise ValueError(
"Received None via the argument `network`."
"Please provide the network as a string, or as a `Callable`, or as a `torch.nn.Module` instance."
)
# Store various environment information
self._env_config = {} if env_config is None else deepcopy(dict(env_config))
self._decrease_rewards_by = 0.0 if decrease_rewards_by is None else float(decrease_rewards_by)
self._alive_bonus_schedule = alive_bonus_schedule
self._action_noise_stdev = None if action_noise_stdev is None else float(action_noise_stdev)
self._observation_normalization = bool(observation_normalization)
self._num_episodes = int(num_episodes)
self._episode_length = None if episode_length is None else int(episode_length)
self._info_keys = dict(cumulative_reward="avg", interaction_count="sum")
self._env: Optional[gym.Env] = None
self._obs_stats: Optional[RunningStat] = None
self._collected_stats: Optional[RunningStat] = None
# Create a temporary environment to read its dimensions
tmp_env = _make_env(self._env_maker, **(self._env_config))
# Store the temporary environment's dimensions
self._obs_length = len(tmp_env.observation_space.low)
if isinstance(tmp_env.action_space, gym.spaces.Discrete):
self._act_length = tmp_env.action_space.n
self._box_act_space = gym.spaces.Box(low=float("-inf"), high=float("inf"), shape=(self._act_length,))
else:
self._act_length = len(tmp_env.action_space.low)
self._box_act_space = tmp_env.action_space
self._act_space = tmp_env.action_space
self._obs_space = tmp_env.observation_space
self._obs_shape = tmp_env.observation_space.low.shape
# Validate the space types of the environment
ensure_space_types(tmp_env)
if self._observation_normalization:
self._obs_stats = RunningStat()
self._collected_stats = RunningStat()
else:
self._obs_stats = None
self._collected_stats = None
self._interaction_count: int = 0
self._episode_count: int = 0
super().__init__(
objective_sense="max", # RL is maximization
network=network, # Using the policy as the network
network_args=network_args,
initial_bounds=initial_bounds,
num_actors=num_actors,
actor_config=actor_config,
subbatch_size=subbatch_size,
device="cpu",
)
self.after_eval_hook.append(self._extra_status)
get_env(self)
¶
get_observation_stats(self)
¶
pop_observation_stats(self)
¶
run(self, policy, *, update_stats=False, visualize=False, num_episodes=None, decrease_rewards_by=None)
¶
Evaluate the policy on the gym environment.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
policy |
Union[torch.nn.modules.module.Module, Iterable] |
The policy to be evaluated. This can be a torch module or a sequence of real numbers representing the parameters of a policy network. |
required |
update_stats |
bool |
Whether or not to update the observation normalization data while running the policy. If observation normalization is not enabled, then this argument will be ignored. |
False |
visualize |
bool |
Whether or not to render the environment while running the policy. |
False |
num_episodes |
Optional[int] |
Over how many episodes will the policy be evaluated.
Expected as None (which is the default), or as an integer.
If given as None, then the |
None |
decrease_rewards_by |
Optional[float] |
How much each reward value should be
decreased. If left as None, the |
None |
Returns:
| Type | Description |
|---|---|
dict |
A dictionary containing the score and the timestep count. |
Source code in evotorch/neuroevolution/gymne.py
def run(
self,
policy: Union[nn.Module, Iterable],
*,
update_stats: bool = False,
visualize: bool = False,
num_episodes: Optional[int] = None,
decrease_rewards_by: Optional[float] = None,
) -> dict:
"""
Evaluate the policy on the gym environment.
Args:
policy: The policy to be evaluated. This can be a torch module
or a sequence of real numbers representing the parameters
of a policy network.
update_stats: Whether or not to update the observation
normalization data while running the policy. If observation
normalization is not enabled, then this argument will be
ignored.
visualize: Whether or not to render the environment while running
the policy.
num_episodes: Over how many episodes will the policy be evaluated.
Expected as None (which is the default), or as an integer.
If given as None, then the `num_episodes` value that was given
while initializing this GymNE will be used.
decrease_rewards_by: How much each reward value should be
decreased. If left as None, the `decrease_rewards_by` value
value that was given while initializing this GymNE will be
used.
Returns:
A dictionary containing the score and the timestep count.
"""
if not isinstance(policy, nn.Module):
policy = self.make_net(policy)
if num_episodes is None:
num_episodes = self._num_episodes
try:
policy.eval()
episode_results = [
self._rollout(
policy=policy,
update_stats=update_stats,
visualize=visualize,
decrease_rewards_by=decrease_rewards_by,
)
for _ in range(num_episodes)
]
results = _accumulate_all_across_dicts(episode_results, self._info_keys)
return results
finally:
policy.train()
save_solution(self, solution, fname)
¶
Save the solution into a pickle file.
Among the saved data within the pickle file are the solution
(as a PyTorch tensor), the policy (as a torch.nn.Module instance),
and observation stats (if any).
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
solution |
Iterable |
The solution to be saved. This can be a PyTorch tensor,
a |
required |
fname |
Union[str, pathlib.Path] |
The file name of the pickle file to be created. |
required |
Source code in evotorch/neuroevolution/gymne.py
def save_solution(self, solution: Iterable, fname: Union[str, Path]):
"""
Save the solution into a pickle file.
Among the saved data within the pickle file are the solution
(as a PyTorch tensor), the policy (as a `torch.nn.Module` instance),
and observation stats (if any).
Args:
solution: The solution to be saved. This can be a PyTorch tensor,
a `Solution` instance, or any `Iterable`.
fname: The file name of the pickle file to be created.
"""
# Convert the solution to a PyTorch tensor on the cpu.
if isinstance(solution, torch.Tensor):
solution = solution.to("cpu")
elif isinstance(solution, Solution):
solution = solution.values.clone().to("cpu")
else:
solution = torch.as_tensor(solution, dtype=torch.float32, device="cpu")
if isinstance(solution, ReadOnlyTensor):
solution = solution.as_subclass(torch.Tensor)
policy = self.to_policy(solution).to("cpu")
# Store the solution and the policy.
result = {
"solution": solution,
"policy": policy,
}
# If available, store the observation stats.
if self.observation_normalization and (self._obs_stats is not None):
result["obs_mean"] = torch.as_tensor(self._obs_stats.mean)
result["obs_stdev"] = torch.as_tensor(self._obs_stats.stdev)
result["obs_sum"] = torch.as_tensor(self._obs_stats.sum)
result["obs_sum_of_squares"] = torch.as_tensor(self._obs_stats.sum_of_squares)
# Some additional data.
result["interaction_count"] = self.interaction_count
result["episode_count"] = self.episode_count
result["time"] = datetime.now()
# If the environment is specified via a string ID, then store that ID.
if isinstance(self._env_maker, str):
result["env"] = self._env_maker
# Save the dictionary which stores the data.
with open(fname, "wb") as f:
pickle.dump(result, f)
set_episode_count(self, n)
¶
set_interaction_count(self, n)
¶
set_observation_stats(self, rs)
¶
to_policy(self, x, *, clip_actions=True)
¶
Convert the given parameter vector to a policy as a PyTorch module.
If the problem is configured to have observation normalization, the PyTorch module also contains an additional normalization layer.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Iterable |
An sequence of real numbers, containing the parameters of a policy. Can be a PyTorch tensor, a numpy array, or a SolutionVector. |
required |
clip_actions |
bool |
Whether or not to add an action clipping layer so that the generated actions will always be within an acceptable range for the environment. |
True |
Returns:
| Type | Description |
|---|---|
Module |
The policy expressed by the parameters. |
Source code in evotorch/neuroevolution/gymne.py
def to_policy(self, x: Iterable, *, clip_actions: bool = True) -> nn.Module:
"""
Convert the given parameter vector to a policy as a PyTorch module.
If the problem is configured to have observation normalization,
the PyTorch module also contains an additional normalization layer.
Args:
x: An sequence of real numbers, containing the parameters
of a policy. Can be a PyTorch tensor, a numpy array,
or a SolutionVector.
clip_actions: Whether or not to add an action clipping layer so
that the generated actions will always be within an
acceptable range for the environment.
Returns:
The policy expressed by the parameters.
"""
policy = self.make_net(x)
if self.observation_normalization and (self._obs_stats.count > 0):
policy = ObsNormWrapperModule(policy, self._obs_stats)
if clip_actions and isinstance(self._get_env().action_space, gym.spaces.Box):
policy = ActClipWrapperModule(policy, self._get_env().action_space)
return policy
update_observation_stats(self, rs)
¶
visualize(self, policy, *, update_stats=False, num_episodes=1, decrease_rewards_by=None)
¶
Evaluate the policy and render its actions in the environment.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
policy |
Union[torch.nn.modules.module.Module, Iterable] |
The policy to be evaluated. This can be a torch module or a sequence of real numbers representing the parameters of a policy network. |
required |
update_stats |
bool |
Whether or not to update the observation normalization data while running the policy. If observation normalization is not enabled, then this argument will be ignored. |
False |
num_episodes |
Optional[int] |
Over how many episodes will the policy be evaluated.
Expected as None (which is the default), or as an integer.
If given as None, then the |
1 |
decrease_rewards_by |
Optional[float] |
How much each reward value should be
decreased. If left as None, the |
None |
Returns:
| Type | Description |
|---|---|
dict |
A dictionary containing the score and the timestep count. |
Source code in evotorch/neuroevolution/gymne.py
def visualize(
self,
policy: Union[nn.Module, Iterable],
*,
update_stats: bool = False,
num_episodes: Optional[int] = 1,
decrease_rewards_by: Optional[float] = None,
) -> dict:
"""
Evaluate the policy and render its actions in the environment.
Args:
policy: The policy to be evaluated. This can be a torch module
or a sequence of real numbers representing the parameters
of a policy network.
update_stats: Whether or not to update the observation
normalization data while running the policy. If observation
normalization is not enabled, then this argument will be
ignored.
num_episodes: Over how many episodes will the policy be evaluated.
Expected as None (which is the default), or as an integer.
If given as None, then the `num_episodes` value that was given
while initializing this GymNE will be used.
decrease_rewards_by: How much each reward value should be
decreased. If left as None, the `decrease_rewards_by` value
value that was given while initializing this GymNE will be
used.
Returns:
A dictionary containing the score and the timestep count.
"""
return self.run(
policy=policy,
update_stats=update_stats,
visualize=True,
num_episodes=num_episodes,
decrease_rewards_by=decrease_rewards_by,
)
neproblem
¶
This namespace contains the NeuroevolutionProblem class.
NEProblem (BaseNEProblem)
¶
Base class for neuro-evolution problems where the goal is to optimize the parameters of a neural network represented as a PyTorch module.
Any problem inheriting from this class is expected to override the method
_evaluate_network(self, net: torch.nn.Module) -> Union[torch.Tensor, float]
where net is the neural network to be evaluated, and the return value
is a scalar or a vector (for multi-objective cases) expressing the
fitness value(s).
Alternatively, this class can be directly instantiated in the following way:
def f(module: MyTorchModuleClass) -> Union[float, torch.Tensor, tuple]:
# Evaluate the given PyTorch module here
fitness = ...
return fitness
problem = NEProblem("min", MyTorchModuleClass, f, ...)
which specifies that the problem's goal is to minimize the return of the
function f.
For multi-objective cases, the fitness returned by f is expected as a
1-dimensional tensor. For when the problem has additional evaluation data,
a two-element tuple can be returned by f instead, where the first
element is the fitness value(s) and the second element is a 1-dimensional
tensor storing the additional data.
Source code in evotorch/neuroevolution/neproblem.py
class NEProblem(BaseNEProblem):
"""
Base class for neuro-evolution problems where the goal is to optimize the
parameters of a neural network represented as a PyTorch module.
Any problem inheriting from this class is expected to override the method
`_evaluate_network(self, net: torch.nn.Module) -> Union[torch.Tensor, float]`
where `net` is the neural network to be evaluated, and the return value
is a scalar or a vector (for multi-objective cases) expressing the
fitness value(s).
Alternatively, this class can be directly instantiated in the following
way:
```python
def f(module: MyTorchModuleClass) -> Union[float, torch.Tensor, tuple]:
# Evaluate the given PyTorch module here
fitness = ...
return fitness
problem = NEProblem("min", MyTorchModuleClass, f, ...)
```
which specifies that the problem's goal is to minimize the return of the
function `f`.
For multi-objective cases, the fitness returned by `f` is expected as a
1-dimensional tensor. For when the problem has additional evaluation data,
a two-element tuple can be returned by `f` instead, where the first
element is the fitness value(s) and the second element is a 1-dimensional
tensor storing the additional data.
"""
def __init__(
self,
objective_sense: ObjectiveSense,
network: Union[str, nn.Module, Callable[[], nn.Module]],
network_eval_func: Optional[Callable] = None,
*,
network_args: Optional[dict] = None,
initial_bounds: Optional[BoundsPairLike] = (-0.00001, 0.00001),
eval_dtype: Optional[DType] = None,
eval_data_length: int = 0,
seed: Optional[int] = None,
num_actors: Optional[Union[int, str]] = "num_devices",
actor_config: Optional[dict] = None,
num_gpus_per_actor: Optional[Union[int, float, str]] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
device: Optional[Device] = None,
):
"""
`__init__(...)`: Initialize the NEProblem.
Args:
objective_sense: The objective sense, expected as "min" or "max"
for single-objective cases, or as a sequence of strings
(each string being "min" or "max") for multi-objective cases.
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
network_eval_func: Optionally a function (or any Callable object)
which receives a PyTorch module as its argument, and returns
either a fitness, or a two-element tuple containing the fitness
and the additional evaluation data. The fitness can be a scalar
(for single-objective cases) or a 1-dimensional tensor (for
multi-objective cases). The additional evaluation data is
expected as a 1-dimensional tensor.
If this argument is left as None, it will be expected that
the method `_evaluate_network(...)` is overriden by the
inheriting class.
network_args: Optionally a dict-like object, storing keyword
arguments to be passed to the network while instantiating it.
initial_bounds: Specifies an interval from which the values of the
initial neural network parameters will be drawn.
eval_dtype: dtype to be used for fitnesses. If not specified, then
`eval_dtype` will be inferred from the dtype of the parameters
of the neural network.
In more details, if the neural network's parameters have a
float dtype, `eval_dtype` will be a compatible float.
Otherwise, it will be "float32".
eval_data_length: Length of the extra evaluation data.
seed: Random number seed. If left as None, this NEProblem instance
will not have its own random generator, and the global random
generator of PyTorch will be used instead.
num_actors: Number of actors to create for parallelized
evaluation of the solutions.
Certain string values are also accepted.
When given as "max" or as "num_cpus", the number of actors
will be equal to the number of all available CPUs in the ray
cluster.
When given as "num_gpus", the number of actors will be
equal to the number of all available GPUs in the ray
cluster, and each actor will be assigned a GPU.
When given as "num_devices", the number of actors will be
equal to the minimum among the number of CPUs and the number
of GPUs available in the cluster (or will be equal to the
number of CPUs if there is no GPU), and each actor will be
assigned a GPU (if available).
If `num_actors` is given as "num_gpus" or "num_devices",
the argument `num_gpus_per_actor` must not be used,
and the `actor_config` dictionary must not contain the
key "num_gpus".
If `num_actors` is given as something other than "num_gpus"
or "num_devices", and if you wish to assign GPUs to each
actor, then please see the argument `num_gpus_per_actor`.
actor_config: A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass `actor_config=dict(num_gpus=1)`.
Can also be given as None (which is the default),
if no such options are to be passed.
num_gpus_per_actor: Number of GPUs to be allocated by each
remote actor.
The default behavior is to NOT allocate any GPU at all
(which is the default behavior of the ray library as well).
When given as a number `n`, each actor will be given
`n` GPUs (where `n` can be an integer, or can be a `float`
for fractional allocation).
When given as a string "max", then the available GPUs
across the entire ray cluster (or within the local computer
in the simplest cases) will be equally distributed among
the actors.
When given as a string "all", then each actor will have
access to all the GPUs (this will be achieved by suppressing
the environment variable `CUDA_VISIBLE_DEVICES` for each
actor).
When the problem is not distributed (i.e. when there are
no actors), this argument is expected to be left as None.
num_subbatches: If `num_subbatches` is None (assuming that
`subbatch_size` is also None), then, when evaluating a
population, the population will be split into n pieces, `n`
being the number of actors, and each actor will evaluate
its assigned piece. If `num_subbatches` is an integer `m`,
then the population will be split into `m` pieces,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
how many sub-batches will be generated, and therefore,
how many gradients will be computed by the remote actors.
subbatch_size: If `subbatch_size` is None (assuming that
`num_subbatches` is also None), then, when evaluating a
population, the population will be split into `n` pieces, `n`
being the number of actors, and each actor will evaluate its
assigned piece. If `subbatch_size` is an integer `m`,
then the population will be split into pieces of size `m`,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
When there can be significant difference across the solutions
in terms of computational requirements, specifying a
`subbatch_size` can be beneficial, because, while one
actor is busy with a subbatch containing computationally
challenging solutions, other actors can accept more
tasks and save time.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
the size of a sub-batch (or sub-population) sampled by a
remote actor for computing a gradient.
In distributed mode, it is expected that the population size
is divisible by `subbatch_size`.
device: Default device in which a new population will be generated
and the neural networks will operate.
If not specified, "cpu" will be used.
"""
# Set the main device of the problem
# Although the operation of setting the main device is done by the main Problem class,
# here we need this at an earlier stage.
if device is None:
device = "cpu"
self._device = torch.device(device)
# Set the network
self._original_network = network
self._network_args = {} if network_args is None else deepcopy(network_args)
if isinstance(self._original_network, nn.Module):
self._original_network = self._original_network.cpu()
# Store the function that will evaluate the network, if available
self._network_eval_func: Optional[Callable] = network_eval_func
self.instantiated_network: nn.Module = None
# Create temporary network
temp_network = self._instantiate_net(self._original_network, device="cpu")
super().__init__(
objective_sense=objective_sense,
initial_bounds=initial_bounds,
bounds=None, # Neuroevolution is an unbounded problem
solution_length=count_parameters(temp_network), # The solution length is inherited from the network passed
dtype=next(temp_network.parameters()).dtype, # The datatype is inherited from the network passed
eval_dtype=eval_dtype,
device=device,
eval_data_length=eval_data_length,
seed=seed,
num_actors=num_actors,
num_gpus_per_actor=num_gpus_per_actor,
actor_config=actor_config,
num_subbatches=num_subbatches,
subbatch_size=subbatch_size,
store_solution_stats=None,
)
@property
def network_device(self) -> Device:
"""The device on which the problem should place data e.g. the network"""
cpu_device = torch.device("cpu")
if self.is_main:
# This is the case where this is the main process (not a remote actor)
if self.device == cpu_device:
# If the main device of the problem is "cpu", then we assume that the network is going to be on the cpu as well
return cpu_device
else:
# If the main device of the problem is some other device, then it is that device into which the network will be put
return self.device
else:
# If this is a remote actor, then the network will be put into the auxiliary device allocated for that actor
return self.aux_device
@property
def _str_network_constants(self) -> dict:
"""
Named constants which will be passed to `str_to_net`.
To be overridden by the user for custom fixed constants for a problem.
"""
return {}
@property
def _network_constants(self) -> dict:
"""
Named constants which will be passed to the network instantiation.
To be overridden by the user for custom fixed constants for a problem.
"""
return {}
def network_constants(self) -> dict:
"""Named constants which can be passed to the network instantiation"""
constants = {}
constants.update(self._network_constants)
constants.update(self._network_args)
return constants
@property
def _nonserialized_attribs(self) -> List[str]:
return ["instantiated_network"]
def _instantiate_net(self, network: Union[str, nn.Module, dict], device: Optional[Device] = None) -> nn.Module:
"""Instantiate the network on the target device, to be overridden by the user for custom behaviour
Returns:
instantiated_network (nn.Module): The network instantiated on the target device
"""
# Branching point determines instantiation of network
if isinstance(network, str):
# Passed argument was a string representation of a torch module
net_consts = {}
net_consts.update(self.network_constants())
net_consts.update(self._str_network_constants)
instantiated_network = str_to_net(network, **net_consts)
elif isinstance(network, nn.Module):
# Passed argument was directly a torch module
instantiated_network = network
else:
# Passed argument was callable yielding network
instantiated_network = pass_info_if_needed(network, self._network_constants)(**self._network_args)
# Map to device
device = self.network_device if device is None else device
instantiated_network = instantiated_network.to(device)
return instantiated_network
def _prepare(self) -> None:
"""Instantiate the network on the target device, if not already done"""
self.instantiated_network = self._instantiate_net(self._original_network)
# Clear reference to original network
self._original_network = None
def make_net(self, parameters: Iterable) -> nn.Module:
"""
Make a new network filled with the provided parameters.
Args:
parameters: Parameters to be used as weights within the network.
Can be a Solution, or any 1-dimensional Iterable that can be
converted to a PyTorch tensor.
Returns:
A new network, as a `torch.Module` instance.
"""
if isinstance(parameters, Solution):
parameters = parameters.access_values(keep_evals=True)
else:
parameters = self.as_tensor(parameters)
with torch.no_grad():
net = deepcopy(self.parameterize_net(parameters))
return net
def parameterize_net(self, parameters: torch.Tensor) -> nn.Module:
"""Parameterize the network with a given set of parameters.
Args:
parameters (torch.Tensor): The parameters with which to instantiate the network
Returns:
instantiated_network (nn.Module): The network instantiated with the parameters
"""
# Check if network exists
if self.instantiated_network is None:
self.instantiated_network = self._instantiate_net(self._original_network)
network = self.instantiated_network
# Move the parameters if needed
if parameters.device != self.network_device:
parameters = parameters.to(self.network_device)
# Fill the network with the parameters
fill_parameters(network, parameters)
# Return the network
return network
@property
def _grad_device(self) -> Device:
"""
Get the device in which new solutions will be made in distributed mode.
In more details, in distributed mode, each actor creates its own
sub-populations, evaluates them, and computes its own gradient
(all such actor gradients eventually being collected by the
distribution-based search algorithm in the main process).
For some problem types, it can make sense for the remote actors to
create their temporary sub-populations on another device
(e.g. on the GPU that is allocated specifically for them).
For such situations, one is encouraged to override this property
and make it return whatever device is to be used.
In the case of NEProblem, this property returns whatever device
is specified by the property `network_device`.
"""
return self.network_device
def _evaluate_network(self, network: nn.Module) -> Union[float, torch.Tensor, tuple]:
"""
Evaluate a network and return the evaluation result(s).
In the case where the `__init__` of `NEProblem` was not given
a network evaluator function (via the argument `network_eval_func`),
it will be expected that the inheriting class overrides this
method and defines how a network should be evaluated.
Args:
network (nn.Module): The network to evaluate
Returns:
fitness: The networks' fitness value(s), as a scalar for
single-objective cases, or as a 1-dimensional tensor
for multi-objective cases. The returned value can also
be a two-element tuple where the first element is the
fitness (as a scalar or as a vector) and the second
element is a 1-dimensional vector storing the extra
evaluation data.
"""
raise NotImplementedError
def _evaluate(self, solution: Solution):
"""
Evaluate a single solution.
This is achieved by parameterising the problem's attribute
named `instantiated_network`, and then evaluating the network
with the method `_evaluate_network(...)`.
Args:
solution (Solution): The solution to evaluate.
"""
parameters = solution.values
if self._network_eval_func is None:
evaluator = self._evaluate_network
else:
evaluator = self._network_eval_func
fitnesses = evaluator(self.parameterize_net(parameters))
if isinstance(fitnesses, tuple):
solution.set_evals(*fitnesses)
else:
solution.set_evals(fitnesses)
network_device: Union[str, torch.device]
property
readonly
¶
The device on which the problem should place data e.g. the network
__init__(self, objective_sense, network, network_eval_func=None, *, network_args=None, initial_bounds=(-1e-05, 1e-05), eval_dtype=None, eval_data_length=0, seed=None, num_actors='num_devices', actor_config=None, num_gpus_per_actor=None, num_subbatches=None, subbatch_size=None, device=None)
special
¶
__init__(...): Initialize the NEProblem.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
objective_sense |
Union[str, Iterable[str]] |
The objective sense, expected as "min" or "max" for single-objective cases, or as a sequence of strings (each string being "min" or "max") for multi-objective cases. |
required |
network |
Union[str, torch.nn.modules.module.Module, Callable[[], torch.nn.modules.module.Module]] |
A network structure string, or a Callable (which can be
a class inheriting from |
required |
network_eval_func |
Optional[Callable] |
Optionally a function (or any Callable object)
which receives a PyTorch module as its argument, and returns
either a fitness, or a two-element tuple containing the fitness
and the additional evaluation data. The fitness can be a scalar
(for single-objective cases) or a 1-dimensional tensor (for
multi-objective cases). The additional evaluation data is
expected as a 1-dimensional tensor.
If this argument is left as None, it will be expected that
the method |
None |
network_args |
Optional[dict] |
Optionally a dict-like object, storing keyword arguments to be passed to the network while instantiating it. |
None |
initial_bounds |
Union[Iterable[Union[float, Iterable[float], torch.Tensor]], evotorch.core.BoundsPair] |
Specifies an interval from which the values of the initial neural network parameters will be drawn. |
(-1e-05, 1e-05) |
eval_dtype |
Union[str, torch.dtype, numpy.dtype, Type] |
dtype to be used for fitnesses. If not specified, then
|
None |
eval_data_length |
int |
Length of the extra evaluation data. |
0 |
seed |
Optional[int] |
Random number seed. If left as None, this NEProblem instance will not have its own random generator, and the global random generator of PyTorch will be used instead. |
None |
num_actors |
Union[int, str] |
Number of actors to create for parallelized
evaluation of the solutions.
Certain string values are also accepted.
When given as "max" or as "num_cpus", the number of actors
will be equal to the number of all available CPUs in the ray
cluster.
When given as "num_gpus", the number of actors will be
equal to the number of all available GPUs in the ray
cluster, and each actor will be assigned a GPU.
When given as "num_devices", the number of actors will be
equal to the minimum among the number of CPUs and the number
of GPUs available in the cluster (or will be equal to the
number of CPUs if there is no GPU), and each actor will be
assigned a GPU (if available).
If |
'num_devices' |
actor_config |
Optional[dict] |
A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass |
None |
num_gpus_per_actor |
Union[int, float, str] |
Number of GPUs to be allocated by each
remote actor.
The default behavior is to NOT allocate any GPU at all
(which is the default behavior of the ray library as well).
When given as a number |
None |
num_subbatches |
Optional[int] |
If |
None |
subbatch_size |
Optional[int] |
If |
None |
device |
Union[str, torch.device] |
Default device in which a new population will be generated and the neural networks will operate. If not specified, "cpu" will be used. |
None |
Source code in evotorch/neuroevolution/neproblem.py
def __init__(
self,
objective_sense: ObjectiveSense,
network: Union[str, nn.Module, Callable[[], nn.Module]],
network_eval_func: Optional[Callable] = None,
*,
network_args: Optional[dict] = None,
initial_bounds: Optional[BoundsPairLike] = (-0.00001, 0.00001),
eval_dtype: Optional[DType] = None,
eval_data_length: int = 0,
seed: Optional[int] = None,
num_actors: Optional[Union[int, str]] = "num_devices",
actor_config: Optional[dict] = None,
num_gpus_per_actor: Optional[Union[int, float, str]] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
device: Optional[Device] = None,
):
"""
`__init__(...)`: Initialize the NEProblem.
Args:
objective_sense: The objective sense, expected as "min" or "max"
for single-objective cases, or as a sequence of strings
(each string being "min" or "max") for multi-objective cases.
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
network_eval_func: Optionally a function (or any Callable object)
which receives a PyTorch module as its argument, and returns
either a fitness, or a two-element tuple containing the fitness
and the additional evaluation data. The fitness can be a scalar
(for single-objective cases) or a 1-dimensional tensor (for
multi-objective cases). The additional evaluation data is
expected as a 1-dimensional tensor.
If this argument is left as None, it will be expected that
the method `_evaluate_network(...)` is overriden by the
inheriting class.
network_args: Optionally a dict-like object, storing keyword
arguments to be passed to the network while instantiating it.
initial_bounds: Specifies an interval from which the values of the
initial neural network parameters will be drawn.
eval_dtype: dtype to be used for fitnesses. If not specified, then
`eval_dtype` will be inferred from the dtype of the parameters
of the neural network.
In more details, if the neural network's parameters have a
float dtype, `eval_dtype` will be a compatible float.
Otherwise, it will be "float32".
eval_data_length: Length of the extra evaluation data.
seed: Random number seed. If left as None, this NEProblem instance
will not have its own random generator, and the global random
generator of PyTorch will be used instead.
num_actors: Number of actors to create for parallelized
evaluation of the solutions.
Certain string values are also accepted.
When given as "max" or as "num_cpus", the number of actors
will be equal to the number of all available CPUs in the ray
cluster.
When given as "num_gpus", the number of actors will be
equal to the number of all available GPUs in the ray
cluster, and each actor will be assigned a GPU.
When given as "num_devices", the number of actors will be
equal to the minimum among the number of CPUs and the number
of GPUs available in the cluster (or will be equal to the
number of CPUs if there is no GPU), and each actor will be
assigned a GPU (if available).
If `num_actors` is given as "num_gpus" or "num_devices",
the argument `num_gpus_per_actor` must not be used,
and the `actor_config` dictionary must not contain the
key "num_gpus".
If `num_actors` is given as something other than "num_gpus"
or "num_devices", and if you wish to assign GPUs to each
actor, then please see the argument `num_gpus_per_actor`.
actor_config: A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass `actor_config=dict(num_gpus=1)`.
Can also be given as None (which is the default),
if no such options are to be passed.
num_gpus_per_actor: Number of GPUs to be allocated by each
remote actor.
The default behavior is to NOT allocate any GPU at all
(which is the default behavior of the ray library as well).
When given as a number `n`, each actor will be given
`n` GPUs (where `n` can be an integer, or can be a `float`
for fractional allocation).
When given as a string "max", then the available GPUs
across the entire ray cluster (or within the local computer
in the simplest cases) will be equally distributed among
the actors.
When given as a string "all", then each actor will have
access to all the GPUs (this will be achieved by suppressing
the environment variable `CUDA_VISIBLE_DEVICES` for each
actor).
When the problem is not distributed (i.e. when there are
no actors), this argument is expected to be left as None.
num_subbatches: If `num_subbatches` is None (assuming that
`subbatch_size` is also None), then, when evaluating a
population, the population will be split into n pieces, `n`
being the number of actors, and each actor will evaluate
its assigned piece. If `num_subbatches` is an integer `m`,
then the population will be split into `m` pieces,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
how many sub-batches will be generated, and therefore,
how many gradients will be computed by the remote actors.
subbatch_size: If `subbatch_size` is None (assuming that
`num_subbatches` is also None), then, when evaluating a
population, the population will be split into `n` pieces, `n`
being the number of actors, and each actor will evaluate its
assigned piece. If `subbatch_size` is an integer `m`,
then the population will be split into pieces of size `m`,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
When there can be significant difference across the solutions
in terms of computational requirements, specifying a
`subbatch_size` can be beneficial, because, while one
actor is busy with a subbatch containing computationally
challenging solutions, other actors can accept more
tasks and save time.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
the size of a sub-batch (or sub-population) sampled by a
remote actor for computing a gradient.
In distributed mode, it is expected that the population size
is divisible by `subbatch_size`.
device: Default device in which a new population will be generated
and the neural networks will operate.
If not specified, "cpu" will be used.
"""
# Set the main device of the problem
# Although the operation of setting the main device is done by the main Problem class,
# here we need this at an earlier stage.
if device is None:
device = "cpu"
self._device = torch.device(device)
# Set the network
self._original_network = network
self._network_args = {} if network_args is None else deepcopy(network_args)
if isinstance(self._original_network, nn.Module):
self._original_network = self._original_network.cpu()
# Store the function that will evaluate the network, if available
self._network_eval_func: Optional[Callable] = network_eval_func
self.instantiated_network: nn.Module = None
# Create temporary network
temp_network = self._instantiate_net(self._original_network, device="cpu")
super().__init__(
objective_sense=objective_sense,
initial_bounds=initial_bounds,
bounds=None, # Neuroevolution is an unbounded problem
solution_length=count_parameters(temp_network), # The solution length is inherited from the network passed
dtype=next(temp_network.parameters()).dtype, # The datatype is inherited from the network passed
eval_dtype=eval_dtype,
device=device,
eval_data_length=eval_data_length,
seed=seed,
num_actors=num_actors,
num_gpus_per_actor=num_gpus_per_actor,
actor_config=actor_config,
num_subbatches=num_subbatches,
subbatch_size=subbatch_size,
store_solution_stats=None,
)
make_net(self, parameters)
¶
Make a new network filled with the provided parameters.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
parameters |
Iterable |
Parameters to be used as weights within the network. Can be a Solution, or any 1-dimensional Iterable that can be converted to a PyTorch tensor. |
required |
Returns:
| Type | Description |
|---|---|
Module |
A new network, as a |
Source code in evotorch/neuroevolution/neproblem.py
def make_net(self, parameters: Iterable) -> nn.Module:
"""
Make a new network filled with the provided parameters.
Args:
parameters: Parameters to be used as weights within the network.
Can be a Solution, or any 1-dimensional Iterable that can be
converted to a PyTorch tensor.
Returns:
A new network, as a `torch.Module` instance.
"""
if isinstance(parameters, Solution):
parameters = parameters.access_values(keep_evals=True)
else:
parameters = self.as_tensor(parameters)
with torch.no_grad():
net = deepcopy(self.parameterize_net(parameters))
return net
network_constants(self)
¶
Named constants which can be passed to the network instantiation
parameterize_net(self, parameters)
¶
Parameterize the network with a given set of parameters.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
parameters |
torch.Tensor |
The parameters with which to instantiate the network |
required |
Returns:
| Type | Description |
|---|---|
instantiated_network (nn.Module) |
The network instantiated with the parameters |
Source code in evotorch/neuroevolution/neproblem.py
def parameterize_net(self, parameters: torch.Tensor) -> nn.Module:
"""Parameterize the network with a given set of parameters.
Args:
parameters (torch.Tensor): The parameters with which to instantiate the network
Returns:
instantiated_network (nn.Module): The network instantiated with the parameters
"""
# Check if network exists
if self.instantiated_network is None:
self.instantiated_network = self._instantiate_net(self._original_network)
network = self.instantiated_network
# Move the parameters if needed
if parameters.device != self.network_device:
parameters = parameters.to(self.network_device)
# Fill the network with the parameters
fill_parameters(network, parameters)
# Return the network
return network
net
special
¶
Utility classes and functions for neural networks
functional
¶
ModuleExpectingFlatParameters
¶
A wrapper which brings a functional interface around a torch module.
For obtaining the functional interface, this class internally uses
the functorch library.
Similar to functorch.FunctionalModule, ModuleExpectingFlatParameters
turns a torch.nn.Module instance to a function which expects a new
leftmost argument representing the parameters of the network.
Unlike functorch.FunctionalModule, a ModuleExpectingFlatParameters
instance, as its name suggests, expects the network parameters to be
given as a 1-dimensional (i.e. flattened) tensor.
Also, unlike functorch.FunctionalModule, an instance of
ModuleExpectingFlatParameters is NOT an instance of torch.nn.Module.
PyTorch modules with buffers can be wrapped by this class, but it is assumed that those buffers are constant. If the wrapped module changes the value(s) of its buffer(s) during its forward passes, most probably things will NOT work right.
As an example, let us consider the following linear layer.
