ParallelEnv¶
- class torchrl.envs.ParallelEnv(*args, **kwargs)[source]¶
Creates one environment per process.
TensorDicts are passed via shared memory or memory map.
Batched environments allow the user to query an arbitrary method / attribute of the environment running remotely.
Those queries will return a list of length equal to the number of workers containing the values resulting from those queries.
>>> env = ParallelEnv(3, my_env_fun) >>> custom_attribute_list = env.custom_attribute >>> custom_method_list = env.custom_method(*args)
- Parameters:
num_workers – number of workers (i.e. env instances) to be deployed simultaneously;
create_env_fn (callable or list of callables) – function (or list of functions) to be used for the environment creation. If a single task is used, a callable should be used and not a list of identical callables: if a list of callable is provided, the environment will be executed as if multiple, diverse tasks were needed, which comes with a slight compute overhead;
- Keyword Arguments:
create_env_kwargs (dict or list of dicts, optional) – kwargs to be used with the environments being created;
share_individual_td (bool, optional) – if
True, a different tensordict is created for every process/worker and a lazy stack is returned. default = None (False if single task);shared_memory (bool) – whether the returned tensordict will be placed in shared memory;
memmap (bool) – whether the returned tensordict will be placed in memory map.
policy_proof (callable, optional) – if provided, it’ll be used to get the list of tensors to return through the
step()andreset()methods, such as"hidden"etc.device (str, int, torch.device) – The device of the batched environment can be passed. If not, it is inferred from the env. In this case, it is assumed that the device of all environments match. If it is provided, it can differ from the sub-environment device(s). In that case, the data will be automatically cast to the appropriate device during collection. This can be used to speed up collection in case casting to device introduces an overhead (eg, numpy-based environents etc.): by using a
"cuda"device for the batched environment but a"cpu"device for the nested environments, one can keep the overhead to a minimum.num_threads (int, optional) – number of threads for this process. Should be equal to one plus the number of processes launched within each subprocess (or one if a single process is launched). Defaults to the number of workers + 1. This parameter has no effect for the
SerialEnvclass.num_sub_threads (int, optional) – number of threads of the subprocesses. Defaults to 1 for safety: if none is indicated, launching multiple workers may charge the cpu load too much and harm performance. This parameter has no effect for the
SerialEnvclass.serial_for_single (bool, optional) – if
True, creating a parallel environment with a single worker will return aSerialEnvinstead. This option has no effect withSerialEnv. Defaults toFalse.non_blocking (bool, optional) – if
True, device moves will be done using thenon_blocking=Trueoption. Defaults toTrue.mp_start_method (str, optional) – the multiprocessing start method. Uses the default start method if not indicated (‘spawn’ by default in TorchRL if not initiated differently before first import). To be used only with
ParallelEnvsubclasses.use_buffers (bool, optional) –
whether communication between workers should occur via circular preallocated memory buffers. Defaults to
Trueunless one of the environment has dynamic specs.Note
Learn more about dynamic specs and environments here.
Note
One can pass keyword arguments to each sub-environments using the following technique: every keyword argument in
reset()will be passed to each environment except for thelist_of_kwargsargument which, if present, should contain a list of the same length as the number of workers with the worker-specific keyword arguments stored in a dictionary. If a partial reset is queried, the element oflist_of_kwargscorresponding to sub-environments that are not reset will be ignored.Examples
>>> from torchrl.envs import GymEnv, ParallelEnv, SerialEnv, EnvCreator >>> make_env = EnvCreator(lambda: GymEnv("Pendulum-v1")) # EnvCreator ensures that the env is sharable. Optional in most cases. >>> env = SerialEnv(2, make_env) # Makes 2 identical copies of the Pendulum env, runs them on the same process serially >>> env = ParallelEnv(2, make_env) # Makes 2 identical copies of the Pendulum env, runs them on dedicated processes >>> from torchrl.envs import DMControlEnv >>> env = ParallelEnv(2, [ ... lambda: DMControlEnv("humanoid", "stand"), ... lambda: DMControlEnv("humanoid", "walk")]) # Creates two independent copies of Humanoid, one that walks one that stands >>> rollout = env.rollout(10) # executes 10 random steps in the environment >>> rollout[0] # data for Humanoid stand TensorDict( fields={ action: Tensor(shape=torch.Size([10, 21]), device=cpu, dtype=torch.float64, is_shared=False), com_velocity: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), extremities: Tensor(shape=torch.Size([10, 12]), device=cpu, dtype=torch.float64, is_shared=False), head_height: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float64, is_shared=False), joint_angles: Tensor(shape=torch.Size([10, 21]), device=cpu, dtype=torch.float64, is_shared=False), next: TensorDict( fields={ com_velocity: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), extremities: Tensor(shape=torch.Size([10, 12]), device=cpu, dtype=torch.float64, is_shared=False), head_height: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float64, is_shared=False), joint_angles: Tensor(shape=torch.Size([10, 21]), device=cpu, dtype=torch.float64, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), torso_vertical: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), velocity: Tensor(shape=torch.Size([10, 27]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), torso_vertical: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), velocity: Tensor(shape=torch.Size([10, 27]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False) >>> rollout[1] # data for Humanoid walk TensorDict( fields={ action: Tensor(shape=torch.Size([10, 21]), device=cpu, dtype=torch.float64, is_shared=False), com_velocity: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), extremities: Tensor(shape=torch.Size([10, 12]), device=cpu, dtype=torch.float64, is_shared=False), head_height: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float64, is_shared=False), joint_angles: Tensor(shape=torch.Size([10, 21]), device=cpu, dtype=torch.float64, is_shared=False), next: TensorDict( fields={ com_velocity: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), extremities: Tensor(shape=torch.Size([10, 12]), device=cpu, dtype=torch.float64, is_shared=False), head_height: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float64, is_shared=False), joint_angles: Tensor(shape=torch.Size([10, 21]), device=cpu, dtype=torch.float64, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), torso_vertical: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), velocity: Tensor(shape=torch.Size([10, 27]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), torso_vertical: Tensor(shape=torch.Size([10, 3]), device=cpu, dtype=torch.float64, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), velocity: Tensor(shape=torch.Size([10, 27]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False) >>> # serial_for_single to avoid creating parallel envs if not necessary >>> env = ParallelEnv(1, make_env, serial_for_single=True) >>> assert isinstance(env, SerialEnv) # serial_for_single allows you to avoid creating parallel envs when not necessary
Warning
TorchRL’s ParallelEnv is quite stringent when it comes to env specs, since these are used to build shared memory buffers for inter-process communication. As such, we encourage users to first run a check of the env specs with
check_env_specs():>>> from torchrl.envs import check_env_specs >>> env = make_env() >>> check_env_specs(env) # if this passes without error you're good to go! >>> penv = ParallelEnv(2, make_env)
In particular, gym-like envs with info-dict readers may be difficult to share across processes if the spec is not properly set, which is hard to do automatically. Check
set_info_dict_reader()for more information. Here is a short example:>>> from torchrl.envs import GymEnv, set_gym_backend, check_env_specs, TransformedEnv, TensorDictPrimer >>> import torch >>> env = GymEnv("HalfCheetah-v4") >>> env.rollout(3) # no info registered, this env passes check_env_specs TensorDict( fields={ action: Tensor(shape=torch.Size([10, 6]), device=cpu, dtype=torch.float32, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([10, 17]), device=cpu, dtype=torch.float64, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([10, 17]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False) >>> check_env_specs(env) # succeeds! >>> env.set_info_dict_reader() # sets the default info_dict reader >>> env.rollout(10) # because the info_dict is empty at reset time, we're missing the root infos! TensorDict( fields={ action: Tensor(shape=torch.Size([10, 6]), device=cpu, dtype=torch.float32, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([10, 17]), device=cpu, dtype=torch.float64, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False), reward_ctrl: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), reward_run: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), x_position: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), x_velocity: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([10, 17]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False) >>> check_env_specs(env) # This check now fails! We should not use an env constructed like this in a parallel env >>> # This ad-hoc fix registers the info-spec for reset. It is wrapped inside `env.auto_register_info_dict()` >>> env_fixed = TransformedEnv(env, TensorDictPrimer(env.info_dict_reader[0].info_spec)) >>> env_fixed.rollout(10) TensorDict( fields={ action: Tensor(shape=torch.Size([10, 6]), device=cpu, dtype=torch.float32, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([10, 17]), device=cpu, dtype=torch.float64, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False), reward_ctrl: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), reward_run: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), x_position: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), x_velocity: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([10, 17]), device=cpu, dtype=torch.float64, is_shared=False), reward_ctrl: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), reward_run: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), x_position: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False), x_velocity: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.float64, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False) >>> check_env_specs(env_fixed) # Succeeds! This env can be used within a parallel env!
