from __future__ import absolute_import, division, print_function import six from six.moves import range from six.moves import cStringIO as StringIO import math from libtbx import adopt_init_args from dials.array_family import flex from dxtbx.model.experiment_list import ExperimentList from cctbx import miller from cctbx.crystal import symmetry from scitbx import matrix from scitbx.math.tests.tst_weighted_correlation import simple_weighted_correlation from cctbx.crystal_orientation import crystal_orientation, basis_type from xfel.merging.application.postrefine.postrefinement_rs import postrefinement_rs, rs_refinery, rs_parameterization, lbfgs_minimizer_base def chosen_weights(observation_set, params): data = observation_set.data() sigmas = observation_set.sigmas() return { "unit": flex.double(len(data),1.), "variance": 1./(sigmas*sigmas), "gentle": flex.pow(flex.sqrt(flex.abs(data))/sigmas,2), "extreme": flex.pow(data/sigmas,2) } [ params.postrefinement.target_weighting ] class postrefinement_rs2(postrefinement_rs): def __init__(self, params, mpi_helper=None, mpi_logger=None): super(postrefinement_rs2, self).__init__(params=params, mpi_helper=mpi_helper, mpi_logger=mpi_logger) def __repr__(self): return 'Postrefinement' def run(self, experiments, reflections): self.logger.log_step_time("POSTREFINEMENT") if (not self.params.postrefinement.enable) or (self.params.scaling.algorithm != "mark0"): # mark1 implies no scaling/post-refinement self.logger.log("No post-refinement was done") if self.mpi_helper.rank == 0: self.logger.main_log("No post-refinement was done") return experiments, reflections target_symm = symmetry(unit_cell = self.params.scaling.unit_cell, space_group_info = self.params.scaling.space_group) i_model = self.params.scaling.i_model miller_set = self.params.scaling.miller_set # Ensure that match_multi_indices() will return identical results # when a frame's observations are matched against the # pre-generated Miller set, miller_set, and the reference # data set, i_model. The implication is that the same match # can be used to map Miller indices to array indices for intensity # accumulation, and for determination of the correlation # coefficient in the presence of a scaling reference. assert len(i_model.indices()) == len(miller_set.indices()) assert (i_model.indices() == miller_set.indices()).count(False) == 0 new_experiments = ExperimentList() new_reflections = flex.reflection_table() experiments_rejected_by_reason = {} # reason:how_many_rejected for expt_id, experiment in enumerate(experiments): exp_reflections = reflections.select(reflections['id'] == expt_id) # Build a miller array with _original_ miller indices of the experiment reflections exp_miller_indices_original = miller.set(target_symm, exp_reflections['miller_index'], not self.params.merging.merge_anomalous) observations_original_index = miller.array(exp_miller_indices_original, exp_reflections['intensity.sum.value'], flex.sqrt(exp_reflections['intensity.sum.variance'])) assert exp_reflections.size() == exp_miller_indices_original.size() assert observations_original_index.size() == exp_miller_indices_original.size() # Build a miller array with _asymmetric_ miller indices of the experiment reflections exp_miller_indices_asu = miller.set(target_symm, exp_reflections['miller_index_asymmetric'], True) observations = miller.array(exp_miller_indices_asu, exp_reflections['intensity.sum.value'], flex.sqrt(exp_reflections['intensity.sum.variance'])) matches = miller.match_multi_indices(miller_indices_unique = miller_set.indices(), miller_indices = observations.indices()) pair1 = flex.int([pair[1] for pair in matches.pairs()]) # refers to the observations pair0 = flex.int([pair[0] for pair in matches.pairs()]) # refers to the model # narrow things down to the set that matches, only observations_pair1_selected = observations.customized_copy(indices = flex.miller_index([observations.indices()[p] for p in pair1]), data = flex.double([observations.data()[p] for p in pair1]), sigmas = flex.double([observations.sigmas()[p] for p in pair1])) observations_original_index_pair1_selected = observations_original_index.customized_copy(indices = flex.miller_index([observations_original_index.indices()[p] for p in pair1]), data = flex.double([observations_original_index.data()[p] for p in pair1]), sigmas = flex.double([observations_original_index.sigmas()[p] for p in pair1])) ################### I_observed = observations_pair1_selected.data() chosen = chosen_weights(observations_pair1_selected, self.params) MILLER = observations_original_index_pair1_selected.indices() ORI = crystal_orientation(experiment.crystal.get_A(), basis_type.reciprocal) Astar = matrix.sqr(ORI.reciprocal_matrix()) Astar_from_experiment = matrix.sqr(experiment.crystal.get_A()) assert Astar == Astar_from_experiment WAVE = experiment.beam.get_wavelength() BEAM = matrix.col((0.0,0.0,-1./WAVE)) BFACTOR = 0. MOSAICITY_DEG = experiment.crystal.get_half_mosaicity_deg() DOMAIN_SIZE_A = experiment.crystal.get_domain_size_ang() # calculation of correlation here I_reference = flex.double([i_model.data()[pair[0]] for pair in matches.pairs()]) I_invalid = flex.bool([i_model.sigmas()[pair[0]] < 0. for pair in matches.pairs()]) use_weights = False # New facility for getting variance-weighted correlation if use_weights: # variance weighting I_weight = flex.double([1./(observations_pair1_selected.sigmas()[pair[1]])**2 for pair in matches.pairs()]) else: I_weight = flex.double(len(observations_pair1_selected.sigmas()), 1.) I_weight.set_selected(I_invalid, 0.) chosen.set_selected(I_invalid,0.) """Explanation of 'include_negatives' semantics as originally implemented in cxi.merge postrefinement: include_negatives = True + and - reflections both used for Rh distribution for initial estimate of RS parameter + and - reflections both used for calc/obs correlation slope for initial estimate of G parameter + and - reflections both passed to the refinery and used in the target function (makes sense if you look at it from a certain point of view) include_negatives = False + and - reflections both used for Rh distribution for initial estimate of RS parameter + reflections only used for calc/obs correlation slope for initial estimate of G parameter + and - reflections both passed to the refinery and used in the target function (makes sense if you look at it from a certain point of view) NOTE: by the new design, "include negatives" is always True """ SWC = simple_weighted_correlation(I_weight, I_reference, I_observed) if self.params.output.log_level == 0: self.logger.log("Old correlation is: %f"%SWC.corr) Rhall = flex.double() for mill in MILLER: H = matrix.col(mill) Xhkl = Astar*H Rh = ( Xhkl + BEAM ).length() - (1./WAVE) Rhall.append(Rh) Rs = math.sqrt(flex.mean(Rhall*Rhall)) RS = 1./10000. # reciprocal effective domain size of 1 micron RS = Rs # try this empirically determined approximate, monochrome, a-mosaic value current = flex.double([SWC.slope, BFACTOR, RS, 0., 0.]) parameterization_class = rs_parameterization refinery = rs2_refinery(ORI=ORI, MILLER=MILLER, BEAM=BEAM, WAVE=WAVE, ICALCVEC = I_reference, IOBSVEC = I_observed, WEIGHTS = chosen) refinery.set_profile_shape(self.params.postrefinement.lineshape) func = refinery.fvec_callable(parameterization_class(current)) functional = flex.sum(func * func) if self.params.output.log_level == 0: self.logger.log("functional: %f"%functional) self.current = current; self.parameterization_class = parameterization_class self.refinery = refinery; self.observations_pair1_selected = observations_pair1_selected; self.observations_original_index_pair1_selected = observations_original_index_pair1_selected