mailRe: Final model-free run with dauvergne_protocol.py


Others Months | Index by Date | Thread Index
>>   [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Header


Content

Posted by Edward d'Auvergne on September 24, 2012 - 10:39:
Hi Nick,


From the relax printouts you gave in the mail, I can see that the

problem lies much higher up in the analysis.  I can see that for all
diffusion tensors, only round 2 was reached.  This is nowhere near
enough number of iterations for the analysis to reach the global
optimisation+model selection minimum.  See figure 2 of d'Auvergne and
Gooley, 2008b (http://dx.doi.org/10.1007/s10858-007-9213-3) for the
iterative protocol concept and how many rounds you should expect and
the d'Auvergne and Gooley 2007 (http://dx.doi.org/10.1039/b702202f)
paper for the description of what a combined optimisation+model
selection minimum for a model-free analysis is.

I'm pretty sure that if the iterative procedure for each diffusion
tensor to find the global combined optimisation+model selection
minimum is completed, that the chi-squared value would significantly
decrease and hence the final AIC model selection between diffusion
tensors will produce quite a different result.  The completion of the
iterations should also eliminate a lot of artificial motions in the
system (see the above papers for details).  Do you know why only 2
iterations for each diffusion tensor was completed?

Regards,

Edward


P. S.  The RelaxError is another problem which may disappear when the
iterative procedure is completed.  This is thoroughly tested in the
relax test suite, so I think this is a corner case problem triggered
by another problem higher up in the analysis.  You would normally
never encounter this.




On 23 September 2012 18:37, Nicolas Doucet <Nicolas.Doucet@xxxxxxxxxxx> wrote:

Edward,

It seems we finally got around the relax halting problem but unfortunately I
don’t know how! Logging didn’t affect anything but our last couple of trials
worked fine and this sudden stop of relax could not be reproduced. Hopefully
it’s not a failing hardware problem like you suggested. I haven't gotten to
it yet but will definitely run relax on another workstation to rule this
out.

In any case, we re-ran each DIFF_MODEL and they all completed without
halting. Still, we get the same error when trying the final run:

-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
relax> value.set(val=-0.00017199999999999998, param='csa', spin_id='@N')

relax> pipe.create(pipe_name='local_tm - mf (Sat Sep 22 16:54:58 2012)',
pipe_type='mf', bundle='mf (Sat Sep 22 16:54:58 2012)')

relax> results.read(file='results',
dir='/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/local_tm/aic')
Opening the file
'/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/local_tm/aic/results.bz2'
for reading.

relax> pipe.create(pipe_name='sphere - mf (Sat Sep 22 16:54:58 2012)',
pipe_type='mf', bundle='mf (Sat Sep 22 16:54:58 2012)')

relax> results.read(file='results',
dir='/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/sphere/round_2/opt')
Opening the file
'/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/sphere/round_2/opt/results.bz2'
for reading.

relax> pipe.create(pipe_name='prolate - mf (Sat Sep 22 16:54:58 2012)',
pipe_type='mf', bundle='mf (Sat Sep 22 16:54:58 2012)')

relax> results.read(file='results',
dir='/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/prolate/round_2/opt')
Opening the file
'/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/prolate/round_2/opt/results.bz2'
for reading.

relax> pipe.create(pipe_name='oblate - mf (Sat Sep 22 16:54:58 2012)',
pipe_type='mf', bundle='mf (Sat Sep 22 16:54:58 2012)')

relax> results.read(file='results',
dir='/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/oblate/round_2/opt')
Opening the file
'/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/oblate/round_2/opt/results.bz2'
for reading.

relax> pipe.create(pipe_name='ellipsoid - mf (Sat Sep 22 16:54:58 2012)',
pipe_type='mf', bundle='mf (Sat Sep 22 16:54:58 2012)')

relax> results.read(file='results',
dir='/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/ellipsoid/round_2/opt')
Opening the file
'/home/dgagne/nmrdata2/DG12-116_15SEP2012_ModelFree_Analysis_of_ECP_Free_Form/dauvergne_protocol_final/ellipsoid/round_2/opt/results.bz2'
for reading.

relax> model_selection(method='AIC', modsel_pipe='final - mf (Sat Sep 22
16:54:58 2012)', bundle='mf (Sat Sep 22 16:54:58 2012)', pipes=['local_tm -
mf (Sat Sep 22 16:54:58 2012)', 'sphere - mf (Sat Sep 22 16:54:58 2012)',
'prolate - mf (Sat Sep 22 16:54:58 2012)', 'oblate - mf (Sat Sep 22 16:54:58
2012)', 'ellipsoid - mf (Sat Sep 22 16:54:58 2012)'])
AIC model selection.

