Re: [bug #22146] Unpacking of R2A and R2B is performed wrong for clustered "full" dispersion modelsRe: [bug #22146] Unpacking of R2A and R2B is performed wrong for clustered "full" dispersion modelsHi,
If this is where the problem is, then I would recommend a full system
test where you perform a grid search where the inc value for all
dimensions is 1 and the upper and lower bounds are equal and set to a
set of parameter values so that a single grid increment is performed,
and then the optimised parameter values matching the lower and upper
bound arguments are unpacked into the spin containers and can be
checked.
As for a separate function for unpacking R20A and R20B in the target
functions, this would normally be a good idea. But as calling
functions has an overhead, for speed it might be better to leave this
as a two line change:
# Unpack the parameter values.
R20A = params[:self.end_index[0]]
R20B = params[self.end_index[0]:self.end_index[1]]
to:
# Unpack the parameter values.
R20 = params[:self.end_index[1]]
R20A, R20B = transpose(R20.reshape(self.num_spins, 2))
This will be sufficient. It would be good to have a system test to
check the before and after implementation. This system test could
then be replicated for all the different 'full' dispersion models.
Then this problem will never reappear.
Cheers,
Edward
On 6 June 2014 11:39, Troels Emtekær Linnet <tlinnet@xxxxxxxxxxxxx> wrote:
There is no doubt that it is the unpacking of the R20A and R20B in the target function. I was thinking of creating a function, which do the the unpacking. This unpacking function could then be tested with a unit test? What do you think? Where should I position such a function? Best Troels 2014-06-06 11:26 GMT+02:00 Edward d'Auvergne <edward@xxxxxxxxxxxxx>:Hi Troels, The best way to handle this is to first create a unit test of the specific_analyses.relax_disp.parameters.disassemble_param_vector() where the problem is likely to be most easily found. I don't understand how this could be a problem as the assemble_param_vector() and disassemble_param_vector() functions both call the same loop_parameters() function for the ordering of the parameter values! Maybe the problem is in the unpacking of the parameter vector in the target functions themselves, for example in the full B14 model: def func_B14_full(self, params): """Target function for the Baldwin (2014) 2-site exact solution model for all time scales. This assumes that pA > pB, and hence this must be implemented as a constraint. @param params: The vector of parameter values. @type params: numpy rank-1 float array @return: The chi-squared value. @rtype: float """ # Scaling. if self.scaling_flag: params = dot(params, self.scaling_matrix) # Unpack the parameter values. R20A = params[:self.end_index[0]] R20B = params[self.end_index[0]:self.end_index[1]] dw = params[self.end_index[1]:self.end_index[2]] pA = params[self.end_index[2]] kex = params[self.end_index[2]+1] # Calculate and return the chi-squared value. return self.calc_B14_chi2(R20A=R20A, R20B=R20B, dw=dw, pA=pA, kex=kex) This R20A and R20B unpacking might be the failure point as this may not match the loop_parameters() function - which it must! In any case, having a unit or system test catch the problem would be very useful for the stability of the dispersion analysis in relax. A code example might be useful: R20_params = array([1, 2, 3, 4]) R20A, R20B = transpose(R20_params.reshape(2, 2) print(R20A) print(R20B) You should see that R20A is [1, 3], and R20B is [2, 4]. This is how the parameters are handled in the loop_parameters() function which defines the parameter vector in all parts of relax. There might be a quicker way to unpack the parameters, but such an idea could be used for the target functions. Cheers, Edward On 6 June 2014 11:08, Troels E. Linnet <NO-REPLY.INVALID-ADDRESS@xxxxxxx> wrote:URL: <http://gna.org/bugs/?22146> Summary: Unpacking of R2A and R2B is performed wrong for clustered "full" dispersion models Project: relax Submitted by: tlinnet Submitted on: Fri 06 Jun 2014 09:08:58 AM UTC Category: relax's source code Specific analysis category: Relaxation dispersion Priority: 9 - Immediate Severity: 4 - Important Status: None Assigned to: None Originator Name: Originator Email: Open/Closed: Open Release: Repository: trunk Discussion Lock: Any Operating System: All systems _______________________________________________________ Details: The unpacking of the R2A and R2B parameters in the "full" model is performed wrong. This will happen performing a clustered analysis, using one of the "full" models. This bug affect all analysis performed running with a "full" model, with clustered residues. The bug is located in the target function: ./target_functions/relax_disp.py For all the "func_MODEL_full", the unpacking of: R20A = params[:self.end_index[0]] R20B = params[self.end_index[0]:self.end_index[1]] This is wrong, since the "params" list, is ordered: [spin, spin, spin, [dw], pA, kex], where spin = [nr_frq*r2a, nr_frq*r2b] This ordering happens in: ./specific_analysis/relax_disp/parameters.py in the loop_parameters.py A possible solutions i shown below. This alter the unpacking of the parameters. An example of profiling_cr72.py is attached. This can be downloaded, and run in base folder of relax: ./profiling_cr72.py . This is with 3 frq, and 3 spins. The current implementations would unpack: ('R20A', array([ 2., 2., 2., 4., 4., 4., 12., 12., 12.]), 9) ('R20B', array([ 14., 14., 14., 22., 22., 22., 24., 24., 24.]), 9) R2A is 2, 12, 22 for the spins 0-3 R2B is, 4, 14, 24 for the spins 0-3 The suggested unpacking loop, unpacks to: ('R20A', array([ 2., 2., 2., 12., 12., 12., 22., 22., 22.]), 9) ('R20B', array([ 4., 4., 4., 14., 14., 14., 24., 24., 24.]), 9) ------- from numpy import array, concatenate, delete, index_exp import numpy p = array([ 1.000000000000000e+01, 1.000000000000000e+01, 1.100000000000000e+01 , 1.100000000000000e+01, 1.000000000000000e+01, 1.000000000000000e+01 , 1.100000000000000e+01, 1.100000000000000e+01, 1.000000000000000e+00 , 1.000000000000000e+00, 9.000000000000000e-01, 1.000000000000000e+03]) e = [4, 8, 10] # Now r2a = p[:e[0]] print r2a r2b = p[e[0]:e[1]] print r2b dw = p[e[1]:e[2]] print dw pA = p[e[2]] print pA kex = p[e[2]+1] print kex print "new" ns = 2 nf = 2 ml = p[:e[1]] R20A = array([]) R20B = array([]) for i in range(0, ns): # Array sorted per [spin, spin, spin], where spin = [nr_frq*r2a, nr_frq*r2b] spin_AB = ml[:nf*2] ml = delete(ml, numpy.s_[:nf*2]) R20A = concatenate([R20A, spin_AB[:nf] ]) R20B = concatenate([R20B, spin_AB[nf:] ]) print R20A print R20B print dw print pA print kex _______________________________________________________ File Attachments: ------------------------------------------------------- Date: Fri 06 Jun 2014 09:08:58 AM UTC Name: profiling_cr72.py Size: 17kB By: tlinnet <http://gna.org/bugs/download.php?file_id=20938> _______________________________________________________ Reply to this item at: <http://gna.org/bugs/?22146> _______________________________________________ Message sent via/by Gna! http://gna.org/ _______________________________________________ relax (http://www.nmr-relax.com) This is the relax-devel mailing list relax-devel@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-devel