Reuse Arpack in order to compute norm of Bk #159
Conversation
|
Please hold off on merging for now, as some tests are affected by #42 . @nathanemac you can try to merge this PR locally with R2N in order to use ADNLPModels |
| @@ -130,7 +130,8 @@ function LMTR( | |||
| JdFk = similar(Fk) # temporary storage | |||
| Jt_Fk = similar(∇fk) # temporary storage | |||
|
|
|||
| σmax = opnorm(Jk) | |||
| σmax, found_σ = opnorm(Jk) | |||
| found_σ || error("operator norm computation failed") | |||
There was a problem hiding this comment.
We should have a backup plan if the computation fails.
There was a problem hiding this comment.
Working on understanding why it may not work ^^
There was a problem hiding this comment.
Done
|
@MohamedLaghdafHABIBOULLAH, using
|
Co-authored-by: Dominique <[email protected]>
|
@nathanemac thanks for the confirmation, we believe that Arrack may be more accurate than TSVD, but still have some bugs... |
Use alphabetical order for external dependencies and remove unused variables.
Enhance robustness of Arpack usage by catching NCV-related errors Catch the XYAUPD_Exception error in Arpack calls and dynamically increase NCV as needed to handle convergence issues. This modification addresses instability in LSR1 computations due to insufficient NCV.
|
This PR should address the problem encountered in this issue #42 |
| while !(have_eig || attempt >= max_attempts) | ||
| attempt += 1 | ||
| try | ||
| # Perform eigendecomposition |
There was a problem hiding this comment.
| # Perform eigendecomposition | |
| # Estimate largest eigenvalue in absolute value |
| catch e | ||
| if occursin("XYAUPD_Exception", string(e)) | ||
| @warn "Arpack error: $e. Increasing NCV to $ncv and retrying." | ||
| ncv = min(2 * ncv, n) # Increase NCV but don't exceed matrix size |
There was a problem hiding this comment.
If we still get an error when ncv == n, we should abort.
| while !(have_svd || attempt >= max_attempts) | ||
| attempt += 1 | ||
| try | ||
| # Perform singular value decomposition |
There was a problem hiding this comment.
| # Perform singular value decomposition | |
| # Estimate largest singular value |
|
Just wondering, @MohamedLaghdafHABIBOULLAH, is it possible to use the frobenius norm instead of the operator norm for your algorithm ? |
|
@MaxenceGollier yes one could and it has been discussed in the paper and in this issue #133 , however I'm afraid that the solver will become slower as the step size is related to 1/|B_k|, in this case the Frobenius norm might represent a loose bound on the operator norm or the largest eig norm. |
|
You can’t compute the Frobenius norm of a quasi-Newton operator without a lot of work. |
|
@MohamedLaghdafHABIBOULLAH can we add a wrapped allocs unit test to be sure that opnorm does not allocate as well ? |
Based on this commit 48e6f01#diff-72ed386c2a0cd1d23c0968297e70023ed98c22490d146dd89fc91f48369bad4dL7