error_detected = False try: self.run_plain() result_observations_original_index, result_observations, result_matches = self.result_for_cxi_merge() assert result_observations_original_index.size() == result_observations.size() assert result_matches.pairs().size() == result_observations_original_index.size() except (AssertionError, ValueError, RuntimeError) as e: error_detected = True reason = repr(e) if not reason: reason = "Unknown error" if not reason in experiments_rejected_by_reason: experiments_rejected_by_reason[reason] = 1 else: experiments_rejected_by_reason[reason] += 1 if not error_detected: new_experiments.append(experiment) new_exp_reflections = flex.reflection_table() new_exp_reflections['miller_index_asymmetric'] = result_observations.indices() new_exp_reflections['intensity.sum.value'] = result_observations.data() new_exp_reflections['intensity.sum.variance'] = flex.pow(result_observations.sigmas(),2) new_exp_reflections['id'] = flex.int(len(new_exp_reflections), len(new_experiments)-1) new_exp_reflections.experiment_identifiers()[len(new_experiments)-1] = experiment.identifier # The original reflection table, i.e. the input to this run() method, has more columns than those used # for the postrefinement ("data" and "sigma" in the miller arrays). The problems is: some of the input reflections may have been rejected by now. # So to bring those extra columns over to the new reflection table, we have to create a subset of the original exp_reflections table, # which would match (by original miller indices) the miller array results of the postrefinement. match_original_indices = miller.match_multi_indices(miller_indices_unique = exp_miller_indices_original.indices(), miller_indices = result_observations_original_index.indices()) exp_reflections_match_results = exp_reflections.select(match_original_indices.pairs().column(0)) assert (exp_reflections_match_results['intensity.sum.value'] == result_observations_original_index.data()).count(False) == 0 new_exp_reflections['intensity.sum.value.unmodified'] = exp_reflections_match_results['intensity.sum.value.unmodified'] new_exp_reflections['intensity.sum.variance.unmodified'] = exp_reflections_match_results['intensity.sum.variance.unmodified'] for key in self.params.input.persistent_refl_cols: if key not in new_exp_reflections.keys(): new_exp_reflections[key] = exp_reflections_match_results[key] new_reflections.extend(new_exp_reflections) # report rejected experiments, reflections experiments_rejected_by_postrefinement = len(experiments) - len(new_experiments) reflections_rejected_by_postrefinement = reflections.size() - new_reflections.size() self.logger.log("Experiments rejected by post-refinement: %d"%experiments_rejected_by_postrefinement) self.logger.log("Reflections rejected by post-refinement: %d"%reflections_rejected_by_postrefinement) all_reasons = [] for reason, count in six.iteritems(experiments_rejected_by_reason): self.logger.log("Experiments rejected due to %s: %d"%(reason,count)) all_reasons.append(reason) comm = self.mpi_helper.comm MPI = self.mpi_helper.MPI # Collect all rejection reasons from all ranks. Use allreduce to let each rank have all reasons. all_reasons = comm.allreduce(all_reasons, MPI.SUM) all_reasons = set(all_reasons) # Now that each rank has all reasons from all ranks, we can treat the reasons in a uniform way. total_experiments_rejected_by_reason = {} for reason in all_reasons: rejected_experiment_count = 0 if reason in experiments_rejected_by_reason: rejected_experiment_count = experiments_rejected_by_reason[reason] total_experiments_rejected_by_reason[reason] = comm.reduce(rejected_experiment_count, MPI.SUM, 0) total_accepted_experiment_count = comm.reduce(len(new_experiments), MPI.SUM, 0) # how many reflections have we rejected due to post-refinement? rejected_reflections = len(reflections) - len(new_reflections); total_rejected_reflections = self.mpi_helper.sum(rejected_reflections) if self.mpi_helper.rank == 0: for reason, count in six.iteritems(total_experiments_rejected_by_reason): self.logger.main_log("Total experiments rejected due to %s: %d"%(reason,count)) self.logger.main_log("Total experiments accepted: %d"%total_accepted_experiment_count) self.logger.main_log("Total reflections rejected due to post-refinement: %d"%total_rejected_reflections) self.logger.log_step_time("POSTREFINEMENT", True) # Do we have any data left? from xfel.merging.application.utils.data_counter import data_counter data_counter(self.params).count(new_experiments, new_reflections) return new_experiments, new_reflections def run_plain(self): out = StringIO() self.MINI = lbfgs_minimizer_derivatives(self.params, current_x = self.current, parameterization = self.parameterization_class, refinery = self.refinery, out = out) self.refined_mini = self.MINI if self.params.output.log_level == 0: self.logger.log("\n" + out.getvalue()) def rs2_parameter_range_assertions(self, values): # New range assertions for refined variables assert 0 < values.G, "G-scale value out of range ( < 0 ) after rs2 refinement" assert -25 < values.BFACTOR and values.BFACTOR < 25, "B-factor value out of range ( |B|>25 ) after rs2 refinement" assert -0.5<180.*values.thetax/math.pi<0.5,"thetax value out of range ( |rotx|>.5 degrees ) after rs2 refinement" assert -0.5<180.*values.thetay/math.pi<0.5,"thetay value out of range ( |roty|>.5 degrees ) after rs2 refinement" def result_for_cxi_merge(self): values = self.get_parameter_values() self.rs2_parameter_range_assertions(values) scaler = self.refinery.scaler_callable(self.parameterization_class(self.MINI.x)) partiality_array = self.refinery.get_partiality_array(values) p_scaler = flex.pow(partiality_array, 0.5*self.params.postrefinement.merge_partiality_exponent) fat_selection = (partiality_array > 0.2) fat_count = fat_selection.count(True) scaler_s = scaler.select(fat_selection) p_scaler_s = p_scaler.select(fat_selection) # reject an experiment with insufficient number of near-full reflections if fat_count < 3: if self.params.output.log_level == 0: self.logger.log("Rejected experiment, because: On total %5d the fat selection is %5d"%(len(self.observations_pair1_selected.indices()), fat_count)) raise ValueError("< 3 near-fulls after refinement") if self.params.output.log_level == 0: self.logger.log("On total %5d the fat selection is %5d"%(len(self.observations_pair1_selected.indices()), fat_count)) observations_original_index = self.observations_original_index_pair1_selected.select(fat_selection) observations = self.observations_pair1_selected.customized_copy( indices = self.observations_pair1_selected.indices().select(fat_selection), data = (self.observations_pair1_selected.data().select(fat_selection)/scaler_s), sigmas = (self.observations_pair1_selected.sigmas().select(fat_selection)/(scaler_s * p_scaler_s)) ) matches = miller.match_multi_indices( miller_indices_unique=self.params.scaling.miller_set.indices(), miller_indices=observations.indices()) I_weight = flex.double(len(observations.sigmas()), 1.) I_reference = flex.double([self.params.scaling.i_model.data()[pair[0]] for pair in matches.pairs()]) I_invalid = flex.bool([self.params.scaling.i_model.sigmas()[pair[0]] < 0. for pair in matches.pairs()]) I_weight.set_selected(I_invalid,0.) SWC = simple_weighted_correlation(I_weight, I_reference, observations.data()) if self.params.output.log_level == 0: self.logger.log("CORR: NEW correlation is: %f"%SWC.corr) self.logger.log("ASTAR: ") self.logger.log(tuple(self.refinery.get_eff_Astar(values))) self.final_corr = SWC.corr self.refined_mini = self.MINI #another range assertion assert self.final_corr > 0.1,"correlation