Related classes and methods:
auto_register_info_dict()anddefault_info_dict_reader.Warning
The choice of the devices where ParallelEnv needs to be executed can drastically influence its performance. The rule of thumbs is:
If the base environment (backend, e.g., Gym) is executed on CPU, the sub-environments should be executed on CPU and the data should be passed via shared physical memory.
If the base environment is (or can be) executed on CUDA, the sub-environments should be placed on CUDA too.
If a CUDA device is available and the policy is to be executed on CUDA, the ParallelEnv device should be set to CUDA.
Therefore, supposing a CUDA device is available, we have the following scenarios:
>>> # The sub-envs are executed on CPU, but the policy is on GPU >>> env = ParallelEnv(N, MyEnv(..., device="cpu"), device="cuda") >>> # The sub-envs are executed on CUDA >>> env = ParallelEnv(N, MyEnv(..., device="cuda"), device="cuda") >>> # this will create the exact same environment >>> env = ParallelEnv(N, MyEnv(..., device="cuda")) >>> # If no cuda device is available >>> env = ParallelEnv(N, MyEnv(..., device="cpu"))
Warning
ParallelEnv disable gradients in all operations (
step(),reset()andstep_and_maybe_reset()) because gradients cannot be passed throughmultiprocessing.Pipeobjects. OnlySerialEnvwill support backpropagation.- property action_key: NestedKey¶
The action key of an environment.
By default, this will be “action”.
If there is more than one action key in the environment, this function will raise an exception.
- property action_keys: List[NestedKey]¶
The action keys of an environment.
By default, there will only be one key named “action”.
Keys are sorted by depth in the data tree.
- property action_spec: TensorSpec¶
The
actionspec.The
action_specis always stored as a composite spec.If the action spec is provided as a simple spec, this will be returned.
>>> env.action_spec = UnboundedContinuousTensorSpec(1) >>> env.action_spec UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous)
If the action spec is provided as a composite spec and contains only one leaf, this function will return just the leaf.
>>> env.action_spec = CompositeSpec({"nested": {"action": UnboundedContinuousTensorSpec(1)}}) >>> env.action_spec UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous)
If the action spec is provided as a composite spec and has more than one leaf, this function will return the whole spec.
>>> env.action_spec = CompositeSpec({"nested": {"action": UnboundedContinuousTensorSpec(1), "another_action": DiscreteTensorSpec(1)}}) >>> env.action_spec CompositeSpec( nested: CompositeSpec( action: UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous), another_action: DiscreteTensorSpec( shape=torch.Size([]), space=DiscreteBox(n=1), device=cpu, dtype=torch.int64, domain=discrete), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
To retrieve the full spec passed, use:
>>> env.input_spec["full_action_spec"]
This property is mutable.
Examples
>>> from torchrl.envs.libs.gym import GymEnv >>> env = GymEnv("Pendulum-v1") >>> env.action_spec BoundedTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous)
- add_module(name: str, module: Optional[Module]) None¶
Add a child module to the current module.
The module can be accessed as an attribute using the given name.
- Parameters:
name (str) – name of the child module. The child module can be accessed from this module using the given name
module (Module) – child module to be added to the module.
- add_truncated_keys()¶
Adds truncated keys to the environment.
- append_transform(transform: 'Transform' | Callable[[TensorDictBase], TensorDictBase]) None¶
Returns a transformed environment where the callable/transform passed is applied.
- Parameters:
transform (Transform or Callable[[TensorDictBase], TensorDictBase]) – the transform to apply to the environment.
Examples
>>> from torchrl.envs import GymEnv >>> import torch >>> env = GymEnv("CartPole-v1") >>> loc = 0.5 >>> scale = 1.0 >>> transform = lambda data: data.set("observation", (data.get("observation") - loc)/scale) >>> env = env.append_transform(transform=transform) >>> print(env) TransformedEnv( env=GymEnv(env=CartPole-v1, batch_size=torch.Size([]), device=cpu), transform=_CallableTransform(keys=[]))
- apply(fn: Callable[[Module], None]) T¶
Apply
fnrecursively to every submodule (as returned by.children()) as well as self.Typical use includes initializing the parameters of a model (see also torch.nn.init).
- Parameters:
fn (
Module-> None) – function to be applied to each submodule- Returns:
self
- Return type:
Module
Example:
>>> @torch.no_grad() >>> def init_weights(m): >>> print(m) >>> if type(m) == nn.Linear: >>> m.weight.fill_(1.0) >>> print(m.weight) >>> net = nn.Sequential(nn.Linear(2, 2), nn.Linear(2, 2)) >>> net.apply(init_weights) Linear(in_features=2, out_features=2, bias=True) Parameter containing: tensor([[1., 1.], [1., 1.]], requires_grad=True) Linear(in_features=2, out_features=2, bias=True) Parameter containing: tensor([[1., 1.], [1., 1.]], requires_grad=True) Sequential( (0): Linear(in_features=2, out_features=2, bias=True) (1): Linear(in_features=2, out_features=2, bias=True) )
- property batch_locked: bool¶
Whether the environment can be used with a batch size different from the one it was initialized with or not.
If True, the env needs to be used with a tensordict having the same batch size as the env. batch_locked is an immutable property.
- property batch_size: Size¶
Number of envs batched in this environment instance organised in a torch.Size() object.
Environment may be similar or different but it is assumed that they have little if not no interactions between them (e.g., multi-task or batched execution in parallel).
- bfloat16() T¶
Casts all floating point parameters and buffers to
bfloat16datatype.Note
This method modifies the module in-place.
- Returns:
self
- Return type:
Module
- buffers(recurse: bool = True) Iterator[Tensor]¶
Return an iterator over module buffers.
- Parameters:
recurse (bool) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module.
- Yields:
torch.Tensor – module buffer
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> for buf in model.buffers(): >>> print(type(buf), buf.size()) <class 'torch.Tensor'> (20L,) <class 'torch.Tensor'> (20L, 1L, 5L, 5L)
- children() Iterator[Module]¶
Return an iterator over immediate children modules.
- Yields:
Module – a child module
- compile(*args, **kwargs)¶
Compile this Module’s forward using
torch.compile().This Module’s __call__ method is compiled and all arguments are passed as-is to
torch.compile().See
torch.compile()for details on the arguments for this function.
- cpu() T¶
Move all model parameters and buffers to the CPU.
Note
This method modifies the module in-place.
- Returns:
self
- Return type:
Module
- cuda(device: Optional[Union[int, device]] = None) T¶
Move all model parameters and buffers to the GPU.
This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on GPU while being optimized.
Note
This method modifies the module in-place.
- Parameters:
device (int, optional) – if specified, all parameters will be copied to that device
- Returns:
self
- Return type:
Module
- property done_key¶
The done key of an environment.
By default, this will be “done”.
If there is more than one done key in the environment, this function will raise an exception.
- property done_keys: List[NestedKey]¶
The done keys of an environment.
By default, there will only be one key named “done”.
Keys are sorted by depth in the data tree.
- property done_keys_groups¶
A list of done keys, grouped as the reset keys.
This is a list of lists. The outer list has the length of reset keys, the inner lists contain the done keys (eg, done and truncated) that can be read to determine a reset when it is absent.
- property done_spec: TensorSpec¶
The
donespec.The
done_specis always stored as a composite spec.If the done spec is provided as a simple spec, this will be returned.
>>> env.done_spec = DiscreteTensorSpec(2, dtype=torch.bool) >>> env.done_spec DiscreteTensorSpec( shape=torch.Size([]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete)
If the done spec is provided as a composite spec and contains only one leaf, this function will return just the leaf.
>>> env.done_spec = CompositeSpec({"nested": {"done": DiscreteTensorSpec(2, dtype=torch.bool)}}) >>> env.done_spec DiscreteTensorSpec( shape=torch.Size([]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete)
If the done spec is provided as a composite spec and has more than one leaf, this function will return the whole spec.
>>> env.done_spec = CompositeSpec({"nested": {"done": DiscreteTensorSpec(2, dtype=torch.bool), "another_done": DiscreteTensorSpec(2, dtype=torch.bool)}}) >>> env.done_spec CompositeSpec( nested: CompositeSpec( done: DiscreteTensorSpec( shape=torch.Size([]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete), another_done: DiscreteTensorSpec( shape=torch.Size([]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
To always retrieve the full spec passed, use:
>>> env.output_spec["full_done_spec"]
This property is mutable.
Examples
>>> from torchrl.envs.libs.gym import GymEnv >>> env = GymEnv("Pendulum-v1") >>> env.done_spec DiscreteTensorSpec( shape=torch.Size([1]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete)
- double() T¶
Casts all floating point parameters and buffers to
doubledatatype.Note
This method modifies the module in-place.