Global model - all diffusion tensor parameters and spin specific model-free
parameters.
Data pipe            Num_params_(k)       Num_data_sets_(n)    Chi2
Criterion
local_tm - mf (Sat Sep 22 16:54:58 2012) 233                  678
374948.04349         375414.04349
sphere - mf (Sat Sep 22 16:54:58 2012) 150                  678
379689.20076         379989.20076
prolate - mf (Sat Sep 22 16:54:58 2012) 155                  678
379992.81969         380302.81969
oblate - mf (Sat Sep 22 16:54:58 2012) 158                  678
379741.98189         380057.98189
ellipsoid - mf (Sat Sep 22 16:54:58 2012) 161                  678
379888.19715         380210.19715
The model from the data pipe 'local_tm - mf (Sat Sep 22 16:54:58 2012)' has
been selected.

relax> monte_carlo.setup(number=500)


relax> monte_carlo.create_data(method='back_calc')
RelaxError: The specific model has not been selected or set up.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

We went back to each individual DIFF_MODEL logs to figure out whether
irregularities could be observed. Here is what shows up in each individual
run (sphere, oblate, etc.) for each individual spin (we have 133 different
@N spins in this pdb file, a few of those we listed in the 'unresolved'
file):

Fitting to spin '#2KB5_mol1:4@N'
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Cannot run a grid search on a model with zero parameters, directly
calculating the function value.
Storing the optimisation results, the optimised chi-squared value is lower
than the current value (6753.10859718 < None).

The following message then appears:

-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
relax> minimise(min_algor='newton', line_search=None, hessian_mod=None,
hessian_type=None, func_tol=1e-25, grad_tol=None, max_iter=10000000,
constraints=True, scaling=True, verbosity=1)


Over-fit spin deselection.

Only the model-free parameters for single spins will be used.


Fitting to spin '#2KB5_mol1:4@N'
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Cannot run optimisation on a model with zero parameters, directly
calculating the function value.

Parameter values: []
Function value:   6701.4699098031615
Iterations:       0
Function calls:   1
Gradient calls:   0
Hessian calls:    0
Warning:          No optimisation

Storing the optimisation results, the optimised chi-squared value is lower
than the current value (6701.4699098 < None).
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Thanks again for your input.

Nick

--
Nicolas Doucet
Assistant Professor
INRS - Institut Armand-Frappier
University of Quebec
Institut Pasteur International Network
531 Boulevard des Prairies
Laval (Quebec) H7V 1B7 CANADA
Phone: (450) 687-5010 #4212
Fax: (450) 686-5501
Email: nicolas.doucet@xxxxxxxxxxx
Web: http://www.profs.inrs.ca/ndoucet/
--



On 2012-09-21, at 9:48 AM, Edward d'Auvergne <edward@xxxxxxxxxxxxx> wrote:

Hi,

I'll answer below:

On Friday, 21 September 2012, Doucet, Nicolas <Nicolas.Doucet@xxxxxxxxxxx>
wrote:

Edward,

Using all the relaxation .out files previously created by relax in one
single directory including the pdb and the dauvergne_protocol.py script, we
simply ran the following command as you previously suggested:

$ relax --log log dauvergne_protocol.py


I would recommend, simply as a test, to run relax without logging:

$ relax dauvergne_protocol.py



We ran it twice identically except for the following selections:

1- With "res_name_col=3", "spin_num_col=4"
2- With "res_name_col-None", "spin_num_col=None"


I'm guessing that the minus is a typo here.  If you have the residue number,
the name is not essential.  And if you have a unique spin name, then a spin
number is not essential.



In both cases the results are identical and yield the same warnings:

1- We see "Over-fit spin deselection" when "grid_search" is starting.


From your log file, I can see that protons are deselected - this is normal
as you are looking at nitrogen relaxation rather than proton relaxation.



2- We always see the string "RelaxWarning: The spin '#2KB5_mol1:123@N' has
been deselected because of missing relaxation data".


