add clarabel #61
add clarabel #61
Conversation
|
Sorry lower |
|
The opposite. atol is higher. i can look into it later.
…On Wed, Aug 30, 2023 at 12:48 AM Parth Nobel ***@***.***> wrote:
Sorry lower atol as in make it smaller? So we're requiring a more
accurate answer?
—
Reply to this email directly, view it on GitHub
<#61 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AB7LUGMKUWEG4EHCCYQKBULXX3AZTANCNFSM6AAAAAA4DXHYI4>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
|
I can't figure it out. did some basic checking, but Clarabel gets a better solution than SCS on both the initial solve and also the perturbed solve, so Im king of scratching my head. @bstellato |
|
How big is the kingdom of head scratching?On 1 Sep 2023, at 00:06, Shane Barratt ***@***.***> wrote:
I can't figure it out. did some basic checking, but Clarabel gets a better solution than SCS on both the initial solve and also the perturbed solve, so Im king of scratching my head. @bstellato
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
|
|
Could it be an issue of the linear system solved to take the derivative? Clarabel returns a higher accuracy solution with smaller complementary slackness residuals. It may cause the linear system to be badly scaled and lsqr to return less accurate derivatives. Just a guess though... |
|
I tried with a variety of backward methods (lsmr, lsqr and dense) and had
similar results
Shane
…On Mon, Sep 4, 2023 at 5:34 PM Bartolomeo Stellato ***@***.***> wrote:
Could it be an issue of the linear system solved to take the derivative?
Clarabel returns a higher accuracy solution with smaller complementary
slackness residuals. It may cause the linear system to be badly scaled and
lsqr to return less accurate derivatives. Just a guess though...
—
Reply to this email directly, view it on GitHub
<#61 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AB7LUGIX3J7HIUXSMXE2LALXYXYIHANCNFSM6AAAAAA4DXHYI4>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
|
There were two issues with Clarabel
It now works! @phschiele @akshayka @sbarratt what do you think? |
|
If it now passes tests then go ahead!
…On Thu, Oct 19, 2023 at 10:26 PM Bartolomeo Stellato < ***@***.***> wrote:
There were two issues with Clarabel
- It was not using the settings passed
- Its equilibration was making it way less precise
It now works! @phschiele <https://github.com/phschiele> @akshayka
<https://github.com/akshayka> @sbarratt <https://github.com/sbarratt>
what do you think?
—
Reply to this email directly, view it on GitHub
<#61 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AB7LUGPLLGZ7V4IQ6CDHFBTYAHONJAVCNFSM6AAAAAA4DXHYI6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZRHE3TMNZRGI>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
There was a problem hiding this comment.
@sbarratt @bstellato Thanks for this PR! I think it is really important to make sure that the main cvxpy extensions support Clarabel to foster adoption in the community.
I suggest to remove two small debug statements introduced during development.
Could it make sense to disable equilibrate by default if it was causing such trouble in the tests?
|
Thanks @phschiele. I have removed the print comments (sorry I did not see them before) and reverted the equilibration settings to its default ones. It now works as any other solver, I would not disable equilibration by default. It should be good to go now. |
|
@bstellato Thanks! Let's wait for @akshayka 's review. I'm updating the build in #63, so we can go ahead an cut a release once this is merged. |
this passes the tests. i had to lower the atol in the exp_cone test a bit; might be worth looking into.