- Returns:
self
- Return type:
Module
- empty_cache()¶
Erases all the cached values.
For regular envs, the key lists (reward, done etc) are cached, but in some cases they may change during the execution of the code (eg, when adding a transform).
- eval() T¶
Set the module in evaluation mode.
This has any effect only on certain modules. See documentations of particular modules for details of their behaviors in training/evaluation mode, if they are affected, e.g.
Dropout,BatchNorm, etc.This is equivalent with
self.train(False).See Locally disabling gradient computation for a comparison between .eval() and several similar mechanisms that may be confused with it.
- Returns:
self
- Return type:
Module
- extra_repr() str¶
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.
- fake_tensordict() TensorDictBase¶
Returns a fake tensordict with key-value pairs that match in shape, device and dtype what can be expected during an environment rollout.
- float() T¶
Casts all floating point parameters and buffers to
floatdatatype.Note
This method modifies the module in-place.
- Returns:
self
- Return type:
Module
- forward(tensordict: TensorDictBase) TensorDictBase¶
Define 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
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- property full_action_spec: CompositeSpec¶
The full action spec.
full_action_specis aCompositeSpec`instance that contains all the action entries.Examples
>>> from torchrl.envs import BraxEnv >>> for envname in BraxEnv.available_envs: ... break >>> env = BraxEnv(envname) >>> env.full_action_spec
- CompositeSpec(
- action: BoundedTensorSpec(
shape=torch.Size([8]), space=ContinuousBox(
low=Tensor(shape=torch.Size([8]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([8]), device=cpu, dtype=torch.float32, contiguous=True)),
device=cpu, dtype=torch.float32, domain=continuous), device=cpu, shape=torch.Size([]))
- property full_done_spec: CompositeSpec¶
The full done spec.
full_done_specis aCompositeSpec`instance that contains all the done entries. It can be used to generate fake data with a structure that mimics the one obtained at runtime.Examples
>>> import gymnasium >>> from torchrl.envs import GymWrapper >>> env = GymWrapper(gymnasium.make("Pendulum-v1")) >>> env.full_done_spec CompositeSpec( done: DiscreteTensorSpec( shape=torch.Size([1]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete), truncated: DiscreteTensorSpec( shape=torch.Size([1]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete), device=cpu, shape=torch.Size([]))
- property full_reward_spec: CompositeSpec¶
The full reward spec.
full_reward_specis aCompositeSpec`instance that contains all the reward entries.Examples
>>> import gymnasium >>> from torchrl.envs import GymWrapper, TransformedEnv, RenameTransform >>> base_env = GymWrapper(gymnasium.make("Pendulum-v1")) >>> env = TransformedEnv(base_env, RenameTransform("reward", ("nested", "reward"))) >>> env.full_reward_spec CompositeSpec( nested: CompositeSpec( reward: UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous), device=None, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
- property full_state_spec: CompositeSpec¶
The full state spec.
full_state_specis aCompositeSpec`instance that contains all the state entries (ie, the input data that is not action).Examples
>>> from torchrl.envs import BraxEnv >>> for envname in BraxEnv.available_envs: ... break >>> env = BraxEnv(envname) >>> env.full_state_spec CompositeSpec( state: CompositeSpec( pipeline_state: CompositeSpec( q: UnboundedContinuousTensorSpec( shape=torch.Size([15]), space=None, device=cpu, dtype=torch.float32, domain=continuous), [...], device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
- get_buffer(target: str) Tensor¶
Return the buffer given by
targetif it exists, otherwise throw an error.See the docstring for
get_submodulefor a more detailed explanation of this method’s functionality as well as how to correctly specifytarget.- Parameters:
target – The fully-qualified string name of the buffer to look for. (See
get_submodulefor how to specify a fully-qualified string.)- Returns:
The buffer referenced by
target- Return type:
- Raises:
AttributeError – If the target string references an invalid path or resolves to something that is not a buffer
- get_extra_state() Any¶
Return any extra state to include in the module’s state_dict.
Implement this and a corresponding
set_extra_state()for your module if you need to store extra state. This function is called when building the module’s state_dict().Note that extra state should be picklable to ensure working serialization of the state_dict. We only provide provide backwards compatibility guarantees for serializing Tensors; other objects may break backwards compatibility if their serialized pickled form changes.
- Returns:
Any extra state to store in the module’s state_dict
- Return type:
object
- get_parameter(target: str) Parameter¶
Return the parameter given by
targetif it exists, otherwise throw an error.See the docstring for
get_submodulefor a more detailed explanation of this method’s functionality as well as how to correctly specifytarget.- Parameters:
target – The fully-qualified string name of the Parameter to look for. (See
get_submodulefor how to specify a fully-qualified string.)- Returns:
The Parameter referenced by
target- Return type:
torch.nn.Parameter
- Raises:
AttributeError – If the target string references an invalid path or resolves to something that is not an
nn.Parameter
- get_submodule(target: str) Module¶
Return the submodule given by
targetif it exists, otherwise throw an error.For example, let’s say you have an
nn.ModuleAthat looks like this:A( (net_b): Module( (net_c): Module( (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2)) ) (linear): Linear(in_features=100, out_features=200, bias=True) ) )(The diagram shows an
nn.ModuleA.Ahas a nested submodulenet_b, which itself has two submodulesnet_candlinear.net_cthen has a submoduleconv.)To check whether or not we have the
linearsubmodule, we would callget_submodule("net_b.linear"). To check whether we have theconvsubmodule, we would callget_submodule("net_b.net_c.conv").The runtime of
get_submoduleis bounded by the degree of module nesting intarget. A query againstnamed_modulesachieves the same result, but it is O(N) in the number of transitive modules. So, for a simple check to see if some submodule exists,get_submoduleshould always be used.- Parameters:
target – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)
- Returns:
The submodule referenced by
target- Return type:
- Raises:
AttributeError – If the target string references an invalid path or resolves to something that is not an
nn.Module
- half() T¶
Casts all floating point parameters and buffers to
halfdatatype.Note
This method modifies the module in-place.
- Returns:
self
- Return type:
Module
- property input_spec: TensorSpec¶
Input spec.
The composite spec containing all specs for data input to the environments.
It contains:
“full_action_spec”: the spec of the input actions
“full_state_spec”: the spec of all other environment inputs
This attibute is locked and should be read-only. Instead, to set the specs contained in it, use the respective properties.
Examples
>>> from torchrl.envs.libs.gym import GymEnv >>> env = GymEnv("Pendulum-v1") >>> env.input_spec CompositeSpec( full_state_spec: None, full_action_spec: CompositeSpec( action: BoundedTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
- ipu(device: Optional[Union[int, device]] = None) T¶
Move all model parameters and buffers to the IPU.
This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on IPU while being optimized.
Note
This method modifies the module in-place.
- Parameters:
device (int, optional) – if specified, all parameters will be copied to that device
- Returns:
self
- Return type:
Module
- load_state_dict(*args, **kwargs)[source]¶
Copy parameters and buffers from
state_dictinto this module and its descendants.If
strictisTrue, then the keys ofstate_dictmust exactly match the keys returned by this module’sstate_dict()function.Warning
If
assignisTruethe optimizer must be created after the call toload_state_dictunlessget_swap_module_params_on_conversion()isTrue.- Parameters:
state_dict (dict) – a dict containing parameters and persistent buffers.
strict (bool, optional) – whether to strictly enforce that the keys in
state_dictmatch the keys returned by this module’sstate_dict()function. Default:Trueassign (bool, optional) – When
False, the properties of the tensors in the current module are preserved while whenTrue, the properties of the Tensors in the state dict are preserved. The only exception is therequires_gradfield ofDefault: ``False`
- Returns:
- missing_keys is a list of str containing any keys that are expected
by this module but missing from the provided
state_dict.
- unexpected_keys is a list of str containing the keys that are not
expected by this module but present in the provided
state_dict.
- Return type:
NamedTuplewithmissing_keysandunexpected_keysfields
Note
If a parameter or buffer is registered as
Noneand its corresponding key exists instate_dict,load_state_dict()will raise aRuntimeError.
- maybe_reset(tensordict: TensorDictBase) TensorDictBase¶
Checks the done keys of the input tensordict and, if needed, resets the environment where it is done.
- Parameters:
tensordict (TensorDictBase) – a tensordict coming from the output of
step_mdp().- Returns:
A tensordict that is identical to the input where the environment was not reset and contains the new reset data where the environment was reset.
- modules() Iterator[Module]¶
Return an iterator over all modules in the network.