The functional counterpart of net can be obtained via:
from evotorch.neuroevolution.net import ModuleExpectingFlatParameters
fnet = ModuleExpectingFlatParameters(net)
Now, fnet is a callable object which expects network parameters
and network inputs. Let us call fnet with randomly generated network
parameters and with a randomly generated input tensor.
param_length = fnet.parameter_length
random_parameters = torch.randn(param_length)
random_input = torch.randn(3)
result = fnet(random_parameters, random_input)
Source code in evotorch/neuroevolution/net/functional.py
class ModuleExpectingFlatParameters:
"""
A wrapper which brings a functional interface around a torch module.
For obtaining the functional interface, this class internally uses
the `functorch` library.
Similar to `functorch.FunctionalModule`, `ModuleExpectingFlatParameters`
turns a `torch.nn.Module` instance to a function which expects a new
leftmost argument representing the parameters of the network.
Unlike `functorch.FunctionalModule`, a `ModuleExpectingFlatParameters`
instance, as its name suggests, expects the network parameters to be
given as a 1-dimensional (i.e. flattened) tensor.
Also, unlike `functorch.FunctionalModule`, an instance of
`ModuleExpectingFlatParameters` is NOT an instance of `torch.nn.Module`.
PyTorch modules with buffers can be wrapped by this class, but it is
assumed that those buffers are constant. If the wrapped module changes
the value(s) of its buffer(s) during its forward passes, most probably
things will NOT work right.
As an example, let us consider the following linear layer.
```python
import torch
from torch import nn
net = nn.Linear(3, 8)
```
The functional counterpart of `net` can be obtained via:
```python
from evotorch.neuroevolution.net import ModuleExpectingFlatParameters
fnet = ModuleExpectingFlatParameters(net)
```
Now, `fnet` is a callable object which expects network parameters
and network inputs. Let us call `fnet` with randomly generated network
parameters and with a randomly generated input tensor.
```python
param_length = fnet.parameter_length
random_parameters = torch.randn(param_length)
random_input = torch.randn(3)
result = fnet(random_parameters, random_input)
```
"""
def __init__(self, module: nn.Module, disable_autograd_tracking: bool = False):
# Declare the variables which will store information regarding the parameters of the module.
self.__param_shapes = []
self.__param_length = 0
self.__param_slices = []
self.__num_params = 0
self.__buffers = []
# Iterate over the parameters of the module and fill the related information.
i = 0
j = 0
for p in module.parameters():
shape = p.shape
self.__param_shapes.append(shape)
length = _shape_length(shape)
self.__param_length += length
j = i + length
self.__param_slices.append(slice(i, j))
i = j
self.__num_params += 1
self.__fmodel, _, self.__buffers = make_functional_with_buffers(
module, disable_autograd_tracking=bool(disable_autograd_tracking)
)
self.__buffers = list(self.__buffers)
def __transfer_buffers(self, x: torch.Tensor):
"""
Transfer the buffer tensors to the device of the given tensor.
Args:
x: The tensor whose device will also store the buffer tensors.
"""
n = len(self.__buffers)
for i in range(n):
self.__buffers[i] = torch.as_tensor(self.__buffers[i], device=x.device)
@property
def buffers(self) -> tuple:
"""Get the stored buffers"""
return self.__buffers
@property
def parameter_length(self) -> int:
return self.__param_length
def __call__(self, parameter_vector: torch.Tensor, x: torch.Tensor, h: Any = None) -> Any:
"""
Call the wrapped module's forward pass procedure.
Args:
parameter_vector: A 1-dimensional tensor which represents the
parameters of the tensor.
x: The inputs.
h: Hidden state(s), in case this is a recurrent network.
Returns:
The result of the forward pass.
"""
if parameter_vector.ndim != 1:
raise ValueError(
f"Expected the parameters as 1 dimensional,"
f" but the received parameter vector has {parameter_vector.ndim} dimensions"
)
if len(parameter_vector) != self.__param_length:
raise ValueError(
f"Expected a parameter vector of length {self.__param_length},"
f" but the received parameter vector's length is {len(parameter_vector)}."
)
state_args = [] if h is None else [h]
params = []
for i in range(self.__num_params):
param_slice = self.__param_slices[i]
param_shape = self.__param_shapes[i]
param = parameter_vector[param_slice].reshape(param_shape)
params.append(param)
# Make sure that the tensors are in the same device with x
self.__transfer_buffers(x)
# Run the functional module and return the results
return self.__fmodel(params, self.__buffers, x, *state_args)
buffers: tuple
property
readonly
¶
Get the stored buffers
__call__(self, parameter_vector, x, h=None)
special
¶
Call the wrapped module's forward pass procedure.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
parameter_vector |
Tensor |
A 1-dimensional tensor which represents the parameters of the tensor. |
required |
x |
Tensor |
The inputs. |
required |
h |
Any |
Hidden state(s), in case this is a recurrent network. |
None |
Returns:
| Type | Description |
|---|---|
Any |
The result of the forward pass. |
Source code in evotorch/neuroevolution/net/functional.py
def __call__(self, parameter_vector: torch.Tensor, x: torch.Tensor, h: Any = None) -> Any:
"""
Call the wrapped module's forward pass procedure.
Args:
parameter_vector: A 1-dimensional tensor which represents the
parameters of the tensor.
x: The inputs.
h: Hidden state(s), in case this is a recurrent network.
Returns:
The result of the forward pass.
"""
if parameter_vector.ndim != 1:
raise ValueError(
f"Expected the parameters as 1 dimensional,"
f" but the received parameter vector has {parameter_vector.ndim} dimensions"
)
if len(parameter_vector) != self.__param_length:
raise ValueError(
f"Expected a parameter vector of length {self.__param_length},"
f" but the received parameter vector's length is {len(parameter_vector)}."
)
state_args = [] if h is None else [h]
params = []
for i in range(self.__num_params):
param_slice = self.__param_slices[i]
param_shape = self.__param_shapes[i]
param = parameter_vector[param_slice].reshape(param_shape)
params.append(param)
# Make sure that the tensors are in the same device with x
self.__transfer_buffers(x)
# Run the functional module and return the results
return self.__fmodel(params, self.__buffers, x, *state_args)
make_functional_module(net)
¶
Wrap a torch module so that it has a functional interface.
For obtaining a functional interface, this function internally uses the
functorch library.
Similar to functorch.make_functional(...), this function turns a
torch.nn.Module instance to a function which expects a new leftmost
argument representing the parameters of the network.
Unlike with functorch.make_functional(...), the parameters of the
network are expected in a 1-dimensional (i.e. flattened) tensor.
PyTorch modules with buffers can be wrapped by this class, but it is assumed that those buffers are constant. If the wrapped module changes the value(s) of its buffer(s) during its forward passes, most probably things will NOT work right.
As an example, let us consider the following linear layer.
The functional counterpart of net can be obtained via:
Now, fnet is a callable object which expects network parameters
and network inputs. Let us call fnet with randomly generated network
parameters and with a randomly generated input tensor.
param_length = fnet.parameter_length
random_parameters = torch.randn(param_length)
random_input = torch.randn(3)
result = fnet(random_parameters, random_input)
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
net |
Module |
The |
required |
Returns:
| Type | Description |
|---|---|
ModuleExpectingFlatParameters |
The functional wrapper, as an instance of
|
Source code in evotorch/neuroevolution/net/functional.py
def make_functional_module(net: nn.Module) -> ModuleExpectingFlatParameters:
"""
Wrap a torch module so that it has a functional interface.
For obtaining a functional interface, this function internally uses the
`functorch` library.
Similar to `functorch.make_functional(...)`, this function turns a
`torch.nn.Module` instance to a function which expects a new leftmost
argument representing the parameters of the network.
Unlike with `functorch.make_functional(...)`, the parameters of the
network are expected in a 1-dimensional (i.e. flattened) tensor.
PyTorch modules with buffers can be wrapped by this class, but it is
assumed that those buffers are constant. If the wrapped module changes
the value(s) of its buffer(s) during its forward passes, most probably
things will NOT work right.
As an example, let us consider the following linear layer.
```python
import torch
from torch import nn
net = nn.Linear(3, 8)
```
The functional counterpart of `net` can be obtained via:
```python
from evotorch.neuroevolution.net import make_functional_module
fnet = make_functional_module(net)
```
Now, `fnet` is a callable object which expects network parameters
and network inputs. Let us call `fnet` with randomly generated network
parameters and with a randomly generated input tensor.
```python
param_length = fnet.parameter_length
random_parameters = torch.randn(param_length)
random_input = torch.randn(3)
result = fnet(random_parameters, random_input)
```
Args:
net: The `torch.nn.Module` instance to be wrapped by a functional
interface.
Returns:
The functional wrapper, as an instance of
`evotorch.neuroevolution.net.ModuleExpectingFlatParameters`.
"""
return ModuleExpectingFlatParameters(net)
layers
¶
Various neural network layer types
Apply (Module)
¶
A torch module for applying an arithmetic operator on an input tensor
Source code in evotorch/neuroevolution/net/layers.py
class Apply(nn.Module):
"""A torch module for applying an arithmetic operator on an input tensor"""
def __init__(self, operator: str, argument: float):
"""`__init__(...)`: Initialize the Apply module.
Args:
operator: Must be '+', '-', '*', '/', or '**'.
Indicates which operation will be done
on the input tensor.
argument: Expected as a float, represents
the right-argument of the operation
(the left-argument being the input
tensor).
"""
nn.Module.__init__(self)
self._operator = str(operator)
assert self._operator in ("+", "-", "*", "/", "**")
self._argument = float(argument)
def forward(self, x):
op = self._operator
arg = self._argument
if op == "+":
return x + arg
elif op == "-":
return x - arg
elif op == "*":
return x * arg
elif op == "/":
return x / arg
elif op == "**":
return x**arg
else:
raise ValueError("Unknown operator:" + repr(op))
def extra_repr(self):
return "operator={}, argument={}".format(repr(self._operator), self._argument)
__init__(self, operator, argument)
special
¶
__init__(...): Initialize the Apply module.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
operator |
str |
Must be '+', '-', '', '/', or '*'. Indicates which operation will be done on the input tensor. |
required |
argument |
float |
Expected as a float, represents the right-argument of the operation (the left-argument being the input tensor). |
required |
Source code in evotorch/neuroevolution/net/layers.py
def __init__(self, operator: str, argument: float):
"""`__init__(...)`: Initialize the Apply module.
Args:
operator: Must be '+', '-', '*', '/', or '**'.
Indicates which operation will be done
on the input tensor.
argument: Expected as a float, represents
the right-argument of the operation
(the left-argument being the input
tensor).
"""
nn.Module.__init__(self)
self._operator = str(operator)
assert self._operator in ("+", "-", "*", "/", "**")
self._argument = float(argument)
extra_repr(self)
¶
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/layers.py
Bin (Module)
¶
A small torch module for binning the values of tensors.
In more details, considering a lower bound value lb, an upper bound value ub, and an input tensor x, each value within x closer to lb will be converted to lb and each value within x closer to ub will be converted to ub.
Source code in evotorch/neuroevolution/net/layers.py
class Bin(nn.Module):
"""A small torch module for binning the values of tensors.
In more details, considering a lower bound value lb,
an upper bound value ub, and an input tensor x,
each value within x closer to lb will be converted to lb
and each value within x closer to ub will be converted to ub.
"""
def __init__(self, lb: float, ub: float):
"""`__init__(...)`: Initialize the Clip operator.
Args:
lb: Lower bound
ub: Upper bound
"""
nn.Module.__init__(self)
self._lb = float(lb)
self._ub = float(ub)
self._interval_size = self._ub - self._lb
self._shrink_amount = self._interval_size / 2.0
self._shift_amount = (self._ub + self._lb) / 2.0
def forward(self, x: torch.Tensor):
x = x - self._shift_amount
x = x / self._shrink_amount
x = torch.sign(x)
x = x * self._shrink_amount
x = x + self._shift_amount
return x
def extra_repr(self):
return "lb={}, ub={}".format(self._lb, self._ub)
__init__(self, lb, ub)
special
¶
__init__(...): Initialize the Clip operator.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
lb |
float |
Lower bound |
required |
ub |
float |
Upper bound |
required |
Source code in evotorch/neuroevolution/net/layers.py
def __init__(self, lb: float, ub: float):
"""`__init__(...)`: Initialize the Clip operator.
Args:
lb: Lower bound
ub: Upper bound
"""
nn.Module.__init__(self)
self._lb = float(lb)
self._ub = float(ub)
self._interval_size = self._ub - self._lb
self._shrink_amount = self._interval_size / 2.0
self._shift_amount = (self._ub + self._lb) / 2.0
extra_repr(self)
¶
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Clip (Module)
¶
A small torch module for clipping the values of tensors
Source code in evotorch/neuroevolution/net/layers.py
class Clip(nn.Module):
"""A small torch module for clipping the values of tensors"""
def __init__(self, lb: float, ub: float):
"""`__init__(...)`: Initialize the Clip operator.
Args:
lb: Lower bound. Values less than this will be clipped.
ub: Upper bound. Values greater than this will be clipped.
"""
nn.Module.__init__(self)
self._lb = float(lb)
self._ub = float(ub)
def forward(self, x: torch.Tensor):
return x.clamp(self._lb, self._ub)
def extra_repr(self):
return "lb={}, ub={}".format(self._lb, self._ub)
__init__(self, lb, ub)
special
¶
__init__(...): Initialize the Clip operator.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
lb |
float |
Lower bound. Values less than this will be clipped. |
required |
ub |
float |
Upper bound. Values greater than this will be clipped. |
required |
Source code in evotorch/neuroevolution/net/layers.py
extra_repr(self)
¶
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
FeedForwardNet (Module)
¶
Representation of a feed forward neural network as a torch Module.
An example initialization of a FeedForwardNet is as follows:
net = drt.FeedForwardNet(4, [(8, 'tanh'), (6, 'tanh')])
which means that we would like to have a network which expects an input vector of length 4 and passes its input through 2 tanh-activated hidden layers (with neurons count 8 and 6, respectively). The output of the last hidden layer (of length 6) is the final output vector.
The string representation of the module obtained via the example above is:
FeedForwardNet(
(layer_0): Linear(in_features=4, out_features=8, bias=True)
(actfunc_0): Tanh()
(layer_1): Linear(in_features=8, out_features=6, bias=True)
(actfunc_1): Tanh()
)
Source code in evotorch/neuroevolution/net/layers.py
class FeedForwardNet(nn.Module):
"""
Representation of a feed forward neural network as a torch Module.
An example initialization of a FeedForwardNet is as follows:
net = drt.FeedForwardNet(4, [(8, 'tanh'), (6, 'tanh')])
which means that we would like to have a network which expects an input
vector of length 4 and passes its input through 2 tanh-activated hidden
layers (with neurons count 8 and 6, respectively).
The output of the last hidden layer (of length 6) is the final
output vector.
The string representation of the module obtained via the example above
is:
FeedForwardNet(
(layer_0): Linear(in_features=4, out_features=8, bias=True)
(actfunc_0): Tanh()
(layer_1): Linear(in_features=8, out_features=6, bias=True)
(actfunc_1): Tanh()
)
"""
LengthActTuple = Tuple[int, Union[str, Callable]]
LengthActBiasTuple = Tuple[int, Union[str, Callable], Union[bool]]
def __init__(self, input_size: int, layers: List[Union[LengthActTuple, LengthActBiasTuple]]):
"""`__init__(...)`: Initialize the FeedForward network.
Args:
input_size: Input size of the network, expected as an int.
layers: Expected as a list of tuples,
where each tuple is either of the form
`(layer_size, activation_function)`
or of the form
`(layer_size, activation_function, bias)`
in which
(i) `layer_size` is an int, specifying the number of neurons;
(ii) `activation_function` is None, or a callable object,
or a string containing the name of the activation function
('relu', 'selu', 'elu', 'tanh', 'hardtanh', or 'sigmoid');
(iii) `bias` is a boolean, specifying whether the layer
is to have a bias or not.
When omitted, bias is set to True.
"""
nn.Module.__init__(self)
for i, layer in enumerate(layers):
if len(layer) == 2:
size, actfunc = layer
bias = True
elif len(layer) == 3:
size, actfunc, bias = layer
else:
assert False, "A layer tuple of invalid size is encountered"
setattr(self, "layer_" + str(i), nn.Linear(input_size, size, bias=bias))
if isinstance(actfunc, str):
if actfunc == "relu":
actfunc = nn.ReLU()
elif actfunc == "selu":
actfunc = nn.SELU()
elif actfunc == "elu":
actfunc = nn.ELU()
elif actfunc == "tanh":
actfunc = nn.Tanh()
elif actfunc == "hardtanh":
actfunc = nn.Hardtanh()
elif actfunc == "sigmoid":
actfunc = nn.Sigmoid()
elif actfunc == "round":
actfunc = Round()
else:
raise ValueError("Unknown activation function: " + repr(actfunc))
setattr(self, "actfunc_" + str(i), actfunc)
input_size = size
def forward(self, x):
i = 0
while hasattr(self, "layer_" + str(i)):
x = getattr(self, "layer_" + str(i))(x)
f = getattr(self, "actfunc_" + str(i))
if f is not None:
x = f(x)
i += 1
return x
__init__(self, input_size, layers)
special
¶
__init__(...): Initialize the FeedForward network.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
input_size |
int |
Input size of the network, expected as an int. |
required |
layers |
List[Union[Tuple[int, Union[str, Callable]], Tuple[int, Union[str, Callable], bool]]] |
Expected as a list of tuples,
where each tuple is either of the form
|
required |
Source code in evotorch/neuroevolution/net/layers.py
def __init__(self, input_size: int, layers: List[Union[LengthActTuple, LengthActBiasTuple]]):
"""`__init__(...)`: Initialize the FeedForward network.
Args:
input_size: Input size of the network, expected as an int.
layers: Expected as a list of tuples,
where each tuple is either of the form
`(layer_size, activation_function)`
or of the form
`(layer_size, activation_function, bias)`
in which
(i) `layer_size` is an int, specifying the number of neurons;
(ii) `activation_function` is None, or a callable object,
or a string containing the name of the activation function
('relu', 'selu', 'elu', 'tanh', 'hardtanh', or 'sigmoid');
(iii) `bias` is a boolean, specifying whether the layer
is to have a bias or not.
When omitted, bias is set to True.
"""
nn.Module.__init__(self)
for i, layer in enumerate(layers):
if len(layer) == 2:
size, actfunc = layer
bias = True
elif len(layer) == 3:
size, actfunc, bias = layer
else:
assert False, "A layer tuple of invalid size is encountered"
setattr(self, "layer_" + str(i), nn.Linear(input_size, size, bias=bias))
if isinstance(actfunc, str):
if actfunc == "relu":
actfunc = nn.ReLU()
elif actfunc == "selu":
actfunc = nn.SELU()
elif actfunc == "elu":
actfunc = nn.ELU()
elif actfunc == "tanh":
actfunc = nn.Tanh()
elif actfunc == "hardtanh":
actfunc = nn.Hardtanh()
elif actfunc == "sigmoid":
actfunc = nn.Sigmoid()
elif actfunc == "round":
actfunc = Round()
else:
raise ValueError("Unknown activation function: " + repr(actfunc))
setattr(self, "actfunc_" + str(i), actfunc)
input_size = size
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
LSTM (Module)
¶
Source code in evotorch/neuroevolution/net/layers.py
class LSTM(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
*,
dtype: torch.dtype = torch.float32,
device: Union[str, torch.device] = "cpu",
):
super().__init__()
input_size = int(input_size)
hidden_size = int(hidden_size)
self.input_size = input_size
self.hidden_size = hidden_size
def input_weight():
return nn.Parameter(torch.randn(self.hidden_size, self.input_size, dtype=dtype, device=device))
def weight():
return nn.Parameter(torch.randn(self.hidden_size, self.hidden_size, dtype=dtype, device=device))
def bias():
return nn.Parameter(torch.zeros(self.hidden_size, dtype=dtype, device=device))
self.W_ii = input_weight()
self.W_if = input_weight()
self.W_ig = input_weight()
self.W_io = input_weight()
self.W_hi = weight()
self.W_hf = weight()
self.W_hg = weight()
self.W_ho = weight()
self.b_ii = bias()
self.b_if = bias()
self.b_ig = bias()
self.b_io = bias()
self.b_hi = bias()
self.b_hf = bias()
self.b_hg = bias()
self.b_ho = bias()
def forward(self, x: torch.Tensor, hidden=None) -> tuple:
sigm = torch.sigmoid
tanh = torch.tanh
if hidden is None:
h_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
c_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
else:
h_prev, c_prev = hidden
i_t = sigm(self.W_ii @ x + self.b_ii + self.W_hi @ h_prev + self.b_hi)
f_t = sigm(self.W_if @ x + self.b_if + self.W_hf @ h_prev + self.b_hf)
g_t = tanh(self.W_ig @ x + self.b_ig + self.W_hg @ h_prev + self.b_hg)
o_t = sigm(self.W_io @ x + self.b_io + self.W_ho @ h_prev + self.b_ho)
c_t = f_t * c_prev + i_t * g_t
h_t = o_t * tanh(c_t)
return h_t, (h_t, c_t)
def __repr__(self) -> str:
clsname = type(self).__name__
return f"{clsname}(input_size={self.input_size}, hidden_size={self.hidden_size})"
forward(self, x, hidden=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/layers.py
def forward(self, x: torch.Tensor, hidden=None) -> tuple:
sigm = torch.sigmoid
tanh = torch.tanh
if hidden is None:
h_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
c_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
else:
h_prev, c_prev = hidden
i_t = sigm(self.W_ii @ x + self.b_ii + self.W_hi @ h_prev + self.b_hi)
f_t = sigm(self.W_if @ x + self.b_if + self.W_hf @ h_prev + self.b_hf)
g_t = tanh(self.W_ig @ x + self.b_ig + self.W_hg @ h_prev + self.b_hg)
o_t = sigm(self.W_io @ x + self.b_io + self.W_ho @ h_prev + self.b_ho)
c_t = f_t * c_prev + i_t * g_t
h_t = o_t * tanh(c_t)
return h_t, (h_t, c_t)
LSTMNet (Module)
¶
Source code in evotorch/neuroevolution/net/layers.py
class LSTM(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
*,
dtype: torch.dtype = torch.float32,
device: Union[str, torch.device] = "cpu",
):
super().__init__()
input_size = int(input_size)
hidden_size = int(hidden_size)
self.input_size = input_size
self.hidden_size = hidden_size
def input_weight():
return nn.Parameter(torch.randn(self.hidden_size, self.input_size, dtype=dtype, device=device))
def weight():
return nn.Parameter(torch.randn(self.hidden_size, self.hidden_size, dtype=dtype, device=device))
def bias():
return nn.Parameter(torch.zeros(self.hidden_size, dtype=dtype, device=device))
self.W_ii = input_weight()
self.W_if = input_weight()
self.W_ig = input_weight()
self.W_io = input_weight()
self.W_hi = weight()
self.W_hf = weight()
self.W_hg = weight()
self.W_ho = weight()
self.b_ii = bias()
self.b_if = bias()
self.b_ig = bias()
self.b_io = bias()
self.b_hi = bias()
self.b_hf = bias()
self.b_hg = bias()
self.b_ho = bias()
def forward(self, x: torch.Tensor, hidden=None) -> tuple:
sigm = torch.sigmoid
tanh = torch.tanh
if hidden is None:
h_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
c_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
else:
h_prev, c_prev = hidden
i_t = sigm(self.W_ii @ x + self.b_ii + self.W_hi @ h_prev + self.b_hi)
f_t = sigm(self.W_if @ x + self.b_if + self.W_hf @ h_prev + self.b_hf)
g_t = tanh(self.W_ig @ x + self.b_ig + self.W_hg @ h_prev + self.b_hg)
o_t = sigm(self.W_io @ x + self.b_io + self.W_ho @ h_prev + self.b_ho)
c_t = f_t * c_prev + i_t * g_t
h_t = o_t * tanh(c_t)
return h_t, (h_t, c_t)
def __repr__(self) -> str:
clsname = type(self).__name__
return f"{clsname}(input_size={self.input_size}, hidden_size={self.hidden_size})"
forward(self, x, hidden=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/layers.py
def forward(self, x: torch.Tensor, hidden=None) -> tuple:
sigm = torch.sigmoid
tanh = torch.tanh
if hidden is None:
h_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
c_prev = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
else:
h_prev, c_prev = hidden
i_t = sigm(self.W_ii @ x + self.b_ii + self.W_hi @ h_prev + self.b_hi)
f_t = sigm(self.W_if @ x + self.b_if + self.W_hf @ h_prev + self.b_hf)
g_t = tanh(self.W_ig @ x + self.b_ig + self.W_hg @ h_prev + self.b_hg)
o_t = sigm(self.W_io @ x + self.b_io + self.W_ho @ h_prev + self.b_ho)
c_t = f_t * c_prev + i_t * g_t
h_t = o_t * tanh(c_t)
return h_t, (h_t, c_t)
LocomotorNet (Module)
¶
This is a control network which consists of two components: one linear, and one non-linear. The non-linear component is an input-independent set of sinusoidals waves whose amplitudes, frequencies and phases are trainable. Upon execution of a forward pass, the output of the non-linear component is the sum of all these sinusoidal waves. The linear component is a linear layer (optionally with bias) whose weights (and biases) are trainable. The final output of the LocomotorNet at the end of a forward pass is the sum of the linear and the non-linear components.
Note that this is a stateful network, where the only state
is the timestep t, which starts from 0 and gets incremented by 1
at the end of each forward pass. The reset() method resets
t back to 0.
Reference
Mario Srouji, Jian Zhang, Ruslan Salakhutdinov (2018). Structured Control Nets for Deep Reinforcement Learning.
Source code in evotorch/neuroevolution/net/layers.py
class LocomotorNet(nn.Module):
"""LocomotorNet: A locomotion-specific structured control net.
This is a control network which consists of two components:
one linear, and one non-linear. The non-linear component
is an input-independent set of sinusoidals waves whose
amplitudes, frequencies and phases are trainable.
Upon execution of a forward pass, the output of the non-linear
component is the sum of all these sinusoidal waves.
The linear component is a linear layer (optionally with bias)
whose weights (and biases) are trainable.
The final output of the LocomotorNet at the end of a forward pass
is the sum of the linear and the non-linear components.
Note that this is a stateful network, where the only state
is the timestep t, which starts from 0 and gets incremented by 1
at the end of each forward pass. The `reset()` method resets
t back to 0.
Reference:
Mario Srouji, Jian Zhang, Ruslan Salakhutdinov (2018).
Structured Control Nets for Deep Reinforcement Learning.
"""
def __init__(self, *, in_features: int, out_features: int, bias: bool = True, num_sinusoids=16):
"""`__init__(...)`: Initialize the LocomotorNet.
Args:
in_features: Length of the input vector
out_features: Length of the output vector
bias: Whether or not the linear component is to have a bias
num_sinusoids: Number of sinusoidal waves
"""
nn.Module.__init__(self)
self._in_features = in_features
self._out_features = out_features
self._bias = bias
self._num_sinusoids = num_sinusoids
self._linear_component = nn.Linear(
in_features=self._in_features, out_features=self._out_features, bias=self._bias
)
self._amplitudes = nn.ParameterList()
self._frequencies = nn.ParameterList()
self._phases = nn.ParameterList()
for _ in range(self._num_sinusoids):
for paramlist in (self._amplitudes, self._frequencies, self._phases):
paramlist.append(nn.Parameter(torch.randn(self._out_features, dtype=torch.float32)))
self.reset()
def reset(self):
"""Set the timestep t to 0"""
self._t = 0
@property
def t(self) -> int:
"""The current timestep t"""
return self._t
@property
def in_features(self) -> int:
"""Get the length of the input vector"""
return self._in_features
@property
def out_features(self) -> int:
"""Get the length of the output vector"""
return self._out_features
@property
def num_sinusoids(self) -> int:
"""Get the number of sinusoidal waves of the non-linear component"""
return self._num_sinusoids
@property
def bias(self) -> bool:
"""Get whether or not the linear component has bias"""
return self._bias
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Execute a forward pass"""
u_linear = self._linear_component(x)
t = self._t
u_nonlinear = torch.zeros(self._out_features)
for i in range(self._num_sinusoids):
A = self._amplitudes[i]
w = self._frequencies[i]
phi = self._phases[i]
u_nonlinear = u_nonlinear + (A * torch.sin(w * t + phi))
self._t += 1
return u_linear + u_nonlinear
bias: bool
property
readonly
¶
Get whether or not the linear component has bias
in_features: int
property
readonly
¶
Get the length of the input vector
num_sinusoids: int
property
readonly
¶
Get the number of sinusoidal waves of the non-linear component
out_features: int
property
readonly
¶
Get the length of the output vector
t: int
property
readonly
¶
The current timestep t
__init__(self, *, in_features, out_features, bias=True, num_sinusoids=16)
special
¶
__init__(...): Initialize the LocomotorNet.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
in_features |
int |
Length of the input vector |
required |
out_features |
int |
Length of the output vector |
required |
bias |
bool |
Whether or not the linear component is to have a bias |
True |
num_sinusoids |
Number of sinusoidal waves |
16 |
Source code in evotorch/neuroevolution/net/layers.py
def __init__(self, *, in_features: int, out_features: int, bias: bool = True, num_sinusoids=16):
"""`__init__(...)`: Initialize the LocomotorNet.
Args:
in_features: Length of the input vector
out_features: Length of the output vector
bias: Whether or not the linear component is to have a bias
num_sinusoids: Number of sinusoidal waves
"""
nn.Module.__init__(self)
self._in_features = in_features
self._out_features = out_features
self._bias = bias
self._num_sinusoids = num_sinusoids
self._linear_component = nn.Linear(
in_features=self._in_features, out_features=self._out_features, bias=self._bias
)
self._amplitudes = nn.ParameterList()
self._frequencies = nn.ParameterList()
self._phases = nn.ParameterList()
for _ in range(self._num_sinusoids):
for paramlist in (self._amplitudes, self._frequencies, self._phases):
paramlist.append(nn.Parameter(torch.randn(self._out_features, dtype=torch.float32)))
self.reset()
forward(self, x)
¶
Execute a forward pass
Source code in evotorch/neuroevolution/net/layers.py
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Execute a forward pass"""
u_linear = self._linear_component(x)
t = self._t
u_nonlinear = torch.zeros(self._out_features)
for i in range(self._num_sinusoids):
A = self._amplitudes[i]
w = self._frequencies[i]
phi = self._phases[i]
u_nonlinear = u_nonlinear + (A * torch.sin(w * t + phi))
self._t += 1
return u_linear + u_nonlinear
reset(self)
¶
RNN (Module)
¶
Source code in evotorch/neuroevolution/net/layers.py
class RNN(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
nonlinearity: str = "tanh",
*,
dtype: torch.dtype = torch.float32,
device: Union[str, torch.device] = "cpu",
):
super().__init__()
input_size = int(input_size)
hidden_size = int(hidden_size)
nonlinearity = str(nonlinearity)
self.W1 = nn.Parameter(torch.randn(hidden_size, input_size, dtype=dtype, device=device))
self.W2 = nn.Parameter(torch.randn(hidden_size, hidden_size, dtype=dtype, device=device))
self.b1 = nn.Parameter(torch.zeros(hidden_size, dtype=dtype, device=device))
self.b2 = nn.Parameter(torch.zeros(hidden_size, dtype=dtype, device=device))
if nonlinearity == "tanh":
self.actfunc = torch.tanh
else:
self.actfunc = getattr(nnf, nonlinearity)
self.nonlinearity = nonlinearity
self.input_size = input_size
self.hidden_size = hidden_size
def forward(self, x: torch.Tensor, h: Optional[torch.Tensor] = None) -> tuple:
if h is None:
h = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
act = self.actfunc
W1 = self.W1
W2 = self.W2
b1 = self.b1.unsqueeze(-1)
b2 = self.b2.unsqueeze(-1)
x = x.unsqueeze(-1)
h = h.unsqueeze(-1)
y = act(((W1 @ x) + b1) + ((W2 @ h) + b2))
y = y.squeeze(-1)
return y, y
def __repr__(self) -> str:
clsname = type(self).__name__
return f"{clsname}(input_size={self.input_size}, hidden_size={self.hidden_size}, nonlinearity={repr(self.nonlinearity)})"
forward(self, x, h=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/layers.py
def forward(self, x: torch.Tensor, h: Optional[torch.Tensor] = None) -> tuple:
if h is None:
h = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
act = self.actfunc
W1 = self.W1
W2 = self.W2
b1 = self.b1.unsqueeze(-1)
b2 = self.b2.unsqueeze(-1)
x = x.unsqueeze(-1)
h = h.unsqueeze(-1)
y = act(((W1 @ x) + b1) + ((W2 @ h) + b2))
y = y.squeeze(-1)
return y, y
RecurrentNet (Module)
¶
Source code in evotorch/neuroevolution/net/layers.py
class RNN(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
nonlinearity: str = "tanh",
*,
dtype: torch.dtype = torch.float32,
device: Union[str, torch.device] = "cpu",
):
super().__init__()
input_size = int(input_size)
hidden_size = int(hidden_size)
nonlinearity = str(nonlinearity)
self.W1 = nn.Parameter(torch.randn(hidden_size, input_size, dtype=dtype, device=device))
self.W2 = nn.Parameter(torch.randn(hidden_size, hidden_size, dtype=dtype, device=device))
self.b1 = nn.Parameter(torch.zeros(hidden_size, dtype=dtype, device=device))
self.b2 = nn.Parameter(torch.zeros(hidden_size, dtype=dtype, device=device))
if nonlinearity == "tanh":
self.actfunc = torch.tanh
else:
self.actfunc = getattr(nnf, nonlinearity)
self.nonlinearity = nonlinearity
self.input_size = input_size
self.hidden_size = hidden_size
def forward(self, x: torch.Tensor, h: Optional[torch.Tensor] = None) -> tuple:
if h is None:
h = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
act = self.actfunc
W1 = self.W1
W2 = self.W2
b1 = self.b1.unsqueeze(-1)
b2 = self.b2.unsqueeze(-1)
x = x.unsqueeze(-1)
h = h.unsqueeze(-1)
y = act(((W1 @ x) + b1) + ((W2 @ h) + b2))
y = y.squeeze(-1)
return y, y
def __repr__(self) -> str:
clsname = type(self).__name__
return f"{clsname}(input_size={self.input_size}, hidden_size={self.hidden_size}, nonlinearity={repr(self.nonlinearity)})"
forward(self, x, h=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/layers.py
def forward(self, x: torch.Tensor, h: Optional[torch.Tensor] = None) -> tuple:
if h is None:
h = torch.zeros(self.hidden_size, dtype=x.dtype, device=x.device)
act = self.actfunc
W1 = self.W1
W2 = self.W2
b1 = self.b1.unsqueeze(-1)
b2 = self.b2.unsqueeze(-1)
x = x.unsqueeze(-1)
h = h.unsqueeze(-1)
y = act(((W1 @ x) + b1) + ((W2 @ h) + b2))
y = y.squeeze(-1)
return y, y
Round (Module)
¶
A small torch module for rounding the values of an input tensor
Source code in evotorch/neuroevolution/net/layers.py
class Round(nn.Module):
"""A small torch module for rounding the values of an input tensor"""
def __init__(self, ndigits: int = 0):
nn.Module.__init__(self)
self._ndigits = int(ndigits)
self._q = 10.0**self._ndigits
def forward(self, x):
x = x * self._q
x = torch.round(x)
x = x / self._q
return x
def extra_repr(self):
return "ndigits=" + str(self._ndigits)
extra_repr(self)
¶
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Slice (Module)
¶
A small torch module for getting the slice of an input tensor
Source code in evotorch/neuroevolution/net/layers.py
class Slice(nn.Module):
"""A small torch module for getting the slice of an input tensor"""
def __init__(self, from_index: int, to_index: int):
"""`__init__(...)`: Initialize the Slice operator.
