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.

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

Jump to bottom

BUG: Fix the implementation of numpy.array_api.vecdot #21928

Merged
merged 2 commits into from Sep 7, 2022
Merged

BUG: Fix the implementation of numpy.array_api.vecdot #21928

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ddee5d9
Fix the implementation of numpy.array_api.vecdot
asmeurer Jul 5, 2022
285084a
Use moveaxis + matmul instead of einsum in vecdot
asmeurer Jul 7, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
13 changes: 12 additions & 1 deletion numpy/array_api/linalg.py
View file
Open in desktop
Expand Up @@ -379,7 +379,18 @@ def trace(x: Array, /, *, offset: int = 0) -> Array:
def vecdot(x1: Array, x2: Array, /, *, axis: int = -1) -> Array:
if x1.dtype not in _numeric_dtypes or x2.dtype not in _numeric_dtypes:
raise TypeError('Only numeric dtypes are allowed in vecdot')
return tensordot(x1, x2, axes=((axis,), (axis,)))
ndim = max(x1.ndim, x2.ndim)
x1_shape = (1,)*(ndim - x1.ndim) + tuple(x1.shape)
x2_shape = (1,)*(ndim - x2.ndim) + tuple(x2.shape)
Comment on lines +382 to +384
Copy link
Member

Choose a reason for hiding this comment

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

This will go away with the change.

Copy link
Member Author

@asmeurer asmeurer Jul 7, 2022 •
edited

Choose a reason for hiding this comment

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

Maybe this can be simplified, but without this, the shape checking is not correct (because size 1 dimensions broadcast).

Also, the broadcasting is necessary because, assuming I am understanding the idea in the spec correctly, nonnegative axis should refer to the broadcasted dimensions.

Copy link
Member

Choose a reason for hiding this comment

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

You can move the checking to later and use -1 as axis. That probably gives you better errors for bad axis values anyway.

Copy link
Member Author

Choose a reason for hiding this comment

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

I'm sorry, I'm not following your suggestion here.

Copy link
Member

Choose a reason for hiding this comment

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

I am saying that you can do the check after the broadcast+moveaxis at the point where the axis is properly normalized to -1.

Copy link
Member

Choose a reason for hiding this comment

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

I took the liberty to do that change, unless you disagree will merge next time around (or earlier with feedback).

if x1_shape[axis] != x2_shape[axis]:
raise ValueError("x1 and x2 must have the same size along the given axis")

x1_, x2_ = np.broadcast_arrays(x1._array, x2._array)
x1_ = np.moveaxis(x1_, axis, -1)
x2_ = np.moveaxis(x2_, axis, -1)

res = x1_[..., None, :] @ x2_[..., None]
return Array._new(res[..., 0, 0])


# Note: the name here is different from norm(). The array API norm is split
Expand Down