- Yields:
Module – a module in the network
Note
Duplicate modules are returned only once. In the following example,
lwill be returned only once.Example:
>>> l = nn.Linear(2, 2) >>> net = nn.Sequential(l, l) >>> for idx, m in enumerate(net.modules()): ... print(idx, '->', m) 0 -> Sequential( (0): Linear(in_features=2, out_features=2, bias=True) (1): Linear(in_features=2, out_features=2, bias=True) ) 1 -> Linear(in_features=2, out_features=2, bias=True)
- mtia(device: Optional[Union[int, device]] = None) T¶
Move all model parameters and buffers to the MTIA.
This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on MTIA while being optimized.
Note
This method modifies the module in-place.
- Parameters:
device (int, optional) – if specified, all parameters will be copied to that device
- Returns:
self
- Return type:
Module
- named_buffers(prefix: str = '', recurse: bool = True, remove_duplicate: bool = True) Iterator[Tuple[str, Tensor]]¶
Return an iterator over module buffers, yielding both the name of the buffer as well as the buffer itself.
- Parameters:
prefix (str) – prefix to prepend to all buffer names.
recurse (bool, optional) – if True, then yields buffers of this module and all submodules. Otherwise, yields only buffers that are direct members of this module. Defaults to True.
remove_duplicate (bool, optional) – whether to remove the duplicated buffers in the result. Defaults to True.
- Yields:
(str, torch.Tensor) – Tuple containing the name and buffer
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> for name, buf in self.named_buffers(): >>> if name in ['running_var']: >>> print(buf.size())
- named_children() Iterator[Tuple[str, Module]]¶
Return an iterator over immediate children modules, yielding both the name of the module as well as the module itself.
- Yields:
(str, Module) – Tuple containing a name and child module
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> for name, module in model.named_children(): >>> if name in ['conv4', 'conv5']: >>> print(module)
- named_modules(memo: Optional[Set[Module]] = None, prefix: str = '', remove_duplicate: bool = True)¶
Return an iterator over all modules in the network, yielding both the name of the module as well as the module itself.
- Parameters:
memo – a memo to store the set of modules already added to the result
prefix – a prefix that will be added to the name of the module
remove_duplicate – whether to remove the duplicated module instances in the result or not
- Yields:
(str, Module) – Tuple of name and module
Note
Duplicate modules are returned only once. In the following example,
lwill be returned only once.Example:
>>> l = nn.Linear(2, 2) >>> net = nn.Sequential(l, l) >>> for idx, m in enumerate(net.named_modules()): ... print(idx, '->', m) 0 -> ('', Sequential( (0): Linear(in_features=2, out_features=2, bias=True) (1): Linear(in_features=2, out_features=2, bias=True) )) 1 -> ('0', Linear(in_features=2, out_features=2, bias=True))
- named_parameters(prefix: str = '', recurse: bool = True, remove_duplicate: bool = True) Iterator[Tuple[str, Parameter]]¶
Return an iterator over module parameters, yielding both the name of the parameter as well as the parameter itself.
- Parameters:
prefix (str) – prefix to prepend to all parameter names.
recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.
remove_duplicate (bool, optional) – whether to remove the duplicated parameters in the result. Defaults to True.
- Yields:
(str, Parameter) – Tuple containing the name and parameter
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> for name, param in self.named_parameters(): >>> if name in ['bias']: >>> print(param.size())
- property observation_spec: CompositeSpec¶
Observation spec.
Must be a
torchrl.data.CompositeSpecinstance. The keys listed in the spec are directly accessible after reset and step.In TorchRL, even though they are not properly speaking “observations” all info, states, results of transforms etc. outputs from the environment are stored in the
observation_spec.Therefore,
"observation_spec"should be thought as a generic data container for environment outputs that are not done or reward data.Examples
>>> from torchrl.envs.libs.gym import GymEnv >>> env = GymEnv("Pendulum-v1") >>> env.observation_spec CompositeSpec( observation: BoundedTensorSpec( shape=torch.Size([3]), space=ContinuousBox( low=Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous), device=cpu, shape=torch.Size([]))
- property output_spec: TensorSpec¶
Output spec.
The composite spec containing all specs for data output from the environments.
It contains:
“full_reward_spec”: the spec of reward
“full_done_spec”: the spec of done
“full_observation_spec”: the spec of all other environment outputs
This attibute is locked and should be read-only. Instead, to set the specs contained in it, use the respective properties.
Examples
>>> from torchrl.envs.libs.gym import GymEnv >>> env = GymEnv("Pendulum-v1") >>> env.output_spec CompositeSpec( full_reward_spec: CompositeSpec( reward: UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=None, device=cpu, dtype=torch.float32, domain=continuous), device=cpu, shape=torch.Size([])), full_observation_spec: CompositeSpec( observation: BoundedTensorSpec( shape=torch.Size([3]), space=ContinuousBox( low=Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous), device=cpu, shape=torch.Size([])), full_done_spec: CompositeSpec( done: DiscreteTensorSpec( shape=torch.Size([1]), space=DiscreteBox(n=2), device=cpu, dtype=torch.bool, domain=discrete), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
- parameters(recurse: bool = True) Iterator[Parameter]¶
Return an iterator over module parameters.
This is typically passed to an optimizer.
- Parameters:
recurse (bool) – if True, then yields parameters of this module and all submodules. Otherwise, yields only parameters that are direct members of this module.
- Yields:
Parameter – module parameter
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> for param in model.parameters(): >>> print(type(param), param.size()) <class 'torch.Tensor'> (20L,) <class 'torch.Tensor'> (20L, 1L, 5L, 5L)
- rand_action(tensordict: Optional[TensorDictBase] = None)¶
Performs a random action given the action_spec attribute.
- Parameters:
tensordict (TensorDictBase, optional) – tensordict where the resulting action should be written.
- Returns:
a tensordict object with the “action” entry updated with a random sample from the action-spec.
- rand_step(tensordict: Optional[TensorDictBase] = None) TensorDictBase¶
Performs a random step in the environment given the action_spec attribute.
- Parameters:
tensordict (TensorDictBase, optional) – tensordict where the resulting info should be written.
- Returns:
a tensordict object with the new observation after a random step in the environment. The action will be stored with the “action” key.
- register_backward_hook(hook: Callable[[Module, Union[Tuple[Tensor, ...], Tensor], Union[Tuple[Tensor, ...], Tensor]], Union[None, Tuple[Tensor, ...], Tensor]]) RemovableHandle¶
Register a backward hook on the module.
This function is deprecated in favor of
register_full_backward_hook()and the behavior of this function will change in future versions.- Returns:
a handle that can be used to remove the added hook by calling
handle.remove()- Return type:
torch.utils.hooks.RemovableHandle
- register_buffer(name: str, tensor: Optional[Tensor], persistent: bool = True) None¶
Add a buffer to the module.
This is typically used to register a buffer that should not to be considered a model parameter. For example, BatchNorm’s
running_meanis not a parameter, but is part of the module’s state. Buffers, by default, are persistent and will be saved alongside parameters. This behavior can be changed by settingpersistenttoFalse. The only difference between a persistent buffer and a non-persistent buffer is that the latter will not be a part of this module’sstate_dict.Buffers can be accessed as attributes using given names.
- Parameters:
name (str) – name of the buffer. The buffer can be accessed from this module using the given name
tensor (Tensor or None) – buffer to be registered. If
None, then operations that run on buffers, such ascuda, are ignored. IfNone, the buffer is not included in the module’sstate_dict.persistent (bool) – whether the buffer is part of this module’s
state_dict.
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> self.register_buffer('running_mean', torch.zeros(num_features))
- register_forward_hook(hook: Union[Callable[[T, Tuple[Any, ...], Any], Optional[Any]], Callable[[T, Tuple[Any, ...], Dict[str, Any], Any], Optional[Any]]], *, prepend: bool = False, with_kwargs: bool = False, always_call: bool = False) RemovableHandle¶
Register a forward hook on the module.
The hook will be called every time after
forward()has computed an output.If
with_kwargsisFalseor not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to theforward. The hook can modify the output. It can modify the input inplace but it will not have effect on forward since this is called afterforward()is called. The hook should have the following signature:hook(module, args, output) -> None or modified output
If
with_kwargsisTrue, the forward hook will be passed thekwargsgiven to the forward function and be expected to return the output possibly modified. The hook should have the following signature:hook(module, args, kwargs, output) -> None or modified output
- Parameters:
hook (Callable) – The user defined hook to be registered.
prepend (bool) – If
True, the providedhookwill be fired before all existingforwardhooks on thistorch.nn.modules.Module. Otherwise, the providedhookwill be fired after all existingforwardhooks on thistorch.nn.modules.Module. Note that globalforwardhooks registered withregister_module_forward_hook()will fire before all hooks registered by this method. Default:Falsewith_kwargs (bool) – If
True, thehookwill be passed the kwargs given to the forward function. Default:Falsealways_call (bool) – If
Truethehookwill be run regardless of whether an exception is raised while calling the Module. Default:False
- Returns:
a handle that can be used to remove the added hook by calling
handle.remove()- Return type:
torch.utils.hooks.RemovableHandle
- register_forward_pre_hook(hook: Union[Callable[[T, Tuple[Any, ...]], Optional[Any]], Callable[[T, Tuple[Any, ...], Dict[str, Any]], Optional[Tuple[Any, Dict[str, Any]]]]], *, prepend: bool = False, with_kwargs: bool = False) RemovableHandle¶
Register a forward pre-hook on the module.