Args:
from_index: The index from which the slice begins.
to_index: The exclusive index at which the slice ends.
"""
nn.Module.__init__(self)
self._from_index = from_index
self._to_index = to_index
def forward(self, x):
return x[self._from_index : self._to_index]
def extra_repr(self):
return "from_index={}, to_index={}".format(self._from_index, self._to_index)
__init__(self, from_index, to_index)
special
¶
__init__(...): Initialize the Slice operator.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
from_index |
int |
The index from which the slice begins. |
required |
to_index |
int |
The exclusive index at which the slice ends. |
required |
Source code in evotorch/neuroevolution/net/layers.py
extra_repr(self)
¶
Set the extra representation of the module
To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
StructuredControlNet (Module)
¶
Structured Control Net.
This is a control network consisting of two components: (i) a non-linear component, which is a feed-forward network; and (ii) a linear component, which is a linear layer. Both components take the input vector provided to the structured control network. The final output is the sum of the outputs of both components.
Reference
Mario Srouji, Jian Zhang, Ruslan Salakhutdinov (2018). Structured Control Nets for Deep Reinforcement Learning.
Source code in evotorch/neuroevolution/net/layers.py
class StructuredControlNet(nn.Module):
"""Structured Control Net.
This is a control network consisting of two components:
(i) a non-linear component, which is a feed-forward network; and
(ii) a linear component, which is a linear layer.
Both components take the input vector provided to the
structured control network.
The final output is the sum of the outputs of both components.
Reference:
Mario Srouji, Jian Zhang, Ruslan Salakhutdinov (2018).
Structured Control Nets for Deep Reinforcement Learning.
"""
def __init__(
self,
*,
in_features: int,
out_features: int,
num_layers: int,
hidden_size: int,
bias: bool = True,
nonlinearity: Union[str, Callable] = "tanh",
):
"""`__init__(...)`: Initialize the structured control net.
Args:
in_features: Length of the input vector
out_features: Length of the output vector
num_layers: Number of hidden layers for the non-linear component
hidden_size: Number of neurons in a hidden layer of the
non-linear component
bias: Whether or not the linear component is to have bias
nonlinearity: Activation function
"""
nn.Module.__init__(self)
self._in_features = in_features
self._out_features = out_features
self._num_layers = num_layers
self._hidden_size = hidden_size
self._bias = bias
self._nonlinearity = nonlinearity
self._linear_component = nn.Linear(
in_features=self._in_features, out_features=self._out_features, bias=self._bias
)
self._nonlinear_component = FeedForwardNet(
input_size=self._in_features,
layers=(
list((self._hidden_size, self._nonlinearity) for _ in range(self._num_layers))
+ [(self._out_features, self._nonlinearity)]
),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""TODO: documentation"""
return self._linear_component(x) + self._nonlinear_component(x)
@property
def in_features(self):
"""TODO: documentation"""
return self._in_features
@property
def out_features(self):
"""TODO: documentation"""
return self._out_features
@property
def num_layers(self):
"""TODO: documentation"""
return self._num_layers
@property
def hidden_size(self):
"""TODO: documentation"""
return self._hidden_size
@property
def bias(self):
"""TODO: documentation"""
return self._bias
@property
def nonlinearity(self):
"""TODO: documentation"""
return self._nonlinearity
bias
property
readonly
¶
hidden_size
property
readonly
¶
in_features
property
readonly
¶
nonlinearity
property
readonly
¶
num_layers
property
readonly
¶
out_features
property
readonly
¶
__init__(self, *, in_features, out_features, num_layers, hidden_size, bias=True, nonlinearity='tanh')
special
¶
__init__(...): Initialize the structured control net.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
in_features |
int |
Length of the input vector |
required |
out_features |
int |
Length of the output vector |
required |
num_layers |
int |
Number of hidden layers for the non-linear component |
required |
hidden_size |
int |
Number of neurons in a hidden layer of the non-linear component |
required |
bias |
bool |
Whether or not the linear component is to have bias |
True |
nonlinearity |
Union[str, Callable] |
Activation function |
'tanh' |
Source code in evotorch/neuroevolution/net/layers.py
def __init__(
self,
*,
in_features: int,
out_features: int,
num_layers: int,
hidden_size: int,
bias: bool = True,
nonlinearity: Union[str, Callable] = "tanh",
):
"""`__init__(...)`: Initialize the structured control net.
Args:
in_features: Length of the input vector
out_features: Length of the output vector
num_layers: Number of hidden layers for the non-linear component
hidden_size: Number of neurons in a hidden layer of the
non-linear component
bias: Whether or not the linear component is to have bias
nonlinearity: Activation function
"""
nn.Module.__init__(self)
self._in_features = in_features
self._out_features = out_features
self._num_layers = num_layers
self._hidden_size = hidden_size
self._bias = bias
self._nonlinearity = nonlinearity
self._linear_component = nn.Linear(
in_features=self._in_features, out_features=self._out_features, bias=self._bias
)
self._nonlinear_component = FeedForwardNet(
input_size=self._in_features,
layers=(
list((self._hidden_size, self._nonlinearity) for _ in range(self._num_layers))
+ [(self._out_features, self._nonlinearity)]
),
)
forward(self, x)
¶
misc
¶
Utilities for reading and for writing neural network parameters
count_parameters(net)
¶
Get the number of parameters the network.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
net |
Module |
The torch module whose parameters will be counted. |
required |
Returns:
| Type | Description |
|---|---|
int |
The number of parameters, as an integer. |
Source code in evotorch/neuroevolution/net/misc.py
device_of_module(m, default=None)
¶
Get the device in which the module exists.
This function looks at the first parameter of the module, and returns its device. This function is not meant to be used on modules whose parameters exist on different devices.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
m |
Module |
The module whose device is being queried. |
required |
default |
Union[str, torch.device] |
The fallback device to return if the module has no parameters. If this is left as None, the fallback device is assumed to be "cpu". |
None |
Returns:
| Type | Description |
|---|---|
device |
The device of the module, determined from its first parameter. |
Source code in evotorch/neuroevolution/net/misc.py
def device_of_module(m: nn.Module, default: Optional[Union[str, torch.device]] = None) -> torch.device:
"""
Get the device in which the module exists.
This function looks at the first parameter of the module, and returns
its device. This function is not meant to be used on modules whose
parameters exist on different devices.
Args:
m: The module whose device is being queried.
default: The fallback device to return if the module has no
parameters. If this is left as None, the fallback device
is assumed to be "cpu".
Returns:
The device of the module, determined from its first parameter.
"""
if default is None:
default = torch.device("cpu")
device = default
for p in m.parameters():
device = p.device
break
return device
fill_parameters(net, vector)
¶
Fill the parameters of a torch module (net) from a vector.
No gradient information is kept.
The vector's length must be exactly the same with the number of parameters of the PyTorch module.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
net |
Module |
The torch module whose parameter values will be filled. |
required |
vector |
Tensor |
A 1-D torch tensor which stores the parameter values. |
required |
Source code in evotorch/neuroevolution/net/misc.py
@torch.no_grad()
def fill_parameters(net: nn.Module, vector: torch.Tensor):
"""Fill the parameters of a torch module (net) from a vector.
No gradient information is kept.
The vector's length must be exactly the same with the number
of parameters of the PyTorch module.
Args:
net: The torch module whose parameter values will be filled.
vector: A 1-D torch tensor which stores the parameter values.
"""
address = 0
for p in net.parameters():
d = p.data.view(-1)
n = len(d)
d[:] = torch.as_tensor(vector[address : address + n], device=d.device)
address += n
if address != len(vector):
raise IndexError("The parameter vector is larger than expected")
parameter_vector(net, *, device=None)
¶
Get all the parameters of a torch module (net) into a vector
No gradient information is kept.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
net |
Module |
The torch module whose parameters will be extracted. |
required |
device |
Union[str, torch.device] |
The device in which the parameter vector will be constructed. If the network has parameter across multiple devices, you can specify this argument so that concatenation of all the parameters will be successful. |
None |
Returns:
| Type | Description |
|---|---|
Tensor |
The parameters of the module in a 1-D tensor. |
Source code in evotorch/neuroevolution/net/misc.py
@torch.no_grad()
def parameter_vector(net: nn.Module, *, device: Optional[Device] = None) -> torch.Tensor:
"""Get all the parameters of a torch module (net) into a vector
No gradient information is kept.
Args:
net: The torch module whose parameters will be extracted.
device: The device in which the parameter vector will be constructed.
If the network has parameter across multiple devices,
you can specify this argument so that concatenation of all the
parameters will be successful.
Returns:
The parameters of the module in a 1-D tensor.
"""
dev_kwarg = {} if device is None else {"device": device}
all_vectors = []
for p in net.parameters():
all_vectors.append(torch.as_tensor(p.data.view(-1), **dev_kwarg))
return torch.cat(all_vectors)
multilayered
¶
MultiLayered (Module)
¶
Source code in evotorch/neuroevolution/net/multilayered.py
class MultiLayered(nn.Module):
def __init__(self, *layers: nn.Module):
super().__init__()
self._submodules = nn.ModuleList(layers)
def forward(self, x: torch.Tensor, h: Optional[dict] = None):
if h is None:
h = {}
new_h = {}
for i, layer in enumerate(self._submodules):
layer_h = h.get(i, None)
if layer_h is None:
layer_result = layer(x)
else:
layer_result = layer(x, h[i])
if isinstance(layer_result, tuple):
if len(layer_result) == 2:
x, layer_new_h = layer_result
else:
raise ValueError(
f"The layer number {i} returned a tuple of length {len(layer_result)}."
f" A tensor or a tuple of two elements was expected."
)
elif isinstance(layer_result, torch.Tensor):
x = layer_result
layer_new_h = None
else:
raise TypeError(
f"The layer number {i} returned an object of type {type(layer_result)}."
f" A tensor or a tuple of two elements was expected."
)
if layer_new_h is not None:
new_h[i] = layer_new_h
if len(new_h) == 0:
return x
else:
return x, new_h
def __iter__(self):
return self._submodules.__iter__()
def __getitem__(self, i):
return self._submodules[i]
def append(self, module: nn.Module):
self._submodules.append(module)
forward(self, x, h=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/multilayered.py
def forward(self, x: torch.Tensor, h: Optional[dict] = None):
if h is None:
h = {}
new_h = {}
for i, layer in enumerate(self._submodules):
layer_h = h.get(i, None)
if layer_h is None:
layer_result = layer(x)
else:
layer_result = layer(x, h[i])
if isinstance(layer_result, tuple):
if len(layer_result) == 2:
x, layer_new_h = layer_result
else:
raise ValueError(
f"The layer number {i} returned a tuple of length {len(layer_result)}."
f" A tensor or a tuple of two elements was expected."
)
elif isinstance(layer_result, torch.Tensor):
x = layer_result
layer_new_h = None
else:
raise TypeError(
f"The layer number {i} returned an object of type {type(layer_result)}."
f" A tensor or a tuple of two elements was expected."
)
if layer_new_h is not None:
new_h[i] = layer_new_h
if len(new_h) == 0:
return x
else:
return x, new_h
parser
¶
Utilities for parsing string representations of neural net policies
NetParsingError (Exception)
¶
Representation of a parsing error
Source code in evotorch/neuroevolution/net/parser.py
class NetParsingError(Exception):
"""
Representation of a parsing error
"""
def __init__(
self,
message: str,
lineno: Optional[int] = None,
col_offset: Optional[int] = None,
original_error: Optional[Exception] = None,
):
"""
`__init__(...)`: Initialize the NetParsingError.
Args:
message: Error message, as string.
lineno: Erroneous line number in the string representation of the
neural network structure.
col_offset: Erroneous column number in the string representation
of the neural network structure.
original_error: If another error caused this parsing error,
that original error can be attached to this `NetParsingError`
instance via this argument.
"""
super().__init__()
self.message = message
self.lineno = lineno
self.col_offset = col_offset
self.original_error = original_error
def _to_string(self) -> str:
parts = []
parts.append(type(self).__name__)
if self.lineno is not None:
parts.append(" at line(")
parts.append(str(self.lineno - 1))
parts.append(")")
if self.col_offset is not None:
parts.append(" at column(")
parts.append(str(self.col_offset + 1))
parts.append(")")
parts.append(": ")
parts.append(self.message)
return "".join(parts)
def __str__(self) -> str:
return self._to_string()
def __repr__(self) -> str:
return self._to_string()
__init__(self, message, lineno=None, col_offset=None, original_error=None)
special
¶
__init__(...): Initialize the NetParsingError.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
message |
str |
Error message, as string. |
required |
lineno |
Optional[int] |
Erroneous line number in the string representation of the neural network structure. |
None |
col_offset |
Optional[int] |
Erroneous column number in the string representation of the neural network structure. |
None |
original_error |
Optional[Exception] |
If another error caused this parsing error,
that original error can be attached to this |
None |
Source code in evotorch/neuroevolution/net/parser.py
def __init__(
self,
message: str,
lineno: Optional[int] = None,
col_offset: Optional[int] = None,
original_error: Optional[Exception] = None,
):
"""
`__init__(...)`: Initialize the NetParsingError.
Args:
message: Error message, as string.
lineno: Erroneous line number in the string representation of the
neural network structure.
col_offset: Erroneous column number in the string representation
of the neural network structure.
original_error: If another error caused this parsing error,
that original error can be attached to this `NetParsingError`
instance via this argument.
"""
super().__init__()
self.message = message
self.lineno = lineno
self.col_offset = col_offset
self.original_error = original_error
str_to_net(s, **constants)
¶
Read a string representation of a neural net structure,
and return a torch.nn.Module instance out of it.
Let us imagine that one wants to describe the following neural network structure:
from torch import nn
net = nn.Sequential(
nn.Linear(8, 16), nn.Tanh(), nn.Linear(16, 4, bias=False), nn.ReLU()
)
By using str_to_net(...) one can construct an equivalent
module via:
from evotorch.neuroevolution.net import str_to_net
net = str_to_net("Linear(8, 16) >> Tanh() >> Linear(16, 4, bias=False) >> ReLU()")
The string can also be multi-line:
One can also define constants for using them in strings:
net = str_to_net(
'''
Linear(input_size, hidden_size)
>> Tanh()
>> Linear(hidden_size, output_size, bias=False)
>> ReLU()
''',
input_size=8,
hidden_size=16,
output_size=4,
)
In the neural net structure string, when one refers to a module type,
say, Linear, first the name Linear is searched for in the namespace
evotorch.neuroevolution.net.layers, and then in the namespace torch.nn.
In the case of Linear, the searched name exists in torch.nn,
and therefore, the layer type to be instantiated is accepted as
torch.nn.Linear.
Instead of Linear, if one had used the name, say,
StructuredControlNet, then, the layer type to be instantiated
would be evotorch.neuroevolution.net.layers.StructuredControlNet.
The namespace evotorch.neuroevolution.net.layers contains its own
implementations for RNN and LSTM. These recurrent layer implementations
work similarly to their counterparts torch.nn.RNN and torch.nn.LSTM,
except that EvoTorch's implementations do not expect the data with extra
leftmost dimensions for batching and for timesteps. Instead, they expect
to receive a single input and a single current hidden state, and produce
a single output and a single new hidden state. These recurrent layer
implementations of EvoTorch can be used within a neural net structure
string. Therefore, the following examples are valid:
rnn1 = str_to_net("RNN(4, 8) >> Linear(8, 2)")
rnn2 = str_to_net(
'''
Linear(4, 10)
>> Tanh()
>> RNN(input_size=10, hidden_size=24, nonlinearity='tanh'
>> Linear(24, 2)
'''
)
lstm1 = str_to_net("LSTM(4, 32) >> Linear(32, 2)")
lstm2 = str_to_net("LSTM(input_size=4, hidden_size=32) >> Linear(32, 2)")
Notes regarding usage with evotorch.neuroevolution.GymNE
or with evotorch.neuroevolution.VecGymNE:
While instantiating a GymNE or a VecGymNE, one can specify a neural
net structure string as the policy. Therefore, while filling the policy
string for a GymNE, all these rules mentioned above apply. Additionally,
while using str_to_net(...) internally, GymNE and VecGymNE define
these extra constants:
obs_length (length of the observation vector),
act_length (length of the action vector for continuous-action
environments, or number of actions for discrete-action
environments), and
obs_shape (shape of the observation as a tuple, assuming that the
observation space is of type gym.spaces.Box, usable within the string
like obs_shape[0], obs_shape[1], etc., or simply obs_shape to refer
to the entire tuple).
Therefore, while instantiating a GymNE or a VecGymNE, one can define a
single-hidden-layered policy via this string:
In the policy string above, one might choose to omit the last Tanh(), as
GymNE and VecGymNE will clip the final output of the policy to conform
to the action boundaries defined by the target reinforcement learning
environment, and such a clipping operation might be seen as using an
activation function similar to hard-tanh anyway.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
s |
str |
The string which expresses the neural net structure. |
required |
Returns:
| Type | Description |
|---|---|
Module |
The PyTorch module of the specified structure. |
Source code in evotorch/neuroevolution/net/parser.py
def str_to_net(s: str, **constants) -> nn.Module:
"""
Read a string representation of a neural net structure,
and return a `torch.nn.Module` instance out of it.
Let us imagine that one wants to describe the following
neural network structure:
```python
from torch import nn
net = nn.Sequential(
nn.Linear(8, 16), nn.Tanh(), nn.Linear(16, 4, bias=False), nn.ReLU()
)
```
By using `str_to_net(...)` one can construct an equivalent
module via:
```python
from evotorch.neuroevolution.net import str_to_net
net = str_to_net("Linear(8, 16) >> Tanh() >> Linear(16, 4, bias=False) >> ReLU()")
```
The string can also be multi-line:
```python
net = str_to_net(
'''
Linear(8, 16)
>> Tanh()
>> Linear(16, 4, bias=False)
>> ReLU()
'''
)
```
One can also define constants for using them in strings:
```python
net = str_to_net(
'''
Linear(input_size, hidden_size)
>> Tanh()
>> Linear(hidden_size, output_size, bias=False)
>> ReLU()
''',
input_size=8,
hidden_size=16,
output_size=4,
)
```
In the neural net structure string, when one refers to a module type,
say, `Linear`, first the name `Linear` is searched for in the namespace
`evotorch.neuroevolution.net.layers`, and then in the namespace `torch.nn`.
In the case of `Linear`, the searched name exists in `torch.nn`,
and therefore, the layer type to be instantiated is accepted as
`torch.nn.Linear`.
Instead of `Linear`, if one had used the name, say,
`StructuredControlNet`, then, the layer type to be instantiated
would be `evotorch.neuroevolution.net.layers.StructuredControlNet`.
The namespace `evotorch.neuroevolution.net.layers` contains its own
implementations for RNN and LSTM. These recurrent layer implementations
work similarly to their counterparts `torch.nn.RNN` and `torch.nn.LSTM`,
except that EvoTorch's implementations do not expect the data with extra
leftmost dimensions for batching and for timesteps. Instead, they expect
to receive a single input and a single current hidden state, and produce
a single output and a single new hidden state. These recurrent layer
implementations of EvoTorch can be used within a neural net structure
string. Therefore, the following examples are valid:
```python
rnn1 = str_to_net("RNN(4, 8) >> Linear(8, 2)")
rnn2 = str_to_net(
'''
Linear(4, 10)
>> Tanh()
>> RNN(input_size=10, hidden_size=24, nonlinearity='tanh'
>> Linear(24, 2)
'''
)
lstm1 = str_to_net("LSTM(4, 32) >> Linear(32, 2)")
lstm2 = str_to_net("LSTM(input_size=4, hidden_size=32) >> Linear(32, 2)")
```
**Notes regarding usage with `evotorch.neuroevolution.GymNE`
or with `evotorch.neuroevolution.VecGymNE`:**
While instantiating a `GymNE` or a `VecGymNE`, one can specify a neural
net structure string as the policy. Therefore, while filling the policy
string for a `GymNE`, all these rules mentioned above apply. Additionally,
while using `str_to_net(...)` internally, `GymNE` and `VecGymNE` define
these extra constants:
`obs_length` (length of the observation vector),
`act_length` (length of the action vector for continuous-action
environments, or number of actions for discrete-action
environments), and
`obs_shape` (shape of the observation as a tuple, assuming that the
observation space is of type `gym.spaces.Box`, usable within the string
like `obs_shape[0]`, `obs_shape[1]`, etc., or simply `obs_shape` to refer
to the entire tuple).
Therefore, while instantiating a `GymNE` or a `VecGymNE`, one can define a
single-hidden-layered policy via this string:
```
"Linear(obs_length, 16) >> Tanh() >> Linear(16, act_length) >> Tanh()"
```
In the policy string above, one might choose to omit the last `Tanh()`, as
`GymNE` and `VecGymNE` will clip the final output of the policy to conform
to the action boundaries defined by the target reinforcement learning
environment, and such a clipping operation might be seen as using an
activation function similar to hard-tanh anyway.
Args:
s: The string which expresses the neural net structure.
Returns:
The PyTorch module of the specified structure.
"""
s = f"(\n{s}\n)"
return _process_expr(ast.parse(s, mode="eval").body, constants=constants)
rl
¶
This namespace provides various reinforcement learning utilities.
ActClipWrapperModule (Module)
¶
Source code in evotorch/neuroevolution/net/rl.py
class ActClipWrapperModule(nn.Module):
def __init__(self, wrapped_module: nn.Module, obs_space: Box):
super().__init__()
device = device_of_module(wrapped_module)
if not isinstance(obs_space, Box):
raise TypeError(f"Unrecognized observation space: {obs_space}")
self.wrapped_module = wrapped_module
self.register_buffer("_low", torch.from_numpy(obs_space.low).to(device))
self.register_buffer("_high", torch.from_numpy(obs_space.high).to(device))
def forward(self, x: torch.Tensor, h: Any = None) -> Union[torch.Tensor, tuple]:
if h is None:
result = self.wrapped_module(x)
else:
result = self.wrapped_module(x, h)
if isinstance(result, tuple):
x, h = result
got_h = True
else:
x = result
h = None
got_h = False
x = torch.max(x, self._low)
x = torch.min(x, self._high)
if got_h:
return x, h
else:
return x
forward(self, x, h=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/rl.py
def forward(self, x: torch.Tensor, h: Any = None) -> Union[torch.Tensor, tuple]:
if h is None:
result = self.wrapped_module(x)
else:
result = self.wrapped_module(x, h)
if isinstance(result, tuple):
x, h = result
got_h = True
else:
x = result
h = None
got_h = False
x = torch.max(x, self._low)
x = torch.min(x, self._high)
if got_h:
return x, h
else:
return x
AliveBonusScheduleWrapper (Wrapper)
¶
A Wrapper which awards the agent for being alive in a scheduled manner This wrapper is meant to be used for non-vectorized environments.
Source code in evotorch/neuroevolution/net/rl.py
class AliveBonusScheduleWrapper(gym.Wrapper):
"""
A Wrapper which awards the agent for being alive in a scheduled manner
This wrapper is meant to be used for non-vectorized environments.
"""
def __init__(self, env: gym.Env, alive_bonus_schedule: tuple, **kwargs):
"""
`__init__(...)`: Initialize the AliveBonusScheduleWrapper.
Args:
env: Environment to wrap.
alive_bonus_schedule: If given as a tuple `(t, b)`, an alive
bonus `b` will be added onto all the rewards beyond the
timestep `t`.
If given as a tuple `(t0, t1, b)`, a partial (linearly
increasing towards `b`) alive bonus will be added onto
all the rewards between the timesteps `t0` and `t1`,
and a full alive bonus (which equals to `b`) will be added
onto all the rewards beyond the timestep `t1`.
kwargs: Expected in the form of additional keyword arguments,
these will be passed to the initialization method of the
superclass.
"""
super().__init__(env, **kwargs)
self.__t: Optional[int] = None
if len(alive_bonus_schedule) == 3:
self.__t0, self.__t1, self.__bonus = (
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[1]),
float(alive_bonus_schedule[2]),
)
elif len(alive_bonus_schedule) == 2:
self.__t0, self.__t1, self.__bonus = (
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[0]),
float(alive_bonus_schedule[1]),
)
else:
raise ValueError(
f"The argument `alive_bonus_schedule` was expected to have 2 or 3 elements."
f" However, its value is {repr(alive_bonus_schedule)} (having {len(alive_bonus_schedule)} elements)."
)
if self.__t1 > self.__t0:
self.__gap = self.__t1 - self.__t0
else:
self.__gap = None
def reset(self, *args, **kwargs):
self.__t = 0
return self.env.reset(*args, **kwargs)
def step(self, action) -> tuple:
step_result = self.env.step(action)
self.__t += 1
observation = step_result[0]
reward = step_result[1]
rest = step_result[2:]
if self.__t >= self.__t1:
reward = reward + self.__bonus
elif (self.__gap is not None) and (self.__t >= self.__t0):
reward = reward + ((self.__t - self.__t0) / self.__gap) * self.__bonus
return (observation, reward) + rest
__init__(self, env, alive_bonus_schedule, **kwargs)
special
¶
__init__(...): Initialize the AliveBonusScheduleWrapper.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env |
Env |
Environment to wrap. |
required |
alive_bonus_schedule |
tuple |
If given as a tuple |
required |
kwargs |
Expected in the form of additional keyword arguments, these will be passed to the initialization method of the superclass. |
{} |
Source code in evotorch/neuroevolution/net/rl.py
def __init__(self, env: gym.Env, alive_bonus_schedule: tuple, **kwargs):
"""
`__init__(...)`: Initialize the AliveBonusScheduleWrapper.
Args:
env: Environment to wrap.
alive_bonus_schedule: If given as a tuple `(t, b)`, an alive
bonus `b` will be added onto all the rewards beyond the
timestep `t`.
If given as a tuple `(t0, t1, b)`, a partial (linearly
increasing towards `b`) alive bonus will be added onto
all the rewards between the timesteps `t0` and `t1`,
and a full alive bonus (which equals to `b`) will be added
onto all the rewards beyond the timestep `t1`.
kwargs: Expected in the form of additional keyword arguments,
these will be passed to the initialization method of the
superclass.
"""
super().__init__(env, **kwargs)
self.__t: Optional[int] = None
if len(alive_bonus_schedule) == 3:
self.__t0, self.__t1, self.__bonus = (
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[1]),
float(alive_bonus_schedule[2]),
)
elif len(alive_bonus_schedule) == 2:
self.__t0, self.__t1, self.__bonus = (
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[0]),
float(alive_bonus_schedule[1]),
)
else:
raise ValueError(
f"The argument `alive_bonus_schedule` was expected to have 2 or 3 elements."
f" However, its value is {repr(alive_bonus_schedule)} (having {len(alive_bonus_schedule)} elements)."
)
if self.__t1 > self.__t0:
self.__gap = self.__t1 - self.__t0
else:
self.__gap = None
reset(self, *args, **kwargs)
¶
step(self, action)
¶
Steps through the environment with action.
Source code in evotorch/neuroevolution/net/rl.py
def step(self, action) -> tuple:
step_result = self.env.step(action)
self.__t += 1
observation = step_result[0]
reward = step_result[1]
rest = step_result[2:]
if self.__t >= self.__t1:
reward = reward + self.__bonus
elif (self.__gap is not None) and (self.__t >= self.__t0):
reward = reward + ((self.__t - self.__t0) / self.__gap) * self.__bonus
return (observation, reward) + rest
ObsNormWrapperModule (Module)
¶
Source code in evotorch/neuroevolution/net/rl.py
class ObsNormWrapperModule(nn.Module):
def __init__(self, wrapped_module: nn.Module, rn: Union[RunningStat, RunningNorm]):
super().__init__()
device = device_of_module(wrapped_module)
self.wrapped_module = wrapped_module
with torch.no_grad():
normalizer = deepcopy(rn.to_layer()).to(device)
self.normalizer = normalizer
def forward(self, x: torch.Tensor, h: Any = None) -> Union[torch.Tensor, tuple]:
x = self.normalizer(x)
if h is None:
result = self.wrapped_module(x)
else:
result = self.wrapped_module(x, h)
if isinstance(result, tuple):
x, h = result
got_h = True
else:
x = result
h = None
got_h = False
if got_h:
return x, h
else:
return x
forward(self, x, h=None)
¶
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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/rl.py
def forward(self, x: torch.Tensor, h: Any = None) -> Union[torch.Tensor, tuple]:
x = self.normalizer(x)
if h is None:
result = self.wrapped_module(x)
else:
result = self.wrapped_module(x, h)
if isinstance(result, tuple):
x, h = result
got_h = True
else:
x = result
h = None
got_h = False
if got_h:
return x, h
else:
return x
reset_env(env)
¶
Reset a gym environment.
For gym 1.0, the plan is to have a reset(...) method which returns
a two-element tuple (observation, info) where info is an object
providing any additional information regarding the initial state of
the agent. However, the old (pre 1.0) gym API (and some environments
which were written with old gym compatibility in mind) has (or have)
a reset(...) method which returns a single object that is the
initial observation.
With the assumption that the observation space of the environment
is NOT tuple, this function can work with both pre-1.0 and (hopefully)
after-1.0 versions of gym, and always returns the initial observation.
Please do not use this function on environments whose observation
spaces or tuples, because then this function cannot distinguish between
environments whose reset(...) methods return a tuple and environments
whose reset(...) methods return a single observation object but that
observation object is a tuple.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env |
Env |
The gym environment which will be reset. |
required |
Returns:
| Type | Description |
|---|---|
Iterable |
The initial observation |
Source code in evotorch/neuroevolution/net/rl.py
def reset_env(env: gym.Env) -> Iterable:
"""
Reset a gym environment.
For gym 1.0, the plan is to have a `reset(...)` method which returns
a two-element tuple `(observation, info)` where `info` is an object
providing any additional information regarding the initial state of
the agent. However, the old (pre 1.0) gym API (and some environments
which were written with old gym compatibility in mind) has (or have)
a `reset(...)` method which returns a single object that is the
initial observation.
With the assumption that the observation space of the environment
is NOT tuple, this function can work with both pre-1.0 and (hopefully)
after-1.0 versions of gym, and always returns the initial observation.
Please do not use this function on environments whose observation
spaces or tuples, because then this function cannot distinguish between
environments whose `reset(...)` methods return a tuple and environments
whose `reset(...)` methods return a single observation object but that
observation object is a tuple.
Args:
env: The gym environment which will be reset.
Returns:
The initial observation
"""
result = env.reset()
if isinstance(result, tuple) and (len(result) == 2):
result = result[0]
return result
take_step_in_env(env, action)
¶
Take a step in the gym environment. Taking a step means performing the action provided via the arguments.
For gym 1.0, the plan is to have a step(...) method which returns a
5-elements tuple containing observation, reward, terminated,
truncated, info where terminated is a boolean indicating whether
or not the episode is terminated because of the actions taken within the
environment, and truncated is a boolean indicating whether or not the
episode is finished because the time limit is reached.
However, the old (pre 1.0) gym API (and some environments which were
written with old gym compatibility in mind) has (or have) a step(...)
method which returns 4 elements: observation, reward, done, info
where done is a boolean indicating whether or not the episode is
"done", either because of termination or because of truncation.
This function can work with both pre-1.0 and (hopefully) after-1.0
versions of gym, and always returns the 4-element tuple as its result.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env |
Env |
The gym environment in which the given action will be performed. |
required |
Returns:
| Type | Description |
|---|---|
tuple |
A tuple in the form |
Source code in evotorch/neuroevolution/net/rl.py
def take_step_in_env(env: gym.Env, action: Iterable) -> tuple:
"""
Take a step in the gym environment.
Taking a step means performing the action provided via the arguments.
For gym 1.0, the plan is to have a `step(...)` method which returns a
5-elements tuple containing `observation`, `reward`, `terminated`,
`truncated`, `info` where `terminated` is a boolean indicating whether
or not the episode is terminated because of the actions taken within the
environment, and `truncated` is a boolean indicating whether or not the
episode is finished because the time limit is reached.
However, the old (pre 1.0) gym API (and some environments which were
written with old gym compatibility in mind) has (or have) a `step(...)`
method which returns 4 elements: `observation`, `reward`, `done`, `info`
where `done` is a boolean indicating whether or not the episode is
"done", either because of termination or because of truncation.
This function can work with both pre-1.0 and (hopefully) after-1.0
versions of gym, and always returns the 4-element tuple as its result.
Args:
env: The gym environment in which the given action will be performed.
Returns:
A tuple in the form `(observation, reward, done, info)` where
`observation` is the observation received after performing the action,
`reward` is the amount of reward gained,
`done` is a boolean value indicating whether or not the episode has
ended, and
`info` is additional information (usually as a dictionary).
"""
result = env.step(action)
if isinstance(result, tuple):
n = len(result)
if n == 4:
observation, reward, done, info = result
elif n == 5:
observation, reward, terminated, truncated, info = result
done = terminated or truncated
else:
raise ValueError(
f"The result of the `step(...)` method of the gym environment"
f" was expected as a tuple of length 4 or 5."
f" However, the received result is {repr(result)}, which is"
f" of length {len(result)}."
)
else:
raise TypeError(
f"The result of the `step(...)` method of the gym environment"
f" was expected as a tuple of length 4 or 5."
f" However, the received result is {repr(result)}, which is"
f" of type {type(result)}."
)
return observation, reward, done, info
runningnorm
¶
CollectedStats (tuple)
¶
ObsNormLayer (Module)
¶
An observation normalizer which behaves as a PyTorch Module.
Source code in evotorch/neuroevolution/net/runningnorm.py
class ObsNormLayer(nn.Module):
"""
An observation normalizer which behaves as a PyTorch Module.
"""
def __init__(
self, mean: torch.Tensor, stdev: torch.Tensor, low: Optional[float] = None, high: Optional[float] = None
) -> None:
"""
`__init__(...)`: Initialize the ObsNormLayer.
Args:
mean: The mean according to which the observations are to be
normalized.
stdev: The standard deviation according to which the observations
are to be normalized.
low: Optionally a real number if the result of the normalization
is to be clipped. Represents the lower bound for the clipping
operation.
high: Optionally a real number if the result of the normalization
is to be clipped. Represents the upper bound for the clipping
operation.
"""
super().__init__()
self.register_buffer("_mean", mean)
self.register_buffer("_stdev", stdev)
self._lb = None if low is None else float(low)
self._ub = None if high is None else float(high)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Normalize an observation or a batch of observations.
Args:
x: The observation(s).
Returns:
The normalized counterpart of the observation(s).
"""
return _clamp((x - self._mean) / self._stdev, self._lb, self._ub)
__init__(self, mean, stdev, low=None, high=None)
special
¶
__init__(...): Initialize the ObsNormLayer.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
mean |
Tensor |
The mean according to which the observations are to be normalized. |
required |
stdev |
Tensor |
The standard deviation according to which the observations are to be normalized. |
required |
low |
Optional[float] |
Optionally a real number if the result of the normalization is to be clipped. Represents the lower bound for the clipping operation. |
None |
high |
Optional[float] |
Optionally a real number if the result of the normalization is to be clipped. Represents the upper bound for the clipping operation. |
None |
Source code in evotorch/neuroevolution/net/runningnorm.py
def __init__(
self, mean: torch.Tensor, stdev: torch.Tensor, low: Optional[float] = None, high: Optional[float] = None
) -> None:
"""
`__init__(...)`: Initialize the ObsNormLayer.
