Skip to content

Avoid division by zero in adagrad #1072

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
mhucka merged 1 commit into from
Jun 5, 2025
Merged

Avoid division by zero in adagrad #1072

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 4 additions & 2 deletions src/openfermion/resource_estimates/thc/utils/adagrad.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,10 +37,12 @@ def init(x0):
m = np.zeros_like(x0)
return x0, g_sq, m

def update(i, g, state):
def update(i, g, state, eps=1e-9):
x, g_sq, m = state
g_sq += np.square(g)
g_sq_inv_sqrt = np.where(g_sq > 0, 1.0 / np.sqrt(g_sq), 0.0)
# Add a small number to avoid division by zero
g_sq_safe = g_sq + eps
Copy link
Contributor

@dstrain115 dstrain115 Jun 4, 2025

Choose a reason for hiding this comment

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

What do you think about doing this only if abs(g_sq<eps)?

Copy link
Contributor Author

@mhucka mhucka Jun 4, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

It may impact performance, but probably not too badly. I've rewritten it like this:

if np.abs(g_sq < eps):
    # Add a small number to avoid division by zero.
    g_sq_inv_sqrt = 1.0 / np.sqrt(g_sq + eps)
else:
    g_sq_inv_sqrt = 1.0 / np.sqrt(g_sq)

Does that seem okay?

I forgot that g_sq is an array, so a test like x < y means testing every value, and that's going to be more time-consuming.

My reading of to handle cases like this is that the original is not a bad approach in a case where gradients are being done, because ultimately the values are not used in the results. @dstrain115 do you think it's going to lead to errors if left in the original form?

g_sq_inv_sqrt = 1.0 / np.sqrt(g_sq_safe)
m = (1.0 - momentum) * (g * g_sq_inv_sqrt) + momentum * m
x = x - step_size(i) * m
return x, g_sq, m
Expand Down
Loading