The hook will be called every time before
forward()is invoked.If
with_kwargsis false or not specified, the input contains only the positional arguments given to the module. Keyword arguments won’t be passed to the hooks and only to theforward. The hook can modify the input. User can either return a tuple or a single modified value in the hook. We will wrap the value into a tuple if a single value is returned (unless that value is already a tuple). The hook should have the following signature:hook(module, args) -> None or modified input
If
with_kwargsis true, the forward pre-hook will be passed the kwargs given to the forward function. And if the hook modifies the input, both the args and kwargs should be returned. The hook should have the following signature:hook(module, args, kwargs) -> None or a tuple of modified input and kwargs
- Parameters:
hook (Callable) – The user defined hook to be registered.
prepend (bool) – If true, the provided
hookwill be fired before all existingforward_prehooks on thistorch.nn.modules.Module. Otherwise, the providedhookwill be fired after all existingforward_prehooks on thistorch.nn.modules.Module. Note that globalforward_prehooks registered withregister_module_forward_pre_hook()will fire before all hooks registered by this method. Default:Falsewith_kwargs (bool) – If true, the
hookwill be passed the kwargs given to the forward function. Default:False
- Returns:
a handle that can be used to remove the added hook by calling
handle.remove()- Return type:
torch.utils.hooks.RemovableHandle
- register_full_backward_hook(hook: Callable[[Module, Union[Tuple[Tensor, ...], Tensor], Union[Tuple[Tensor, ...], Tensor]], Union[None, Tuple[Tensor, ...], Tensor]], prepend: bool = False) RemovableHandle¶
Register a backward hook on the module.
The hook will be called every time the gradients with respect to a module are computed, i.e. the hook will execute if and only if the gradients with respect to module outputs are computed. The hook should have the following signature:
hook(module, grad_input, grad_output) -> tuple(Tensor) or None
The
grad_inputandgrad_outputare tuples that contain the gradients with respect to the inputs and outputs respectively. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the input that will be used in place ofgrad_inputin subsequent computations.grad_inputwill only correspond to the inputs given as positional arguments and all kwarg arguments are ignored. Entries ingrad_inputandgrad_outputwill beNonefor all non-Tensor arguments.For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.
Warning
Modifying inputs or outputs inplace is not allowed when using backward hooks and will raise an error.
- Parameters:
hook (Callable) – The user-defined hook to be registered.
prepend (bool) – If true, the provided
hookwill be fired before all existingbackwardhooks on thistorch.nn.modules.Module. Otherwise, the providedhookwill be fired after all existingbackwardhooks on thistorch.nn.modules.Module. Note that globalbackwardhooks registered withregister_module_full_backward_hook()will fire before all hooks registered by this method.
- Returns:
a handle that can be used to remove the added hook by calling
handle.remove()- Return type:
torch.utils.hooks.RemovableHandle
- register_full_backward_pre_hook(hook: Callable[[Module, Union[Tuple[Tensor, ...], Tensor]], Union[None, Tuple[Tensor, ...], Tensor]], prepend: bool = False) RemovableHandle¶
Register a backward pre-hook on the module.
The hook will be called every time the gradients for the module are computed. The hook should have the following signature:
hook(module, grad_output) -> tuple[Tensor] or None
The
grad_outputis a tuple. The hook should not modify its arguments, but it can optionally return a new gradient with respect to the output that will be used in place ofgrad_outputin subsequent computations. Entries ingrad_outputwill beNonefor all non-Tensor arguments.For technical reasons, when this hook is applied to a Module, its forward function will receive a view of each Tensor passed to the Module. Similarly the caller will receive a view of each Tensor returned by the Module’s forward function.
Warning
Modifying inputs inplace is not allowed when using backward hooks and will raise an error.
- Parameters:
hook (Callable) – The user-defined hook to be registered.
prepend (bool) – If true, the provided
hookwill be fired before all existingbackward_prehooks on thistorch.nn.modules.Module. Otherwise, the providedhookwill be fired after all existingbackward_prehooks on thistorch.nn.modules.Module. Note that globalbackward_prehooks registered withregister_module_full_backward_pre_hook()will fire before all hooks registered by this method.
- Returns:
a handle that can be used to remove the added hook by calling
handle.remove()- Return type:
torch.utils.hooks.RemovableHandle
- classmethod register_gym(id: str, *, entry_point: Callable | None = None, transform: 'Transform' | None = None, info_keys: List[NestedKey] | None = None, backend: str = None, to_numpy: bool = False, reward_threshold: float | None = None, nondeterministic: bool = False, max_episode_steps: int | None = None, order_enforce: bool = True, autoreset: bool = False, disable_env_checker: bool = False, apply_api_compatibility: bool = False, **kwargs)¶
Registers an environment in gym(nasium).
This method is designed with the following scopes in mind:
Incorporate a TorchRL-first environment in a framework that uses Gym;
Incorporate another environment (eg, DeepMind Control, Brax, Jumanji, …) in a framework that uses Gym.
- Parameters:
id (str) – the name of the environment. Should follow the gym naming convention.
- Keyword Arguments:
entry_point (callable, optional) –
the entry point to build the environment. If none is passed, the parent class will be used as entry point. Typically, this is used to register an environment that does not necessarily inherit from the base being used:
>>> from torchrl.envs import DMControlEnv >>> DMControlEnv.register_gym("DMC-cheetah-v0", env_name="cheetah", task="run") >>> # equivalently >>> EnvBase.register_gym("DMC-cheetah-v0", entry_point=DMControlEnv, env_name="cheetah", task="run")
transform (torchrl.envs.Transform) – a transform (or list of transforms within a
torchrl.envs.Composeinstance) to be used with the env. This arg can be passed during a call tomake()(see example below).info_keys (List[NestedKey], optional) –
if provided, these keys will be used to build the info dictionary and will be excluded from the observation keys. This arg can be passed during a call to
make()(see example below).Warning
It may be the case that using
info_keysmakes a spec empty because the content has been moved to the info dictionary. Gym does not like emptyDictin the specs, so this empty content should be removed withRemoveEmptySpecs.backend (str, optional) – the backend. Can be either “gym” or “gymnasium” or any other backend compatible with
set_gym_backend.to_numpy (bool, optional) – if
True, the result of calls to step and reset will be mapped to numpy arrays. Defaults toFalse(results are tensors). This arg can be passed during a call tomake()(see example below).reward_threshold (float, optional) – [Gym kwarg] The reward threshold considered to have learnt an environment.
nondeterministic (bool, optional) – [Gym kwarg If the environment is nondeterministic (even with knowledge of the initial seed and all actions). Defaults to
False.max_episode_steps (int, optional) – [Gym kwarg] The maximum number of episodes steps before truncation. Used by the Time Limit wrapper.
order_enforce (bool, optional) – [Gym >= 0.14] Whether the order enforcer wrapper should be applied to ensure users run functions in the correct order. Defaults to
True.autoreset (bool, optional) – [Gym >= 0.14] Whether the autoreset wrapper should be added such that reset does not need to be called. Defaults to
False.disable_env_checker – [Gym >= 0.14] Whether the environment checker should be disabled for the environment. Defaults to
False.apply_api_compatibility – [Gym >= 0.26] If to apply the StepAPICompatibility wrapper. Defaults to
False.**kwargs – arbitrary keyword arguments which are passed to the environment constructor.