Args:
mean: The mean according to which the observations are to be
normalized.
stdev: The standard deviation according to which the observations
are to be normalized.
low: Optionally a real number if the result of the normalization
is to be clipped. Represents the lower bound for the clipping
operation.
high: Optionally a real number if the result of the normalization
is to be clipped. Represents the upper bound for the clipping
operation.
"""
super().__init__()
self.register_buffer("_mean", mean)
self.register_buffer("_stdev", stdev)
self._lb = None if low is None else float(low)
self._ub = None if high is None else float(high)
forward(self, x)
¶
Normalize an observation or a batch of observations.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Tensor |
The observation(s). |
required |
Returns:
| Type | Description |
|---|---|
Tensor |
The normalized counterpart of the observation(s). |
Source code in evotorch/neuroevolution/net/runningnorm.py
RunningNorm
¶
An online observation normalization tool
Source code in evotorch/neuroevolution/net/runningnorm.py
class RunningNorm:
"""
An online observation normalization tool
"""
def __init__(
self,
*,
shape: Union[tuple, int],
dtype: DType,
device: Optional[Device] = None,
min_variance: float = 1e-2,
clip: Optional[tuple] = None,
) -> None:
"""
`__init__(...)`: Initialize the RunningNorm
Args:
shape: Observation shape. Can be an integer or a tuple.
dtype: The dtype of the observations.
device: The device in which the observation stats are held.
If left as None, the device is assumed to be "cpu".
min_variance: A lower bound for the variance to be used in
the normalization computations.
In other words, if the computed variance according to the
collected observations ends up lower than `min_variance`,
this `min_variance` will be used instead (in an elementwise
manner) while computing the normalized observations.
As in Salimans et al. (2017), the default is 1e-2.
clip: Can be left as None (which is the default), or can be
given as a pair of real numbers.
This is used for clipping the observations after the
normalization operation.
In Salimans et al. (2017), (-5.0, +5.0) was used.
"""
# Make sure that the shape is stored as a torch.Size object.
if isinstance(shape, Iterable):
self._shape = torch.Size(shape)
else:
self._shape = torch.Size([int(shape)])
# Store the number of dimensions
self._ndim = len(self._shape)
# Store the dtype and the device
self._dtype = to_torch_dtype(dtype)
self._device = "cpu" if device is None else device
# Initialize the internally stored data as empty
self._sum: Optional[torch.Tensor] = None
self._sum_of_squares: Optional[torch.Tensor] = None
self._count: int = 0
# Store the minimum variance
self._min_variance = float(min_variance)
if clip is not None:
# If a clip tuple was provided, store the specified lower and upper bounds
lb, ub = clip
self._lb = float(lb)
self._ub = float(ub)
else:
# If a clip tuple was not provided the bounds are stored as None
self._lb = None
self._ub = None
def to(self, device: Device) -> "RunningNorm":
"""
If the target device is a different device, then make a copy of this
RunningNorm instance on the target device.
If the target device is the same with this RunningNorm's device, then
return this RunningNorm itself.
Args:
device: The target device.
Returns:
The RunningNorm on the target device. This can be a copy, or the
original RunningNorm instance itself.
"""
if torch.device(device) == torch.device(self.device):
return self
else:
new_running_norm = object.__new__(type(self))
already_handled = {"_sum", "_sum_of_squares", "_device"}
new_running_norm._sum = self._sum.to(device)
new_running_norm._sum_of_squares = self._sum_of_squares.to(device)
new_running_norm._device = device
for k, v in self.__dict__.items():
if k not in already_handled:
setattr(new_running_norm, k, deepcopy(v))
return new_running_norm
@property
def device(self) -> Device:
"""
The device in which the observation stats are held
"""
return self._device
@property
def dtype(self) -> DType:
"""
The dtype of the stored observation stats
"""
return self._dtype
@property
def shape(self) -> tuple:
"""
Observation shape
"""
return self._shape
@property
def min_variance(self) -> float:
"""
Minimum variance
"""
return self._min_variance
@property
def low(self) -> Optional[float]:
"""
The lower component of the bounds given in the `clip` tuple.
If `clip` was initialized as None, this is also None.
"""
return self._lb
@property
def high(self) -> Optional[float]:
"""
The higher (upper) component of the bounds given in the `clip` tuple.
If `clip` was initialized as None, this is also None.
"""
return self._ub
def _like_its_own(self, x: Iterable) -> torch.Tensor:
return torch.as_tensor(x, dtype=self._dtype, device=self._device)
def _verify(self, x: Iterable) -> torch.Tensor:
x = self._like_its_own(x)
if x.ndim == self._ndim:
if x.shape != self._shape:
raise ValueError(
f"This RunningNorm instance was initialized with shape: {self._shape}."
f" However, the provided tensor has an incompatible shape: {x._shape}."
)
elif x.ndim == (self._ndim + 1):
if x.shape[1:] != self._shape:
raise ValueError(
f"This RunningNorm instance was initialized with shape: {self._shape}."
f" The provided tensor is shaped {x.shape}."
f" Accepting the tensor's leftmost dimension as the batch size,"
f" the remaining shape is incompatible: {x.shape[1:]}"
)
else:
raise ValueError(
f"This RunningNorm instance was initialized with shape: {self._shape}."
f" The provided tensor is shaped {x.shape}."
f" The number of dimensions of the given tensor is incompatible."
)
return x
def _has_no_data(self) -> bool:
return (self._sum is None) and (self._sum_of_squares is None) and (self._count == 0)
def _has_data(self) -> bool:
return (self._sum is not None) and (self._sum_of_squares is not None) and (self._count > 0)
def reset(self):
"""
Remove all the collected observation data.
"""
self._sum = None
self._sum_of_squares = None
self._count = 0
@torch.no_grad()
def update(self, x: Union[Iterable, "RunningNorm"], mask: Optional[Iterable] = None, *, verify: bool = True):
"""
Update the stored stats with new observation data.
Args:
x: The new observation(s), as a PyTorch tensor, or any Iterable
that can be converted to a PyTorch tensor, or another
RunningNorm instance.
If given as a tensor or as an Iterable, the shape of `x` can
be the same with observation shape, or it can be augmented
with an extra leftmost dimension.
In the case of augmented dimension, `x` is interpreted not as
a single observation, but as a batch of observations.
If `x` is another RunningNorm instance, the stats stored by
this RunningNorm instance will be updated with all the data
stored by `x`.
mask: Can be given as a 1-dimensional Iterable of booleans ONLY
if `x` represents a batch of observations.
If a `mask` is provided, the i-th observation within the
observation batch `x` will be taken into account only if
the i-th item of the `mask` is True.
verify: Whether or not to verify the shape of the given Iterable
objects. The default is True.
"""
if isinstance(x, RunningNorm):
# If we are to update our stats according to another RunningNorm instance
if x._count > 0:
# We bother only if x is non-empty
if mask is not None:
# We were given another RunningNorm, not a batch of observations.
# So, we do not expect to receive a mask tensor.
# If a mask was provided, then this is an unexpected way of calling this function.
# We therefore raise an error.
raise ValueError(
"The `mask` argument is expected as None if the first argument is a RunningNorm."
" However, `mask` is found as something other than None."
)
if self._shape != x._shape:
# If the shapes of this RunningNorm and of the other RunningNorm
# do not match, then we cannot use `x` for updating our stats.
# It might be the case that `x` was initialized for another
# task, with differently sized observations.
# We therefore raise an error.
raise ValueError(
f"The RunningNorm to be updated has the shape {self._shape}"
f" The other RunningNorm has the shape {self._shape}"
f" These shapes are incompatible."
)
if self._has_no_data():
# If this RunningNorm has no data at all, then we clone the
# data of x.
self._sum = self._like_its_own(x._sum.clone())
self._sum_of_squares = self._like_its_own(x._sum_of_squares.clone())
self._count = x._count
elif self._has_data():
# If this RunningNorm has its own data, then we update the
# stored data with the data stored by x.
self._sum += self._like_its_own(x._sum)
self._sum_of_squares += self._like_its_own(x._sum_of_squares)
self._count += x._count
else:
assert False, "RunningNorm is in an invalid state! This might be a bug."
else:
# This is the case where the received argument x is not a
# RunningNorm object, but an Iterable.
if verify:
# If we have the `verify` flag, then we make sure that
# x is a tensor of the correct shape
x = self._verify(x)
if x.ndim == self._ndim:
# If the shape of x is exactly the same with the observation shape
# then we assume that x represents a single observation, and not a
# batch of observations.
if mask is not None:
# Since we are dealing with a single observation,
# we do not expect to receive a mask argument.
# If the mask argument was provided, then this is an unexpected
# usage of this function.
# We therefore raise an error.
raise ValueError(
"The `mask` argument is expected as None if the first argument is a single observation"
" (i.e. not a batch of observations, with an extra leftmost dimension)."
" However, `mask` is found as something other than None."
)
# Since x is a single observation,
# the sum of observations extracted from x is x itself,
# and the sum of squared observations extracted from x is
# the square of x itself.
sum_of_x = x
sum_of_x_squared = x.square()
# We extracted a single observation from x
n = 1
elif x.ndim == (self._ndim + 1):
# If the number of dimensions of x is one more than the number
# of dimensions of this RunningNorm, then we assume that x is a batch
# of observations.
if mask is not None:
# If a mask is provided, then we first make sure that it is a tensor
# of dtype bool in the correct device.
mask = torch.as_tensor(mask, dtype=torch.bool, device=self._device)
if mask.ndim != 1:
# We expect the mask to be 1-dimensional.
# If not, we raise an error.
raise ValueError(
f"The `mask` tensor was expected as a 1-dimensional tensor."
f" However, its shape is {mask.shape}."
)
if len(mask) != x.shape[0]:
# If the length of the mask is not the batch size of x,
# then there is a mismatch.
# We therefore raise an error.
raise ValueError(
f"The shape of the given tensor is {x.shape}."
f" Therefore, the batch size of observations is {x.shape[0]}."
f" However, the given `mask` tensor does not has an incompatible length: {len(mask)}."
)
# We compute how many True items we have in the mask.
# This integer gives us how many observations we extract from x.
n = int(torch.sum(torch.as_tensor(mask, dtype=torch.int64, device=self._device)))
# We now re-cast the mask as the observation dtype (so that True items turn to 1.0
# and False items turn to 0.0), and then increase its number of dimensions so that
# it can operate directly with x.
mask = self._like_its_own(mask).reshape(torch.Size([x.shape[0]] + ([1] * (x.ndim - 1))))
# Finally, we multiply x with the mask. This means that the observations with corresponding
# mask values as False are zeroed out.
x = x * mask
else:
# This is the case where we did not receive a mask.
# We can simply say that the number of observations to extract from x
# is the size of its leftmost dimension, i.e. the batch size.
n = x.shape[0]
# With or without a mask, we are now ready to extract the sum and sum of squares
# from x.
sum_of_x = torch.sum(x, dim=0)
sum_of_x_squared = torch.sum(x.square(), dim=0)
else:
# This is the case where the number of dimensions of x is unrecognized.
# This case is actually already checked by the _verify(...) method earlier.
# This defensive fallback case is only for when verify=False and it turned out
# that the ndim is invalid.
raise ValueError(f"Invalid shape: {x.shape}")
# At this point, we handled all the valid cases regarding the Iterable x,
# and we have our sum_of_x (sum of all observations), sum_of_squares
# (sum of all squared observations), and n (number of observations extracted
# from x).
if self._has_no_data():
# If our RunningNorm is empty, the observation data we extracted from x
# become our RunningNorm's new data.
self._sum = sum_of_x
self._sum_of_squares = sum_of_x_squared
self._count = n
elif self._has_data():
# If our RunningNorm is not empty, the stored data is updated with the
# data extracted from x.
self._sum += sum_of_x
self._sum_of_squares += sum_of_x_squared
self._count += n
else:
# This is an erroneous state where the internal data looks neither
# existent nor completely empty.
# This might be the result of a bug, or maybe this instance's
# protected variables were tempered with from the outside.
assert False, "RunningNorm is in an invalid state! This might be a bug."
@property
@torch.no_grad()
def stats(self) -> CollectedStats:
"""
The collected data's mean and standard deviation (stdev) in a tuple
"""
# Using the internally stored sum, sum_of_squares, and count,
# compute E[x] and E[x^2]
E_x = self._sum / self._count
E_x2 = self._sum_of_squares / self._count
# The mean is E[x]
mean = E_x
# The variance is E[x^2] - (E[x])^2, elementwise clipped such that
# it cannot go below min_variance
variance = _clamp(E_x2 - E_x.square(), self._min_variance, None)
# Standard deviation is finally computed as the square root of the variance
stdev = torch.sqrt(variance)
# Return the stats in a named tuple
return CollectedStats(mean=mean, stdev=stdev)
@property
def mean(self) -> torch.Tensor:
"""
The collected data's mean
"""
return self._sum / self._count
@property
def stdev(self) -> torch.Tensor:
"""
The collected data's standard deviation
"""
return self.stats.stdev
@property
def sum(self) -> torch.Tensor:
"""
The collected data's sum
"""
return self._sum
@property
def sum_of_squares(self) -> torch.Tensor:
"""
Sum of squares of the collected data
"""
return self._sum_of_squares
@property
def count(self) -> int:
"""
Number of observations encountered
"""
return self._count
@torch.no_grad()
def normalize(self, x: Iterable, *, result_as_numpy: Optional[bool] = None, verify: bool = True) -> Iterable:
"""
Normalize the given observation x.
Args:
x: The observation(s), as a PyTorch tensor, or any Iterable
that is convertable to a PyTorch tensor.
`x` can be a single observation, or it can be a batch
of observations (with an extra leftmost dimension).
result_as_numpy: Whether or not to return the normalized
observation as a numpy array.
If left as None (which is the default), then the returned
type depends on x: a PyTorch tensor is returned if x is a
PyTorch tensor, and a numpy array is returned otherwise.
If True, the result is always a numpy array.
If False, the result is always a PyTorch tensor.
verify: Whether or not to check the type and dimensions of x.
This is True by default.
Note that, if `verify` is False, this function will not
properly check the type of `x` and will assume that `x`
is a PyTorch tensor.
Returns:
The normalized observation, as a PyTorch tensor or a numpy array.
"""
if self._count == 0:
# If this RunningNorm instance has no data yet,
# then we do not know how to do the normalization.
# We therefore raise an error.
raise ValueError("Cannot do normalization because no data is collected yet.")
if verify:
# Here we verify the type and shape of x.
if result_as_numpy is None:
# If there is not an explicit request about the return type,
# we infer the return type from the type of x:
# if x is a tensor, we return a tensor;
# otherwise, we assume x to be a CPU-bound iterable, and
# therefore we return a numpy array.
result_as_numpy = not isinstance(x, torch.Tensor)
else:
result_as_numpy = bool(result_as_numpy)
# We call _verify() to make sure that x is of correct shape
# and is properly converted to a PyTorch tensor.
x = self._verify(x)
# We get the mean and stdev of the collected data
mean, stdev = self.stats
# Now we compute the normalized observation, clipped according to the
# lower and upper bounds expressed by the `clip` tuple, if exists.
result = _clamp((x - mean) / stdev, self._lb, self._ub)
if result_as_numpy:
# If we are to return the result as a numpy array, we do the
# necessary conversion.
result = result.cpu().numpy()
# Finally, return the result
return result
@torch.no_grad()
def update_and_normalize(self, x: Iterable, mask: Optional[Iterable] = None) -> Iterable:
"""
Update the observation stats according to x, then normalize x.
Args:
x: The observation(s), as a PyTorch tensor, or as an Iterable
which can be converted to a PyTorch tensor.
The shape of x can be the same with the observaiton shape,
or it can be augmented with an extra leftmost dimension
to express a batch of observations.
mask: Can be given as a 1-dimensional Iterable of booleans ONLY
if `x` represents a batch of observations.
If a `mask` is provided, the i-th observation within the
observation batch `x` will be taken into account only if
the the i-th item of the `mask` is True.
Returns:
The normalized counterpart of the observation(s) expressed by x.
"""
result_as_numpy = not isinstance(x, torch.Tensor)
x = self._verify(x)
self.update(x, mask, verify=False)
result = self.normalize(x, verify=False)
if result_as_numpy:
result = result.cpu().numpy()
return result
def to_layer(self) -> "ObsNormLayer":
"""
Make a PyTorch module which normalizes the its inputs.
Returns:
An ObsNormLayer instance.
"""
mean, stdev = self.stats
low = self.low
high = self.high
return ObsNormLayer(mean=mean, stdev=stdev, low=low, high=high)
def __repr__(self) -> str:
return f"<{self.__class__.__name__}, count: {self.count}>"
def __copy__(self) -> "RunningNorm":
return deepcopy(self)
count: int
property
readonly
¶
Number of observations encountered
device: Union[str, torch.device]
property
readonly
¶
The device in which the observation stats are held
dtype: Union[str, torch.dtype, numpy.dtype, Type]
property
readonly
¶
The dtype of the stored observation stats
high: Optional[float]
property
readonly
¶
The higher (upper) component of the bounds given in the clip tuple.
If clip was initialized as None, this is also None.
low: Optional[float]
property
readonly
¶
The lower component of the bounds given in the clip tuple.
If clip was initialized as None, this is also None.
mean: Tensor
property
readonly
¶
The collected data's mean
min_variance: float
property
readonly
¶
Minimum variance
shape: tuple
property
readonly
¶
Observation shape
stats: CollectedStats
property
readonly
¶
The collected data's mean and standard deviation (stdev) in a tuple
stdev: Tensor
property
readonly
¶
The collected data's standard deviation
sum: Tensor
property
readonly
¶
The collected data's sum
sum_of_squares: Tensor
property
readonly
¶
Sum of squares of the collected data
__init__(self, *, shape, dtype, device=None, min_variance=0.01, clip=None)
special
¶
__init__(...): Initialize the RunningNorm
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
shape |
Union[tuple, int] |
Observation shape. Can be an integer or a tuple. |
required |
dtype |
Union[str, torch.dtype, numpy.dtype, Type] |
The dtype of the observations. |
required |
device |
Union[str, torch.device] |
The device in which the observation stats are held. If left as None, the device is assumed to be "cpu". |
None |
min_variance |
float |
A lower bound for the variance to be used in
the normalization computations.
In other words, if the computed variance according to the
collected observations ends up lower than |
0.01 |
clip |
Optional[tuple] |
Can be left as None (which is the default), or can be given as a pair of real numbers. This is used for clipping the observations after the normalization operation. In Salimans et al. (2017), (-5.0, +5.0) was used. |
None |
Source code in evotorch/neuroevolution/net/runningnorm.py
def __init__(
self,
*,
shape: Union[tuple, int],
dtype: DType,
device: Optional[Device] = None,
min_variance: float = 1e-2,
clip: Optional[tuple] = None,
) -> None:
"""
`__init__(...)`: Initialize the RunningNorm
Args:
shape: Observation shape. Can be an integer or a tuple.
dtype: The dtype of the observations.
device: The device in which the observation stats are held.
If left as None, the device is assumed to be "cpu".
min_variance: A lower bound for the variance to be used in
the normalization computations.
In other words, if the computed variance according to the
collected observations ends up lower than `min_variance`,
this `min_variance` will be used instead (in an elementwise
manner) while computing the normalized observations.
As in Salimans et al. (2017), the default is 1e-2.
clip: Can be left as None (which is the default), or can be
given as a pair of real numbers.
This is used for clipping the observations after the
normalization operation.
In Salimans et al. (2017), (-5.0, +5.0) was used.
"""
# Make sure that the shape is stored as a torch.Size object.
if isinstance(shape, Iterable):
self._shape = torch.Size(shape)
else:
self._shape = torch.Size([int(shape)])
# Store the number of dimensions
self._ndim = len(self._shape)
# Store the dtype and the device
self._dtype = to_torch_dtype(dtype)
self._device = "cpu" if device is None else device
# Initialize the internally stored data as empty
self._sum: Optional[torch.Tensor] = None
self._sum_of_squares: Optional[torch.Tensor] = None
self._count: int = 0
# Store the minimum variance
self._min_variance = float(min_variance)
if clip is not None:
# If a clip tuple was provided, store the specified lower and upper bounds
lb, ub = clip
self._lb = float(lb)
self._ub = float(ub)
else:
# If a clip tuple was not provided the bounds are stored as None
self._lb = None
self._ub = None
normalize(self, x, *, result_as_numpy=None, verify=True)
¶
Normalize the given observation x.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Iterable |
The observation(s), as a PyTorch tensor, or any Iterable
that is convertable to a PyTorch tensor.
|
required |
result_as_numpy |
Optional[bool] |
Whether or not to return the normalized observation as a numpy array. If left as None (which is the default), then the returned type depends on x: a PyTorch tensor is returned if x is a PyTorch tensor, and a numpy array is returned otherwise. If True, the result is always a numpy array. If False, the result is always a PyTorch tensor. |
None |
verify |
bool |
Whether or not to check the type and dimensions of x.
This is True by default.
Note that, if |
True |
Returns:
| Type | Description |
|---|---|
Iterable |
The normalized observation, as a PyTorch tensor or a numpy array. |
Source code in evotorch/neuroevolution/net/runningnorm.py
@torch.no_grad()
def normalize(self, x: Iterable, *, result_as_numpy: Optional[bool] = None, verify: bool = True) -> Iterable:
"""
Normalize the given observation x.
Args:
x: The observation(s), as a PyTorch tensor, or any Iterable
that is convertable to a PyTorch tensor.
`x` can be a single observation, or it can be a batch
of observations (with an extra leftmost dimension).
result_as_numpy: Whether or not to return the normalized
observation as a numpy array.
If left as None (which is the default), then the returned
type depends on x: a PyTorch tensor is returned if x is a
PyTorch tensor, and a numpy array is returned otherwise.
If True, the result is always a numpy array.
If False, the result is always a PyTorch tensor.
verify: Whether or not to check the type and dimensions of x.
This is True by default.
Note that, if `verify` is False, this function will not
properly check the type of `x` and will assume that `x`
is a PyTorch tensor.
Returns:
The normalized observation, as a PyTorch tensor or a numpy array.
"""
if self._count == 0:
# If this RunningNorm instance has no data yet,
# then we do not know how to do the normalization.
# We therefore raise an error.
raise ValueError("Cannot do normalization because no data is collected yet.")
if verify:
# Here we verify the type and shape of x.
if result_as_numpy is None:
# If there is not an explicit request about the return type,
# we infer the return type from the type of x:
# if x is a tensor, we return a tensor;
# otherwise, we assume x to be a CPU-bound iterable, and
# therefore we return a numpy array.
result_as_numpy = not isinstance(x, torch.Tensor)
else:
result_as_numpy = bool(result_as_numpy)
# We call _verify() to make sure that x is of correct shape
# and is properly converted to a PyTorch tensor.
x = self._verify(x)
# We get the mean and stdev of the collected data
mean, stdev = self.stats
# Now we compute the normalized observation, clipped according to the
# lower and upper bounds expressed by the `clip` tuple, if exists.
result = _clamp((x - mean) / stdev, self._lb, self._ub)
if result_as_numpy:
# If we are to return the result as a numpy array, we do the
# necessary conversion.
result = result.cpu().numpy()
# Finally, return the result
return result
reset(self)
¶
to(self, device)
¶
If the target device is a different device, then make a copy of this RunningNorm instance on the target device. If the target device is the same with this RunningNorm's device, then return this RunningNorm itself.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
device |
Union[str, torch.device] |
The target device. |
required |
Returns:
| Type | Description |
|---|---|
RunningNorm |
The RunningNorm on the target device. This can be a copy, or the original RunningNorm instance itself. |
Source code in evotorch/neuroevolution/net/runningnorm.py
def to(self, device: Device) -> "RunningNorm":
"""
If the target device is a different device, then make a copy of this
RunningNorm instance on the target device.
If the target device is the same with this RunningNorm's device, then
return this RunningNorm itself.
Args:
device: The target device.
Returns:
The RunningNorm on the target device. This can be a copy, or the
original RunningNorm instance itself.
"""
if torch.device(device) == torch.device(self.device):
return self
else:
new_running_norm = object.__new__(type(self))
already_handled = {"_sum", "_sum_of_squares", "_device"}
new_running_norm._sum = self._sum.to(device)
new_running_norm._sum_of_squares = self._sum_of_squares.to(device)
new_running_norm._device = device
for k, v in self.__dict__.items():
if k not in already_handled:
setattr(new_running_norm, k, deepcopy(v))
return new_running_norm
to_layer(self)
¶
Make a PyTorch module which normalizes the its inputs.
Returns:
| Type | Description |
|---|---|
ObsNormLayer |
An ObsNormLayer instance. |
Source code in evotorch/neuroevolution/net/runningnorm.py
update(self, x, mask=None, *, verify=True)
¶
Update the stored stats with new observation data.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Union[Iterable, RunningNorm] |
The new observation(s), as a PyTorch tensor, or any Iterable
that can be converted to a PyTorch tensor, or another
RunningNorm instance.
If given as a tensor or as an Iterable, the shape of |
required |
mask |
Optional[Iterable] |
Can be given as a 1-dimensional Iterable of booleans ONLY
if |
None |
verify |
bool |
Whether or not to verify the shape of the given Iterable objects. The default is True. |
True |
Source code in evotorch/neuroevolution/net/runningnorm.py
@torch.no_grad()
def update(self, x: Union[Iterable, "RunningNorm"], mask: Optional[Iterable] = None, *, verify: bool = True):
"""
Update the stored stats with new observation data.
Args:
x: The new observation(s), as a PyTorch tensor, or any Iterable
that can be converted to a PyTorch tensor, or another
RunningNorm instance.
If given as a tensor or as an Iterable, the shape of `x` can
be the same with observation shape, or it can be augmented
with an extra leftmost dimension.
In the case of augmented dimension, `x` is interpreted not as
a single observation, but as a batch of observations.
If `x` is another RunningNorm instance, the stats stored by
this RunningNorm instance will be updated with all the data
stored by `x`.
mask: Can be given as a 1-dimensional Iterable of booleans ONLY
if `x` represents a batch of observations.
If a `mask` is provided, the i-th observation within the
observation batch `x` will be taken into account only if
the i-th item of the `mask` is True.
verify: Whether or not to verify the shape of the given Iterable
objects. The default is True.
"""
if isinstance(x, RunningNorm):
# If we are to update our stats according to another RunningNorm instance
if x._count > 0:
# We bother only if x is non-empty
if mask is not None:
# We were given another RunningNorm, not a batch of observations.
# So, we do not expect to receive a mask tensor.
# If a mask was provided, then this is an unexpected way of calling this function.
# We therefore raise an error.
raise ValueError(
"The `mask` argument is expected as None if the first argument is a RunningNorm."
" However, `mask` is found as something other than None."
)
if self._shape != x._shape:
# If the shapes of this RunningNorm and of the other RunningNorm
# do not match, then we cannot use `x` for updating our stats.
# It might be the case that `x` was initialized for another
# task, with differently sized observations.
# We therefore raise an error.
raise ValueError(
f"The RunningNorm to be updated has the shape {self._shape}"
f" The other RunningNorm has the shape {self._shape}"
f" These shapes are incompatible."
)
if self._has_no_data():
# If this RunningNorm has no data at all, then we clone the
# data of x.
self._sum = self._like_its_own(x._sum.clone())
self._sum_of_squares = self._like_its_own(x._sum_of_squares.clone())
self._count = x._count
elif self._has_data():
# If this RunningNorm has its own data, then we update the
# stored data with the data stored by x.
self._sum += self._like_its_own(x._sum)
self._sum_of_squares += self._like_its_own(x._sum_of_squares)
self._count += x._count
else:
assert False, "RunningNorm is in an invalid state! This might be a bug."
else:
# This is the case where the received argument x is not a
# RunningNorm object, but an Iterable.
if verify:
# If we have the `verify` flag, then we make sure that
# x is a tensor of the correct shape
x = self._verify(x)
if x.ndim == self._ndim:
# If the shape of x is exactly the same with the observation shape
# then we assume that x represents a single observation, and not a
# batch of observations.
if mask is not None:
# Since we are dealing with a single observation,
# we do not expect to receive a mask argument.
# If the mask argument was provided, then this is an unexpected
# usage of this function.
# We therefore raise an error.
raise ValueError(
"The `mask` argument is expected as None if the first argument is a single observation"
" (i.e. not a batch of observations, with an extra leftmost dimension)."
" However, `mask` is found as something other than None."
)
# Since x is a single observation,
# the sum of observations extracted from x is x itself,
# and the sum of squared observations extracted from x is
# the square of x itself.
sum_of_x = x
sum_of_x_squared = x.square()
# We extracted a single observation from x
n = 1
elif x.ndim == (self._ndim + 1):
# If the number of dimensions of x is one more than the number
# of dimensions of this RunningNorm, then we assume that x is a batch
# of observations.
if mask is not None:
# If a mask is provided, then we first make sure that it is a tensor
# of dtype bool in the correct device.
mask = torch.as_tensor(mask, dtype=torch.bool, device=self._device)
if mask.ndim != 1:
# We expect the mask to be 1-dimensional.
# If not, we raise an error.
raise ValueError(
f"The `mask` tensor was expected as a 1-dimensional tensor."
f" However, its shape is {mask.shape}."
)
if len(mask) != x.shape[0]:
# If the length of the mask is not the batch size of x,
# then there is a mismatch.
# We therefore raise an error.
raise ValueError(
f"The shape of the given tensor is {x.shape}."
f" Therefore, the batch size of observations is {x.shape[0]}."
f" However, the given `mask` tensor does not has an incompatible length: {len(mask)}."
)
# We compute how many True items we have in the mask.
# This integer gives us how many observations we extract from x.
n = int(torch.sum(torch.as_tensor(mask, dtype=torch.int64, device=self._device)))
# We now re-cast the mask as the observation dtype (so that True items turn to 1.0
# and False items turn to 0.0), and then increase its number of dimensions so that
# it can operate directly with x.
mask = self._like_its_own(mask).reshape(torch.Size([x.shape[0]] + ([1] * (x.ndim - 1))))
# Finally, we multiply x with the mask. This means that the observations with corresponding
# mask values as False are zeroed out.
x = x * mask
else:
# This is the case where we did not receive a mask.
# We can simply say that the number of observations to extract from x
# is the size of its leftmost dimension, i.e. the batch size.
n = x.shape[0]
# With or without a mask, we are now ready to extract the sum and sum of squares
# from x.
sum_of_x = torch.sum(x, dim=0)
sum_of_x_squared = torch.sum(x.square(), dim=0)
else:
# This is the case where the number of dimensions of x is unrecognized.
# This case is actually already checked by the _verify(...) method earlier.
# This defensive fallback case is only for when verify=False and it turned out
# that the ndim is invalid.
raise ValueError(f"Invalid shape: {x.shape}")
# At this point, we handled all the valid cases regarding the Iterable x,
# and we have our sum_of_x (sum of all observations), sum_of_squares
# (sum of all squared observations), and n (number of observations extracted
# from x).
if self._has_no_data():
# If our RunningNorm is empty, the observation data we extracted from x
# become our RunningNorm's new data.
self._sum = sum_of_x
self._sum_of_squares = sum_of_x_squared
self._count = n
elif self._has_data():
# If our RunningNorm is not empty, the stored data is updated with the
# data extracted from x.
self._sum += sum_of_x
self._sum_of_squares += sum_of_x_squared
self._count += n
else:
# This is an erroneous state where the internal data looks neither
# existent nor completely empty.
# This might be the result of a bug, or maybe this instance's
# protected variables were tempered with from the outside.
assert False, "RunningNorm is in an invalid state! This might be a bug."
update_and_normalize(self, x, mask=None)
¶
Update the observation stats according to x, then normalize x.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Iterable |
The observation(s), as a PyTorch tensor, or as an Iterable which can be converted to a PyTorch tensor. The shape of x can be the same with the observaiton shape, or it can be augmented with an extra leftmost dimension to express a batch of observations. |
required |
mask |
Optional[Iterable] |
Can be given as a 1-dimensional Iterable of booleans ONLY
if |
None |
Returns:
| Type | Description |
|---|---|
Iterable |
The normalized counterpart of the observation(s) expressed by x. |
Source code in evotorch/neuroevolution/net/runningnorm.py
@torch.no_grad()
def update_and_normalize(self, x: Iterable, mask: Optional[Iterable] = None) -> Iterable:
"""
Update the observation stats according to x, then normalize x.
Args:
x: The observation(s), as a PyTorch tensor, or as an Iterable
which can be converted to a PyTorch tensor.
The shape of x can be the same with the observaiton shape,
or it can be augmented with an extra leftmost dimension
to express a batch of observations.
mask: Can be given as a 1-dimensional Iterable of booleans ONLY
if `x` represents a batch of observations.
If a `mask` is provided, the i-th observation within the
observation batch `x` will be taken into account only if
the the i-th item of the `mask` is True.
Returns:
The normalized counterpart of the observation(s) expressed by x.
"""
result_as_numpy = not isinstance(x, torch.Tensor)
x = self._verify(x)
self.update(x, mask, verify=False)
result = self.normalize(x, verify=False)
if result_as_numpy:
result = result.cpu().numpy()
return result
runningstat
¶
RunningStat
¶
Tool for efficiently computing the mean and stdev of arrays. The arrays themselves are not stored separately, instead, they are accumulated.
This RunningStat is implemented as a wrapper around RunningNorm. The difference is that the interface of RunningStat is simplified to expect only numpy arrays, and expect only non-vectorized observations. With this simplified interface, RunningStat is meant to be used by GymNE, on classical non-vectorized gym tasks.
Source code in evotorch/neuroevolution/net/runningstat.py
class RunningStat:
"""
Tool for efficiently computing the mean and stdev of arrays.
The arrays themselves are not stored separately,
instead, they are accumulated.
This RunningStat is implemented as a wrapper around RunningNorm.
The difference is that the interface of RunningStat is simplified
to expect only numpy arrays, and expect only non-vectorized
observations.
With this simplified interface, RunningStat is meant to be used
by GymNE, on classical non-vectorized gym tasks.
"""
def __init__(self):
"""
`__init__(...)`: Initialize the RunningStat.
"""
self._rn: Optional[RunningNorm] = None
self.reset()
def reset(self):
"""
Reset the RunningStat to its initial state.
"""
self._rn = None
@property
def count(self) -> int:
"""
Get the number of arrays accumulated.
"""
if self._rn is None:
return 0
else:
return self._rn.count
@property
def sum(self) -> np.ndarray:
"""
Get the sum of all accumulated arrays.
"""
return self._rn.sum.numpy()
@property
def sum_of_squares(self) -> np.ndarray:
"""
Get the sum of squares of all accumulated arrays.
"""
return self._rn.sum_of_squares.numpy()
@property
def mean(self) -> np.ndarray:
"""
Get the mean of all accumulated arrays.
"""
return self._rn.mean.numpy()
@property
def stdev(self) -> np.ndarray:
"""
Get the standard deviation of all accumulated arrays.
"""
return self._rn.stdev.numpy()
def update(self, x: Union[np.ndarray, "RunningStat"]):
"""
Accumulate more data into the RunningStat object.
If the argument is an array, that array is added
as one more data element.
If the argument is another RunningStat instance,
all the stats accumulated by that RunningStat object
are added into this RunningStat object.
"""
if isinstance(x, RunningStat):
if x.count > 0:
if self._rn is None:
self._rn = deepcopy(x._rn)
else:
self._rn.update(x._rn)
else:
if self._rn is None:
x = np.array(x, dtype="float32")
self._rn = RunningNorm(shape=x.shape, dtype="float32", device="cpu")
self._rn.update(x)
def normalize(self, x: Union[np.ndarray, list]) -> np.ndarray:
"""
Normalize the array x according to the accumulated stats.
