Low Rank dispatch rules #57
Conversation
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## main #57 +/- ##
==========================================
+ Coverage 78.44% 78.78% +0.33%
==========================================
Files 36 36
Lines 2974 2988 +14
==========================================
+ Hits 2333 2354 +21
+ Misses 641 634 -7
☔ View full report in Codecov by Sentry. |
Fr0do
force-pushed
the
lowrank_dispatch
branch
from
ca09dd6 to
0ae5c03
Compare
There was a problem hiding this comment.
Great work @Fr0do 🙏, this is a nice little addition showcasing some of the more complicated dispatch rules 🚀
Let's see if we can generalize this even further with some additional types (U,V need not be dense, Diagonal can be made more general). If you don't get to it, that's fine too, I can finish up with some additional cases later this week
| U, V = A.Ms[0].Ms | ||
| D_inv = inv(A.Ms[1], **kwargs) | ||
| I = Identity(shape=(V.shape[0], U.shape[1]), dtype=V.dtype) | ||
| return D_inv - D_inv @ U @ inv(V @ D_inv @ U + I, **kwargs) @ V @ D_inv |
|
|
||
|
|
||
| @dispatch | ||
| def inv(A: Sum[Product[Dense,Dense], Diagonal], **kwargs): |
There was a problem hiding this comment.
We should be able to generalize this to
Sum[Product[LinearOperator,LinearOperator] right?
Also I believe we should be able to generalize Diagonal to efficiently_invertible = Union[Diagonal, Identity, ScalarMul], we could extend that further to Triangular, Permutation.
|
I didn't realize you're on a fork, I made some edits and here is a diff: Though the tests pass (with the adjustment to take into account that |
Is there a way to work collaboratively? I can add mainainers of CoLA to my fork as collaborators. Or maybe there is a way I can create a branch in your repo? |
Low-rank dispatch rules for Woodbury Identity case of
invand computationally effectivetraceof low-rank product, resolves #48