[6fe5100] | 1 | """ |
---|
| 2 | BumpsFitting module runs the bumps optimizer. |
---|
| 3 | """ |
---|
[249a7c6] | 4 | import os |
---|
[35086c3] | 5 | from datetime import timedelta, datetime |
---|
| 6 | |
---|
[6fe5100] | 7 | import numpy |
---|
| 8 | |
---|
| 9 | from bumps import fitters |
---|
[7945367] | 10 | try: |
---|
| 11 | from bumps.options import FIT_CONFIG |
---|
| 12 | # Default bumps to use the Levenberg-Marquardt optimizer |
---|
| 13 | FIT_CONFIG.selected_id = fitters.LevenbergMarquardtFit.id |
---|
| 14 | def get_fitter(): |
---|
| 15 | return FIT_CONFIG.selected_fitter, FIT_CONFIG.selected_values |
---|
| 16 | except: |
---|
| 17 | # CRUFT: Bumps changed its handling of fit options around 0.7.5.6 |
---|
| 18 | # Default bumps to use the Levenberg-Marquardt optimizer |
---|
| 19 | fitters.FIT_DEFAULT = 'lm' |
---|
| 20 | def get_fitter(): |
---|
| 21 | fitopts = fitters.FIT_OPTIONS[fitters.FIT_DEFAULT] |
---|
| 22 | return fitopts.fitclass, fitopts.options.copy() |
---|
| 23 | |
---|
| 24 | |
---|
[249a7c6] | 25 | from bumps.mapper import SerialMapper, MPMapper |
---|
[e3efa6b3] | 26 | from bumps import parameter |
---|
| 27 | from bumps.fitproblem import FitProblem |
---|
[7945367] | 28 | |
---|
[6fe5100] | 29 | |
---|
[fd5ac0d] | 30 | from sas.fit.AbstractFitEngine import FitEngine |
---|
| 31 | from sas.fit.AbstractFitEngine import FResult |
---|
| 32 | from sas.fit.expression import compile_constraints |
---|
[6fe5100] | 33 | |
---|
[35086c3] | 34 | class Progress(object): |
---|
| 35 | def __init__(self, history, max_step, pars, dof): |
---|
| 36 | remaining_time = int(history.time[0]*(float(max_step)/history.step[0]-1)) |
---|
| 37 | # Depending on the time remaining, either display the expected |
---|
| 38 | # time of completion, or the amount of time remaining. Use precision |
---|
| 39 | # appropriate for the duration. |
---|
| 40 | if remaining_time >= 1800: |
---|
| 41 | completion_time = datetime.now() + timedelta(seconds=remaining_time) |
---|
| 42 | if remaining_time >= 36000: |
---|
| 43 | time = completion_time.strftime('%Y-%m-%d %H:%M') |
---|
| 44 | else: |
---|
| 45 | time = completion_time.strftime('%H:%M') |
---|
| 46 | else: |
---|
| 47 | if remaining_time >= 3600: |
---|
| 48 | time = '%dh %dm'%(remaining_time//3600, (remaining_time%3600)//60) |
---|
| 49 | elif remaining_time >= 60: |
---|
| 50 | time = '%dm %ds'%(remaining_time//60, remaining_time%60) |
---|
| 51 | else: |
---|
| 52 | time = '%ds'%remaining_time |
---|
| 53 | chisq = "%.3g"%(2*history.value[0]/dof) |
---|
| 54 | step = "%d of %d"%(history.step[0], max_step) |
---|
| 55 | header = "=== Steps: %s chisq: %s ETA: %s\n"%(step, chisq, time) |
---|
| 56 | parameters = ["%15s: %-10.3g%s"%(k,v,("\n" if i%3==2 else " | ")) |
---|
| 57 | for i,(k,v) in enumerate(zip(pars,history.point[0]))] |
---|
| 58 | self.msg = "".join([header]+parameters) |
---|
| 59 | |
---|
| 60 | def __str__(self): |
---|
| 61 | return self.msg |
---|
| 62 | |
---|
| 63 | |
---|
[85f17f6] | 64 | class BumpsMonitor(object): |
---|
[35086c3] | 65 | def __init__(self, handler, max_step, pars, dof): |
---|
[85f17f6] | 66 | self.handler = handler |
---|
| 67 | self.max_step = max_step |
---|
[35086c3] | 68 | self.pars = pars |
---|
| 69 | self.dof = dof |
---|
[ed4aef2] | 70 | |
---|
[85f17f6] | 71 | def config_history(self, history): |
---|
| 72 | history.requires(time=1, value=2, point=1, step=1) |
---|
[ed4aef2] | 73 | |
---|
[85f17f6] | 74 | def __call__(self, history): |
---|
[e3efa6b3] | 75 | if self.handler is None: return |
---|
[35086c3] | 76 | self.handler.set_result(Progress(history, self.max_step, self.pars, self.dof)) |
---|
[85f17f6] | 77 | self.handler.progress(history.step[0], self.max_step) |
---|