"""
if self._rn is None:
return x
else:
x = np.array(x, dtype="float32")
return self._rn.normalize(x)
def __copy__(self):
return deepcopy(self)
def __repr__(self) -> str:
return f"<{self.__class__.__name__}, count: {self.count}>"
def to(self, device: Union[str, torch.device]) -> "RunningStat":
"""
If the target device is cpu, return this RunningStat instance itself.
A RunningStat object is meant to work with numpy arrays. Therefore,
any device other than the cpu will trigger an error.
Args:
device: The target device. Only cpu is supported.
Returns:
The original RunningStat.
"""
if torch.device(device) == torch.device("cpu"):
return self
else:
raise ValueError(
f"The received target device is {repr(device)}. However, RunningStat can only work on a cpu."
)
def to_layer(self) -> nn.Module:
"""
Make a PyTorch module which normalizes the its inputs.
Returns:
An ObsNormLayer instance.
"""
return self._rn.to_layer()
count: int
property
readonly
¶
Get the number of arrays accumulated.
mean: ndarray
property
readonly
¶
Get the mean of all accumulated arrays.
stdev: ndarray
property
readonly
¶
Get the standard deviation of all accumulated arrays.
sum: ndarray
property
readonly
¶
Get the sum of all accumulated arrays.
sum_of_squares: ndarray
property
readonly
¶
Get the sum of squares of all accumulated arrays.
__init__(self)
special
¶
normalize(self, x)
¶
Normalize the array x according to the accumulated stats.
reset(self)
¶
to(self, device)
¶
If the target device is cpu, return this RunningStat instance itself. A RunningStat object is meant to work with numpy arrays. Therefore, any device other than the cpu will trigger an error.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
device |
Union[str, torch.device] |
The target device. Only cpu is supported. |
required |
Returns:
| Type | Description |
|---|---|
RunningStat |
The original RunningStat. |
Source code in evotorch/neuroevolution/net/runningstat.py
def to(self, device: Union[str, torch.device]) -> "RunningStat":
"""
If the target device is cpu, return this RunningStat instance itself.
A RunningStat object is meant to work with numpy arrays. Therefore,
any device other than the cpu will trigger an error.
Args:
device: The target device. Only cpu is supported.
Returns:
The original RunningStat.
"""
if torch.device(device) == torch.device("cpu"):
return self
else:
raise ValueError(
f"The received target device is {repr(device)}. However, RunningStat can only work on a cpu."
)
to_layer(self)
¶
Make a PyTorch module which normalizes the its inputs.
Returns:
| Type | Description |
|---|---|
Module |
An ObsNormLayer instance. |
update(self, x)
¶
Accumulate more data into the RunningStat object. If the argument is an array, that array is added as one more data element. If the argument is another RunningStat instance, all the stats accumulated by that RunningStat object are added into this RunningStat object.
Source code in evotorch/neuroevolution/net/runningstat.py
def update(self, x: Union[np.ndarray, "RunningStat"]):
"""
Accumulate more data into the RunningStat object.
If the argument is an array, that array is added
as one more data element.
If the argument is another RunningStat instance,
all the stats accumulated by that RunningStat object
are added into this RunningStat object.
"""
if isinstance(x, RunningStat):
if x.count > 0:
if self._rn is None:
self._rn = deepcopy(x._rn)
else:
self._rn.update(x._rn)
else:
if self._rn is None:
x = np.array(x, dtype="float32")
self._rn = RunningNorm(shape=x.shape, dtype="float32", device="cpu")
self._rn.update(x)
statefulmodule
¶
StatefulModule (Module)
¶
A wrapper that provides a stateful interface for recurrent torch modules.
If the torch module to be wrapped is non-recurrent and its forward method has a single input (the input tensor) and a single output (the output tensor), then this wrapper module acts as a no-op wrapper.
If the torch module to be wrapped is recurrent and its forward method has
two inputs (the input tensor and an optional second argument for the hidden
state) and two outputs (the output tensor and the new hidden state), then
this wrapper brings a new forward-passing interface. In this new interface,
the forward method has a single input (the input tensor) and a single
output (the output tensor). The hidden states, instead of being
explicitly requested via a second argument and returned as a second
result, are stored and used by the wrapper.
When a new series of inputs is to be used, one has to call the reset()
method of this wrapper.
Source code in evotorch/neuroevolution/net/statefulmodule.py
class StatefulModule(nn.Module):
"""
A wrapper that provides a stateful interface for recurrent torch modules.
If the torch module to be wrapped is non-recurrent and its forward method
has a single input (the input tensor) and a single output (the output
tensor), then this wrapper module acts as a no-op wrapper.
If the torch module to be wrapped is recurrent and its forward method has
two inputs (the input tensor and an optional second argument for the hidden
state) and two outputs (the output tensor and the new hidden state), then
this wrapper brings a new forward-passing interface. In this new interface,
the forward method has a single input (the input tensor) and a single
output (the output tensor). The hidden states, instead of being
explicitly requested via a second argument and returned as a second
result, are stored and used by the wrapper.
When a new series of inputs is to be used, one has to call the `reset()`
method of this wrapper.
"""
def __init__(self, wrapped_module: nn.Module):
"""
`__init__(...)`: Initialize the StatefulModule.
Args:
wrapped_module: The `torch.nn.Module` instance to wrap.
"""
super().__init__()
# Declare the variable that will store the hidden state of wrapped_module, if any.
self._hidden: Any = None
# Store the module that is wrapped.
self.wrapped_module = wrapped_module
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self._hidden is None:
# If there is no stored hidden state, then only pass the input tensor to the wrapped module.
out = self.wrapped_module(x)
else:
# If there is a hidden state saved from the previous call to this `forward(...)` method, then pass the
# input tensor and this stored hidden state.
out = self.wrapped_module(x, self._hidden)
if isinstance(out, tuple):
# If the result of the wrapped module is a tuple, then we assume that the wrapped module returned an
# output tensor and a hidden state. We assume the first element of this tuple as the output tensor,
# and the second element as the new hidden state.
# We set the variable y to the output tensor, and we store the new hidden state via the attribute
# `_hidden`.
y, self._hidden = out
else:
# If the result of the wrapped module is not a tuple, then we assume that the wrapped module returned
# only the output tensor. We set the variable y to the output tensor, and set the attribute `_hidden`
# as None to indicate that there was no hidden state received.
y = out
self._hidden = None
# We return y, which stores the output received by the wrapped module.
return y
def reset(self):
"""
Reset the hidden state, if any.
"""
self._hidden = None
__init__(self, wrapped_module)
special
¶
__init__(...): Initialize the StatefulModule.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
wrapped_module |
Module |
The |
required |
Source code in evotorch/neuroevolution/net/statefulmodule.py
def __init__(self, wrapped_module: nn.Module):
"""
`__init__(...)`: Initialize the StatefulModule.
Args:
wrapped_module: The `torch.nn.Module` instance to wrap.
"""
super().__init__()
# Declare the variable that will store the hidden state of wrapped_module, if any.
self._hidden: Any = None
# Store the module that is wrapped.
self.wrapped_module = wrapped_module
forward(self, 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 :class:Module instance afterwards
instead of this since the former takes care of running the
registered hooks while the latter silently ignores them.
Source code in evotorch/neuroevolution/net/statefulmodule.py
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self._hidden is None:
# If there is no stored hidden state, then only pass the input tensor to the wrapped module.
out = self.wrapped_module(x)
else:
# If there is a hidden state saved from the previous call to this `forward(...)` method, then pass the
# input tensor and this stored hidden state.
out = self.wrapped_module(x, self._hidden)
if isinstance(out, tuple):
# If the result of the wrapped module is a tuple, then we assume that the wrapped module returned an
# output tensor and a hidden state. We assume the first element of this tuple as the output tensor,
# and the second element as the new hidden state.
# We set the variable y to the output tensor, and we store the new hidden state via the attribute
# `_hidden`.
y, self._hidden = out
else:
# If the result of the wrapped module is not a tuple, then we assume that the wrapped module returned
# only the output tensor. We set the variable y to the output tensor, and set the attribute `_hidden`
# as None to indicate that there was no hidden state received.
y = out
self._hidden = None
# We return y, which stores the output received by the wrapped module.
return y
reset(self)
¶
ensure_stateful(net)
¶
Ensure that a module is wrapped by StatefulModule.
If the given module is already wrapped by StatefulModule, then the module itself is returned. If the given module is not wrapped by StatefulModule, then this function first wraps the module via a new StatefulModule instance, and then this new wrapper is returned.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
net |
Module |
The |
required |
Returns:
| Type | Description |
|---|---|
The module |
Source code in evotorch/neuroevolution/net/statefulmodule.py
def ensure_stateful(net: nn.Module):
"""
Ensure that a module is wrapped by StatefulModule.
If the given module is already wrapped by StatefulModule, then the
module itself is returned.
If the given module is not wrapped by StatefulModule, then this function
first wraps the module via a new StatefulModule instance, and then this
new wrapper is returned.
Args:
net: The `torch.nn.Module` to be wrapped by StatefulModule (if it is
not already wrapped by it).
Returns:
The module `net`, wrapped by StatefulModule.
"""
if not isinstance(net, StatefulModule):
return StatefulModule(net)
vecrl
¶
This namespace provides various vectorized reinforcement learning utilities.
Policy
¶
A Policy for deciding the actions for a reinforcement learning environment.
This can be seen as a stateful wrapper around a PyTorch module.
Let us assume that we have the following PyTorch module:
which has 48 parameters (when all the parameters are flattened).
Let us randomly generate a parameter vector for our module net:
We can now prepare a policy:
If we generate a random observation:
We can receive our action as follows:
If the PyTorch module that we wish to wrap is a recurrent network (i.e. a network which expects an optional second argument for the hidden state, and returns a second value which represents the updated hidden state), then, the hidden state is automatically managed by the Policy instance.
Let us assume that we have a recurrent network named recnet.
In this case, because the hidden state of the network is internally managed, the usage is still the same with our previous non-recurrent
Examples:
When using a recurrent module on multiple episodes, it is important to reset the hidden state of the network. This is achieved by the reset method:
policy.reset()
action1 = policy(observation1)
# action2 will be computed with the hidden state generated by the
# previous forward-pass.
action2 = policy(observation2)
policy.reset()
# action3 will be computed according to the renewed hidden state.
action3 = policy(observation3)
Both for non-recurrent and recurrent networks, it is possible to perform vectorized operations. For now, let us return to our first non-recurrent example:
Instead of generating only one parameter vector, we now generate a batch of parameter vectors. Let us say that our batch size is 10:
Like we did in the non-batched examples, we can do:
Because we are now in the batched mode, policy now expects a batch
of observations and will return a batch of actions:
When doing vectorized reinforcement learning with a recurrent module,
it can be the case that only some of the environments are finished,
and therefore it is necessary to reset the hidden states associated
with those environments only. The reset(...) method of Policy
has a second argument to specify which of the recurrent network
instances are to be reset. For example, if the episodes of the
environments with indices 2 and 5 are about to restart (and therefore
we wish to reset the states of the networks with indices 2 and 5),
then, we can do:
Source code in evotorch/neuroevolution/net/vecrl.py
class Policy:
"""
A Policy for deciding the actions for a reinforcement learning environment.
This can be seen as a stateful wrapper around a PyTorch module.
Let us assume that we have the following PyTorch module:
```python
from torch import nn
net = nn.Linear(5, 8)
```
which has 48 parameters (when all the parameters are flattened).
Let us randomly generate a parameter vector for our module `net`:
```python
parameters = torch.randn(48)
```
We can now prepare a policy:
```python
policy = Policy(net)
policy.set_parameters(parameters)
```
If we generate a random observation:
```python
observation = torch.randn(5)
```
We can receive our action as follows:
```python
action = policy(observation)
```
If the PyTorch module that we wish to wrap is a recurrent network (i.e.
a network which expects an optional second argument for the hidden state,
and returns a second value which represents the updated hidden state),
then, the hidden state is automatically managed by the Policy instance.
Let us assume that we have a recurrent network named `recnet`.
```python
policy = Policy(recnet)
policy.set_parameters(parameters_of_recnet)
```
In this case, because the hidden state of the network is internally
managed, the usage is still the same with our previous non-recurrent
example:
```python
action = policy(observation)
```
When using a recurrent module on multiple episodes, it is important
to reset the hidden state of the network. This is achieved by the
reset method:
```python
policy.reset()
action1 = policy(observation1)
# action2 will be computed with the hidden state generated by the
# previous forward-pass.
action2 = policy(observation2)
policy.reset()
# action3 will be computed according to the renewed hidden state.
action3 = policy(observation3)
```
Both for non-recurrent and recurrent networks, it is possible to
perform vectorized operations. For now, let us return to our
first non-recurrent example:
```python
net = nn.Linear(5, 8)
```
Instead of generating only one parameter vector, we now generate
a batch of parameter vectors. Let us say that our batch size is 10:
```python
batch_of_parameters = torch.randn(10, 48)
```
Like we did in the non-batched examples, we can do:
```python
policy = Policy(net)
policy.set_parameters(batch_of_parameters)
```
Because we are now in the batched mode, `policy` now expects a batch
of observations and will return a batch of actions:
```python
batch_of_observations = torch.randn(10, 5)
batch_of_actions = policy(batch_of_observations)
```
When doing vectorized reinforcement learning with a recurrent module,
it can be the case that only some of the environments are finished,
and therefore it is necessary to reset the hidden states associated
with those environments only. The `reset(...)` method of Policy
has a second argument to specify which of the recurrent network
instances are to be reset. For example, if the episodes of the
environments with indices 2 and 5 are about to restart (and therefore
we wish to reset the states of the networks with indices 2 and 5),
then, we can do:
```python
policy.reset(torch.tensor([2, 5]))
```
"""
def __init__(self, net: Union[str, Callable, nn.Module], **kwargs):
"""
`__init__(...)`: Initialize the Policy.
Args:
net: The network to be wrapped by the Policy object.
This can be a string, a Callable (e.g. a `torch.nn.Module`
subclass), or a `torch.nn.Module` instance.
When this argument is a string, the network will be
created with the help of the function
`evotorch.neuroevolution.net.str_to_net(...)` and then
wrapped. Please see the `str_to_net(...)` function's
documentation for details regarding how a network structure
can be expressed via strings.
kwargs: Expected in the form of additional keyword arguments,
these keyword arguments will be passed to the provided
Callable object (if the argument `net` is a Callable)
or to `str_to_net(...)` (if the argument `net` is a string)
at the moment of generating the network.
If the argument `net` is a `torch.nn.Module` instance,
having any additional keyword arguments will trigger an
error, because the network is already instantiated and
therefore, it is not possible to pass these keyword arguments.
"""
from ..net import str_to_net
from ..net.functional import ModuleExpectingFlatParameters, make_functional_module
if isinstance(net, str):
self.__module = str_to_net(net, **kwargs)
elif isinstance(net, nn.Module):
if len(kwargs) > 0:
raise ValueError(
f"When the network is given as an `nn.Module` instance, extra network arguments cannot be used"
f" (because the network is already instantiated)."
f" However, these extra keyword arguments were received: {kwargs}."
)
self.__module = net
elif isinstance(net, Callable):
self.__module = net(**kwargs)
else:
raise TypeError(
f"The class `Policy` expected a string or an `nn.Module` instance, or a Callable, but received {net}"
f" (whose type is {type(net)})."
)
self.__fmodule: ModuleExpectingFlatParameters = make_functional_module(self.__module)
self.__state: Any = None
self.__parameters: Optional[torch.Tensor] = None
def set_parameters(self, parameters: torch.Tensor, indices: Optional[MaskOrIndices] = None, *, reset: bool = True):
"""
Set the parameters of the policy.
Args:
parameters: A 1-dimensional or a 2-dimensional tensor containing
the flattened parameters to be used with the neural network.
If the given parameters are two-dimensional, then, given that
the leftmost size of the parameter tensor is `n`, the
observations will be expected in a batch with leftmost size
`n`, and the returned actions will also be in a batch,
again with the leftmost size `n`.
indices: For when the parameters were previously given via a
2-dimensional tensor, provide this argument if you would like
to change only some rows of the previously given parameters.
For example, if `indices` is given as `torch.tensor([2, 4])`
and the argument `parameters` is given as a 2-dimensional
tensor with leftmost size 2, then the rows with indices
2 and 4 will be replaced by these new parameters provided
via the argument `parameters`.
reset: If given as True, the hidden states of the networks whose
parameters just changed will be reset. If `indices` was not
provided at all, then this means that the parameters of all
networks are modified, in which case, all the hidden states
will be reset.
If given as False, no such resetting will be done.
"""
if self.__parameters is None:
if indices is not None:
raise ValueError(
"The argument `indices` can be used only if network parameters were previously specified."
" However, it seems that the method `set_parameters(...)` was not called before."
)
self.__parameters = parameters
else:
if indices is None:
self.__parameters = parameters
else:
self.__parameters[indices] = parameters
if reset:
self.reset(indices)
def __call__(self, x: torch.Tensor) -> torch.Tensor:
"""
Pass the given observations through the network.
Args:
x: The observations, as a PyTorch tensor.
If the parameters were given (via the method
`set_parameters(...)`) as a 1-dimensional tensor, then this
argument is expected to store a single observation.
If the parameters were given as a 2-dimensional tensor,
then, this argument is expected to store a batch of
observations, and the leftmost size of this observation
tensor must match with the leftmost size of the parameter
tensor.
Returns:
The output tensor, which represents the action to take.
"""
if self.__parameters is None:
raise ValueError("Please use the method `set_parameters(...)` before calling the policy.")
if self.__state is None:
further_args = (x,)
else:
further_args = (x, self.__state)
parameters = self.__parameters
ndim = parameters.ndim
if ndim == 1:
result = self.__fmodule(parameters, *further_args)
elif ndim == 2:
vmapped = vmap(self.__fmodule)
result = vmapped(parameters, *further_args)
else:
raise ValueError(
f"Expected the parameters as a 1 or 2 dimensional tensor."
f" However, the received parameters tensor has {ndim} dimensions."
)
if isinstance(result, torch.Tensor):
return result
elif isinstance(result, tuple):
result, state = result
self.__state = state
return result
else:
raise TypeError(f"The torch module used by the Policy returned an unexpected object: {result}")
def reset(self, indices: Optional[MaskOrIndices] = None, *, copy: bool = True):
"""
Reset the hidden states, if the contained module is a recurrent network.
Args:
indices: Optionally a sequence of integers or a sequence of
booleans, specifying which networks' states will be
reset. If left as None, then the states of all the networks
will be reset.
copy: When `indices` is given as something other than None,
if `copy` is given as True, then the resetting will NOT
be done in-place. Instead, a new copy of the hidden state
will first be created, and then the specified regions
of this new copy will be cleared, and then finally this
modified copy will be declared as the new hidden state.
It is a common practice for recurrent neural network
implementations to return the same tensor both as its
output and as (part of) its hidden state. With `copy=False`,
the resetting would be done in-place, and the action
tensor could be involuntarily reset as well.
This in-place modification could cause silent bugs
if the unintended modification on the action tensor
happens BEFORE the action is sent to the reinforcement
learning environment.
To prevent such situations, the default value for the argument
`copy` is True.
"""
if indices is None:
self.__state = None
else:
if self.__state is not None:
with torch.no_grad():
if copy:
self.__state = deepcopy(self.__state)
reset_tensors(self.__state, indices)
@property
def parameters(self) -> torch.Tensor:
"""
The currently used parameters.
"""
return self.__parameters
@property
def h(self) -> Optional[torch.Tensor]:
"""
The hidden state of the contained recurrent network, if any.
If the contained recurrent network did not generate a hidden state
yet, or if the contained network is not recurrent, then the result
will be None.
"""
return self.__state
@property
def parameter_length(self) -> int:
"""
Length of the parameter tensor.
"""
return self.__fmodule.parameter_length
@property
def wrapped_module(self) -> nn.Module:
"""
The wrapped `torch.nn.Module` instance.
"""
return self.__module
def to_torch_module(self, parameter_vector: torch.Tensor) -> nn.Module:
"""
Get a copy of the contained network, parameterized as specified.
Args:
parameter_vector: The parameters to be used by the new network.
Returns:
Copy of the contained network, as a `torch.nn.Module` instance.
"""
with torch.no_grad():
net = deepcopy(self.__module).to(parameter_vector.device)
nnu.vector_to_parameters(parameter_vector, net.parameters())
return net
h: Optional[torch.Tensor]
property
readonly
¶
The hidden state of the contained recurrent network, if any.
If the contained recurrent network did not generate a hidden state yet, or if the contained network is not recurrent, then the result will be None.
parameter_length: int
property
readonly
¶
Length of the parameter tensor.
parameters: Tensor
property
readonly
¶
The currently used parameters.
wrapped_module: Module
property
readonly
¶
The wrapped torch.nn.Module instance.
__call__(self, x)
special
¶
Pass the given observations through the network.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Tensor |
The observations, as a PyTorch tensor.
If the parameters were given (via the method
|
required |
Returns:
| Type | Description |
|---|---|
Tensor |
The output tensor, which represents the action to take. |
Source code in evotorch/neuroevolution/net/vecrl.py
def __call__(self, x: torch.Tensor) -> torch.Tensor:
"""
Pass the given observations through the network.
Args:
x: The observations, as a PyTorch tensor.
If the parameters were given (via the method
`set_parameters(...)`) as a 1-dimensional tensor, then this
argument is expected to store a single observation.
If the parameters were given as a 2-dimensional tensor,
then, this argument is expected to store a batch of
observations, and the leftmost size of this observation
tensor must match with the leftmost size of the parameter
tensor.
Returns:
The output tensor, which represents the action to take.
"""
if self.__parameters is None:
raise ValueError("Please use the method `set_parameters(...)` before calling the policy.")
if self.__state is None:
further_args = (x,)
else:
further_args = (x, self.__state)
parameters = self.__parameters
ndim = parameters.ndim
if ndim == 1:
result = self.__fmodule(parameters, *further_args)
elif ndim == 2:
vmapped = vmap(self.__fmodule)
result = vmapped(parameters, *further_args)
else:
raise ValueError(
f"Expected the parameters as a 1 or 2 dimensional tensor."
f" However, the received parameters tensor has {ndim} dimensions."
)
if isinstance(result, torch.Tensor):
return result
elif isinstance(result, tuple):
result, state = result
self.__state = state
return result
else:
raise TypeError(f"The torch module used by the Policy returned an unexpected object: {result}")
__init__(self, net, **kwargs)
special
¶
__init__(...): Initialize the Policy.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
net |
Union[str, Callable, torch.nn.modules.module.Module] |
The network to be wrapped by the Policy object.
This can be a string, a Callable (e.g. a |
required |
kwargs |
Expected in the form of additional keyword arguments,
these keyword arguments will be passed to the provided
Callable object (if the argument |
{} |
Source code in evotorch/neuroevolution/net/vecrl.py
def __init__(self, net: Union[str, Callable, nn.Module], **kwargs):
"""
`__init__(...)`: Initialize the Policy.
Args:
net: The network to be wrapped by the Policy object.
This can be a string, a Callable (e.g. a `torch.nn.Module`
subclass), or a `torch.nn.Module` instance.
When this argument is a string, the network will be
created with the help of the function
`evotorch.neuroevolution.net.str_to_net(...)` and then
wrapped. Please see the `str_to_net(...)` function's
documentation for details regarding how a network structure
can be expressed via strings.
kwargs: Expected in the form of additional keyword arguments,
these keyword arguments will be passed to the provided
Callable object (if the argument `net` is a Callable)
or to `str_to_net(...)` (if the argument `net` is a string)
at the moment of generating the network.
If the argument `net` is a `torch.nn.Module` instance,
having any additional keyword arguments will trigger an
error, because the network is already instantiated and
therefore, it is not possible to pass these keyword arguments.
"""
from ..net import str_to_net
from ..net.functional import ModuleExpectingFlatParameters, make_functional_module
if isinstance(net, str):
self.__module = str_to_net(net, **kwargs)
elif isinstance(net, nn.Module):
if len(kwargs) > 0:
raise ValueError(
f"When the network is given as an `nn.Module` instance, extra network arguments cannot be used"
f" (because the network is already instantiated)."
f" However, these extra keyword arguments were received: {kwargs}."
)
self.__module = net
elif isinstance(net, Callable):
self.__module = net(**kwargs)
else:
raise TypeError(
f"The class `Policy` expected a string or an `nn.Module` instance, or a Callable, but received {net}"
f" (whose type is {type(net)})."
)
self.__fmodule: ModuleExpectingFlatParameters = make_functional_module(self.__module)
self.__state: Any = None
self.__parameters: Optional[torch.Tensor] = None
reset(self, indices=None, *, copy=True)
¶
Reset the hidden states, if the contained module is a recurrent network.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
indices |
Union[int, Iterable] |
Optionally a sequence of integers or a sequence of booleans, specifying which networks' states will be reset. If left as None, then the states of all the networks will be reset. |
None |
copy |
bool |
When |
True |
Source code in evotorch/neuroevolution/net/vecrl.py
def reset(self, indices: Optional[MaskOrIndices] = None, *, copy: bool = True):
"""
Reset the hidden states, if the contained module is a recurrent network.
Args:
indices: Optionally a sequence of integers or a sequence of
booleans, specifying which networks' states will be
reset. If left as None, then the states of all the networks
will be reset.
copy: When `indices` is given as something other than None,
if `copy` is given as True, then the resetting will NOT
be done in-place. Instead, a new copy of the hidden state
will first be created, and then the specified regions
of this new copy will be cleared, and then finally this
modified copy will be declared as the new hidden state.
It is a common practice for recurrent neural network
implementations to return the same tensor both as its
output and as (part of) its hidden state. With `copy=False`,
the resetting would be done in-place, and the action
tensor could be involuntarily reset as well.
This in-place modification could cause silent bugs
if the unintended modification on the action tensor
happens BEFORE the action is sent to the reinforcement
learning environment.
To prevent such situations, the default value for the argument
`copy` is True.
"""
if indices is None:
self.__state = None
else:
if self.__state is not None:
with torch.no_grad():
if copy:
self.__state = deepcopy(self.__state)
reset_tensors(self.__state, indices)
set_parameters(self, parameters, indices=None, *, reset=True)
¶
Set the parameters of the policy.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
parameters |
Tensor |
A 1-dimensional or a 2-dimensional tensor containing
the flattened parameters to be used with the neural network.
If the given parameters are two-dimensional, then, given that
the leftmost size of the parameter tensor is |
required |
indices |
Union[int, Iterable] |
For when the parameters were previously given via a
2-dimensional tensor, provide this argument if you would like
to change only some rows of the previously given parameters.
For example, if |
None |
reset |
bool |
If given as True, the hidden states of the networks whose
parameters just changed will be reset. If |
True |
Source code in evotorch/neuroevolution/net/vecrl.py
def set_parameters(self, parameters: torch.Tensor, indices: Optional[MaskOrIndices] = None, *, reset: bool = True):
"""
Set the parameters of the policy.
Args:
parameters: A 1-dimensional or a 2-dimensional tensor containing
the flattened parameters to be used with the neural network.
If the given parameters are two-dimensional, then, given that
the leftmost size of the parameter tensor is `n`, the
observations will be expected in a batch with leftmost size
`n`, and the returned actions will also be in a batch,
again with the leftmost size `n`.
indices: For when the parameters were previously given via a
2-dimensional tensor, provide this argument if you would like
to change only some rows of the previously given parameters.
For example, if `indices` is given as `torch.tensor([2, 4])`
and the argument `parameters` is given as a 2-dimensional
tensor with leftmost size 2, then the rows with indices
2 and 4 will be replaced by these new parameters provided
via the argument `parameters`.
reset: If given as True, the hidden states of the networks whose
parameters just changed will be reset. If `indices` was not
provided at all, then this means that the parameters of all
networks are modified, in which case, all the hidden states
will be reset.
If given as False, no such resetting will be done.
"""
if self.__parameters is None:
if indices is not None:
raise ValueError(
"The argument `indices` can be used only if network parameters were previously specified."
" However, it seems that the method `set_parameters(...)` was not called before."
)
self.__parameters = parameters
else:
if indices is None:
self.__parameters = parameters
else:
self.__parameters[indices] = parameters
if reset:
self.reset(indices)
to_torch_module(self, parameter_vector)
¶
Get a copy of the contained network, parameterized as specified.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
parameter_vector |
Tensor |
The parameters to be used by the new network. |
required |
Returns:
| Type | Description |
|---|---|
Module |
Copy of the contained network, as a |
Source code in evotorch/neuroevolution/net/vecrl.py
def to_torch_module(self, parameter_vector: torch.Tensor) -> nn.Module:
"""
Get a copy of the contained network, parameterized as specified.
Args:
parameter_vector: The parameters to be used by the new network.
Returns:
Copy of the contained network, as a `torch.nn.Module` instance.
"""
with torch.no_grad():
net = deepcopy(self.__module).to(parameter_vector.device)
nnu.vector_to_parameters(parameter_vector, net.parameters())
return net
TorchWrapper (Wrapper)
¶
A gym wrapper which ensures that the actions, observations, rewards, and the 'done' values are expressed as PyTorch tensors.
Source code in evotorch/neuroevolution/net/vecrl.py
class TorchWrapper(gym.Wrapper):
"""
A gym wrapper which ensures that the actions, observations, rewards, and
the 'done' values are expressed as PyTorch tensors.
"""
def __init__(
self,
env: Union[gym.Env],
*,
force_classic_api: bool = False,
discrete_to_continuous_act: bool = False,
clip_actions: bool = False,
**kwargs,
):
"""
`__init__(...)`: Initialize the TorchWrapper.
Args:
env: The gym environment to be wrapped.
force_classic_api: Set this as True if you would like to enable
the classic API. In the classic API, the `reset(...)` method
returns only the observation and the `step(...)` method
returns 4 elements (not 5).
discrete_to_continuous_act: When this is set as True and the
wrapped environment has a Discrete action space, this wrapper
will transform the action space to Box. A Discrete-action
environment with `n` actions will be converted to a Box-action
environment where the action length is `n`.
The index of the largest value within the action vector will
be applied to the underlying environment.
clip_actions: Set this as True if you would like to clip the given
actions so that they conform to the declared boundaries of the
action space.
kwargs: Expected in the form of additional keyword arguments.
These additional keyword arguments are passed to the
superclass.
"""
super().__init__(env, **kwargs)
# Declare the variable that will store the array type of the underlying environment.
self.__array_type: Optional[str] = None
if hasattr(env, "single_observation_space"):
# If the underlying environment has the attribute "single_observation_space",
# then this is a vectorized environment.
self.__vectorized = True
# Get the observation and action spaces.
obs_space = env.single_observation_space
act_space = env.single_action_space
else:
# If the underlying environment has the attribute "single_observation_space",
# then this is a non-vectorized environment.
self.__vectorized = False
# Get the observation and action spaces.
obs_space = env.observation_space
act_space = env.action_space
# Ensure that the observation and action spaces are supported.
_must_be_supported_space(obs_space)
_must_be_supported_space(act_space)
# Store the choice of the user regarding "force_classic_api".
self.__force_classic_api = bool(force_classic_api)
if isinstance(act_space, Discrete) and discrete_to_continuous_act:
# The underlying action space is Discrete and `discrete_to_continuous_act` is given as True.
# Therefore, we convert the action space to continuous (to Box).
# Take the shape and the dtype of the discrete action space.
single_action_shape = (act_space.n,)
single_action_dtype = torch.from_numpy(np.array([], dtype=act_space.dtype)).dtype
# We store the integer dtype of the environment.
self.__discrete_dtype = single_action_dtype
if self.__vectorized:
# If the environment is vectorized, we declare the new `action_space` and the `single_action_space`
# for the enviornment.
action_shape = (env.num_envs,) + single_action_shape
self.single_action_space = Box(float("-inf"), float("inf"), shape=single_action_shape, dtype=np.float32)
self.action_space = Box(float("-inf"), float("inf"), shape=action_shape, dtype=np.float32)
else:
# If the environment is not vectorized, we declare the new `action_space` for the environment.
self.action_space = Box(float("-inf"), float("inf"), shape=single_action_shape, dtype=np.float32)
else:
# This is the case where we do not transform the action space.
# The discrete dtype will not be used, so, we set it as None.
self.__discrete_dtype = None
if isinstance(act_space, Box) and clip_actions:
# If the action space is Box and the wrapper is configured to clip the actions, then we store the lower
# and the upper bounds for the actions.
self.__act_lb = torch.from_numpy(act_space.low)
self.__act_ub = torch.from_numpy(act_space.high)
else:
# If there will not be any action clipping, then we store the lower and the upper bounds as None.
self.__act_lb = None
self.__act_ub = None
@property
def array_type(self) -> Optional[str]:
"""
Get the array type of the wrapped environment.
This can be "jax", "torch", or "numpy".
"""
return self.__array_type
def __infer_array_type(self, observation):
if self.__array_type is None:
# If the array type is not determined yet, set it as the array type of the received observation.
# If the observation has an unrecognized type, set the array type as "numpy".
self.__array_type = array_type(observation, "numpy")
def reset(self, *args, **kwargs):
"""Reset the environment"""
# Call the reset method of the wrapped environment.
reset_result = self.env.reset(*args, **kwargs)
if isinstance(reset_result, tuple):
# If we received a tuple of two elements, then we assume that this is the new gym API.
# We note that we received an info dictionary.
got_info = True
# We keep the received observation and info.
observation, info = reset_result
else:
# If we did not receive a tuple, then we assume that this is the old gym API.
# We note that we did not receive an info dictionary.
got_info = False
# We keep the received observation.
observation = reset_result
# We did not receive an info dictionary, so, we set it as an empty dictionary.
info = {}
# We understand the array type of the underlying environment from the first observation.
self.__infer_array_type(observation)
# Convert the observation to a PyTorch tensor.
observation = convert_to_torch(observation)
if self.__force_classic_api:
# If the option `force_classic_api` was set as True, then we only return the observation.
return observation
else:
# Here we handle the case where `force_classic_api` was set as False.
if got_info:
# If we got an additional info dictionary, we return it next to the observation.
return observation, info
else:
# If we did not get any info dictionary, we return only the observation.
return observation
def step(self, action, *args, **kwargs):
"""Take a step in the environment"""
if self.__array_type is None:
# If the array type is not known yet, then probably `reset()` has not been called yet.
# We raise an error.
raise ValueError(
"Could not understand what type of array this environment works with."
" Perhaps the `reset()` method has not been called yet?"
