Skip to content

refactor(autogram): Switch to ComputeModuleJacobians #452

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

Merged
ValerianRey merged 4 commits into from
Oct 12, 2025
Merged

refactor(autogram): Switch to ComputeModuleJacobians #452

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
1d92c6e
Make `AccumulateJacobian` return the jacobian w.r.t. the parameters o…
PierreQuinton Oct 12, 2025
c5c1d60
Fix type hint of in_dims
ValerianRey Oct 12, 2025
782a480
Create variable for out_dims in vmap
ValerianRey Oct 12, 2025
63e6e42
Change returned out_dims to None
ValerianRey Oct 12, 2025
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
25 changes: 11 additions & 14 deletions src/torchjd/autogram/_module_hook_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -206,52 +206,49 @@ def backward(ctx, *grad_outputs: Tensor) -> tuple:
if not ctx.gramian_accumulation_phase:
return None, None, None, None, None, None, *grad_outputs

AccumulateJacobian.apply(
path_jacobians = ComputeModuleJacobians.apply(
ctx.vjp,
ctx.args,
ctx.kwargs,
ctx.gramian_accumulator,
ctx.module,
*grad_outputs,
)
ctx.gramian_accumulator.accumulate_path_jacobians(path_jacobians)

return None, None, None, None, None, None, *grad_outputs


class AccumulateJacobian(torch.autograd.Function):
class ComputeModuleJacobians(torch.autograd.Function):

@staticmethod
def forward(
vjp: VJP,
args: tuple[PyTree, ...],
kwargs: dict[str, PyTree],
gramian_accumulator: GramianAccumulator,
module: nn.Module,
*grad_outputs: Tensor,
) -> None:
) -> dict[Tensor, Tensor]:
# There is no non-batched dimension
generalized_jacobians = vjp(grad_outputs, args, kwargs)
path_jacobians = AccumulateJacobian._make_path_jacobians(module, generalized_jacobians)
gramian_accumulator.accumulate_path_jacobians(path_jacobians)
path_jacobians = ComputeModuleJacobians._make_path_jacobians(module, generalized_jacobians)
return path_jacobians

@staticmethod
def vmap(
_,
in_dims: tuple, # tuple[None, tuple[PyTree, ...], dict[str, PyTree], None, None, *tuple[int | None, ...]]
in_dims: tuple, # tuple[None, tuple[PyTree, ...], dict[str, PyTree], None, *tuple[int | None, ...]]
vjp: VJP,
args: tuple[PyTree, ...],
kwargs: dict[str, PyTree],
gramian_accumulator: GramianAccumulator,
module: nn.Module,
*jac_outputs: Tensor,
) -> tuple[None, None]:
) -> tuple[dict[Tensor, Tensor], None]:
# There is a non-batched dimension
# We do not vmap over the args for the non-batched dimension
in_dims = (in_dims[5:], tree_map(lambda _: None, args), tree_map(lambda _: None, kwargs))
in_dims = (in_dims[4:], tree_map(lambda _: None, args), tree_map(lambda _: None, kwargs))
generalized_jacobians = torch.vmap(vjp, in_dims=in_dims)(jac_outputs, args, kwargs)
path_jacobians = AccumulateJacobian._make_path_jacobians(module, generalized_jacobians)
gramian_accumulator.accumulate_path_jacobians(path_jacobians)
return None, None
path_jacobians = ComputeModuleJacobians._make_path_jacobians(module, generalized_jacobians)
return path_jacobians, None

@staticmethod
def _make_path_jacobians(
Expand Down
Loading