Skip to content

Add a bias_correction_v flag to scale_by_amsgrad to align with the original AMSGrad paper and Pytorch/tensorflow impl #1423

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
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Add a bias_correction_v flag to scale_by_amsgrad to align with the original AMSGrad paper and Pytorch/tensorflow impl #1423

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b4ee246
fix amsgrad: add option to disable bias correction on second moment (…
vvsvictor Sep 26, 2025
f858c76
remove trailing whitespace in alias.py
vvsvictor Sep 26, 2025
e7abe46
ci retrigger actions
vvsvictor Sep 26, 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
8 changes: 7 additions & 1 deletion optax/_src/alias.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1028,6 +1028,7 @@ def amsgrad(
eps: float = 1e-8,
eps_root: float = 0.0,
mu_dtype: Optional[Any] = None,
bias_correction_v: bool = True
) -> base.GradientTransformationExtraArgs:
"""The AMSGrad optimizer.

Expand Down Expand Up @@ -1077,7 +1078,12 @@ def amsgrad(
"""
return combine.chain(
transform.scale_by_amsgrad(
b1=b1, b2=b2, eps=eps, eps_root=eps_root, mu_dtype=mu_dtype
b1=b1,
b2=b2,
eps=eps,
eps_root=eps_root,
mu_dtype=mu_dtype,
bias_correction_v=bias_correction_v
Copy link
Collaborator

@vroulet vroulet Oct 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bias_correction_v is not a great name.

bias_correction_nu is already better

debias_nu may even be better but "bias_correction" as a boolean argument is already used in e.g. rmsprop (shame on me for that naming).

@rdyro what do you think?

),
transform.scale_by_learning_rate(learning_rate),
)
Expand Down
11 changes: 9 additions & 2 deletions optax/_src/transform.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -323,6 +323,7 @@ def scale_by_amsgrad(
eps: float = 1e-8,
eps_root: float = 0.0,
mu_dtype: Optional[chex.ArrayDType] = None,
bias_correction_v: bool = True
) -> base.GradientTransformation:
"""Rescale updates according to the AMSGrad algorithm.

Expand All @@ -336,6 +337,9 @@ def scale_by_amsgrad(
numerical stability when backpropagating gradients through the rescaling.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
bias_correction_v: Whether to apply bias correction to the second moment
estimate before taking the elementwise maximum (``nu_max``). Set to
``False`` to match the original AMSGrad paper and PyTorch/Keras behavior.
Copy link
Collaborator

@vroulet vroulet Oct 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"match pytorch behavior" -> no, see the conversation.

So just say "set to False to match original AMSGrad paper"


Returns:
A :class:`optax.GradientTransformation` object.
Expand All @@ -357,8 +361,11 @@ def update_fn(updates, state, params=None):
nu = optax.tree.update_moment_per_elem_norm(updates, state.nu, b2, 2)
count_inc = numerics.safe_increment(state.count)
mu_hat = optax.tree.bias_correction(mu, b1, count_inc)
nu_hat = optax.tree.bias_correction(nu, b2, count_inc)
nu_max = jax.tree.map(jnp.maximum, state.nu_max, nu_hat)
if bias_correction_v:
nu_eff = optax.tree.bias_correction(nu, b2, count_inc)
else:
nu_eff = nu
nu_max = jax.tree.map(jnp.maximum, state.nu_max, nu_eff)
updates = jax.tree.map(
lambda m, v: None if m is None else m / (jnp.sqrt(v + eps_root) + eps),
mu_hat,
Expand Down
Loading