)
if self.__discrete_dtype is not None:
# If the wrapped environment is discrete-actioned, then we take the integer counterpart of the action.
action = torch.argmax(action, dim=-1).to(dtype=self.__discrete_dtype)
if self.__act_lb is not None:
# The internal variable `__act_lb` having a value other than None means that the initialization argument
# `clip_actions` was given as True.
# Therefore, we clip the actions.
self.__act_lb = self.__act_lb.to(action.device)
self.__act_ub = self.__act_ub.to(action.device)
action = torch.max(action, self.__act_lb)
action = torch.min(action, self.__act_ub)
# Convert the action tensor to the expected array type of the underlying environment.
action = convert_from_torch(action, self.__array_type)
# Perform the step and get the result.
result = self.env.step(action, *args, **kwargs)
if not isinstance(result, tuple):
# If the `step(...)` method returned anything other than tuple, we raise an error.
raise TypeError(f"Expected a tuple as the result of the `step()` method, but received a {type(result)}")
if len(result) == 5:
# If the result is a tuple of 5 elements, then we note that we are using the new API.
using_new_api = True
# Take the observation, reward, two boolean variables done and done2 indicating that the episode(s)
# has/have ended, and additional info.
# `done` indicates whether or not the episode(s) reached terminal state(s).
# `done2` indicates whether or not the episode(s) got truncated because of the timestep limit.
observation, reward, done, done2, info = result
elif len(result) == 4:
# If the result is a tuple of 5 elements, then we note that we are not using the new API.
using_new_api = False
# Take the observation, reward, the done boolean flag, and additional info.
observation, reward, done, info = result
done2 = None
else:
raise ValueError(f"Unexpected number of elements were returned from step(): {len(result)}")
# Convert the observation, reward, and done variables to PyTorch tensors.
observation = convert_to_torch(observation)
reward = convert_to_torch(reward)
done = convert_to_torch_bool(done)
if done2 is not None:
done2 = convert_to_torch_bool(done2)
if self.__force_classic_api:
# This is the case where the initialization argument `force_classic_api` was set as True.
if done2 is not None:
# We combine the terminal state and truncation signals into a single boolean tensor indicating
# whether or not the episode(s) ended.
done = done | done2
# Return 4 elements, compatible with the classic gym API.
return observation, reward, done, info
else:
# This is the case where the initialization argument `force_classic_api` was set as False.
if using_new_api:
# If we are using the new API, then we return the 5-element result.
return observation, reward, done, done2, info
else:
# If we are using the new API, then we return the 4-element result.
return observation, reward, done, info
array_type: Optional[str]
property
readonly
¶
Get the array type of the wrapped environment. This can be "jax", "torch", or "numpy".
__init__(self, env, *, force_classic_api=False, discrete_to_continuous_act=False, clip_actions=False, **kwargs)
special
¶
__init__(...): Initialize the TorchWrapper.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env |
Env |
The gym environment to be wrapped. |
required |
force_classic_api |
bool |
Set this as True if you would like to enable
the classic API. In the classic API, the |
False |
discrete_to_continuous_act |
bool |
When this is set as True and the
wrapped environment has a Discrete action space, this wrapper
will transform the action space to Box. A Discrete-action
environment with |
False |
clip_actions |
bool |
Set this as True if you would like to clip the given actions so that they conform to the declared boundaries of the action space. |
False |
kwargs |
Expected in the form of additional keyword arguments. These additional keyword arguments are passed to the superclass. |
{} |
Source code in evotorch/neuroevolution/net/vecrl.py
def __init__(
self,
env: Union[gym.Env],
*,
force_classic_api: bool = False,
discrete_to_continuous_act: bool = False,
clip_actions: bool = False,
**kwargs,
):
"""
`__init__(...)`: Initialize the TorchWrapper.
Args:
env: The gym environment to be wrapped.
force_classic_api: Set this as True if you would like to enable
the classic API. In the classic API, the `reset(...)` method
returns only the observation and the `step(...)` method
returns 4 elements (not 5).
discrete_to_continuous_act: When this is set as True and the
wrapped environment has a Discrete action space, this wrapper
will transform the action space to Box. A Discrete-action
environment with `n` actions will be converted to a Box-action
environment where the action length is `n`.
The index of the largest value within the action vector will
be applied to the underlying environment.
clip_actions: Set this as True if you would like to clip the given
actions so that they conform to the declared boundaries of the
action space.
kwargs: Expected in the form of additional keyword arguments.
These additional keyword arguments are passed to the
superclass.
"""
super().__init__(env, **kwargs)
# Declare the variable that will store the array type of the underlying environment.
self.__array_type: Optional[str] = None
if hasattr(env, "single_observation_space"):
# If the underlying environment has the attribute "single_observation_space",
# then this is a vectorized environment.
self.__vectorized = True
# Get the observation and action spaces.
obs_space = env.single_observation_space
act_space = env.single_action_space
else:
# If the underlying environment has the attribute "single_observation_space",
# then this is a non-vectorized environment.
self.__vectorized = False
# Get the observation and action spaces.
obs_space = env.observation_space
act_space = env.action_space
# Ensure that the observation and action spaces are supported.
_must_be_supported_space(obs_space)
_must_be_supported_space(act_space)
# Store the choice of the user regarding "force_classic_api".
self.__force_classic_api = bool(force_classic_api)
if isinstance(act_space, Discrete) and discrete_to_continuous_act:
# The underlying action space is Discrete and `discrete_to_continuous_act` is given as True.
# Therefore, we convert the action space to continuous (to Box).
# Take the shape and the dtype of the discrete action space.
single_action_shape = (act_space.n,)
single_action_dtype = torch.from_numpy(np.array([], dtype=act_space.dtype)).dtype
# We store the integer dtype of the environment.
self.__discrete_dtype = single_action_dtype
if self.__vectorized:
# If the environment is vectorized, we declare the new `action_space` and the `single_action_space`
# for the enviornment.
action_shape = (env.num_envs,) + single_action_shape
self.single_action_space = Box(float("-inf"), float("inf"), shape=single_action_shape, dtype=np.float32)
self.action_space = Box(float("-inf"), float("inf"), shape=action_shape, dtype=np.float32)
else:
# If the environment is not vectorized, we declare the new `action_space` for the environment.
self.action_space = Box(float("-inf"), float("inf"), shape=single_action_shape, dtype=np.float32)
else:
# This is the case where we do not transform the action space.
# The discrete dtype will not be used, so, we set it as None.
self.__discrete_dtype = None
if isinstance(act_space, Box) and clip_actions:
# If the action space is Box and the wrapper is configured to clip the actions, then we store the lower
# and the upper bounds for the actions.
self.__act_lb = torch.from_numpy(act_space.low)
self.__act_ub = torch.from_numpy(act_space.high)
else:
# If there will not be any action clipping, then we store the lower and the upper bounds as None.
self.__act_lb = None
self.__act_ub = None
reset(self, *args, **kwargs)
¶
Reset the environment
Source code in evotorch/neuroevolution/net/vecrl.py
def reset(self, *args, **kwargs):
"""Reset the environment"""
# Call the reset method of the wrapped environment.
reset_result = self.env.reset(*args, **kwargs)
if isinstance(reset_result, tuple):
# If we received a tuple of two elements, then we assume that this is the new gym API.
# We note that we received an info dictionary.
got_info = True
# We keep the received observation and info.
observation, info = reset_result
else:
# If we did not receive a tuple, then we assume that this is the old gym API.
# We note that we did not receive an info dictionary.
got_info = False
# We keep the received observation.
observation = reset_result
# We did not receive an info dictionary, so, we set it as an empty dictionary.
info = {}
# We understand the array type of the underlying environment from the first observation.
self.__infer_array_type(observation)
# Convert the observation to a PyTorch tensor.
observation = convert_to_torch(observation)
if self.__force_classic_api:
# If the option `force_classic_api` was set as True, then we only return the observation.
return observation
else:
# Here we handle the case where `force_classic_api` was set as False.
if got_info:
# If we got an additional info dictionary, we return it next to the observation.
return observation, info
else:
# If we did not get any info dictionary, we return only the observation.
return observation
step(self, action, *args, **kwargs)
¶
Take a step in the environment
Source code in evotorch/neuroevolution/net/vecrl.py
def step(self, action, *args, **kwargs):
"""Take a step in the environment"""
if self.__array_type is None:
# If the array type is not known yet, then probably `reset()` has not been called yet.
# We raise an error.
raise ValueError(
"Could not understand what type of array this environment works with."
" Perhaps the `reset()` method has not been called yet?"
)
if self.__discrete_dtype is not None:
# If the wrapped environment is discrete-actioned, then we take the integer counterpart of the action.
action = torch.argmax(action, dim=-1).to(dtype=self.__discrete_dtype)
if self.__act_lb is not None:
# The internal variable `__act_lb` having a value other than None means that the initialization argument
# `clip_actions` was given as True.
# Therefore, we clip the actions.
self.__act_lb = self.__act_lb.to(action.device)
self.__act_ub = self.__act_ub.to(action.device)
action = torch.max(action, self.__act_lb)
action = torch.min(action, self.__act_ub)
# Convert the action tensor to the expected array type of the underlying environment.
action = convert_from_torch(action, self.__array_type)
# Perform the step and get the result.
result = self.env.step(action, *args, **kwargs)
if not isinstance(result, tuple):
# If the `step(...)` method returned anything other than tuple, we raise an error.
raise TypeError(f"Expected a tuple as the result of the `step()` method, but received a {type(result)}")
if len(result) == 5:
# If the result is a tuple of 5 elements, then we note that we are using the new API.
using_new_api = True
# Take the observation, reward, two boolean variables done and done2 indicating that the episode(s)
# has/have ended, and additional info.
# `done` indicates whether or not the episode(s) reached terminal state(s).
# `done2` indicates whether or not the episode(s) got truncated because of the timestep limit.
observation, reward, done, done2, info = result
elif len(result) == 4:
# If the result is a tuple of 5 elements, then we note that we are not using the new API.
using_new_api = False
# Take the observation, reward, the done boolean flag, and additional info.
observation, reward, done, info = result
done2 = None
else:
raise ValueError(f"Unexpected number of elements were returned from step(): {len(result)}")
# Convert the observation, reward, and done variables to PyTorch tensors.
observation = convert_to_torch(observation)
reward = convert_to_torch(reward)
done = convert_to_torch_bool(done)
if done2 is not None:
done2 = convert_to_torch_bool(done2)
if self.__force_classic_api:
# This is the case where the initialization argument `force_classic_api` was set as True.
if done2 is not None:
# We combine the terminal state and truncation signals into a single boolean tensor indicating
# whether or not the episode(s) ended.
done = done | done2
# Return 4 elements, compatible with the classic gym API.
return observation, reward, done, info
else:
# This is the case where the initialization argument `force_classic_api` was set as False.
if using_new_api:
# If we are using the new API, then we return the 5-element result.
return observation, reward, done, done2, info
else:
# If we are using the new API, then we return the 4-element result.
return observation, reward, done, info
array_type(x, fallback=None)
¶
Get the type of an array as a string ("jax", "torch", or "numpy"). If the type of the array cannot be determined and a fallback is provided, then the fallback value will be returned.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Any |
The array whose type will be determined. |
required |
fallback |
Optional[str] |
Fallback value, as a string, which will be returned if the array type cannot be determined. |
None |
Returns:
| Type | Description |
|---|---|
str |
The array type as a string ("jax", "torch", or "numpy"). |
Exceptions:
| Type | Description |
|---|---|
TypeError |
if the array type cannot be determined and a fallback value is not provided. |
Source code in evotorch/neuroevolution/net/vecrl.py
def array_type(x: Any, fallback: Optional[str] = None) -> str:
"""
Get the type of an array as a string ("jax", "torch", or "numpy").
If the type of the array cannot be determined and a fallback is provided,
then the fallback value will be returned.
Args:
x: The array whose type will be determined.
fallback: Fallback value, as a string, which will be returned if the
array type cannot be determined.
Returns:
The array type as a string ("jax", "torch", or "numpy").
Raises:
TypeError: if the array type cannot be determined and a fallback
value is not provided.
"""
if is_jax_array(x):
return "jax"
elif isinstance(x, torch.Tensor):
return "torch"
elif isinstance(x, np.ndarray):
return "numpy"
elif fallback is not None:
return fallback
else:
raise TypeError(f"The object has an unrecognized type: {type(x)}")
convert_from_torch(x, array_type)
¶
Convert the given PyTorch tensor to an array of the specified type.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Tensor |
The PyTorch array that will be converted. |
required |
array_type |
str |
Type to which the PyTorch tensor will be converted. Expected as one of these strings: "jax", "torch", "numpy". |
required |
Source code in evotorch/neuroevolution/net/vecrl.py
def convert_from_torch(x: torch.Tensor, array_type: str) -> Any:
"""
Convert the given PyTorch tensor to an array of the specified type.
Args:
x: The PyTorch array that will be converted.
array_type: Type to which the PyTorch tensor will be converted.
Expected as one of these strings: "jax", "torch", "numpy".
Returns:
"""
if array_type == "torch":
return x
elif array_type == "jax":
return torch_to_jax(x)
elif array_type == "numpy":
return x.cpu().numpy()
else:
raise ValueError(f"Unrecognized array type: {array_type}")
convert_to_torch(x)
¶
Convert the given array to PyTorch tensor.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Any |
Array to be converted. Can be a JAX array, a numpy array, a PyTorch tensor (in which case the input tensor will be returned as it is) or any Iterable object. |
required |
Returns:
| Type | Description |
|---|---|
Tensor |
The PyTorch counterpart of the given array. |
Source code in evotorch/neuroevolution/net/vecrl.py
def convert_to_torch(x: Any) -> torch.Tensor:
"""
Convert the given array to PyTorch tensor.
Args:
x: Array to be converted. Can be a JAX array, a numpy array,
a PyTorch tensor (in which case the input tensor will be
returned as it is) or any Iterable object.
Returns:
The PyTorch counterpart of the given array.
"""
if isinstance(x, torch.Tensor):
return x
elif is_jax_array(x):
return jax_to_torch(x)
elif isinstance(x, np.ndarray):
return torch.from_numpy(x)
else:
return torch.as_tensor(x)
convert_to_torch_bool(x)
¶
Convert the given array to a PyTorch tensor of bools.
If the given object is an array of floating point numbers, then, values that are near to 0.0 (with a tolerance of 1e-4) will be converted to False, and the others will be converted to True. If the given object is an array of integers, then zero values will be converted to False, and non-zero values will be converted to True. If the given object is an array of booleans, then no change will be made to those boolean values.
The given object can be a JAX array, a numpy array, or a PyTorch tensor. The result will always be a PyTorch tensor.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Any |
Array to be converted. |
required |
Returns:
| Type | Description |
|---|---|
Tensor |
The array converted to a PyTorch tensor with its dtype set as bool. |
Source code in evotorch/neuroevolution/net/vecrl.py
def convert_to_torch_bool(x: Any) -> torch.Tensor:
"""
Convert the given array to a PyTorch tensor of bools.
If the given object is an array of floating point numbers, then, values
that are near to 0.0 (with a tolerance of 1e-4) will be converted to
False, and the others will be converted to True.
If the given object is an array of integers, then zero values will be
converted to False, and non-zero values will be converted to True.
If the given object is an array of booleans, then no change will be made
to those boolean values.
The given object can be a JAX array, a numpy array, or a PyTorch tensor.
The result will always be a PyTorch tensor.
Args:
x: Array to be converted.
Returns:
The array converted to a PyTorch tensor with its dtype set as bool.
"""
x = convert_to_torch(x)
if x.dtype == torch.bool:
pass # nothing to do
elif "float" in str(x.dtype):
x = torch.abs(x) > 1e-4
else:
x = torch.as_tensor(x, dtype=torch.bool)
return x
make_brax_env(env_name, *, force_classic_api=False, num_envs=None, discrete_to_continuous_act=False, clip_actions=False, **kwargs)
¶
Make a brax environment and wrap it via TorchWrapper.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env_name |
str |
Name of the brax environment, as string (e.g. "humanoid"). |
required |
force_classic_api |
bool |
Whether or not the classic gym API is to be used. |
False |
num_envs |
Optional[int] |
Batch size for the vectorized environment. |
None |
discrete_to_continuous_act |
bool |
Whether or not the the discrete action space of the environment is to be converted to a continuous one. This does nothing if the environment's action space is not discrete. |
False |
clip_actions |
bool |
Whether or not the actions should be explicitly clipped so that they stay within the declared action boundaries. |
False |
kwargs |
Expected in the form of additional keyword arguments, these are passed to the environment. |
{} |
Returns:
| Type | Description |
|---|---|
TorchWrapper |
The brax environment, wrapped by TorchWrapper. |
Source code in evotorch/neuroevolution/net/vecrl.py
def make_brax_env(
env_name: str,
*,
force_classic_api: bool = False,
num_envs: Optional[int] = None,
discrete_to_continuous_act: bool = False,
clip_actions: bool = False,
**kwargs,
) -> TorchWrapper:
"""
Make a brax environment and wrap it via TorchWrapper.
Args:
env_name: Name of the brax environment, as string (e.g. "humanoid").
force_classic_api: Whether or not the classic gym API is to be used.
num_envs: Batch size for the vectorized environment.
discrete_to_continuous_act: Whether or not the the discrete action
space of the environment is to be converted to a continuous one.
This does nothing if the environment's action space is not
discrete.
clip_actions: Whether or not the actions should be explicitly clipped
so that they stay within the declared action boundaries.
kwargs: Expected in the form of additional keyword arguments, these
are passed to the environment.
Returns:
The brax environment, wrapped by TorchWrapper.
"""
if brax is not None:
config = {}
config.update(kwargs)
if num_envs is not None:
config["batch_size"] = num_envs
env = brax.envs.create_gym_env(env_name, **config)
env = TorchWrapper(
env,
force_classic_api=force_classic_api,
discrete_to_continuous_act=discrete_to_continuous_act,
clip_actions=clip_actions,
)
return env
else:
_brax_is_missing()
make_gym_env(env_name, *, force_classic_api=False, num_envs=None, discrete_to_continuous_act=False, clip_actions=False, **kwargs)
¶
Make gym environments and wrap them via a SyncVectorEnv and a TorchWrapper.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env_name |
str |
Name of the gym environment, as string (e.g. "Humanoid-v4"). |
required |
force_classic_api |
bool |
Whether or not the classic gym API is to be used. |
False |
num_envs |
Optional[int] |
Batch size for the vectorized environment. |
None |
discrete_to_continuous_act |
bool |
Whether or not the the discrete action space of the environment is to be converted to a continuous one. This does nothing if the environment's action space is not discrete. |
False |
clip_actions |
bool |
Whether or not the actions should be explicitly clipped so that they stay within the declared action boundaries. |
False |
kwargs |
Expected in the form of additional keyword arguments, these are passed to the environment. |
{} |
Returns:
| Type | Description |
|---|---|
TorchWrapper |
The gym environments, wrapped by a TorchWrapper. |
Source code in evotorch/neuroevolution/net/vecrl.py
def make_gym_env(
env_name: str,
*,
force_classic_api: bool = False,
num_envs: Optional[int] = None,
discrete_to_continuous_act: bool = False,
clip_actions: bool = False,
**kwargs,
) -> TorchWrapper:
"""
Make gym environments and wrap them via a SyncVectorEnv and a TorchWrapper.
Args:
env_name: Name of the gym environment, as string (e.g. "Humanoid-v4").
force_classic_api: Whether or not the classic gym API is to be used.
num_envs: Batch size for the vectorized environment.
discrete_to_continuous_act: Whether or not the the discrete action
space of the environment is to be converted to a continuous one.
This does nothing if the environment's action space is not
discrete.
clip_actions: Whether or not the actions should be explicitly clipped
so that they stay within the declared action boundaries.
kwargs: Expected in the form of additional keyword arguments, these
are passed to the environment.
Returns:
The gym environments, wrapped by a TorchWrapper.
"""
def make_the_env():
return gym.make(env_name, **kwargs)
env_fns = [make_the_env for _ in range(num_envs)]
vec_env = TorchWrapper(
SyncVectorEnv(env_fns),
force_classic_api=force_classic_api,
discrete_to_continuous_act=discrete_to_continuous_act,
clip_actions=clip_actions,
)
return vec_env
make_vector_env(env_name, *, force_classic_api=False, num_envs=None, discrete_to_continuous_act=False, clip_actions=False, **kwargs)
¶
Make a new vectorized environment and wrap it via TorchWrapper.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env_name |
str |
Name of the gym environment, as string. If the string starts with "gym::" (e.g. "gym::Humanoid-v4", etc.), then it is assumed that the target environment is a classical gym environment which will first be wrapped via a SyncVectorEnv and then via a TorchWrapper. If the string starts with "brax::" (e.g. "brax::humanoid", etc.), then it is assumed that the target environment is a brax environment which will be wrapped via TorchWrapper. If the string does not contain "::" at all (e.g. "Humanoid-v4"), then it is assumed that the target environment is a classical gym environment. Therefore, "gym::Humanoid-v4" and "Humanoid-v4" are equivalent. |
required |
force_classic_api |
bool |
Whether or not the classic gym API is to be used. |
False |
num_envs |
Optional[int] |
Batch size for the vectorized environment. |
None |
discrete_to_continuous_act |
bool |
Whether or not the the discrete action space of the environment is to be converted to a continuous one. This does nothing if the environment's action space is not discrete. |
False |
clip_actions |
bool |
Whether or not the actions should be explicitly clipped so that they stay within the declared action boundaries. |
False |
kwargs |
Expected in the form of additional keyword arguments, these are passed to the environment. |
{} |
Returns:
| Type | Description |
|---|---|
TorchWrapper |
The gym environments, wrapped by a TorchWrapper. |
Source code in evotorch/neuroevolution/net/vecrl.py
def make_vector_env(
env_name: str,
*,
force_classic_api: bool = False,
num_envs: Optional[int] = None,
discrete_to_continuous_act: bool = False,
clip_actions: bool = False,
**kwargs,
) -> TorchWrapper:
"""
Make a new vectorized environment and wrap it via TorchWrapper.
Args:
env_name: Name of the gym environment, as string.
If the string starts with "gym::" (e.g. "gym::Humanoid-v4", etc.),
then it is assumed that the target environment is a classical gym
environment which will first be wrapped via a SyncVectorEnv and
then via a TorchWrapper.
If the string starts with "brax::" (e.g. "brax::humanoid", etc.),
then it is assumed that the target environment is a brax
environment which will be wrapped via TorchWrapper.
If the string does not contain "::" at all (e.g. "Humanoid-v4"),
then it is assumed that the target environment is a classical gym
environment. Therefore, "gym::Humanoid-v4" and "Humanoid-v4"
are equivalent.
force_classic_api: Whether or not the classic gym API is to be used.
num_envs: Batch size for the vectorized environment.
discrete_to_continuous_act: Whether or not the the discrete action
space of the environment is to be converted to a continuous one.
This does nothing if the environment's action space is not
discrete.
clip_actions: Whether or not the actions should be explicitly clipped
so that they stay within the declared action boundaries.
kwargs: Expected in the form of additional keyword arguments, these
are passed to the environment.
Returns:
The gym environments, wrapped by a TorchWrapper.
"""
env_parts = str(env_name).split("::", maxsplit=1)
if len(env_parts) == 0:
raise ValueError(f"Invalid value for `env_name`: {repr(env_name)}")
elif len(env_parts) == 1:
fn = make_gym_env
elif len(env_parts) == 2:
env_name = env_parts[1]
if env_parts[0] == "gym":
fn = make_gym_env
elif env_parts[0] == "brax":
fn = make_brax_env
else:
invalid_value = env_parts[0] + "::"
raise ValueError(
f"The argument `env_name` starts with {repr(invalid_value)}, implying that the environment is stored"
f" in a registry named {repr(env_parts[0])}."
f" However, the registry {repr(env_parts[0])} is not recognized."
f" Supported environment registries are: 'gym', 'brax'."
)
else:
assert False, "Unexpected value received from len(env_parts)"
return fn(
env_name,
force_classic_api=force_classic_api,
num_envs=num_envs,
discrete_to_continuous_act=discrete_to_continuous_act,
clip_actions=clip_actions,
**kwargs,
)
reset_tensors(x, indices)
¶
Reset the specified regions of the given tensor(s) as 0.
Note that the resetting is performed in-place, which means, the provided tensors are modified.
The regions are determined by the argument indices, which can be a sequence of booleans (in which case it is
interpreted as a mask), or a sequence of integers (in which case it is interpreted as the list of indices).
For example, let us imagine that we have the following tensor:
import torch
x = torch.tensor(
[
[0, 1, 2, 3],
[4, 5, 6, 7],
[8, 9, 10, 11],
[12, 13, 14, 15],
],
dtype=torch.float32,
)
If we wish to reset the rows with indices 0 and 2, we could use:
The new value of x would then be:
torch.tensor(
[
[0, 0, 0, 0],
[4, 5, 6, 7],
[0, 0, 0, 0],
[12, 13, 14, 15],
],
dtype=torch.float32,
)
The first argument does not have to be a single tensor. Instead, it can be a container (i.e. a dictionary-like object or an iterable) that stores tensors. In this case, each tensor stored by the container will be subject to resetting. In more details, each tensor within the iterable(s) and each tensor within the value part of the dictionary-like object(s) will be reset.
As an example, let us assume that we have the following collection:
a = torch.tensor(
[
[0, 1],
[2, 3],
[4, 5],
],
dtype=torch.float32,
)
b = torch.tensor(
[
[0, 10, 20],
[30, 40, 50],
[60, 70, 80],
],
dtype=torch.float32,
)
c = torch.tensor(
[
[100],
[200],
[300],
],
dtype=torch.float32,
)
d = torch.tensor([-1, -2, -3], dtype=torch.float32)
my_tensors = [a, {"1": b, "2": (c, d)}]
To clear the regions with indices, e.g, (1, 2), we could do:
and the result would be:
>>> print(a)
torch.tensor(
[
[0, 1],
[0, 0],
[0, 0],
],
dtype=torch.float32,
)
>>> print(b)
torch.tensor(
[
[0, 10, 20],
[0, 0, 0],
[0, 0, 0],
],
dtype=torch.float32,
)
>>> print(c)
c = torch.tensor(
[
[100],
[0],
[0],
],
dtype=torch.float32,
)
>>> print(d)
torch.tensor([-1, 0, 0], dtype=torch.float32)
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
x |
Any |
A tensor or a collection of tensors, whose values are subject to resetting. |
required |
indices |
Union[int, Iterable] |
A sequence of integers or booleans, specifying which regions of the tensor(s) will be reset. |
required |
Source code in evotorch/neuroevolution/net/vecrl.py
def reset_tensors(x: Any, indices: MaskOrIndices):
"""
Reset the specified regions of the given tensor(s) as 0.
Note that the resetting is performed in-place, which means, the provided tensors are modified.
The regions are determined by the argument `indices`, which can be a sequence of booleans (in which case it is
interpreted as a mask), or a sequence of integers (in which case it is interpreted as the list of indices).
For example, let us imagine that we have the following tensor:
```python
import torch
x = torch.tensor(
[
[0, 1, 2, 3],
[4, 5, 6, 7],
[8, 9, 10, 11],
[12, 13, 14, 15],
],
dtype=torch.float32,
)
```
If we wish to reset the rows with indices 0 and 2, we could use:
```python
reset_tensors(x, [0, 2])
```
The new value of `x` would then be:
```
torch.tensor(
[
[0, 0, 0, 0],
[4, 5, 6, 7],
[0, 0, 0, 0],
[12, 13, 14, 15],
],
dtype=torch.float32,
)
```
The first argument does not have to be a single tensor.
Instead, it can be a container (i.e. a dictionary-like object or an iterable) that stores tensors.
In this case, each tensor stored by the container will be subject to resetting.
In more details, each tensor within the iterable(s) and each tensor within the value part of the dictionary-like
object(s) will be reset.
As an example, let us assume that we have the following collection:
```python
a = torch.tensor(
[
[0, 1],
[2, 3],
[4, 5],
],
dtype=torch.float32,
)
b = torch.tensor(
[
[0, 10, 20],
[30, 40, 50],
[60, 70, 80],
],
dtype=torch.float32,
)
c = torch.tensor(
[
[100],
[200],
[300],
],
dtype=torch.float32,
)
d = torch.tensor([-1, -2, -3], dtype=torch.float32)
my_tensors = [a, {"1": b, "2": (c, d)}]
```
To clear the regions with indices, e.g, (1, 2), we could do:
```python
reset_tensors(my_tensors, [1, 2])
```
and the result would be:
```
>>> print(a)
torch.tensor(
[
[0, 1],
[0, 0],
[0, 0],
],
dtype=torch.float32,
)
>>> print(b)
torch.tensor(
[
[0, 10, 20],
[0, 0, 0],
[0, 0, 0],
],
dtype=torch.float32,
)
>>> print(c)
c = torch.tensor(
[
[100],
[0],
[0],
],
dtype=torch.float32,
)
>>> print(d)
torch.tensor([-1, 0, 0], dtype=torch.float32)
```
Args:
x: A tensor or a collection of tensors, whose values are subject to resetting.
indices: A sequence of integers or booleans, specifying which regions of the tensor(s) will be reset.
"""
if isinstance(x, torch.Tensor):
# If the first argument is a tensor, then we clear it according to the indices we received.
x[indices] = 0
elif isinstance(x, (str, bytes, bytearray)):
# str, bytes, and bytearray are the types of `Iterable` that we do not wish to process.
# Therefore, we explicitly add a condition for them here, and explicitly state that nothing should be done
# when instances of them are encountered.
pass
elif isinstance(x, Mapping):
# If the first argument is a Mapping (i.e. a dictionary-like object), then, for each value part of the
# Mapping instance, we call this function itself.
for key, value in x.items():
reset_tensors(value, indices)
elif isinstance(x, Iterable):
# If the first argument is an Iterable (e.g. a list, a tuple, etc.), then, for each value contained by this
# Iterable instance, we call this function itself.
for value in x:
reset_tensors(value, indices)
supervisedne
¶
SupervisedNE (NEProblem)
¶
Representation of a neuro-evolution problem where the goal is to minimize a loss function in a supervised learning setting.
A supervised learning problem can be defined via subclassing this class
and overriding the methods
_loss(y_hat, y) (which is to define how the loss is computed)
and _make_dataloader() (which is to define how a new DataLoader is
created).
Alternatively, this class can be directly instantiated as follows:
def my_loss_function(output_of_network, desired_output):
loss = ... # compute the loss here
return loss
problem = SupervisedNE(
my_dataset, MyTorchModuleClass, my_loss_function, minibatch_size=..., ...
)
Source code in evotorch/neuroevolution/supervisedne.py
class SupervisedNE(NEProblem):
"""
Representation of a neuro-evolution problem where the goal is to minimize
a loss function in a supervised learning setting.
A supervised learning problem can be defined via subclassing this class
and overriding the methods
`_loss(y_hat, y)` (which is to define how the loss is computed)
and `_make_dataloader()` (which is to define how a new DataLoader is
created).
Alternatively, this class can be directly instantiated as follows:
```python
def my_loss_function(output_of_network, desired_output):
loss = ... # compute the loss here
return loss
problem = SupervisedNE(
my_dataset, MyTorchModuleClass, my_loss_function, minibatch_size=..., ...
)
```
"""
def __init__(
self,
dataset: Dataset,
network: Union[str, nn.Module, Callable[[], nn.Module]],
loss_func: Optional[Callable] = None,
*,
network_args: Optional[dict] = None,
initial_bounds: Optional[BoundsPairLike] = (-0.00001, 0.00001),
minibatch_size: Optional[int] = None,
num_minibatches: Optional[int] = None,
num_actors: Optional[Union[int, str]] = "num_devices",
common_minibatch: bool = True,
num_gpus_per_actor: Optional[Union[int, float, str]] = None,
actor_config: Optional[dict] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
device: Optional[Device] = None,
):
"""
`__init__(...)`: Initialize the SupervisedNE.
Args:
dataset: The Dataset from which the minibatches will be pulled
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
loss_func: Optionally a function (or a Callable object) which
receives `y_hat` (the output generated by the neural network)
and `y` (the desired output), and returns the loss as a
scalar.
This argument can also be left as None, in which case it will
be expected that the method `_loss(self, y_hat, y)` is
overridden by the inheriting class.
network_args: Optionally a dict-like object, storing keyword
arguments to be passed to the network while instantiating it.
initial_bounds: Specifies an interval from which the values of the
initial neural network parameters will be drawn.
minibatch_size: Optionally an integer, describing the size of a
minibatch when pulling data from the dataset.
Can also be left as None, in which case it will be expected
that the inheriting class overrides the method
`_make_dataloader()` and defines how a new DataLoader is to be
made.
num_minibatches: An integer, specifying over how many minibatches
will a single neural network be evaluated.
If not specified, it will be assumed that the desired number
of minibatches per network evaluation is 1.
num_actors: Number of actors to create for parallelized
evaluation of the solutions.
Certain string values are also accepted.
When given as "max" or as "num_cpus", the number of actors
will be equal to the number of all available CPUs in the ray
cluster.
When given as "num_gpus", the number of actors will be
equal to the number of all available GPUs in the ray
cluster, and each actor will be assigned a GPU.
When given as "num_devices", the number of actors will be
equal to the minimum among the number of CPUs and the number
of GPUs available in the cluster (or will be equal to the
number of CPUs if there is no GPU), and each actor will be
assigned a GPU (if available).
If `num_actors` is given as "num_gpus" or "num_devices",
the argument `num_gpus_per_actor` must not be used,
and the `actor_config` dictionary must not contain the
key "num_gpus".
If `num_actors` is given as something other than "num_gpus"
or "num_devices", and if you wish to assign GPUs to each
actor, then please see the argument `num_gpus_per_actor`.
common_minibatch: Whether the same minibatches will be
used when evaluating the solutions or not.
actor_config: A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass `actor_config=dict(num_gpus=1)`.
Can also be given as None (which is the default),
if no such options are to be passed.
num_gpus_per_actor: Number of GPUs to be allocated by each
remote actor.
The default behavior is to NOT allocate any GPU at all
(which is the default behavior of the ray library as well).
When given as a number `n`, each actor will be given
`n` GPUs (where `n` can be an integer, or can be a `float`
for fractional allocation).
When given as a string "max", then the available GPUs
across the entire ray cluster (or within the local computer
in the simplest cases) will be equally distributed among
the actors.
When given as a string "all", then each actor will have
access to all the GPUs (this will be achieved by suppressing
the environment variable `CUDA_VISIBLE_DEVICES` for each
actor).
When the problem is not distributed (i.e. when there are
no actors), this argument is expected to be left as None.
num_subbatches: If `num_subbatches` is None (assuming that
`subbatch_size` is also None), then, when evaluating a
population, the population will be split into n pieces, `n`
being the number of actors, and each actor will evaluate
its assigned piece. If `num_subbatches` is an integer `m`,
then the population will be split into `m` pieces,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
how many sub-batches will be generated, and therefore,
how many gradients will be computed by the remote actors.
subbatch_size: If `subbatch_size` is None (assuming that
`num_subbatches` is also None), then, when evaluating a
population, the population will be split into `n` pieces, `n`
being the number of actors, and each actor will evaluate its
assigned piece. If `subbatch_size` is an integer `m`,
then the population will be split into pieces of size `m`,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
When there can be significant difference across the solutions
in terms of computational requirements, specifying a
`subbatch_size` can be beneficial, because, while one
actor is busy with a subbatch containing computationally
challenging solutions, other actors can accept more
tasks and save time.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
the size of a sub-batch (or sub-population) sampled by a
remote actor for computing a gradient.
In distributed mode, it is expected that the population size
is divisible by `subbatch_size`.
device: Default device in which a new population will be generated
and the neural networks will operate.
If not specified, "cpu" will be used.
"""
super().__init__(
objective_sense="min",
network=network,
network_args=network_args,
initial_bounds=initial_bounds,
num_actors=num_actors,
num_gpus_per_actor=num_gpus_per_actor,
actor_config=actor_config,
num_subbatches=num_subbatches,
subbatch_size=subbatch_size,
device=device,
)
self.dataset = dataset
self.dataloader: DataLoader = None
self._loss_func = loss_func
self._minibatch_size = None if minibatch_size is None else int(minibatch_size)
self._num_minibatches = 1 if num_minibatches is None else int(num_minibatches)
self._common_minibatch = common_minibatch
self._current_minibatches: Optional[list] = None
def _make_dataloader(self) -> DataLoader:
"""
Make a new DataLoader.
This method, in its default state, does not contain an implementation.
In the case where the `__init__` of `SupervisedNE` is not provided
with a minibatch size, it will be expected that this method is
overridden by the inheriting class and that the operation of creating
a new DataLoader is defined here.
Returns:
The new DataLoader.
"""
raise NotImplementedError
def make_dataloader(self) -> DataLoader:
"""
Make a new DataLoader.