How many data points do you have for this residue 123 nitrogen?



3- In "fitting to spin (Grid search)", "fk" values are identical in both
runs.
4- In "fitting to spin (Grid search)", the output message is the
following: "Storing the optimisation results, the optimised chi-squared
value is lower than the current value (12620.7348396 < None)". Is this what
one would expect to see?


This is normal.  This message makes more sense in the multiple-processor
mode whereby the grid search is broken into pieces and each processor
produces this message or the alternative saying that the chi-squared value
is higher.



Still, the major issue is the halting of relax in the middle of the tm8
grid search. I will try to install relax 2.1.1 on a 64-bit dual processor
later today and run everything again.


This is strange!  Maybe without the logging you will see more.  Otherwise
type 'dmesg' and see if the OS system log says more.  I really have only
seen such problems on failing hardware.  But as this occurs in the same
place each time, there must be some other problem!  Your ram size of 3 Gb
(https://gna.org/bugs/?20177) should be sufficient.  We need to get to the
bottom of this problem.

Regards,

Edward



Nick

--
Nicolas Doucet
Assistant Professor
INRS - Institut Armand-Frappier
University of Quebec
Institut Pasteur International Network
531 Boulevard des Prairies
Laval (Quebec) H7V 1B7 CANADA
Phone: (450) 687-5010 #4212
Fax: (450) 686-5501
Email: nicolas.doucet@xxxxxxxxxxx
Web: http://www.profs.inrs.ca/ndoucet/
--



-----Original Message-----
From: edward.dauvergne@xxxxxxxxx on behalf of Edward d'Auvergne
Sent: Fri 9/21/2012 4:35 AM
To: Doucet, Nicolas
Cc: relax-users@xxxxxxx
Subject: Re: Final model-free run with dauvergne_protocol.py

Hi Nick,

How did you run relax?  Could you copy and paste the entire command
used?  This will help me work out if you are using Gary Thompson's
multiprocessor and how IO redirection is occurring.  Do you get the
same messages without IO redirection?  It is quite peculiar that relax
just stops like that without giving any output.  I have seen this
before, but it was due to a dying CPU (or harddrive, I don't remember
exactly).  I would guess that without the multiprocessor or IO
redirection, you should see a much more informative error message.

Cheers,

Edward



On 20 September 2012 21:33, Nicolas Doucet <nicolas.doucet@xxxxxxxxxxx>
wrote:

Edward,

We started the dauvergne_protocol.py again from scratch and we do see a
few
warning messages in the initial local_tm run. Here is what we see right
before each tm grid search (the example is from the log file for tm0):

relax> pipe.copy(pipe_from='origin - mf (Tue Sep 18 14:19:13 2012)',
pipe_to='tm0 - mf (Tue Sep 18 14:19:13 2012)', bundle_to='mf (Tue Sep 18
14:19:13 2012)')

relax> pipe.switch(pipe_name='tm0 - mf (Tue Sep 18 14:19:13 2012)')

relax> model_free.select_model(model='tm0', spin_id=None)

relax> grid_search(lower=None, upper=None, inc=11, constraints=True,
verbosity=1)


Over-fit spin deselection.

RelaxWarning: The spin '#2KB5_mol1:2@N' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:3@N' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:4@H' has been deselected because of an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:4@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:5@H' has been deselected because of an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:5@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:6@H' has been deselected because of an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:6@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:7@H' has been deselected because of an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:7@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:8@H' has been deselected because of an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:8@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:9@H' has been deselected because of an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:9@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:10@H' has been deselected because of
an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:10@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:11@H' has been deselected because of
an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:11@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:12@H' has been deselected because of
an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:12@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:13@H' has been deselected because of
an
absence of relaxation mechanisms.
RelaxWarning: The spin '#2KB5_mol1:13@H' has been deselected because of
missing relaxation data.
RelaxWarning: The spin '#2KB5_mol1:14@H' has been deselected because of
an
absence of relaxation mechanisms.
RelaxWarning: T




_______________________________________________
relax (http://www.nmr-relax.com)

This is the relax-users mailing list
relax-users@xxxxxxx

To unsubscribe from this list, get a password
reminder, or change your subscription options,
visit the list information page at
https://mail.gna.org/listinfo/relax-users

Related Messages


Powered by MHonArc, Updated Mon Sep 24 23:20:08 2012