Note
TorchRL’s environment do not have the concept of an
"info"dictionary, asTensorDictoffers all the storage requirements deemed necessary in most training settings. Still, you can use theinfo_keysargument to have a fine grained control over what is deemed to be considered as an observation and what should be seen as info.Examples
>>> # Register the "cheetah" env from DMControl with the "run" task >>> from torchrl.envs import DMControlEnv >>> import torch >>> DMControlEnv.register_gym("DMC-cheetah-v0", to_numpy=False, backend="gym", env_name="cheetah", task_name="run") >>> import gym >>> envgym = gym.make("DMC-cheetah-v0") >>> envgym.seed(0) >>> torch.manual_seed(0) >>> envgym.reset() ({'position': tensor([-0.0855, 0.0215, -0.0881, -0.0412, -0.1101, 0.0080, 0.0254, 0.0424], dtype=torch.float64), 'velocity': tensor([ 1.9609e-02, -1.9776e-04, -1.6347e-03, 3.3842e-02, 2.5338e-02, 3.3064e-02, 1.0381e-04, 7.6656e-05, 1.0204e-02], dtype=torch.float64)}, {}) >>> envgym.step(envgym.action_space.sample()) ({'position': tensor([-0.0833, 0.0275, -0.0612, -0.0770, -0.1256, 0.0082, 0.0186, 0.0476], dtype=torch.float64), 'velocity': tensor([ 0.2221, 0.2256, 0.5930, 2.6937, -3.5865, -1.5479, 0.0187, -0.6825, 0.5224], dtype=torch.float64)}, tensor([0.0018], dtype=torch.float64), tensor([False]), tensor([False]), {}) >>> # same environment with observation stacked >>> from torchrl.envs import CatTensors >>> envgym = gym.make("DMC-cheetah-v0", transform=CatTensors(in_keys=["position", "velocity"], out_key="observation")) >>> envgym.reset() ({'observation': tensor([-0.1005, 0.0335, -0.0268, 0.0133, -0.0627, 0.0074, -0.0488, -0.0353, -0.0075, -0.0069, 0.0098, -0.0058, 0.0033, -0.0157, -0.0004, -0.0381, -0.0452], dtype=torch.float64)}, {}) >>> # same environment with numpy observations >>> envgym = gym.make("DMC-cheetah-v0", transform=CatTensors(in_keys=["position", "velocity"], out_key="observation"), to_numpy=True) >>> envgym.reset() ({'observation': array([-0.11355747, 0.04257728, 0.00408397, 0.04155852, -0.0389733 , -0.01409826, -0.0978704 , -0.08808327, 0.03970837, 0.00535434, -0.02353762, 0.05116226, 0.02788907, 0.06848346, 0.05154399, 0.0371798 , 0.05128025])}, {}) >>> # If gymnasium is installed, we can register the environment there too. >>> DMControlEnv.register_gym("DMC-cheetah-v0", to_numpy=False, backend="gymnasium", env_name="cheetah", task_name="run") >>> import gymnasium >>> envgym = gymnasium.make("DMC-cheetah-v0") >>> envgym.seed(0) >>> torch.manual_seed(0) >>> envgym.reset() ({'position': tensor([-0.0855, 0.0215, -0.0881, -0.0412, -0.1101, 0.0080, 0.0254, 0.0424], dtype=torch.float64), 'velocity': tensor([ 1.9609e-02, -1.9776e-04, -1.6347e-03, 3.3842e-02, 2.5338e-02, 3.3064e-02, 1.0381e-04, 7.6656e-05, 1.0204e-02], dtype=torch.float64)}, {})
Note
This feature also works for stateless environments (eg,
BraxEnv).>>> import gymnasium >>> import torch >>> from tensordict import TensorDict >>> from torchrl.envs import BraxEnv, SelectTransform >>> >>> # get action for dydactic purposes >>> env = BraxEnv("ant", batch_size=[2]) >>> env.set_seed(0) >>> torch.manual_seed(0) >>> td = env.rollout(10) >>> >>> actions = td.get("action") >>> >>> # register env >>> env.register_gym("Brax-Ant-v0", env_name="ant", batch_size=[2], info_keys=["state"]) >>> gym_env = gymnasium.make("Brax-Ant-v0") >>> gym_env.seed(0) >>> torch.manual_seed(0) >>> >>> gym_env.reset() >>> obs = [] >>> for i in range(10): ... obs, reward, terminated, truncated, info = gym_env.step(td[..., i].get("action"))
- register_load_state_dict_post_hook(hook)¶
Register a post-hook to be run after module’s
load_state_dict()is called.- It should have the following signature::
hook(module, incompatible_keys) -> None
The
moduleargument is the current module that this hook is registered on, and theincompatible_keysargument is aNamedTupleconsisting of attributesmissing_keysandunexpected_keys.missing_keysis alistofstrcontaining the missing keys andunexpected_keysis alistofstrcontaining the unexpected keys.The given incompatible_keys can be modified inplace if needed.
Note that the checks performed when calling
load_state_dict()withstrict=Trueare affected by modifications the hook makes tomissing_keysorunexpected_keys, as expected. Additions to either set of keys will result in an error being thrown whenstrict=True, and clearing out both missing and unexpected keys will avoid an error.- Returns:
a handle that can be used to remove the added hook by calling
handle.remove()- Return type:
torch.utils.hooks.RemovableHandle
- register_load_state_dict_pre_hook(hook)¶
Register a pre-hook to be run before module’s
load_state_dict()is called.- It should have the following signature::
hook(module, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) -> None # noqa: B950
- Parameters:
hook (Callable) – Callable hook that will be invoked before loading the state dict.
- register_module(name: str, module: Optional[Module]) None¶
Alias for
add_module().
- register_parameter(name: str, param: Optional[Parameter]) None¶
Add a parameter to the module.
The parameter can be accessed as an attribute using given name.
- Parameters:
name (str) – name of the parameter. The parameter can be accessed from this module using the given name
param (Parameter or None) – parameter to be added to the module. If
None, then operations that run on parameters, such ascuda, are ignored. IfNone, the parameter is not included in the module’sstate_dict.
- register_state_dict_post_hook(hook)¶
Register a post-hook for the
state_dict()method.- It should have the following signature::
hook(module, state_dict, prefix, local_metadata) -> None
The registered hooks can modify the
state_dictinplace.
- register_state_dict_pre_hook(hook)¶
Register a pre-hook for the
state_dict()method.- It should have the following signature::
hook(module, prefix, keep_vars) -> None
The registered hooks can be used to perform pre-processing before the
state_dictcall is made.
- requires_grad_(requires_grad: bool = True) T¶
Change if autograd should record operations on parameters in this module.
This method sets the parameters’
requires_gradattributes in-place.This method is helpful for freezing part of the module for finetuning or training parts of a model individually (e.g., GAN training).
See Locally disabling gradient computation for a comparison between .requires_grad_() and several similar mechanisms that may be confused with it.
- Parameters:
requires_grad (bool) – whether autograd should record operations on parameters in this module. Default:
True.- Returns:
self
- Return type:
Module
- reset(tensordict: Optional[TensorDictBase] = None, **kwargs) TensorDictBase¶
Resets the environment.
As for step and _step, only the private method
_resetshould be overwritten by EnvBase subclasses.- Parameters:
tensordict (TensorDictBase, optional) – tensordict to be used to contain the resulting new observation. In some cases, this input can also be used to pass argument to the reset function.
kwargs (optional) – other arguments to be passed to the native reset function.
- Returns:
a tensordict (or the input tensordict, if any), modified in place with the resulting observations.
- property reset_keys: List[NestedKey]¶
Returns a list of reset keys.
Reset keys are keys that indicate partial reset, in batched, multitask or multiagent settings. They are structured as
(*prefix, "_reset")whereprefixis a (possibly empty) tuple of strings pointing to a tensordict location where a done state can be found.Keys are sorted by depth in the data tree.
- property reward_key¶
The reward key of an environment.
By default, this will be “reward”.
If there is more than one reward key in the environment, this function will raise an exception.
- property reward_keys: List[NestedKey]¶
The reward keys of an environment.
By default, there will only be one key named “reward”.
Keys are sorted by depth in the data tree.
- property reward_spec: TensorSpec¶
The
rewardspec.The
reward_specis always stored as a composite spec.If the reward spec is provided as a simple spec, this will be returned.
>>> env.reward_spec = UnboundedContinuousTensorSpec(1) >>> env.reward_spec UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous)
If the reward spec is provided as a composite spec and contains only one leaf, this function will return just the leaf.
>>> env.reward_spec = CompositeSpec({"nested": {"reward": UnboundedContinuousTensorSpec(1)}}) >>> env.reward_spec UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous)
If the reward spec is provided as a composite spec and has more than one leaf, this function will return the whole spec.
>>> env.reward_spec = CompositeSpec({"nested": {"reward": UnboundedContinuousTensorSpec(1), "another_reward": DiscreteTensorSpec(1)}}) >>> env.reward_spec CompositeSpec( nested: CompositeSpec( reward: UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=ContinuousBox( low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True), high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)), device=cpu, dtype=torch.float32, domain=continuous), another_reward: DiscreteTensorSpec( shape=torch.Size([]), space=DiscreteBox(n=1), device=cpu, dtype=torch.int64, domain=discrete), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
To retrieve the full spec passed, use:
>>> env.output_spec["full_reward_spec"]
This property is mutable.
