Skip to content

Feature: optional forward diff mode #14

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.

Sign up for GitHub

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

Open
mattia-spider wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

Feature: optional forward diff mode #14

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9bc8aff
Feature: optional forward diff mode
mattia-spider Jun 19, 2026
6fcedb7
Fix: deleted value = base.as_array(value) if fwd mode in _diff.py bec…
mattia-spider Jun 22, 2026
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
2 changes: 1 addition & 1 deletion pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ dependencies = [
'jraph',
'orbax-checkpoint',
'pytray',
'e3nn-jax',
'e3nn-jax>=0.21.1',
"equinox>=0.12.0",
"reax>=0.6.10,<0.7",
"rich",
Expand Down
30 changes: 17 additions & 13 deletions src/tensorial/gcnn/_diff.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,6 @@
DERIV_DELIMITER: Final[str] = ","
ArgumentSpecifier = Union[int, "gcnn.typing.TreePath"]


class DerivableGraphFunction(Protocol):
def __call__(
self, graph: jraph.GraphsTuple, *args: jt.PyTree
Expand Down Expand Up @@ -129,14 +128,15 @@ def evaluator(
argnum: int = 0,
scale: float = 1.0,
at: dict | None = None,
mode: str = "rev",
) -> "Evaluator":
if at is not None:
at = flax.core.FrozenDict(at)
return Evaluator(func, self, return_graph, argnum, scale=scale, at=at)
return Evaluator(func, self, return_graph, argnum, scale=scale, at=at, mode=mode)

@abc.abstractmethod
def build_derivative_fn(
self, func: DerivableGraphFunction, return_graph: bool, argnum: int
self, func: DerivableGraphFunction, return_graph: bool, argnum: int, mode: str = "rev"
) -> DerivableGraphFunction:
"""Get evaluate function from derivative"""

Expand Down Expand Up @@ -277,7 +277,7 @@ def __truediv__(self, other: "GraphEntrySpecLike | SingleDerivative") -> "MultiD
return MultiDerivative((self, SingleDerivative.create(self.out, wrt)))

def build_derivative_fn(
self, func: DerivableGraphFunction, return_graph: bool, argnum: int
self, func: DerivableGraphFunction, return_graph: bool, argnum: int, mode: str = "rev"
) -> DerivableGraphFunction:
if not argnum >= 0:
raise ValueError(f"argnum must be >= 0, got: {argnum}")
Expand All @@ -286,16 +286,15 @@ def build_derivative_fn(
# Scalar valued
diff_fn = jax.grad
else:
# Vector valued
# note: we could use forward mode in cases where dim inputs << outputs
# but the memory use is often exterme
diff_fn = jax.jacrev
# Vector valued; forward mode is more memory-efficient when dim(input) << dim(output)
diff_fn = jax.jacfwd if mode == "fwd" else jax.jacrev

def _diff_and_pre_process(
graph: jraph.GraphsTuple, *args: jt.PyTree
) -> tuple[jt.Array, jraph.GraphsTuple]:
value, graph = func(graph, *args)
value, graph = self._pre_process(value, graph)

return value, graph

do_diff = diff_fn(_diff_and_pre_process, argnums=1 + argnum, has_aux=True)
Expand Down Expand Up @@ -436,7 +435,7 @@ def __iter__(self):
yield from self.parts.__iter__()

def build_derivative_fn(
self, func: DerivableGraphFunction, return_graph: bool, argnum: int
self, func: DerivableGraphFunction, return_graph: bool, argnum: int, mode: str = "rev"
) -> DerivableGraphFunction:
# Work our way from right to left creating the derivative evaluators
argnums = []
Expand All @@ -447,9 +446,9 @@ def build_derivative_fn(
else:
argnums.append(self.graph_tuple_paths[wrt_path])

func = self[0].build_derivative_fn(func, return_graph=return_graph, argnum=argnums[0])
func = self[0].build_derivative_fn(func, return_graph=return_graph, argnum=argnums[0], mode=mode)
for part, argnum_ in zip(self[1:], argnums[1:]):
func = part.build_derivative_fn(func, return_graph=return_graph, argnum=argnum_)
func = part.build_derivative_fn(func, return_graph=return_graph, argnum=argnum_, mode=mode)

return func

Expand Down Expand Up @@ -478,6 +477,7 @@ class Evaluator:
argnum: int
scale: float = 1.0
at: flax.core.FrozenDict[str, jt.PyTree] | None = dataclasses.field(default=None, hash=False)
mode: str = "rev"

# will be set in __post_init__
_evaluate_at: DerivableGraphFunction = dataclasses.field(init=False)
Expand All @@ -486,7 +486,7 @@ def __post_init__(self):
object.__setattr__(
self,
"_evaluate_at",
self.spec.build_derivative_fn(self.func, return_graph=True, argnum=self.argnum),
self.spec.build_derivative_fn(self.func, return_graph=True, argnum=self.argnum, mode=self.mode),
)

def __call__(
Expand Down Expand Up @@ -526,6 +526,7 @@ def diff(
scale: float = 1.0,
at: dict | None = None,
return_graph=False,
mode: str = "rev",
) -> Evaluator:
"""
Constructs a JAX-compatible evaluator for computing single or multiple derivatives
Expand Down Expand Up @@ -568,6 +569,9 @@ def diff(
If True, the derivative tensor is packaged into a new Graph object under the
name specified by the 'out' argument. If False, the function returns the
raw derivative tensor. Defaults to False.
mode (str):
"rev" (default) uses reverse-mode AD (jax.jacrev). "fwd" uses forward-mode AD
(jax.jacfwd), which is more memory-efficient when dim(input) << dim(output).

Returns:
Evaluator: A callable object that takes a Graph and returns the computed
Expand Down Expand Up @@ -599,7 +603,7 @@ def diff(
else:
deriv = MultiDerivative.create(of, wrt, out)

return deriv.evaluator(deriv.adapt(func), return_graph, scale=scale, at=at)
return deriv.evaluator(deriv.adapt(func), return_graph, scale=scale, at=at, mode=mode)


def ordered_unique_indices(lst):
Expand Down
Loading