gh-405: array API support for glass.core.algorithm
#423
Conversation
|
Perhaps, before merging any changes, we should figure out a way to
|
glass/core/algorithm.py
Outdated
| msg = "input arrays should belong to the same array library" | ||
| raise ValueError(msg) | ||
|
|
||
| xp = a.__array_namespace__() |
There was a problem hiding this comment.
Was talking to @matt-graham and he wondered if we should be using array-api-compat here. That should solve the cupy issue. I believe only NumPy and JAX are stable at this point.
There was a problem hiding this comment.
The CuPy support will become official in their next major release (v14) where they will start supporting NumPy v2. I hope they push the release in the next 3 months, so that we don't have to experiment with array-api-compat.
Saransh-cpp
force-pushed
the
saransh/array-api-core
branch
2 times, most recently
from
6e8ef45 to
761795c
Compare
Saransh-cpp
force-pushed
the
saransh/array-api-test
branch
from
fa75b7a to
1b0d92b
Compare
Saransh-cpp
force-pushed
the
saransh/array-api-core
branch
from
ee54338 to
536f1ab
Compare
Saransh-cpp
force-pushed
the
saransh/array-api-core
branch
from
536f1ab to
038bbbf
Compare
| def nnls( | ||
| a: NDArray[np.float64], | ||
| b: NDArray[np.float64], | ||
| a: NDArray[np.float64] | ArrayLike, |
| a = rng.uniform(low=-10, high=10, size=[50, 10]) | ||
| b = np.abs(rng.uniform(low=-2, high=2, size=[10])) | ||
| b = xp.abs(rng.uniform(low=-2, high=2, size=[10])) |
There was a problem hiding this comment.
The functions might not use RNGs but their tests might, so we do need to come up with a consensus on RNGs.
There was a problem hiding this comment.
I guess for tests we can always do -
xp.asarray(rng(...)) # numpy rng output casted to xp.arrayThere was a problem hiding this comment.
Use the workaround (NumPy-like - rng-jax) for now, and see if we need a more detailed (JAX-like) solution in the future.
| p[m] = True | ||
| while True: | ||
| ap = a[:, p] | ||
| xp = x[p] | ||
| sp = np.linalg.solve(ap.T @ ap, b @ ap) | ||
| x_new = x[p] | ||
| sp = xp.linalg.solve(ap.T @ ap, b @ ap) | ||
| t = sp <= 0 | ||
| if not np.any(t): | ||
| if not xp.any(t): | ||
| break | ||
| alpha = -np.min(xp[t] / (xp[t] - sp[t])) | ||
| alpha = -xp.min(x_new[t] / (x_new[t] - sp[t])) | ||
| x[p] += alpha * (sp - xp) | ||
| p[x <= 0] = False |
| if a.__array_namespace__() != b.__array_namespace__(): | ||
| msg = "input arrays should belong to the same array library" | ||
| raise ValueError(msg) | ||
|
|
||
| xp = a.__array_namespace__() |
There was a problem hiding this comment.
Should we make this into a generic helper? Seeing as all functions will need it
There was a problem hiding this comment.
That sounds good. I'll open a new issue and take this up.
|
@paddyroddy I think I've now done this in #643 |
Agreed, sorry @Saransh-cpp |
4 participants
WIP (not draft to test array API in the CI)
Has some caveats that I am working on, but opening this PR for visibility.
Closes: #405