coefficient out of range (<= 0.1) after rs2 refinement" return observations_original_index, observations, matches def get_parameter_values(self): return self.refined_mini.parameterization(self.refined_mini.x) class rs2_refinery(rs_refinery): def set_profile_shape(self, shape): self.profile_shape = shape self.get_partiality_array = { "lorentzian":super(rs2_refinery, self).get_partiality_array, "gaussian": self.get_gaussian_partiality_array }[shape] def get_gaussian_partiality_array(self,values): rs = values.RS Rh = self.get_Rh_array(values) immersion = Rh/rs gaussian = flex.exp(-2. * math.log(2) * (immersion*immersion)) return gaussian def jacobian_callable(self,values): PB = self.get_partiality_array(values) EXP = flex.exp(-2.*values.BFACTOR*self.DSSQ) G_terms = (EXP * PB * self.ICALCVEC) B_terms = (values.G * EXP * PB * self.ICALCVEC)*(-2.*self.DSSQ) P_terms = (values.G * EXP * self.ICALCVEC) thetax = values.thetax; thetay = values.thetay; Rx = matrix.col((1,0,0)).axis_and_angle_as_r3_rotation_matrix(thetax) dRx_dthetax = matrix.col((1,0,0)).axis_and_angle_as_r3_derivative_wrt_angle(thetax) Ry = matrix.col((0,1,0)).axis_and_angle_as_r3_rotation_matrix(thetay) dRy_dthetay = matrix.col((0,1,0)).axis_and_angle_as_r3_derivative_wrt_angle(thetay) ref_ori = matrix.sqr(self.ORI.reciprocal_matrix()) miller_vec = self.MILLER.as_vec3_double() ds1_dthetax = flex.mat3_double(len(self.MILLER),Ry * dRx_dthetax * ref_ori) * miller_vec ds1_dthetay = flex.mat3_double(len(self.MILLER),dRy_dthetay * Rx * ref_ori) * miller_vec s1vec = self.get_s1_array(values) s1lenvec = flex.sqrt(s1vec.dot(s1vec)) dRh_dthetax = s1vec.dot(ds1_dthetax)/s1lenvec dRh_dthetay = s1vec.dot(ds1_dthetay)/s1lenvec rs = values.RS Rh = self.get_Rh_array(values) rs_sq = rs*rs denomin = (2. * Rh * Rh + rs_sq) dPB_dRh = { "lorentzian": -PB * 4. * Rh / denomin, "gaussian": -PB * 4. * math.log(2) * Rh / rs_sq }[self.profile_shape] dPB_dthetax = dPB_dRh * dRh_dthetax dPB_dthetay = dPB_dRh * dRh_dthetay Px_terms = P_terms * dPB_dthetax; Py_terms = P_terms * dPB_dthetay return [G_terms,B_terms,0,Px_terms,Py_terms] class lbfgs_minimizer_derivatives(lbfgs_minimizer_base): def __init__(self, params, current_x=None, parameterization=None, refinery=None, out=None, min_iterations=0, max_calls=1000, max_drop_eps=1.e-5): adopt_init_args(self, locals()) self.n = current_x.size() self.x = current_x from scitbx import lbfgs self.minimizer = lbfgs.run( target_evaluator=self, termination_params=lbfgs.termination_parameters( traditional_convergence_test=True, drop_convergence_test_max_drop_eps=max_drop_eps, min_iterations=min_iterations, max_iterations = None, max_calls=max_calls), exception_handling_params=lbfgs.exception_handling_parameters( ignore_line_search_failed_rounding_errors=True, ignore_line_search_failed_step_at_lower_bound=True,#the only change from default ignore_line_search_failed_step_at_upper_bound=False, ignore_line_search_failed_maxfev=False, ignore_line_search_failed_xtol=False, ignore_search_direction_not_descent=False) ) def compute_functional_and_gradients(self): values = self.parameterization(self.x) assert -150. < values.BFACTOR < 150,"B-factor out of range (+/-150) within rs2 functional and gradients" self.func = self.refinery.fvec_callable(values) functional = flex.sum(self.refinery.WEIGHTS*self.func*self.func) self.f = functional jacobian = self.refinery.jacobian_callable(values) self.g = flex.double(self.n) for ix in range(self.n): self.g[ix] = flex.sum(2. * self.refinery.WEIGHTS * self.func * jacobian[ix]) print("rms %10.3f"%math.sqrt(flex.sum(self.refinery.WEIGHTS*self.func*self.func)/flex.sum(self.refinery.WEIGHTS)), file=self.out, end='') values.show(self.out) return self.f, self.g if __name__ == '__main__': from xfel.merging.application.worker import exercise_worker exercise_worker(postrefinement_rs2)