If the `__init__` of `SupervisedNE` was provided with a minibatch size
via the argument `minibatch_size`, then a new DataLoader will be made
with that minibatch size.
Otherwise, it will be expected that the method `_make_dataloader(...)`
was overridden to contain details regarding how the DataLoader should be
created, and that method will be executed.
Returns:
The created DataLoader.
"""
if self._minibatch_size is None:
return self._make_dataloader()
else:
return DataLoader(self.dataset, shuffle=True, batch_size=self._minibatch_size)
def _evaluate_using_minibatch(self, network: nn.Module, batch: Any) -> Union[float, torch.Tensor]:
"""
Pass a minibatch through a network, and compute the loss.
Args:
network: The network using which the loss will be computed.
batch: The minibatch that will be used as data.
Returns:
The loss.
"""
with torch.no_grad():
x, y = batch
yhat = network(x)
return self.loss(yhat, y)
def _loss(self, y_hat: Any, y: Any) -> Union[float, torch.Tensor]:
"""
The loss function.
This method, in its default state, does not contain an implementation.
In the case where `__init__` of `SupervisedNE` class was not given
a loss function via the argument `loss_func`, it will be expected
that this method is overridden by the inheriting class and that the
operation of computing the loss is defined here.
Args:
y_hat: The output estimated by the network
y: The desired output
Returns:
A scalar, representing the loss
"""
raise NotImplementedError
def loss(self, y_hat: Any, y: Any) -> Union[float, torch.Tensor]:
"""
Run the loss function and return the loss.
If the `__init__` of `SupervisedNE` class was given a loss
function via the argument `loss_func`, then that loss function
will be used. Otherwise, it will be expected that the method
`_loss(...)` is overriden with a loss definition, and that method
will be used to compute the loss.
The computed loss will be returned.
Args:
y_hat: The output estimated by the network
y: The desired output
Returns:
A scalar, representing the loss
"""
if self._loss_func is None:
return self._loss(y_hat, y)
else:
return self._loss_func(y_hat, y)
def _prepare(self) -> None:
self.dataloader = self.make_dataloader()
def get_minibatch(self) -> Any:
"""
Get the next minibatch from the DataLoader.
"""
if self.dataloader is None:
self._prepare()
try:
batch = next(self.dataloader_iterator)
if batch is None:
self.dataloader_iterator = iter(self.dataloader)
batch = self.get_minibatch()
else:
batch = batch
except Exception:
self.dataloader_iterator = iter(self.dataloader)
batch = self.get_minibatch()
# Move batch to device of network
return [var.to(self.network_device) for var in batch]
def _evaluate_network(self, network: nn.Module) -> torch.Tensor:
loss = 0.0
for batch_idx in range(self._num_minibatches):
if not self._common_minibatch:
self._current_minibatch = self.get_minibatch()
else:
self._current_minibatch = self._current_minibatches[batch_idx]
loss += self._evaluate_using_minibatch(network, self._current_minibatch) / self._num_minibatches
return loss
def _evaluate_batch(self, batch: SolutionBatch):
if self._common_minibatch:
# If using a common data batch, generate them now and use them for the entire batch of solutions
self._current_minibatches = [self.get_minibatch() for _ in range(self._num_minibatches)]
return super()._evaluate_batch(batch)
__init__(self, dataset, network, loss_func=None, *, network_args=None, initial_bounds=(-1e-05, 1e-05), minibatch_size=None, num_minibatches=None, num_actors='num_devices', common_minibatch=True, num_gpus_per_actor=None, actor_config=None, num_subbatches=None, subbatch_size=None, device=None)
special
¶
__init__(...): Initialize the SupervisedNE.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
dataset |
Dataset |
The Dataset from which the minibatches will be pulled |
required |
network |
Union[str, torch.nn.modules.module.Module, Callable[[], torch.nn.modules.module.Module]] |
A network structure string, or a Callable (which can be
a class inheriting from |
required |
loss_func |
Optional[Callable] |
Optionally a function (or a Callable object) which
receives |
None |
network_args |
Optional[dict] |
Optionally a dict-like object, storing keyword arguments to be passed to the network while instantiating it. |
None |
initial_bounds |
Union[Iterable[Union[float, Iterable[float], torch.Tensor]], evotorch.core.BoundsPair] |
Specifies an interval from which the values of the initial neural network parameters will be drawn. |
(-1e-05, 1e-05) |
minibatch_size |
Optional[int] |
Optionally an integer, describing the size of a
minibatch when pulling data from the dataset.
Can also be left as None, in which case it will be expected
that the inheriting class overrides the method
|
None |
num_minibatches |
Optional[int] |
An integer, specifying over how many minibatches will a single neural network be evaluated. If not specified, it will be assumed that the desired number of minibatches per network evaluation is 1. |
None |
num_actors |
Union[int, str] |
Number of actors to create for parallelized
evaluation of the solutions.
Certain string values are also accepted.
When given as "max" or as "num_cpus", the number of actors
will be equal to the number of all available CPUs in the ray
cluster.
When given as "num_gpus", the number of actors will be
equal to the number of all available GPUs in the ray
cluster, and each actor will be assigned a GPU.
When given as "num_devices", the number of actors will be
equal to the minimum among the number of CPUs and the number
of GPUs available in the cluster (or will be equal to the
number of CPUs if there is no GPU), and each actor will be
assigned a GPU (if available).
If |
'num_devices' |
common_minibatch |
bool |
Whether the same minibatches will be used when evaluating the solutions or not. |
True |
actor_config |
Optional[dict] |
A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass |
None |
num_gpus_per_actor |
Union[int, float, str] |
Number of GPUs to be allocated by each
remote actor.
The default behavior is to NOT allocate any GPU at all
(which is the default behavior of the ray library as well).
When given as a number |
None |
num_subbatches |
Optional[int] |
If |
None |
subbatch_size |
Optional[int] |
If |
None |
device |
Union[str, torch.device] |
Default device in which a new population will be generated and the neural networks will operate. If not specified, "cpu" will be used. |
None |
Source code in evotorch/neuroevolution/supervisedne.py
def __init__(
self,
dataset: Dataset,
network: Union[str, nn.Module, Callable[[], nn.Module]],
loss_func: Optional[Callable] = None,
*,
network_args: Optional[dict] = None,
initial_bounds: Optional[BoundsPairLike] = (-0.00001, 0.00001),
minibatch_size: Optional[int] = None,
num_minibatches: Optional[int] = None,
num_actors: Optional[Union[int, str]] = "num_devices",
common_minibatch: bool = True,
num_gpus_per_actor: Optional[Union[int, float, str]] = None,
actor_config: Optional[dict] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
device: Optional[Device] = None,
):
"""
`__init__(...)`: Initialize the SupervisedNE.
Args:
dataset: The Dataset from which the minibatches will be pulled
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
loss_func: Optionally a function (or a Callable object) which
receives `y_hat` (the output generated by the neural network)
and `y` (the desired output), and returns the loss as a
scalar.
This argument can also be left as None, in which case it will
be expected that the method `_loss(self, y_hat, y)` is
overridden by the inheriting class.
network_args: Optionally a dict-like object, storing keyword
arguments to be passed to the network while instantiating it.
initial_bounds: Specifies an interval from which the values of the
initial neural network parameters will be drawn.
minibatch_size: Optionally an integer, describing the size of a
minibatch when pulling data from the dataset.
Can also be left as None, in which case it will be expected
that the inheriting class overrides the method
`_make_dataloader()` and defines how a new DataLoader is to be
made.
num_minibatches: An integer, specifying over how many minibatches
will a single neural network be evaluated.
If not specified, it will be assumed that the desired number
of minibatches per network evaluation is 1.
num_actors: Number of actors to create for parallelized
evaluation of the solutions.
Certain string values are also accepted.
When given as "max" or as "num_cpus", the number of actors
will be equal to the number of all available CPUs in the ray
cluster.
When given as "num_gpus", the number of actors will be
equal to the number of all available GPUs in the ray
cluster, and each actor will be assigned a GPU.
When given as "num_devices", the number of actors will be
equal to the minimum among the number of CPUs and the number
of GPUs available in the cluster (or will be equal to the
number of CPUs if there is no GPU), and each actor will be
assigned a GPU (if available).
If `num_actors` is given as "num_gpus" or "num_devices",
the argument `num_gpus_per_actor` must not be used,
and the `actor_config` dictionary must not contain the
key "num_gpus".
If `num_actors` is given as something other than "num_gpus"
or "num_devices", and if you wish to assign GPUs to each
actor, then please see the argument `num_gpus_per_actor`.
common_minibatch: Whether the same minibatches will be
used when evaluating the solutions or not.
actor_config: A dictionary, representing the keyword arguments
to be passed to the options(...) used when creating the
ray actor objects. To be used for explicitly allocating
resources per each actor.
For example, for declaring that each actor is to use a GPU,
one can pass `actor_config=dict(num_gpus=1)`.
Can also be given as None (which is the default),
if no such options are to be passed.
num_gpus_per_actor: Number of GPUs to be allocated by each
remote actor.
The default behavior is to NOT allocate any GPU at all
(which is the default behavior of the ray library as well).
When given as a number `n`, each actor will be given
`n` GPUs (where `n` can be an integer, or can be a `float`
for fractional allocation).
When given as a string "max", then the available GPUs
across the entire ray cluster (or within the local computer
in the simplest cases) will be equally distributed among
the actors.
When given as a string "all", then each actor will have
access to all the GPUs (this will be achieved by suppressing
the environment variable `CUDA_VISIBLE_DEVICES` for each
actor).
When the problem is not distributed (i.e. when there are
no actors), this argument is expected to be left as None.
num_subbatches: If `num_subbatches` is None (assuming that
`subbatch_size` is also None), then, when evaluating a
population, the population will be split into n pieces, `n`
being the number of actors, and each actor will evaluate
its assigned piece. If `num_subbatches` is an integer `m`,
then the population will be split into `m` pieces,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
how many sub-batches will be generated, and therefore,
how many gradients will be computed by the remote actors.
subbatch_size: If `subbatch_size` is None (assuming that
`num_subbatches` is also None), then, when evaluating a
population, the population will be split into `n` pieces, `n`
being the number of actors, and each actor will evaluate its
assigned piece. If `subbatch_size` is an integer `m`,
then the population will be split into pieces of size `m`,
and actors will continually accept the next unevaluated
piece as they finish their current tasks.
When there can be significant difference across the solutions
in terms of computational requirements, specifying a
`subbatch_size` can be beneficial, because, while one
actor is busy with a subbatch containing computationally
challenging solutions, other actors can accept more
tasks and save time.
The arguments `num_subbatches` and `subbatch_size` cannot
be given values other than None at the same time.
While using a distributed algorithm, this argument determines
the size of a sub-batch (or sub-population) sampled by a
remote actor for computing a gradient.
In distributed mode, it is expected that the population size
is divisible by `subbatch_size`.
device: Default device in which a new population will be generated
and the neural networks will operate.
If not specified, "cpu" will be used.
"""
super().__init__(
objective_sense="min",
network=network,
network_args=network_args,
initial_bounds=initial_bounds,
num_actors=num_actors,
num_gpus_per_actor=num_gpus_per_actor,
actor_config=actor_config,
num_subbatches=num_subbatches,
subbatch_size=subbatch_size,
device=device,
)
self.dataset = dataset
self.dataloader: DataLoader = None
self._loss_func = loss_func
self._minibatch_size = None if minibatch_size is None else int(minibatch_size)
self._num_minibatches = 1 if num_minibatches is None else int(num_minibatches)
self._common_minibatch = common_minibatch
self._current_minibatches: Optional[list] = None
get_minibatch(self)
¶
Get the next minibatch from the DataLoader.
Source code in evotorch/neuroevolution/supervisedne.py
def get_minibatch(self) -> Any:
"""
Get the next minibatch from the DataLoader.
"""
if self.dataloader is None:
self._prepare()
try:
batch = next(self.dataloader_iterator)
if batch is None:
self.dataloader_iterator = iter(self.dataloader)
batch = self.get_minibatch()
else:
batch = batch
except Exception:
self.dataloader_iterator = iter(self.dataloader)
batch = self.get_minibatch()
# Move batch to device of network
return [var.to(self.network_device) for var in batch]
loss(self, y_hat, y)
¶
Run the loss function and return the loss.
If the __init__ of SupervisedNE class was given a loss
function via the argument loss_func, then that loss function
will be used. Otherwise, it will be expected that the method
_loss(...) is overriden with a loss definition, and that method
will be used to compute the loss.
The computed loss will be returned.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
y_hat |
Any |
The output estimated by the network |
required |
y |
Any |
The desired output |
required |
Returns:
| Type | Description |
|---|---|
Union[float, torch.Tensor] |
A scalar, representing the loss |
Source code in evotorch/neuroevolution/supervisedne.py
def loss(self, y_hat: Any, y: Any) -> Union[float, torch.Tensor]:
"""
Run the loss function and return the loss.
If the `__init__` of `SupervisedNE` class was given a loss
function via the argument `loss_func`, then that loss function
will be used. Otherwise, it will be expected that the method
`_loss(...)` is overriden with a loss definition, and that method
will be used to compute the loss.
The computed loss will be returned.
Args:
y_hat: The output estimated by the network
y: The desired output
Returns:
A scalar, representing the loss
"""
if self._loss_func is None:
return self._loss(y_hat, y)
else:
return self._loss_func(y_hat, y)
make_dataloader(self)
¶
Make a new DataLoader.
If the __init__ of SupervisedNE was provided with a minibatch size
via the argument minibatch_size, then a new DataLoader will be made
with that minibatch size.
Otherwise, it will be expected that the method _make_dataloader(...)
was overridden to contain details regarding how the DataLoader should be
created, and that method will be executed.
Returns:
| Type | Description |
|---|---|
DataLoader |
The created DataLoader. |
Source code in evotorch/neuroevolution/supervisedne.py
def make_dataloader(self) -> DataLoader:
"""
Make a new DataLoader.
If the `__init__` of `SupervisedNE` was provided with a minibatch size
via the argument `minibatch_size`, then a new DataLoader will be made
with that minibatch size.
Otherwise, it will be expected that the method `_make_dataloader(...)`
was overridden to contain details regarding how the DataLoader should be
created, and that method will be executed.
Returns:
The created DataLoader.
"""
if self._minibatch_size is None:
return self._make_dataloader()
else:
return DataLoader(self.dataset, shuffle=True, batch_size=self._minibatch_size)
vecgymne
¶
VecGymNE (BaseNEProblem)
¶
An EvoTorch problem for solving vectorized gym environments
Source code in evotorch/neuroevolution/vecgymne.py
class VecGymNE(BaseNEProblem):
"""
An EvoTorch problem for solving vectorized gym environments
"""
def __init__(
self,
env: Union[str, Callable],
network: Union[str, Callable, nn.Module],
*,
env_config: Optional[Mapping] = None,
max_num_envs: Optional[int] = None,
network_args: Optional[Mapping] = None,
observation_normalization: bool = False,
decrease_rewards_by: Optional[float] = None,
alive_bonus_schedule: Optional[tuple] = None,
action_noise_stdev: Optional[float] = None,
num_episodes: int = 1,
device: Optional[Device] = None,
num_actors: Optional[Union[int, str]] = "num_devices",
num_gpus_per_actor: Optional[int] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
actor_config: Optional[Mapping] = None,
):
"""
Initialize the VecGymNE.
Args:
env: Environment to be solved.
If this is given as a string starting with "gym::" (e.g.
"gym::Humanoid-v4", etc.), then it is assumed that the target
environment is a classical gym environment.
If this is given as a string starting with "brax::" (e.g.
"brax::humanoid", etc.), then it is assumed that the target
environment is a brax environment.
If this is given as a string which does not contain "::" at
all (e.g. "Humanoid-v4", etc.), then it is assumed that the
target environment is a classical gym environment. Therefore,
"gym::Humanoid-v4" and "Humanoid-v4" are equivalent.
If this argument is given as a Callable (maybe a function or a
class), then, with the assumption that this Callable expects
a keyword argument `num_envs: int`, this Callable is called
and its result (expected as a `gym.vector.VectorEnv` instance)
is used as the environment.
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
Note that this network can be a recurrent network.
When the network's `forward(...)` method can optionally accept
an additional positional argument for the hidden state of the
network and returns an additional value for its next state,
then the policy is treated as a recurrent one.
When the network is given as a callable object (e.g.
a subclass of `nn.Module` or a function) and this callable
object is decorated via `evotorch.decorators.pass_info`,
the following keyword arguments will be passed:
(i) `obs_length` (the length of the observation vector),
(ii) `act_length` (the length of the action vector),
(iii) `obs_shape` (the shape tuple of the observation space),
(iv) `act_shape` (the shape tuple of the action space),
(v) `obs_space` (the Box object specifying the observation
space, and
(vi) `act_space` (the Box object specifying the action
space). Note that `act_space` will always be given as a
`gym.spaces.Box` instance, even when the actual gym
environment has a discrete action space. This because
`VecGymNE` always expects the neural network to return
a tensor of floating-point numbers.
env_config: Keyword arguments to pass to the environment while
it is being created.
max_num_envs: Maximum number of environments to be instantiated.
By default, this is None, which means that the number of
environments can go up to the population size (or up to the
number of solutions that a remote actor receives, if the
problem object is configured to have parallelization).
For situations where the current reinforcement learning task
requires large amount of resources (e.g. memory), allocating
environments as much as the number of solutions might not
be feasible. In such cases, one can set `max_num_envs` as an
integer to bring an upper bound (in total, across all the
remote actors, for when the problem is parallelized) to how
many environments can be allocated.
network_args: Any additional keyword argument to be used when
instantiating the network can be specified via `network_args`
as a dictionary. If there are no such additional keyword
arguments, then `network_args` can be left as None.
Note that the argument `network_args` is expected to be None
when the network is specified as a `torch.nn.Module` instance.
observation_normalization: Whether or not online normalization
will be done on the encountered observations.
decrease_rewards_by: If given as a float, each reward will be
decreased by this amount. For example, if the environment's
reward function has a constant "alive bonus" (i.e. a bonus
that is constantly added onto the reward as long as the
agent is alive), and if you wish to negate this bonus,
you can set `decrease_rewards_by` to this bonus amount,
and the bonus will be nullified.
If you do not wish to affect the rewards in this manner,
keep this as None.
alive_bonus_schedule: Use this to add a customized amount of
alive bonus.
If left as None (which is the default), additional alive
bonus will not be added.
If given as a tuple `(t, b)`, an alive bonus `b` will be
added onto all the rewards beyond the timestep `t`.
If given as a tuple `(t0, t1, b)`, a partial (linearly
increasing towards `b`) alive bonus will be added onto
all the rewards between the timesteps `t0` and `t1`,
and a full alive bonus (which equals to `b`) will be added
onto all the rewards beyond the timestep `t1`.
action_noise_stdev: If given as a real number `s`, then, for
each generated action, Gaussian noise with standard
deviation `s` will be sampled, and then this sampled noise
will be added onto the action.
If action noise is not desired, then this argument can be
left as None.
For sampling the noise, the global random number generator
of PyTorch on the simulator's device will be used.
num_episodes: Number of episodes over which each policy will
be evaluated. The default is 1.
device: The device in which the population will be kept.
If you wish to do a single-GPU evolution, we recommend
to set this as "cuda" (or "cuda:0", or "cuda:1", etc.),
assuming that the simulator will also instantiate itself
on that same device.
Alternatively, if you wish to do a multi-GPU evolution,
we recommend to leave this as None or set this as "cpu",
so that the main population will be kept on the cpu
and the remote actors will perform their evaluations on
the GPUs that are assigned to them.
num_actors: Number of actors.
This is, by default, "num_devices" which creates multiple
actors in such a way that each actor will be assigned
an entire GPU.
This default setting considers the multi-GPU scenario
(and requires multiple GPUs to work correctly).
If you wish to perform the entire evolution on a single
GPU, we recommend to set `num_actors` as 0 (so that there
will be no remote actors) and set `device` as "cuda"
(or "cuda:0", or "cuda:1", etc., the device in which the
simulator will be instantiated).
If you do not have any GPUs the special value "num_devices"
will allocate n actors, n being the number of CPUs.
You can also set `num_actors` as an integer to determine
the exact number of subprocesses which will evaluate the
solutions in parallel. Note that, however, when you set
`num_actors` as an integer, you will also have to explicitly
specify the number of GPUs each actor will be assigned
via the argument `num_gpus_per_actor`.
num_gpus_per_actor: Number of GPUs to be assigned to each
actor. This can be an integer or a float (for when you
wish to assign fractional amounts of GPUs to actors).
When `num_actors` has the special value "num_devices",
the argument `num_gpus_per_actor` is expected to be left as
None.
num_subbatches: For when there are multiple actors, you can
set this to an integer n if you wish the population
to be divided exactly into n sub-batches. The actors, as they
finish their currently assigned sub-batch of solutions,
will pick the next un-evaluated sub-batch.
If you specify too large numbers for this argument, then
each sub-batch will be smaller.
When working with vectorized simulators on GPU, having too
many and too small sub-batches can hurt the performance.
This argument can be left as None, in which case, assuming
that `subbatch_size` is also None, the population will be
split to m sub-batches, m being the number of actors.
subbatch_size: For when there are multiple actors, you can
set this to an integer n if you wish the population to be
divided into sub-batches in such a way that each sub-batch
will consist of exactly n solutions. The actors, as they
finish their currently assigned sub-batch of solutions,
will pick the next un-evaluated sub-batch.
If you specify too small numbers for this argument, then
there will be many sub-batches, each sub-batch having a
small number of solutions.
When working with vectorized simulators on GPU, having too
many and too small sub-batches can hurt the performance.
This argument can be left as None, in which case, assuming
that `num_subbatches` is also None, the population will be
split to m sub-batches, m being the number of actors.
actor_config: Additional configuration to be used when creating
each actor with the help of `ray` library.
Can be left as None if additional configuration is not needed.
"""
# Store the string or the Callable that will be used to generate the reinforcement learning environment.
self._env_maker = env
# Declare the variable which will store the environment.
self._env: Optional[TorchWrapper] = None
# Declare the variable which will store the batch size of the vectorized environment.
self._num_envs: Optional[int] = None
# Store the upper bound (if any) regarding how many environments can exist at the same time.
self._max_num_envs: Optional[int] = None if max_num_envs is None else int(max_num_envs)
# Actor-specific upper bound regarding how many environments can exist at the same time.
# This variable will be filled by the `_parallelize(...)` method.
self._actor_max_num_envs: Optional[int] = None
# Declare the variable which stores whether or not we properly initialized the `_actor_max_num_envs` variable.
self._actor_max_num_envs_ready: bool = False
# Store the additional configurations to be used as keyword arguments while instantiating the environment.
self._env_config: dict = {} if env_config is None else dict(env_config)
# Declare the variable that will store the device of the simulator.
# This variable will be filled when the first observation is received from the environment.
# The device of the observation array received from the environment will determine the value of this variable.
self._simulator_device: Optional[torch.device] = None
# Store the neural network architecture (that might be a string or an `nn.Module` instance).
self._architecture = network
if network_args is None:
# If `network_args` is given as None, change it to an empty dictionary
network_args = {}
if isinstance(network, str):
# If the network is given as a string, then we will need the values for the constants `obs_length`,
# `act_length`, and `obs_space`. To obtain those values, we use our helper function
# `_env_constants_for_str_net(...)` which temporarily instantiates the specified environment and returns
# its needed constants.
env_constants = _env_constants_for_str_net(self._env_maker, **(self._env_config))
elif isinstance(network, nn.Module):
# If the network is an already instantiated nn.Module, then we do not prepare any pre-defined constants.
env_constants = {}
else:
# If the network is given as a Callable, then we will need the values for the constants `obs_length`,
# `act_length`, and `obs_space`. To obtain those values, we use our helper function
# `_env_constants_for_callable_net(...)` which temporarily instantiates the specified environment and
# returns its needed constants.
env_constants = _env_constants_for_callable_net(self._env_maker, **(self._env_config))
# Build a `Policy` instance according to the given architecture, and store it.
if isinstance(network, str):
instantiated_net = str_to_net(network, **{**env_constants, **network_args})
elif isinstance(network, nn.Module):
instantiated_net = network
else:
instantiated_net = pass_info_if_needed(network, env_constants)(**network_args)
self._policy = Policy(instantiated_net)
# Store the boolean which indicates whether or not there will be observation normalization.
self._observation_normalization = bool(observation_normalization)
# Declare the variables that will store the observation-related stats if observation normalization is enabled.
self._obs_stats: Optional[RunningNorm] = None
self._collected_stats: Optional[RunningNorm] = None
# Store the number of episodes configuration given by the user.
self._num_episodes = int(num_episodes)
# Store the `decrease_rewards_by` configuration given by the user.
self._decrease_rewards_by = None if decrease_rewards_by is None else float(decrease_rewards_by)
if alive_bonus_schedule is None:
# If `alive_bonus_schedule` argument is None, then we store it as None as well.
self._alive_bonus_schedule = None
else:
# This is the case where the user has specified an `alive_bonus_schedule`.
alive_bonus_schedule = list(alive_bonus_schedule)
alive_bonus_schedule_length = len(alive_bonus_schedule)
if alive_bonus_schedule_length == 2:
# If `alive_bonus_schedule` was given as a 2-element sequence (t, b), then store it as (t, t, b).
# This means that the partial alive bonus time window starts and ends at t, therefore, there will
# be no alive bonus until t, and beginning with t, there will be full alive bonus.
self._alive_bonus_schedule = [
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[0]),
float(alive_bonus_schedule[1]),
]
elif alive_bonus_schedule_length == 3:
# If `alive_bonus_schedule` was given as a 3-element sequence (t0, t1, b), then store those 3
# elements.
self._alive_bonus_schedule = [
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[1]),
float(alive_bonus_schedule[2]),
]
else:
# `alive_bonus_schedule` sequences with unrecognized lengths trigger an error.
raise ValueError(
f"Received invalid number elements as the alive bonus schedule."
f" Expected 2 or 3 items, but got these: {self._alive_bonus_schedule}"
f" (having a length of {len(self._alive_bonus_schedule)})."
)
# If `action_noise_stdev` is specified, store it.
self._action_noise_stdev = None if action_noise_stdev is None else float(action_noise_stdev)
# Initialize the counters for the number of simulator interactions and the number of episodes.
self._interaction_count: int = 0
self._episode_count: int = 0
# Call the superclass
super().__init__(
objective_sense="max",
initial_bounds=(-0.00001, 0.00001),
solution_length=self._policy.parameter_length,
device=device,
dtype=torch.float32,
num_actors=num_actors,
num_gpus_per_actor=num_gpus_per_actor,
actor_config=actor_config,
num_subbatches=num_subbatches,
subbatch_size=subbatch_size,
)
def _parallelize(self):
super()._parallelize()
if self.is_main:
if not self._actor_max_num_envs_ready:
if self._actors is None:
self._actor_max_num_envs = self._max_num_envs
else:
if self._max_num_envs is not None:
max_num_envs_per_actor = split_workload(self._max_num_envs, len(self._actors))
for i_actor, actor in enumerate(self._actors):
actor.call.remote("_set_actor_max_num_envs", max_num_envs_per_actor[i_actor])
self._actor_max_num_envs_ready = True
def _set_actor_max_num_envs(self, n: int):
self._actor_max_num_envs = n
self._actor_max_num_envs_ready = True
@property
def observation_normalization(self) -> bool:
return self._observation_normalization
def set_episode_count(self, n: int):
"""
Set the episode count manually.
"""
self._episode_count = int(n)
def set_interaction_count(self, n: int):
"""
Set the interaction count manually.
"""
self._interaction_count = int(n)
@property
def interaction_count(self) -> int:
"""
Get the total number of simulator interactions made.
"""
return self._interaction_count
@property
def episode_count(self) -> int:
"""
Get the total number of episodes completed.
"""
return self._episode_count
def _get_local_episode_count(self) -> int:
return self.episode_count
def _get_local_interaction_count(self) -> int:
return self.interaction_count
def _get_env(self, num_policies: int) -> TorchWrapper:
# Get the existing environment instance stored by this VecGymNE, after (re)building it if needed.
if (self._env is None) or (num_policies > self._num_envs):
# If this VecGymNE does not have its environment ready yet (i.e. the `_env` attribute is None)
# or it the batch size of the previously instantiated environment is not enough to deal with
# the number of policies (i.e. the `_num_envs` attribute is less than `num_policies`), then
# we (re)build the environment.
# Keyword arguments to pass to the TorchWrapper.
torch_wrapper_cfg = dict(
force_classic_api=True,
discrete_to_continuous_act=True,
clip_actions=True,
)
if isinstance(self._env_maker, str):
# If the environment is specified via a string, then we use our `make_vector_env` function.
self._env = make_vector_env(
self._env_maker, num_envs=num_policies, **torch_wrapper_cfg, **(self._env_config)
)
else:
# If the environment is specified via a Callable, then we call it.
# We expect this Callable to accept a keyword argument named `num_envs`, and additionally, we pass
# the environment configuration dictionary as keyword arguments.
self._env = self._env_maker(num_envs=num_policies, **(self._env_config))
if not isinstance(self._env, gym.vector.VectorEnv):
# If what is returned by the Callable is not a vectorized environment, then we trigger an error.
raise TypeError("This is not a vectorized environment")
# We wrap the returned vectorized environment with a TorchWrapper, so that the actions that we send
# and the observations and rewards that we receive are PyTorch tensors.
self._env = TorchWrapper(self._env, **torch_wrapper_cfg)
if self._env.num_envs != num_policies:
# If the finally obtained vectorized environment has a different number of batch size, then we trigger
# an error.
raise ValueError("Incompatible number of environments")
# We update the batch size of the created environment.
self._num_envs = num_policies
if not isinstance(self._env.single_observation_space, Box):
# If the observation space is not Box, then we trigger an error.
raise TypeError(
f"Unsupported observation type: {self._env.single_observation_space}."
f" Only Box-typed observation spaces are supported."
)
try:
# If possible, use the `seed(...)` method to explicitly randomize the environment.
# Although the new gym API removed the seed method, some environments define their own `seed(...)`
# method for randomization.
new_seed = random.randint(0, (2**32) - 1)
self._env.seed(new_seed)
except Exception:
# Our attempt at manually seeding the environment has failed.
# This could be because the environment does not have a `seed(...)` method.
# Nothing to do.
pass
return self._env
@property
def _nonserialized_attribs(self):
# Call the `_nonserialized_attribs` property implementation of the superclass to receive the base list
# of non-serialized attributes, then add "_env" to this base list, and then return the resulting list.
return super()._nonserialized_attribs + ["_env"]
@property
def _grad_device(self) -> torch.device:
# For distributed mode, this property determines the device in which the temporary populations will be made
# for gradient computation.
if self._simulator_device is None:
# If the simulator device is not known yet, then we return the cpu device.
return torch.device("cpu")
else:
# If the simulator device is known, then we return that device.
return self._simulator_device
def _make_running_norm(self, observation: torch.Tensor) -> RunningNorm:
# Make a new RunningNorm instance according to the observation tensor.
# The dtype and the device of the new RunningNorm is taken from the observation.
# This new RunningNorm is empty (i.e. does not contain any stats yet).
return RunningNorm(shape=observation.shape[1:], dtype=observation.dtype, device=observation.device)
def _transfer_running_norm(self, rn: RunningNorm, observation: torch.Tensor) -> RunningNorm:
# Transfer (if necessary) the RunningNorm to the device of the observation tensor.
# The returned RunningNorm may be the RunningNorm itself (if the device did not change)
# or a new copy (if the device did change).
if torch.device(rn.device) != torch.device(observation.device):
rn = rn.to(observation.device)
return rn
def _normalize_observation(
self, observation: torch.Tensor, *, mask: Optional[torch.Tensor] = None, update_stats: bool = True
) -> torch.Tensor:
# This function normalizes the received observation batch.
# If a mask is given (as a tensor of booleans), only observations with corresponding mask value set as True
# will be taken into consideration.
# If `update_stats` is given as True and observation normalization is enabled, then we will update the
# RunningNorm instances as well.
if self._observation_normalization:
# This is the case where observation normalization is enabled.
if self._obs_stats is None:
# If we do not have observation stats yet, we build a new one (according to the dtype and device
# of the observation).
self._obs_stats = self._make_running_norm(observation)
else:
# If we already have observation stats, we make sure that it is in the correct device.
self._obs_stats = self._transfer_running_norm(self._obs_stats, observation)
if update_stats:
# This is the case where the `update_stats` argument was encountered as True.
if self._collected_stats is None:
# If the RunningNorm responsible to collect new stats is not built yet, we build it here
# (according to the dtype and device of the observation).
self._collected_stats = self._make_running_norm(observation)
else:
# If the RunningNorm responsible to collect new stats already exists, then we make sure
# that it is in the correct device.
self._collected_stats = self._transfer_running_norm(self._collected_stats, observation)
# We first update the RunningNorm responsible for collecting the new stats.
self._collected_stats.update(observation, mask)
# We now update the RunningNorm which stores all the stats, and return the normalized observation.
result = self._obs_stats.update_and_normalize(observation, mask)
else:
# This is the case where the `update_stats` argument was encountered as False.
# Here we normalize the observation but do not update our existing RunningNorm instances.
result = self._obs_stats.update(observation, mask)
return result
else:
# This is the case where observation normalization is disabled.
# In this case, we just return the observation as it is.
return observation
def _ensure_obsnorm(self):
if not self.observation_normalization:
raise ValueError("This feature can only be used when observation_normalization=True.")
