Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Feature] TensorDictPrimer with single default_value callable #2732

Open
wants to merge 2 commits into
base: gh/vmoens/84/base
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 13 additions & 4 deletions torchrl/envs/custom/pendulum.py
Original file line number Diff line number Diff line change
Expand Up @@ -269,11 +269,20 @@ def _reset(self, tensordict):
batch_size = (
tensordict.batch_size if tensordict is not None else self.batch_size
)
if tensordict is None or tensordict.is_empty():
if tensordict is None or "params" not in tensordict:
# if no ``tensordict`` is passed, we generate a single set of hyperparameters
# Otherwise, we assume that the input ``tensordict`` contains all the relevant
# parameters to get started.
tensordict = self.gen_params(batch_size=batch_size, device=self.device)
elif "th" in tensordict and "thdot" in tensordict:
# we can hard-reset the env too
return tensordict
out = self._reset_random_data(
tensordict.shape, batch_size, tensordict["params"]
)
return out

def _reset_random_data(self, shape, batch_size, params):

high_th = torch.tensor(self.DEFAULT_X, device=self.device)
high_thdot = torch.tensor(self.DEFAULT_Y, device=self.device)
Expand All @@ -284,20 +293,20 @@ def _reset(self, tensordict):
# of simulators run simultaneously. In other contexts, the initial
# random state's shape will depend upon the environment batch-size instead.
th = (
torch.rand(tensordict.shape, generator=self.rng, device=self.device)
torch.rand(shape, generator=self.rng, device=self.device)
* (high_th - low_th)
+ low_th
)
thdot = (
torch.rand(tensordict.shape, generator=self.rng, device=self.device)
torch.rand(shape, generator=self.rng, device=self.device)
* (high_thdot - low_thdot)
+ low_thdot
)
out = TensorDict(
{
"th": th,
"thdot": thdot,
"params": tensordict["params"],
"params": params,
},
batch_size=batch_size,
)
Expand Down
34 changes: 30 additions & 4 deletions torchrl/envs/transforms/transforms.py
Original file line number Diff line number Diff line change
Expand Up @@ -5674,14 +5674,20 @@ class TensorDictPrimer(Transform):
Defaults to `False`.
default_value (float, Callable, Dict[NestedKey, float], Dict[NestedKey, Callable], optional): If non-random
filling is chosen, `default_value` will be used to populate the tensors. If `default_value` is a float,
all elements of the tensors will be set to that value. If it is a callable, this callable is expected to
return a tensor fitting the specs, and it will be used to generate the tensors. Finally, if `default_value`
is a dictionary of tensors or a dictionary of callables with keys matching those of the specs, these will
be used to generate the corresponding tensors. Defaults to `0.0`.
all elements of the tensors will be set to that value.
If it is a callable and `single_default_value=False` (default), this callable is expected to return a tensor
fitting the specs (ie, ``default_value()`` will be called independently for each leaf spec). If it is a
callable and ``single_default_value=True``, then the callable will be called just once and it is expected
that the structure of its returned TensorDict instance or equivalent will match the provided specs.
Finally, if `default_value` is a dictionary of tensors or a dictionary of callables with keys matching
those of the specs, these will be used to generate the corresponding tensors. Defaults to `0.0`.
reset_key (NestedKey, optional): the reset key to be used as partial
reset indicator. Must be unique. If not provided, defaults to the
only reset key of the parent environment (if it has only one)
and raises an exception otherwise.
single_default_value (bool, optional): if ``True`` and `default_value` is a callable, it will be expected that
``default_value`` returns a single tensordict matching the specs. If `False`, `default_value()` will be
called independently for each leaf. Defaults to ``False``.
**kwargs: each keyword argument corresponds to a key in the tensordict.
The corresponding value has to be a TensorSpec instance indicating
what the value must be.
Expand Down Expand Up @@ -5781,6 +5787,7 @@ def __init__(
| Dict[NestedKey, Callable] = None,
reset_key: NestedKey | None = None,
expand_specs: bool = None,
single_default_value: bool = False,
**kwargs,
):
self.device = kwargs.pop("device", None)
Expand Down Expand Up @@ -5821,10 +5828,13 @@ def __init__(
raise ValueError(
"If a default_value dictionary is provided, it must match the primers keys."
)
elif single_default_value:
pass
else:
default_value = {
key: default_value for key in self.primers.keys(True, True)
}
self.single_default_value = single_default_value
self.default_value = default_value
self._validated = False
self.reset_key = reset_key
Expand Down Expand Up @@ -5937,6 +5947,14 @@ def _validate_value_tensor(self, value, spec):
return True

def forward(self, tensordict: TensorDictBase) -> TensorDictBase:
if self.single_default_value and callable(self.default_value):
tensordict.update(self.default_value())
for key, spec in self.primers.items(True, True):
if not self._validated:
self._validate_value_tensor(tensordict.get(key), spec)
if not self._validated:
self._validated = True
return tensordict
for key, spec in self.primers.items(True, True):
if spec.shape[: len(tensordict.shape)] != tensordict.shape:
raise RuntimeError(
Expand Down Expand Up @@ -5991,6 +6009,14 @@ def _reset(
):
self.primers = self._expand_shape(self.primers)
if _reset.any():
if self.single_default_value and callable(self.default_value):
tensordict_reset.update(self.default_value())
for key, spec in self.primers.items(True, True):
if not self._validated:
self._validate_value_tensor(tensordict_reset.get(key), spec)
self._validated = True
return tensordict_reset

for key, spec in self.primers.items(True, True):
if self.random:
shape = (
Expand Down
Loading