Examples
>>> from torchrl.envs.libs.gym import GymEnv >>> env = GymEnv("Pendulum-v1") >>> env.reward_spec UnboundedContinuousTensorSpec( shape=torch.Size([1]), space=None, device=cpu, dtype=torch.float32, domain=continuous)
- rollout(max_steps: int, policy: Optional[Callable[[TensorDictBase], TensorDictBase]] = None, callback: Optional[Callable[[TensorDictBase, ...], Any]] = None, auto_reset: bool = True, auto_cast_to_device: bool = False, break_when_any_done: bool = True, return_contiguous: bool = True, tensordict: Optional[TensorDictBase] = None, set_truncated: bool = False, out=None)¶
Executes a rollout in the environment.
The function will stop as soon as one of the contained environments returns done=True.
- Parameters:
max_steps (int) – maximum number of steps to be executed. The actual number of steps can be smaller if the environment reaches a done state before max_steps have been executed.
policy (callable, optional) – callable to be called to compute the desired action. If no policy is provided, actions will be called using
env.rand_step(). The policy can be any callable that reads either a tensordict or the entire sequence of observation entries __sorted as__ theenv.observation_spec.keys(). Defaults to None.callback (Callable[[TensorDict], Any], optional) – function to be called at each iteration with the given TensorDict. Defaults to
None. The output ofcallbackwill not be collected, it is the user responsibility to save any result within the callback call if data needs to be carried over beyond the call torollout.auto_reset (bool, optional) – if
True, resets automatically the environment if it is in a done state when the rollout is initiated. Default isTrue.auto_cast_to_device (bool, optional) – if
True, the device of the tensordict is automatically cast to the policy device before the policy is used. Default isFalse.break_when_any_done (bool) – breaks if any of the done state is True. If False, a reset() is called on the sub-envs that are done. Default is True.
return_contiguous (bool) – if False, a LazyStackedTensorDict will be returned. Default is True.
tensordict (TensorDict, optional) – if
auto_resetis False, an initial tensordict must be provided. Rollout will check if this tensordict has done flags and reset the environment in those dimensions (if needed). This normally should not occur iftensordictis the output of a reset, but can occur iftensordictis the last step of a previous rollout. Atensordictcan also be provided whenauto_reset=Trueif metadata need to be passed to theresetmethod, such as a batch-size or a device for stateless environments.set_truncated (bool, optional) – if
True,"truncated"and"done"keys will be set toTrueafter completion of the rollout. If no"truncated"is found within thedone_spec, an exception is raised. Truncated keys can be set throughenv.add_truncated_keys. Defaults toFalse.
- Returns:
TensorDict object containing the resulting trajectory.
The data returned will be marked with a “time” dimension name for the last dimension of the tensordict (at the
env.ndimindex).rolloutis quite handy to display what the data structure of the environment looks like.Examples
>>> # Using rollout without a policy >>> from torchrl.envs.libs.gym import GymEnv >>> from torchrl.envs.transforms import TransformedEnv, StepCounter >>> env = TransformedEnv(GymEnv("Pendulum-v1"), StepCounter(max_steps=20)) >>> rollout = env.rollout(max_steps=1000) >>> print(rollout) TensorDict( fields={ action: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.float32, is_shared=False), done: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([20, 3]), device=cpu, dtype=torch.float32, is_shared=False), reward: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.float32, is_shared=False), step_count: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.int64, is_shared=False), truncated: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([20]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([20, 3]), device=cpu, dtype=torch.float32, is_shared=False), step_count: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.int64, is_shared=False), truncated: Tensor(shape=torch.Size([20, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([20]), device=cpu, is_shared=False) >>> print(rollout.names) ['time'] >>> # with envs that contain more dimensions >>> from torchrl.envs import SerialEnv >>> env = SerialEnv(3, lambda: TransformedEnv(GymEnv("Pendulum-v1"), StepCounter(max_steps=20))) >>> rollout = env.rollout(max_steps=1000) >>> print(rollout) TensorDict( fields={ action: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.float32, is_shared=False), done: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([3, 20, 3]), device=cpu, dtype=torch.float32, is_shared=False), reward: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.float32, is_shared=False), step_count: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.int64, is_shared=False), truncated: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([3, 20]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([3, 20, 3]), device=cpu, dtype=torch.float32, is_shared=False), step_count: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.int64, is_shared=False), truncated: Tensor(shape=torch.Size([3, 20, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([3, 20]), device=cpu, is_shared=False) >>> print(rollout.names) [None, 'time']
Using a policy (a regular
Moduleor aTensorDictModule) is also easy:Examples
>>> from torch import nn >>> env = GymEnv("CartPole-v1", categorical_action_encoding=True) >>> class ArgMaxModule(nn.Module): ... def forward(self, values): ... return values.argmax(-1) >>> n_obs = env.observation_spec["observation"].shape[-1] >>> n_act = env.action_spec.n >>> # A deterministic policy >>> policy = nn.Sequential( ... nn.Linear(n_obs, n_act), ... ArgMaxModule()) >>> env.rollout(max_steps=10, policy=policy) TensorDict( fields={ action: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.int64, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([10, 4]), device=cpu, dtype=torch.float32, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([10, 4]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False) >>> # Under the hood, rollout will wrap the policy in a TensorDictModule >>> # To speed things up we can do that ourselves >>> from tensordict.nn import TensorDictModule >>> policy = TensorDictModule(policy, in_keys=list(env.observation_spec.keys()), out_keys=["action"]) >>> env.rollout(max_steps=10, policy=policy) TensorDict( fields={ action: Tensor(shape=torch.Size([10]), device=cpu, dtype=torch.int64, is_shared=False), done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([10, 4]), device=cpu, dtype=torch.float32, is_shared=False), reward: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([10, 4]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([10, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([10]), device=cpu, is_shared=False)
In some instances, contiguous tensordict cannot be obtained because they cannot be stacked. This can happen when the data returned at each step may have a different shape, or when different environments are executed together. In that case,
return_contiguous=Falsewill cause the returned tensordict to be a lazy stack of tensordicts:- Examples of non-contiguous rollout:
>>> rollout = env.rollout(4, return_contiguous=False) >>> print(rollout) LazyStackedTensorDict( fields={ action: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.float32, is_shared=False), done: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: LazyStackedTensorDict( fields={ done: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([3, 4, 3]), device=cpu, dtype=torch.float32, is_shared=False), reward: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.float32, is_shared=False), step_count: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.int64, is_shared=False), truncated: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([3, 4]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([3, 4, 3]), device=cpu, dtype=torch.float32, is_shared=False), step_count: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.int64, is_shared=False), truncated: Tensor(shape=torch.Size([3, 4, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([3, 4]), device=cpu, is_shared=False) >>> print(rollout.names) [None, 'time']
Rollouts can be used in a loop to emulate data collection. To do so, you need to pass as input the last tensordict coming from the previous rollout after calling
step_mdp()on it.- Examples of data collection rollouts:
>>> from torchrl.envs import GymEnv, step_mdp >>> env = GymEnv("CartPole-v1") >>> epochs = 10 >>> input_td = env.reset() >>> for i in range(epochs): ... rollout_td = env.rollout( ... max_steps=100, ... break_when_any_done=False, ... auto_reset=False, ... tensordict=input_td, ... ) ... input_td = step_mdp( ... rollout_td[..., -1], ... )
- set_extra_state(state: Any) None¶
Set extra state contained in the loaded state_dict.
This function is called from
load_state_dict()to handle any extra state found within the state_dict. Implement this function and a correspondingget_extra_state()for your module if you need to store extra state within its state_dict.- Parameters:
state (dict) – Extra state from the state_dict
- set_seed(*args, **kwargs)[source]¶
Sets the seed of the environment and returns the next seed to be used (which is the input seed if a single environment is present).
- Parameters:
seed (int) – seed to be set. The seed is set only locally in the environment. To handle the global seed, see
manual_seed().static_seed (bool, optional) – if
True, the seed is not incremented. Defaults to False
- Returns:
i.e. the seed that should be used for another environment if created concomitantly to this environment.
- Return type:
integer representing the “next seed”
- set_submodule(target: str, module: Module) None¶
Set the submodule given by
targetif it exists, otherwise throw an error.For example, let’s say you have an
nn.ModuleAthat looks like this:A( (net_b): Module( (net_c): Module( (conv): Conv2d(16, 33, kernel_size=(3, 3), stride=(2, 2)) ) (linear): Linear(in_features=100, out_features=200, bias=True) ) )(The diagram shows an
nn.ModuleA.Ahas a nested submodulenet_b, which itself has two submodulesnet_candlinear.net_cthen has a submoduleconv.)To overide the
Conv2dwith a new submoduleLinear, you would callset_submodule("net_b.net_c.conv", nn.Linear(33, 16)).- Parameters:
target – The fully-qualified string name of the submodule to look for. (See above example for how to specify a fully-qualified string.)
module – The module to set the submodule to.
- Raises:
ValueError – If the target string is empty
AttributeError – If the target string references an invalid path or resolves to something that is not an
nn.Module
- property shape¶
Equivalent to
batch_size.
