Gradient optimization: ROUNDING_ERROR

[Yogesh1-11]

hi,

with plmc usage
my calculation ran to 90 steps but then terminated with the following error.
Found focus OsNIP2-1 as sequence 1
255 valid sequences out of 255
298 sites out of 882
Effective number of samples: 135.2 (80% identical neighborhood = 1.000 samples)
iter time cond fx -loglk ||h|| ||e||
1 0.8 39.06 44305.5 44222.6 90.5 1.0
..................................................................................
89 88.7 0.07 15355.1 7713.3 124.5 21.8
90 89.8 0.04 15355.0 7711.4 124.5 21.8
Gradient optimization: ROUNDING_ERROR

can someone specify what does the error says?

my command was
plmc -c score_couple -le 16.0 -lh 0.01 -m 100 -g -f uniprot_id2041 aaa.a2m

[aaronkollasch]

Looking at lbfgs.c, it looks like this error happens when stp escapes from its bounds of stmin and stmax, which prevents further progress.

plmc/src/lib/lbfgs.c

Lines 909 to 913 in 18c9e55

	/* Test for errors and convergence. */
	if (brackt && ((*stp <= stmin || stmax <= *stp) || uinfo != 0)) {
	/* Rounding errors prevent further progress. */
	return LBFGSERR_ROUNDING_ERROR;
	}

Did you compile plmc in double precision mode, or single precision?

It's possible that plmc had already converged before encountering a rounding error, so you could try just reducing the number of steps to below 90 in this case, rather than the default 100.

[Yogesh1-11]

Hi Aaron Thankyou for the quick response. Yes when i am reducing the iterations i am getting "maximum iterations" . Also Is it possible to use the generated coupling scores directly in evcouplings. And how do one choose values for le, lh etc?

…

On Tue, 28 Mar, 2023, 6:58 pm Aaron Kollasch, ***@***.***> wrote: Looking at lbfgs.c, it looks like this error happens when stp escapes from its bounds of stmin and stmax, which prevents further progress. https://github.com/debbiemarkslab/plmc/blob/18c9e55e3bd2f14f4968be19a807b401996c929a/src/lib/lbfgs.c#L909-L913 Did you compile plmc in double precision mode, or single precision? It's possible that plmc had already converged before encountering a rounding error, so you could try just reducing the number of steps to below 90 in this case, rather than the default 100. — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZNDTOTI7YQ6KNHF74ORXPLW6LRQDANCNFSM6AAAAAAWKRDZKE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[aaronkollasch]

It is possible to use a .model file in evcouplings, but it's not the standard way to use it so you may have to use the code directly. What are you intending to do with evcouplings?

The CouplingsModel package https://github.com/debbiemarkslab/EVcouplings/blob/af27842c5fc72c291831261c9a19849a3e313efd/evcouplings/couplings/model.py#L236 can load the .model file and predict sequence likelihoods.

If you modify the ec_file and model_file in the config, and remove the "couplings" section, you may be able to get the protocol to work with an arbitrary model file. https://github.com/debbiemarkslab/EVcouplings/blob/develop/config/sample_config_monomer.txt

The other option is to run the evcouplings pipeline with the "existing" alignment protocol to use your a2m file. https://github.com/debbiemarkslab/EVcouplings/blob/af27842c5fc72c291831261c9a19849a3e313efd/config/sample_config_monomer.txt#L126-L139
Or, run the full evcouplings pipeline.

I would leave the lambda penalties as their default values unless you have a reason to change them.