def get_observation_stats(self) -> RunningNorm:
"""Get the observation stats"""
self._ensure_obsnorm()
return self._obs_stats
def _make_sync_data_for_actors(self) -> Any:
if self.observation_normalization:
obs_stats = self.get_observation_stats()
if obs_stats is not None:
obs_stats = obs_stats.to("cpu")
return dict(obs_stats=obs_stats)
else:
return None
def set_observation_stats(self, rn: RunningNorm):
"""Set the observation stats"""
self._ensure_obsnorm()
self._obs_stats = rn
def _use_sync_data_from_main(self, received: dict):
for k, v in received.items():
if k == "obs_stats":
self.set_observation_stats(v)
def pop_observation_stats(self) -> RunningNorm:
"""Get and clear the collected observation stats"""
self._ensure_obsnorm()
result = self._collected_stats
self._collected_stats = None
return result
def _make_sync_data_for_main(self) -> Any:
result = dict(episode_count=self.episode_count, interaction_count=self.interaction_count)
if self.observation_normalization:
collected = self.pop_observation_stats()
if collected is not None:
collected = collected.to("cpu")
result["obs_stats_delta"] = collected
return result
def update_observation_stats(self, rn: RunningNorm):
"""Update the observation stats via another RunningNorm instance"""
self._ensure_obsnorm()
if self._obs_stats is None:
self._obs_stats = rn
else:
self._obs_stats.update(rn)
def _use_sync_data_from_actors(self, received: list):
total_episode_count = 0
total_interaction_count = 0
for data in received:
data: dict
total_episode_count += data["episode_count"]
total_interaction_count += data["interaction_count"]
if self.observation_normalization:
self.update_observation_stats(data["obs_stats_delta"])
self.set_episode_count(total_episode_count)
self.set_interaction_count(total_interaction_count)
def _make_pickle_data_for_main(self) -> dict:
# For when the main Problem object (the non-remote one) gets pickled,
# this function returns the counters of this remote Problem instance,
# to be sent to the main one.
return dict(interaction_count=self.interaction_count, episode_count=self.episode_count)
def _use_pickle_data_from_main(self, state: dict):
# For when a newly unpickled Problem object gets (re)parallelized,
# this function restores the inner states specific to this remote
# worker. In the case of GymNE, those inner states are episode
# and interaction counters.
for k, v in state.items():
if k == "episode_count":
self.set_episode_count(v)
elif k == "interaction_count":
self.set_interaction_count(v)
else:
raise ValueError(f"When restoring the inner state of a remote worker, unrecognized state key: {k}")
def _evaluate_batch(self, batch: SolutionBatch):
if self._actor_max_num_envs is None:
self._evaluate_subbatch(batch)
else:
subbatches = batch.split(max_size=self._actor_max_num_envs)
for subbatch in subbatches:
self._evaluate_subbatch(subbatch)
def _evaluate_subbatch(self, batch: SolutionBatch):
# Get the number of solutions and the solution batch from the shape of the batch.
num_solutions, solution_length = batch.values_shape
# Get (possibly after (re)building) the environment object.
env = self._get_env(num_solutions)
# Reset the environment and receive the first observation batch.
obs_per_env = env.reset()
# Update the simulator device according to the device of the observation batch received.
self._simulator_device = obs_per_env.device
# Get the number of environments.
num_envs = obs_per_env.shape[0]
# Transfer (if necessary) the solutions (which are the network parameters) to the simulator device.
batch_values = batch.values.to(self._simulator_device)
if num_solutions == num_envs:
# If the number of solutions is equal to the number of environments, then we declare all of the solutions
# as the network parameters, and we declare all of these environments active.
params_per_env = batch_values
active_per_env = torch.ones(num_solutions, dtype=torch.bool, device=self._simulator_device)
elif num_solutions < num_envs:
# If the number of solutions is less than the number of environments, then we allocate a new empty
# tensor to represent the network parameters.
params_per_env = torch.empty((num_envs, solution_length), dtype=batch.dtype, device=self._simulator_device)
# The first `num_solutions` rows of this new parameters tensor is filled with the values of the solutions.
params_per_env[:num_solutions, :] = batch_values
# The remaining parameters become the clones of the first solution.
params_per_env[num_solutions:, :] = batch_values[0]
# At first, all the environments are declared as inactive.
active_per_env = torch.zeros(num_envs, dtype=torch.bool, device=self._simulator_device)
# Now, the first `num_solutions` amount of environments is declared as active.
# The remaining ones remain inactive.
active_per_env[:num_solutions] = True
else:
assert False, "Received incompatible number of environments"
# We get the policy and fill it with the parameters stored by the solutions.
policy = self._policy
policy.set_parameters(params_per_env)
# Declare the counter which stores the total timesteps encountered during this evaluation.
total_timesteps = 0
# Declare the counters (one for each environment) storing the number of episodes completed.
num_eps_per_env = torch.zeros(num_envs, dtype=torch.int64, device=self._simulator_device)
# Declare the scores (one for each environment).
score_per_env = torch.zeros(num_envs, dtype=torch.float32, device=self._simulator_device)
if self._alive_bonus_schedule is not None:
# If an alive_bonus_schedule was provided, then we extract the timesteps.
# bonus_t0 is the timestep where the partial alive bonus will start.
# bonus_t1 is the timestep where the full alive bonus will start.
# alive_bonus is the amount that will be added to reward if the agent is alive.
bonus_t0, bonus_t1, alive_bonus = self._alive_bonus_schedule
if bonus_t1 > bonus_t0:
# If bonus_t1 is bigger than bonus_t0, then we have a partial alive bonus time window.
add_partial_alive_bonus = True
# We compute and store the length of the time window.
bonus_t_gap_as_float = float(bonus_t1 - bonus_t0)
else:
# If bonus_t1 is NOT bigger than bonus_t0, then we do NOT have a partial alive bonus time window.
add_partial_alive_bonus = False
# To properly give the alive bonus for each solution, we need to keep track of the timesteps for all
# the running solutions. So, we declare the following variable.
t_per_env = torch.zeros(num_envs, dtype=torch.int64, device=self._simulator_device)
# We normalize the initial observation.
obs_per_env = self._normalize_observation(obs_per_env, mask=active_per_env)
while True:
# Pass the observations through the policy and get the actions to perform.
action_per_env = policy(torch.as_tensor(obs_per_env, dtype=params_per_env.dtype))
if self._action_noise_stdev is not None:
# If we are to apply action noise, we sample from a Gaussian distribution and add the noise onto
# the actions.
action_per_env = action_per_env + (torch.rand_like(action_per_env) * self._action_noise_stdev)
# Apply the actions, get the observations, rewards, and the 'done' flags.
obs_per_env, reward_per_env, done_per_env, _ = env.step(action_per_env)
if self._decrease_rewards_by is not None:
# We decrease the rewards, if we have the configuration to do so.
reward_per_env = reward_per_env - self._decrease_rewards_by
if self._alive_bonus_schedule is not None:
# Here we handle the alive bonus schedule.
# For each environment, increment the timestep.
t_per_env[active_per_env] += 1
# For those who are within the full alive bonus time region, increase the scores by the full amount.
in_full_bonus_t_per_env = active_per_env & (t_per_env >= bonus_t1)
score_per_env[in_full_bonus_t_per_env] += alive_bonus
if add_partial_alive_bonus:
# Here we handle the partial alive bonus time window.
# We first determine which environments are in the partial alive bonus time window.
in_partial_bonus_t_per_env = active_per_env & (t_per_env >= bonus_t0) & (t_per_env < bonus_t1)
# Here we compute the partial alive bonuses and add those bonuses to the scores.
score_per_env[in_partial_bonus_t_per_env] += alive_bonus * (
torch.as_tensor(t_per_env[in_partial_bonus_t_per_env] - bonus_t0, dtype=torch.float32)
/ bonus_t_gap_as_float
)
# Determine which environments just finished their episodes.
just_finished_per_env = active_per_env & done_per_env
# Reset the timestep counters of the environments which are just finished.
t_per_env[just_finished_per_env] = 0
# For each active environment, increase the score by the reward received.
score_per_env[active_per_env] += reward_per_env[active_per_env]
# Update the total timesteps counter.
total_timesteps += int(torch.sum(active_per_env))
# Reset the policies whose episodes are done (so that their hidden states become 0).
policy.reset(done_per_env)
# Update the number of episodes counter for each environment.
num_eps_per_env[done_per_env] += 1
# Solutions with number of completed episodes larger than the number of allowed episodes become inactive.
active_per_env[:num_solutions] = num_eps_per_env[:num_solutions] < self._num_episodes
if not torch.any(active_per_env[:num_solutions]):
# If there is not a single active solution left, then we exit this loop.
break
# For the next iteration of this loop, we normalize the observation.
obs_per_env = self._normalize_observation(obs_per_env, mask=active_per_env)
# Update the interaction count and the episode count stored by this VecGymNE instance.
self._interaction_count += total_timesteps
self._episode_count += num_solutions * self._num_episodes
# Compute the fitnesses
fitnesses = score_per_env[:num_solutions]
if self._num_episodes > 1:
fitnesses /= self._num_episodes
# Assign the scores to the solutions as fitnesses.
batch.set_evals(fitnesses)
def get_env(self) -> Optional[gym.Env]:
"""
Get the gym environment.
Returns:
The gym environment if it is built. If not built yet, None.
"""
return self._env
def to_policy(self, solution: Iterable, *, with_wrapper_modules: bool = True) -> nn.Module:
"""
Convert the given solution to a policy.
Args:
solution: A solution which can be given as a `torch.Tensor`, as a
`Solution`, or as any `Iterable`.
with_wrapper_modules: Whether or not to wrap the policy module
with helper modules so that observations are normalized
and actions are clipped to be within the correct boundaries.
The default and the recommended value is True.
Returns:
The policy, as a `torch.nn.Module` instance.
"""
# Get the gym environment
env = self._get_env(1)
# Get the observation space, its lower and higher bounds.
obs_space = env.single_action_space
low = obs_space.low
high = obs_space.high
# If the lower and higher bounds are not -inf and +inf respectively, then this environment needs clipping.
needs_clipping = _numpy_arrays_specify_bounds(low, high)
# Convert the solution to a PyTorch tensor on cpu.
if isinstance(solution, torch.Tensor):
solution = solution.to("cpu")
elif isinstance(solution, Solution):
solution = solution.values.clone().to("cpu")
else:
solution = torch.as_tensor(solution, dtype=torch.float32, device="cpu")
# Convert the internally stored policy to a PyTorch module.
result = self._policy.to_torch_module(solution)
if with_wrapper_modules:
if self.observation_normalization and (self._obs_stats is not None):
# If observation normalization is needed and there are collected observation stats, then we wrap the
# policy with an ObsNormWrapperModule.
result = ObsNormWrapperModule(result, self._obs_stats)
if needs_clipping:
# If clipping is needed, then we wrap the policy with an ActClipWrapperModule
result = ActClipWrapperModule(result, obs_space)
return result
def save_solution(self, solution: Iterable, fname: Union[str, Path]):
"""
Save the solution into a pickle file.
Among the saved data within the pickle file are the solution
(as a PyTorch tensor), the policy (as a `torch.nn.Module` instance),
and observation stats (if any).
Args:
solution: The solution to be saved. This can be a PyTorch tensor,
a `Solution` instance, or any `Iterable`.
fname: The file name of the pickle file to be created.
"""
# Convert the solution to a PyTorch tensor on the cpu.
if isinstance(solution, torch.Tensor):
solution = solution.to("cpu")
elif isinstance(solution, Solution):
solution = solution.values.clone().to("cpu")
else:
solution = torch.as_tensor(solution, dtype=torch.float32, device="cpu")
if isinstance(solution, ReadOnlyTensor):
solution = solution.as_subclass(torch.Tensor)
# Store the solution and the policy.
result = {
"solution": solution,
"policy": self.to_policy(solution),
}
# If available, store the observation stats.
if self.observation_normalization and (self._obs_stats is not None):
result["obs_mean"] = self._obs_stats.mean.to("cpu")
result["obs_stdev"] = self._obs_stats.stdev.to("cpu")
result["obs_sum"] = self._obs_stats.sum.to("cpu")
result["obs_sum_of_squares"] = self._obs_stats.sum_of_squares.to("cpu")
# Some additional data.
result["interaction_count"] = self.interaction_count
result["episode_count"] = self.episode_count
result["time"] = datetime.now()
if isinstance(self._env_maker, str):
# If the environment was specified via a string, store the string.
result["env"] = self._env_maker
# Store the network architecture.
result["architecture"] = self._architecture
# Save the dictionary which stores the data.
with open(fname, "wb") as f:
pickle.dump(result, f)
@property
def max_num_envs(self) -> Optional[int]:
"""
Maximum number of environments to be allocated.
If a maximum number of environments is not set, then None is returned.
If this problem instance is the main one, then the overall maximum
number of environments is returned.
If this problem instance is a remote one (i.e. is on a remote actor)
then the maximum number of environments for that actor is returned.
"""
if self.is_main:
return self._max_num_envs
else:
return self._actor_max_num_envs
def make_net(self, solution: Iterable) -> nn.Module:
"""
Make a new policy network parameterized by the given solution.
Note that this parameterized network assumes that the observation
is already normalized, and it does not do action clipping to ensure
that the generated actions are within valid bounds.
To have a policy network which has its own observation normalization
and action clipping layers, please see the method `to_policy(...)`.
Args:
solution: The solution which stores the parameters.
This can be a Solution instance, or a 1-dimensional tensor,
or any Iterable of real numbers.
Returns:
The policy network, as a PyTorch module.
"""
return self.to_policy(solution, with_wrapper_modules=False)
@property
def network_device(self) -> Optional[Device]:
"""
The device on which the policy networks will operate.
Specific to VecGymNE, the network device is determined only
after receiving the first observation from the reinforcement
learning environment. Until then, this property has the value
None.
"""
return self._simulator_device
episode_count: int
property
readonly
¶
Get the total number of episodes completed.
interaction_count: int
property
readonly
¶
Get the total number of simulator interactions made.
max_num_envs: Optional[int]
property
readonly
¶
Maximum number of environments to be allocated.
If a maximum number of environments is not set, then None is returned. If this problem instance is the main one, then the overall maximum number of environments is returned. If this problem instance is a remote one (i.e. is on a remote actor) then the maximum number of environments for that actor is returned.
network_device: Union[str, torch.device]
property
readonly
¶
The device on which the policy networks will operate.
Specific to VecGymNE, the network device is determined only after receiving the first observation from the reinforcement learning environment. Until then, this property has the value None.
__init__(self, env, network, *, env_config=None, max_num_envs=None, network_args=None, observation_normalization=False, decrease_rewards_by=None, alive_bonus_schedule=None, action_noise_stdev=None, num_episodes=1, device=None, num_actors='num_devices', num_gpus_per_actor=None, num_subbatches=None, subbatch_size=None, actor_config=None)
special
¶
Initialize the VecGymNE.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
env |
Union[str, Callable] |
Environment to be solved.
If this is given as a string starting with "gym::" (e.g.
"gym::Humanoid-v4", etc.), then it is assumed that the target
environment is a classical gym environment.
If this is given as a string starting with "brax::" (e.g.
"brax::humanoid", etc.), then it is assumed that the target
environment is a brax environment.
If this is given as a string which does not contain "::" at
all (e.g. "Humanoid-v4", etc.), then it is assumed that the
target environment is a classical gym environment. Therefore,
"gym::Humanoid-v4" and "Humanoid-v4" are equivalent.
If this argument is given as a Callable (maybe a function or a
class), then, with the assumption that this Callable expects
a keyword argument |
required |
network |
Union[str, Callable, torch.nn.modules.module.Module] |
A network structure string, or a Callable (which can be
a class inheriting from |
required |
env_config |
Optional[collections.abc.Mapping] |
Keyword arguments to pass to the environment while it is being created. |
None |
max_num_envs |
Optional[int] |
Maximum number of environments to be instantiated.
By default, this is None, which means that the number of
environments can go up to the population size (or up to the
number of solutions that a remote actor receives, if the
problem object is configured to have parallelization).
For situations where the current reinforcement learning task
requires large amount of resources (e.g. memory), allocating
environments as much as the number of solutions might not
be feasible. In such cases, one can set |
None |
network_args |
Optional[collections.abc.Mapping] |
Any additional keyword argument to be used when
instantiating the network can be specified via |
None |
observation_normalization |
bool |
Whether or not online normalization will be done on the encountered observations. |
False |
decrease_rewards_by |
Optional[float] |
If given as a float, each reward will be
decreased by this amount. For example, if the environment's
reward function has a constant "alive bonus" (i.e. a bonus
that is constantly added onto the reward as long as the
agent is alive), and if you wish to negate this bonus,
you can set |
None |
alive_bonus_schedule |
Optional[tuple] |
Use this to add a customized amount of
alive bonus.
If left as None (which is the default), additional alive
bonus will not be added.
If given as a tuple |
None |
action_noise_stdev |
Optional[float] |
If given as a real number |
None |
num_episodes |
int |
Number of episodes over which each policy will be evaluated. The default is 1. |
1 |
device |
Union[str, torch.device] |
The device in which the population will be kept. If you wish to do a single-GPU evolution, we recommend to set this as "cuda" (or "cuda:0", or "cuda:1", etc.), assuming that the simulator will also instantiate itself on that same device. Alternatively, if you wish to do a multi-GPU evolution, we recommend to leave this as None or set this as "cpu", so that the main population will be kept on the cpu and the remote actors will perform their evaluations on the GPUs that are assigned to them. |
None |
num_actors |
Union[int, str] |
Number of actors.
This is, by default, "num_devices" which creates multiple
actors in such a way that each actor will be assigned
an entire GPU.
This default setting considers the multi-GPU scenario
(and requires multiple GPUs to work correctly).
If you wish to perform the entire evolution on a single
GPU, we recommend to set |
'num_devices' |
num_gpus_per_actor |
Optional[int] |
Number of GPUs to be assigned to each
actor. This can be an integer or a float (for when you
wish to assign fractional amounts of GPUs to actors).
When |
None |
num_subbatches |
Optional[int] |
For when there are multiple actors, you can
set this to an integer n if you wish the population
to be divided exactly into n sub-batches. The actors, as they
finish their currently assigned sub-batch of solutions,
will pick the next un-evaluated sub-batch.
If you specify too large numbers for this argument, then
each sub-batch will be smaller.
When working with vectorized simulators on GPU, having too
many and too small sub-batches can hurt the performance.
This argument can be left as None, in which case, assuming
that |
None |
subbatch_size |
Optional[int] |
For when there are multiple actors, you can
set this to an integer n if you wish the population to be
divided into sub-batches in such a way that each sub-batch
will consist of exactly n solutions. The actors, as they
finish their currently assigned sub-batch of solutions,
will pick the next un-evaluated sub-batch.
If you specify too small numbers for this argument, then
there will be many sub-batches, each sub-batch having a
small number of solutions.
When working with vectorized simulators on GPU, having too
many and too small sub-batches can hurt the performance.
This argument can be left as None, in which case, assuming
that |
None |
actor_config |
Optional[collections.abc.Mapping] |
Additional configuration to be used when creating
each actor with the help of |
None |
Source code in evotorch/neuroevolution/vecgymne.py
def __init__(
self,
env: Union[str, Callable],
network: Union[str, Callable, nn.Module],
*,
env_config: Optional[Mapping] = None,
max_num_envs: Optional[int] = None,
network_args: Optional[Mapping] = None,
observation_normalization: bool = False,
decrease_rewards_by: Optional[float] = None,
alive_bonus_schedule: Optional[tuple] = None,
action_noise_stdev: Optional[float] = None,
num_episodes: int = 1,
device: Optional[Device] = None,
num_actors: Optional[Union[int, str]] = "num_devices",
num_gpus_per_actor: Optional[int] = None,
num_subbatches: Optional[int] = None,
subbatch_size: Optional[int] = None,
actor_config: Optional[Mapping] = None,
):
"""
Initialize the VecGymNE.
Args:
env: Environment to be solved.
If this is given as a string starting with "gym::" (e.g.
"gym::Humanoid-v4", etc.), then it is assumed that the target
environment is a classical gym environment.
If this is given as a string starting with "brax::" (e.g.
"brax::humanoid", etc.), then it is assumed that the target
environment is a brax environment.
If this is given as a string which does not contain "::" at
all (e.g. "Humanoid-v4", etc.), then it is assumed that the
target environment is a classical gym environment. Therefore,
"gym::Humanoid-v4" and "Humanoid-v4" are equivalent.
If this argument is given as a Callable (maybe a function or a
class), then, with the assumption that this Callable expects
a keyword argument `num_envs: int`, this Callable is called
and its result (expected as a `gym.vector.VectorEnv` instance)
is used as the environment.
network: A network structure string, or a Callable (which can be
a class inheriting from `torch.nn.Module`, or a function
which returns a `torch.nn.Module` instance), or an instance
of `torch.nn.Module`.
The object provided here determines the structure of the
neural network whose parameters will be evolved.
A network structure string is a string which can be processed
by `evotorch.neuroevolution.net.str_to_net(...)`.
Please see the documentation of the function
`evotorch.neuroevolution.net.str_to_net(...)` to see how such
a neural network structure string looks like.
Note that this network can be a recurrent network.
When the network's `forward(...)` method can optionally accept
an additional positional argument for the hidden state of the
network and returns an additional value for its next state,
then the policy is treated as a recurrent one.
When the network is given as a callable object (e.g.
a subclass of `nn.Module` or a function) and this callable
object is decorated via `evotorch.decorators.pass_info`,
the following keyword arguments will be passed:
(i) `obs_length` (the length of the observation vector),
(ii) `act_length` (the length of the action vector),
(iii) `obs_shape` (the shape tuple of the observation space),
(iv) `act_shape` (the shape tuple of the action space),
(v) `obs_space` (the Box object specifying the observation
space, and
(vi) `act_space` (the Box object specifying the action
space). Note that `act_space` will always be given as a
`gym.spaces.Box` instance, even when the actual gym
environment has a discrete action space. This because
`VecGymNE` always expects the neural network to return
a tensor of floating-point numbers.
env_config: Keyword arguments to pass to the environment while
it is being created.
max_num_envs: Maximum number of environments to be instantiated.
By default, this is None, which means that the number of
environments can go up to the population size (or up to the
number of solutions that a remote actor receives, if the
problem object is configured to have parallelization).
For situations where the current reinforcement learning task
requires large amount of resources (e.g. memory), allocating
environments as much as the number of solutions might not
be feasible. In such cases, one can set `max_num_envs` as an
integer to bring an upper bound (in total, across all the
remote actors, for when the problem is parallelized) to how
many environments can be allocated.
network_args: Any additional keyword argument to be used when
instantiating the network can be specified via `network_args`
as a dictionary. If there are no such additional keyword
arguments, then `network_args` can be left as None.
Note that the argument `network_args` is expected to be None
when the network is specified as a `torch.nn.Module` instance.
observation_normalization: Whether or not online normalization
will be done on the encountered observations.
decrease_rewards_by: If given as a float, each reward will be
decreased by this amount. For example, if the environment's
reward function has a constant "alive bonus" (i.e. a bonus
that is constantly added onto the reward as long as the
agent is alive), and if you wish to negate this bonus,
you can set `decrease_rewards_by` to this bonus amount,
and the bonus will be nullified.
If you do not wish to affect the rewards in this manner,
keep this as None.
alive_bonus_schedule: Use this to add a customized amount of
alive bonus.
If left as None (which is the default), additional alive
bonus will not be added.
If given as a tuple `(t, b)`, an alive bonus `b` will be
added onto all the rewards beyond the timestep `t`.
If given as a tuple `(t0, t1, b)`, a partial (linearly
increasing towards `b`) alive bonus will be added onto
all the rewards between the timesteps `t0` and `t1`,
and a full alive bonus (which equals to `b`) will be added
onto all the rewards beyond the timestep `t1`.
action_noise_stdev: If given as a real number `s`, then, for
each generated action, Gaussian noise with standard
deviation `s` will be sampled, and then this sampled noise
will be added onto the action.
If action noise is not desired, then this argument can be
left as None.
For sampling the noise, the global random number generator
of PyTorch on the simulator's device will be used.
num_episodes: Number of episodes over which each policy will
be evaluated. The default is 1.
device: The device in which the population will be kept.
If you wish to do a single-GPU evolution, we recommend
to set this as "cuda" (or "cuda:0", or "cuda:1", etc.),
assuming that the simulator will also instantiate itself
on that same device.
Alternatively, if you wish to do a multi-GPU evolution,
we recommend to leave this as None or set this as "cpu",
so that the main population will be kept on the cpu
and the remote actors will perform their evaluations on
the GPUs that are assigned to them.
num_actors: Number of actors.
This is, by default, "num_devices" which creates multiple
actors in such a way that each actor will be assigned
an entire GPU.
This default setting considers the multi-GPU scenario
(and requires multiple GPUs to work correctly).
If you wish to perform the entire evolution on a single
GPU, we recommend to set `num_actors` as 0 (so that there
will be no remote actors) and set `device` as "cuda"
(or "cuda:0", or "cuda:1", etc., the device in which the
simulator will be instantiated).
If you do not have any GPUs the special value "num_devices"
will allocate n actors, n being the number of CPUs.
You can also set `num_actors` as an integer to determine
the exact number of subprocesses which will evaluate the
solutions in parallel. Note that, however, when you set
`num_actors` as an integer, you will also have to explicitly
specify the number of GPUs each actor will be assigned
via the argument `num_gpus_per_actor`.
num_gpus_per_actor: Number of GPUs to be assigned to each
actor. This can be an integer or a float (for when you
wish to assign fractional amounts of GPUs to actors).
When `num_actors` has the special value "num_devices",
the argument `num_gpus_per_actor` is expected to be left as
None.
num_subbatches: For when there are multiple actors, you can
set this to an integer n if you wish the population
to be divided exactly into n sub-batches. The actors, as they
finish their currently assigned sub-batch of solutions,
will pick the next un-evaluated sub-batch.
If you specify too large numbers for this argument, then
each sub-batch will be smaller.
When working with vectorized simulators on GPU, having too
many and too small sub-batches can hurt the performance.
This argument can be left as None, in which case, assuming
that `subbatch_size` is also None, the population will be
split to m sub-batches, m being the number of actors.
subbatch_size: For when there are multiple actors, you can
set this to an integer n if you wish the population to be
divided into sub-batches in such a way that each sub-batch
will consist of exactly n solutions. The actors, as they
finish their currently assigned sub-batch of solutions,
will pick the next un-evaluated sub-batch.
If you specify too small numbers for this argument, then
there will be many sub-batches, each sub-batch having a
small number of solutions.
When working with vectorized simulators on GPU, having too
many and too small sub-batches can hurt the performance.
This argument can be left as None, in which case, assuming
that `num_subbatches` is also None, the population will be
split to m sub-batches, m being the number of actors.
actor_config: Additional configuration to be used when creating
each actor with the help of `ray` library.
Can be left as None if additional configuration is not needed.
"""
# Store the string or the Callable that will be used to generate the reinforcement learning environment.
self._env_maker = env
# Declare the variable which will store the environment.
self._env: Optional[TorchWrapper] = None
# Declare the variable which will store the batch size of the vectorized environment.
self._num_envs: Optional[int] = None
# Store the upper bound (if any) regarding how many environments can exist at the same time.
self._max_num_envs: Optional[int] = None if max_num_envs is None else int(max_num_envs)
# Actor-specific upper bound regarding how many environments can exist at the same time.
# This variable will be filled by the `_parallelize(...)` method.
self._actor_max_num_envs: Optional[int] = None
# Declare the variable which stores whether or not we properly initialized the `_actor_max_num_envs` variable.
self._actor_max_num_envs_ready: bool = False
# Store the additional configurations to be used as keyword arguments while instantiating the environment.
self._env_config: dict = {} if env_config is None else dict(env_config)
# Declare the variable that will store the device of the simulator.
# This variable will be filled when the first observation is received from the environment.
# The device of the observation array received from the environment will determine the value of this variable.
self._simulator_device: Optional[torch.device] = None
# Store the neural network architecture (that might be a string or an `nn.Module` instance).
self._architecture = network
if network_args is None:
# If `network_args` is given as None, change it to an empty dictionary
network_args = {}
if isinstance(network, str):
# If the network is given as a string, then we will need the values for the constants `obs_length`,
# `act_length`, and `obs_space`. To obtain those values, we use our helper function
# `_env_constants_for_str_net(...)` which temporarily instantiates the specified environment and returns
# its needed constants.
env_constants = _env_constants_for_str_net(self._env_maker, **(self._env_config))
elif isinstance(network, nn.Module):
# If the network is an already instantiated nn.Module, then we do not prepare any pre-defined constants.
env_constants = {}
else:
# If the network is given as a Callable, then we will need the values for the constants `obs_length`,
# `act_length`, and `obs_space`. To obtain those values, we use our helper function
# `_env_constants_for_callable_net(...)` which temporarily instantiates the specified environment and
# returns its needed constants.
env_constants = _env_constants_for_callable_net(self._env_maker, **(self._env_config))
# Build a `Policy` instance according to the given architecture, and store it.
if isinstance(network, str):
instantiated_net = str_to_net(network, **{**env_constants, **network_args})
elif isinstance(network, nn.Module):
instantiated_net = network
else:
instantiated_net = pass_info_if_needed(network, env_constants)(**network_args)
self._policy = Policy(instantiated_net)
# Store the boolean which indicates whether or not there will be observation normalization.
self._observation_normalization = bool(observation_normalization)
# Declare the variables that will store the observation-related stats if observation normalization is enabled.
self._obs_stats: Optional[RunningNorm] = None
self._collected_stats: Optional[RunningNorm] = None
# Store the number of episodes configuration given by the user.
self._num_episodes = int(num_episodes)
# Store the `decrease_rewards_by` configuration given by the user.
self._decrease_rewards_by = None if decrease_rewards_by is None else float(decrease_rewards_by)
if alive_bonus_schedule is None:
# If `alive_bonus_schedule` argument is None, then we store it as None as well.
self._alive_bonus_schedule = None
else:
# This is the case where the user has specified an `alive_bonus_schedule`.
alive_bonus_schedule = list(alive_bonus_schedule)
alive_bonus_schedule_length = len(alive_bonus_schedule)
if alive_bonus_schedule_length == 2:
# If `alive_bonus_schedule` was given as a 2-element sequence (t, b), then store it as (t, t, b).
# This means that the partial alive bonus time window starts and ends at t, therefore, there will
# be no alive bonus until t, and beginning with t, there will be full alive bonus.
self._alive_bonus_schedule = [
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[0]),
float(alive_bonus_schedule[1]),
]
elif alive_bonus_schedule_length == 3:
# If `alive_bonus_schedule` was given as a 3-element sequence (t0, t1, b), then store those 3
# elements.
self._alive_bonus_schedule = [
int(alive_bonus_schedule[0]),
int(alive_bonus_schedule[1]),
float(alive_bonus_schedule[2]),
]
else:
# `alive_bonus_schedule` sequences with unrecognized lengths trigger an error.
raise ValueError(
f"Received invalid number elements as the alive bonus schedule."
f" Expected 2 or 3 items, but got these: {self._alive_bonus_schedule}"
f" (having a length of {len(self._alive_bonus_schedule)})."
)
# If `action_noise_stdev` is specified, store it.
self._action_noise_stdev = None if action_noise_stdev is None else float(action_noise_stdev)
# Initialize the counters for the number of simulator interactions and the number of episodes.
self._interaction_count: int = 0
self._episode_count: int = 0
# Call the superclass
super().__init__(
objective_sense="max",
initial_bounds=(-0.00001, 0.00001),
solution_length=self._policy.parameter_length,
device=device,
dtype=torch.float32,
num_actors=num_actors,
num_gpus_per_actor=num_gpus_per_actor,
actor_config=actor_config,
num_subbatches=num_subbatches,
subbatch_size=subbatch_size,
)
get_env(self)
¶
Get the gym environment.
Returns:
| Type | Description |
|---|---|
Optional[gym.core.Env] |
The gym environment if it is built. If not built yet, None. |
get_observation_stats(self)
¶
make_net(self, solution)
¶
Make a new policy network parameterized by the given solution. Note that this parameterized network assumes that the observation is already normalized, and it does not do action clipping to ensure that the generated actions are within valid bounds.
To have a policy network which has its own observation normalization
and action clipping layers, please see the method to_policy(...).
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
solution |
Iterable |
The solution which stores the parameters. This can be a Solution instance, or a 1-dimensional tensor, or any Iterable of real numbers. |
required |
Returns:
| Type | Description |
|---|---|
Module |
The policy network, as a PyTorch module. |
Source code in evotorch/neuroevolution/vecgymne.py
def make_net(self, solution: Iterable) -> nn.Module:
"""
Make a new policy network parameterized by the given solution.
Note that this parameterized network assumes that the observation
is already normalized, and it does not do action clipping to ensure
that the generated actions are within valid bounds.
To have a policy network which has its own observation normalization
and action clipping layers, please see the method `to_policy(...)`.
Args:
solution: The solution which stores the parameters.
This can be a Solution instance, or a 1-dimensional tensor,
or any Iterable of real numbers.
Returns:
The policy network, as a PyTorch module.
"""
return self.to_policy(solution, with_wrapper_modules=False)
pop_observation_stats(self)
¶
save_solution(self, solution, fname)
¶
Save the solution into a pickle file.
Among the saved data within the pickle file are the solution
(as a PyTorch tensor), the policy (as a torch.nn.Module instance),
and observation stats (if any).
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
solution |
Iterable |
The solution to be saved. This can be a PyTorch tensor,
a |
required |
fname |
Union[str, pathlib.Path] |
The file name of the pickle file to be created. |
required |
Source code in evotorch/neuroevolution/vecgymne.py
def save_solution(self, solution: Iterable, fname: Union[str, Path]):
"""
Save the solution into a pickle file.
Among the saved data within the pickle file are the solution
(as a PyTorch tensor), the policy (as a `torch.nn.Module` instance),
and observation stats (if any).
Args:
solution: The solution to be saved. This can be a PyTorch tensor,
a `Solution` instance, or any `Iterable`.
fname: The file name of the pickle file to be created.
"""
# Convert the solution to a PyTorch tensor on the cpu.
if isinstance(solution, torch.Tensor):
solution = solution.to("cpu")
elif isinstance(solution, Solution):
solution = solution.values.clone().to("cpu")
else:
solution = torch.as_tensor(solution, dtype=torch.float32, device="cpu")
if isinstance(solution, ReadOnlyTensor):
solution = solution.as_subclass(torch.Tensor)
# Store the solution and the policy.
result = {
"solution": solution,
"policy": self.to_policy(solution),
}
# If available, store the observation stats.
if self.observation_normalization and (self._obs_stats is not None):
result["obs_mean"] = self._obs_stats.mean.to("cpu")
result["obs_stdev"] = self._obs_stats.stdev.to("cpu")
result["obs_sum"] = self._obs_stats.sum.to("cpu")
result["obs_sum_of_squares"] = self._obs_stats.sum_of_squares.to("cpu")
# Some additional data.
result["interaction_count"] = self.interaction_count
result["episode_count"] = self.episode_count
result["time"] = datetime.now()
if isinstance(self._env_maker, str):
# If the environment was specified via a string, store the string.
result["env"] = self._env_maker
# Store the network architecture.
result["architecture"] = self._architecture
# Save the dictionary which stores the data.
with open(fname, "wb") as f:
pickle.dump(result, f)
set_episode_count(self, n)
¶
set_interaction_count(self, n)
¶
set_observation_stats(self, rn)
¶
to_policy(self, solution, *, with_wrapper_modules=True)
¶
Convert the given solution to a policy.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
solution |
Iterable |
A solution which can be given as a |
required |
with_wrapper_modules |
bool |
Whether or not to wrap the policy module with helper modules so that observations are normalized and actions are clipped to be within the correct boundaries. The default and the recommended value is True. |
True |
Returns:
| Type | Description |
|---|---|
Module |
The policy, as a |
Source code in evotorch/neuroevolution/vecgymne.py
def to_policy(self, solution: Iterable, *, with_wrapper_modules: bool = True) -> nn.Module:
"""
Convert the given solution to a policy.
Args:
solution: A solution which can be given as a `torch.Tensor`, as a
`Solution`, or as any `Iterable`.
with_wrapper_modules: Whether or not to wrap the policy module
with helper modules so that observations are normalized
and actions are clipped to be within the correct boundaries.
The default and the recommended value is True.
Returns:
The policy, as a `torch.nn.Module` instance.
"""
# Get the gym environment
env = self._get_env(1)
# Get the observation space, its lower and higher bounds.
obs_space = env.single_action_space
low = obs_space.low
high = obs_space.high
# If the lower and higher bounds are not -inf and +inf respectively, then this environment needs clipping.
needs_clipping = _numpy_arrays_specify_bounds(low, high)
# Convert the solution to a PyTorch tensor on cpu.
if isinstance(solution, torch.Tensor):
solution = solution.to("cpu")
elif isinstance(solution, Solution):
solution = solution.values.clone().to("cpu")
else:
solution = torch.as_tensor(solution, dtype=torch.float32, device="cpu")
# Convert the internally stored policy to a PyTorch module.
result = self._policy.to_torch_module(solution)
if with_wrapper_modules:
if self.observation_normalization and (self._obs_stats is not None):
# If observation normalization is needed and there are collected observation stats, then we wrap the
# policy with an ObsNormWrapperModule.
result = ObsNormWrapperModule(result, self._obs_stats)
if needs_clipping:
# If clipping is needed, then we wrap the policy with an ActClipWrapperModule
result = ActClipWrapperModule(result, obs_space)
return result
update_observation_stats(self, rn)
¶
Update the observation stats via another RunningNorm instance