- property specs: CompositeSpec¶
Returns a Composite container where all the environment are present.
This feature allows one to create an environment, retrieve all of the specs in a single data container and then erase the environment from the workspace.
- state_dict(*args, **kwargs)[source]¶
Return a dictionary containing references to the whole state of the module.
Both parameters and persistent buffers (e.g. running averages) are included. Keys are corresponding parameter and buffer names. Parameters and buffers set to
Noneare not included.Note
The returned object is a shallow copy. It contains references to the module’s parameters and buffers.
Warning
Currently
state_dict()also accepts positional arguments fordestination,prefixandkeep_varsin order. However, this is being deprecated and keyword arguments will be enforced in future releases.Warning
Please avoid the use of argument
destinationas it is not designed for end-users.- Parameters:
destination (dict, optional) – If provided, the state of module will be updated into the dict and the same object is returned. Otherwise, an
OrderedDictwill be created and returned. Default:None.prefix (str, optional) – a prefix added to parameter and buffer names to compose the keys in state_dict. Default:
''.keep_vars (bool, optional) – by default the
Tensors returned in the state dict are detached from autograd. If it’s set toTrue, detaching will not be performed. Default:False.
- Returns:
a dictionary containing a whole state of the module
- Return type:
dict
Example:
>>> # xdoctest: +SKIP("undefined vars") >>> module.state_dict().keys() ['bias', 'weight']
- property state_keys: List[NestedKey]¶
The state keys of an environment.
By default, there will only be one key named “state”.
Keys are sorted by depth in the data tree.
- property state_spec: CompositeSpec¶
State spec.
Must be a
torchrl.data.CompositeSpecinstance. The keys listed here should be provided as input alongside actions to the environment.In TorchRL, even though they are not properly speaking “state” all inputs to the environment that are not actions are stored in the
state_spec.Therefore,
"state_spec"should be thought as a generic data container for environment inputs that are not action data.Examples
>>> from torchrl.envs import BraxEnv >>> for envname in BraxEnv.available_envs: ... break >>> env = BraxEnv(envname) >>> env.state_spec CompositeSpec( state: CompositeSpec( pipeline_state: CompositeSpec( q: UnboundedContinuousTensorSpec( shape=torch.Size([15]), space=None, device=cpu, dtype=torch.float32, domain=continuous), [...], device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
- step(tensordict: TensorDictBase) TensorDictBase¶
Makes a step in the environment.
Step accepts a single argument, tensordict, which usually carries an ‘action’ key which indicates the action to be taken. Step will call an out-place private method, _step, which is the method to be re-written by EnvBase subclasses.
- Parameters:
tensordict (TensorDictBase) – Tensordict containing the action to be taken. If the input tensordict contains a
"next"entry, the values contained in it will prevail over the newly computed values. This gives a mechanism to override the underlying computations.- Returns:
the input tensordict, modified in place with the resulting observations, done state and reward (+ others if needed).
- step_and_maybe_reset(*args, **kwargs)[source]¶
Runs a step in the environment and (partially) resets it if needed.
- Parameters:
tensordict (TensorDictBase) – an input data structure for the
step()method.
This method allows to easily code non-stopping rollout functions.
Examples
>>> from torchrl.envs import ParallelEnv, GymEnv >>> def rollout(env, n): ... data_ = env.reset() ... result = [] ... for i in range(n): ... data, data_ = env.step_and_maybe_reset(data_) ... result.append(data) ... return torch.stack(result) >>> env = ParallelEnv(2, lambda: GymEnv("CartPole-v1")) >>> print(rollout(env, 2)) TensorDict( fields={ done: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False), next: TensorDict( fields={ done: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False), observation: Tensor(shape=torch.Size([2, 2, 4]), device=cpu, dtype=torch.float32, is_shared=False), reward: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([2, 2]), device=cpu, is_shared=False), observation: Tensor(shape=torch.Size([2, 2, 4]), device=cpu, dtype=torch.float32, is_shared=False), terminated: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False), truncated: Tensor(shape=torch.Size([2, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False)}, batch_size=torch.Size([2, 2]), device=cpu, is_shared=False)
- to(device: Union[device, str, int])[source]¶
Move and/or cast the parameters and buffers.
This can be called as
- to(device=None, dtype=None, non_blocking=False)[source]
- to(dtype, non_blocking=False)[source]
- to(tensor, non_blocking=False)[source]
- to(memory_format=torch.channels_last)[source]
Its signature is similar to
torch.Tensor.to(), but only accepts floating point or complexdtypes. In addition, this method will only cast the floating point or complex parameters and buffers todtype(if given). The integral parameters and buffers will be moveddevice, if that is given, but with dtypes unchanged. Whennon_blockingis set, it tries to convert/move asynchronously with respect to the host if possible, e.g., moving CPU Tensors with pinned memory to CUDA devices.See below for examples.
Note
This method modifies the module in-place.
- Parameters:
device (
torch.device) – the desired device of the parameters and buffers in this moduledtype (
torch.dtype) – the desired floating point or complex dtype of the parameters and buffers in this moduletensor (torch.Tensor) – Tensor whose dtype and device are the desired dtype and device for all parameters and buffers in this module
memory_format (
torch.memory_format) – the desired memory format for 4D parameters and buffers in this module (keyword only argument)
- Returns:
self
- Return type:
Module
Examples:
>>> # xdoctest: +IGNORE_WANT("non-deterministic") >>> linear = nn.Linear(2, 2) >>> linear.weight Parameter containing: tensor([[ 0.1913, -0.3420], [-0.5113, -0.2325]]) >>> linear.to(torch.double) Linear(in_features=2, out_features=2, bias=True) >>> linear.weight Parameter containing: tensor([[ 0.1913, -0.3420], [-0.5113, -0.2325]], dtype=torch.float64) >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CUDA1) >>> gpu1 = torch.device("cuda:1") >>> linear.to(gpu1, dtype=torch.half, non_blocking=True) Linear(in_features=2, out_features=2, bias=True) >>> linear.weight Parameter containing: tensor([[ 0.1914, -0.3420], [-0.5112, -0.2324]], dtype=torch.float16, device='cuda:1') >>> cpu = torch.device("cpu") >>> linear.to(cpu) Linear(in_features=2, out_features=2, bias=True) >>> linear.weight Parameter containing: tensor([[ 0.1914, -0.3420], [-0.5112, -0.2324]], dtype=torch.float16) >>> linear = nn.Linear(2, 2, bias=None).to(torch.cdouble) >>> linear.weight Parameter containing: tensor([[ 0.3741+0.j, 0.2382+0.j], [ 0.5593+0.j, -0.4443+0.j]], dtype=torch.complex128) >>> linear(torch.ones(3, 2, dtype=torch.cdouble)) tensor([[0.6122+0.j, 0.1150+0.j], [0.6122+0.j, 0.1150+0.j], [0.6122+0.j, 0.1150+0.j]], dtype=torch.complex128)
- to_empty(*, device: Optional[Union[device, str, int]], recurse: bool = True) T¶
Move the parameters and buffers to the specified device without copying storage.
- Parameters:
device (
torch.device) – The desired device of the parameters and buffers in this module.recurse (bool) – Whether parameters and buffers of submodules should be recursively moved to the specified device.
- Returns:
self
- Return type:
Module
- train(mode: bool = True) T¶
Set the module in training mode.
This has any effect only on certain modules. See documentations of particular modules for details of their behaviors in training/evaluation mode, if they are affected, e.g.
Dropout,BatchNorm, etc.- Parameters:
mode (bool) – whether to set training mode (
True) or evaluation mode (False). Default:True.- Returns:
self
- Return type:
Module
- type(dst_type: Union[dtype, str]) T¶
Casts all parameters and buffers to
dst_type.Note
This method modifies the module in-place.
- Parameters:
dst_type (type or string) – the desired type
- Returns:
self
- Return type:
Module
- update_kwargs(kwargs: Union[dict, List[dict]]) None¶
Updates the kwargs of each environment given a dictionary or a list of dictionaries.
- Parameters:
kwargs (dict or list of dict) – new kwargs to use with the environments
- xpu(device: Optional[Union[int, device]] = None) T¶
Move all model parameters and buffers to the XPU.
This also makes associated parameters and buffers different objects. So it should be called before constructing optimizer if the module will live on XPU while being optimized.
Note
This method modifies the module in-place.
- Parameters:
device (int, optional) – if specified, all parameters will be copied to that device
- Returns:
self
- Return type:
Module
- zero_grad(set_to_none: bool = True) None¶
Reset gradients of all model parameters.
See similar function under
torch.optim.Optimizerfor more context.- Parameters:
set_to_none (bool) – instead of setting to zero, set the grads to None. See
torch.optim.Optimizer.zero_grad()for details.
