Truncated SVD #74
Truncated SVD #74
Conversation
Draft truncated SVD implementation
|
I changed naming to be consistent with Julia conventions. I fix the inline docs, and I messed a bit with one of the test files. It really needs more tests I think but it would be ok to merge this now so it becomes usable. @wcwitt -- are you happy for us to merge this and tag a new version? |
test/test_linearsolvers.jl
Outdated
| A = randn(Nobs, Nfeat) / sqrt(Nobs) | ||
| y = randn(Nobs) | ||
| A1 = randn(Nobs, Nfeat) / sqrt(Nobs) | ||
| U, S1, V = svd(A) |
There was a problem hiding this comment.
This line breaks the tests because of a typo. I would fix it myself, but I have a question. The changes in this block affect the tests for ALL solvers by redefining A to be something slightly different. I'm not opposed to this in principle, but are you sure it's what you want?
There was a problem hiding this comment.
I want to see how changing regularisation changes the solution. In the original all singluar values are very close to 1, so for methods like RRQR and truncated SVD varying the regularisation has essentially zero effect.
There was a problem hiding this comment.
I'll fix the typo, sorry I wasn't paying attention there.
|
Thanks - looks interesting! Once the previous comment is resolved (and the tests are passing) I'm happy for you to merge. |
|
The CI ran successfully, so happy for you to merge and tag. |
|
registered as v"0.1.5" |
continue Jerry's PR for adding the truncated SVD