| 78 | if len(history.step)>1 and history.step[1] > history.step[0]: |
---|
| 79 | self.handler.improvement() |
---|
| 80 | self.handler.update_fit() |
---|
| 81 | |
---|
[ed4aef2] | 82 | class ConvergenceMonitor(object): |
---|
| 83 | """ |
---|
| 84 | ConvergenceMonitor contains population summary statistics to show progress |
---|
| 85 | of the fit. This is a list [ (best, 0%, 25%, 50%, 75%, 100%) ] or |
---|
| 86 | just a list [ (best, ) ] if population size is 1. |
---|
| 87 | """ |
---|
| 88 | def __init__(self): |
---|
| 89 | self.convergence = [] |
---|
| 90 | |
---|
| 91 | def config_history(self, history): |
---|
| 92 | history.requires(value=1, population_values=1) |
---|
| 93 | |
---|
| 94 | def __call__(self, history): |
---|
| 95 | best = history.value[0] |
---|
| 96 | try: |
---|
| 97 | p = history.population_values[0] |
---|
| 98 | n,p = len(p), numpy.sort(p) |
---|
| 99 | QI,Qmid, = int(0.2*n),int(0.5*n) |
---|
| 100 | self.convergence.append((best, p[0],p[QI],p[Qmid],p[-1-QI],p[-1])) |
---|
| 101 | except: |
---|
[e3efa6b3] | 102 | self.convergence.append((best, best,best,best,best,best)) |
---|
[ed4aef2] | 103 | |
---|
[e3efa6b3] | 104 | |
---|
[4e9f227] | 105 | # Note: currently using bumps parameters for each parameter object so that |
---|
| 106 | # a SasFitness can be used directly in bumps with the usual semantics. |
---|
| 107 | # The disadvantage of this technique is that we need to copy every parameter |
---|
| 108 | # back into the model each time the function is evaluated. We could instead |
---|
[fd5ac0d] | 109 | # define reference parameters for each sas parameter, but then we would not |
---|
[4e9f227] | 110 | # be able to express constraints using python expressions in the usual way |
---|
| 111 | # from bumps, and would instead need to use string expressions. |
---|
[e3efa6b3] | 112 | class SasFitness(object): |
---|
[6fe5100] | 113 | """ |
---|
[e3efa6b3] | 114 | Wrap SAS model as a bumps fitness object |
---|
[6fe5100] | 115 | """ |
---|
[5044543] | 116 | def __init__(self, model, data, fitted=[], constraints={}, |
---|
| 117 | initial_values=None, **kw): |
---|
[4e9f227] | 118 | self.name = model.name |
---|
| 119 | self.model = model.model |
---|
[6fe5100] | 120 | self.data = data |
---|
[9f7fbd9] | 121 | if self.data.smearer is not None: |
---|
| 122 | self.data.smearer.model = self.model |
---|
[e3efa6b3] | 123 | self._define_pars() |
---|
| 124 | self._init_pars(kw) |
---|
[5044543] | 125 | if initial_values is not None: |
---|
| 126 | self._reset_pars(fitted, initial_values) |
---|
[4e9f227] | 127 | self.constraints = dict(constraints) |
---|
[e3efa6b3] | 128 | self.set_fitted(fitted) |
---|
[5044543] | 129 | self.update() |
---|
| 130 | |
---|
| 131 | def _reset_pars(self, names, values): |
---|
| 132 | for k,v in zip(names, values): |
---|
| 133 | self._pars[k].value = v |
---|
[e3efa6b3] | 134 | |
---|
| 135 | def _define_pars(self): |
---|
| 136 | self._pars = {} |
---|
| 137 | for k in self.model.getParamList(): |
---|
| 138 | name = ".".join((self.name,k)) |
---|
| 139 | value = self.model.getParam(k) |
---|
| 140 | bounds = self.model.details.get(k,["",None,None])[1:3] |
---|
| 141 | self._pars[k] = parameter.Parameter(value=value, bounds=bounds, |
---|
| 142 | fixed=True, name=name) |
---|
[4e9f227] | 143 | #print parameter.summarize(self._pars.values()) |
---|
[e3efa6b3] | 144 | |
---|
| 145 | def _init_pars(self, kw): |
---|
| 146 | for k,v in kw.items(): |
---|
| 147 | # dispersion parameters initialized with _field instead of .field |
---|
| 148 | if k.endswith('_width'): k = k[:-6]+'.width' |
---|
| 149 | elif k.endswith('_npts'): k = k[:-5]+'.npts' |
---|
| 150 | elif k.endswith('_nsigmas'): k = k[:-7]+'.nsigmas' |
---|
| 151 | elif k.endswith('_type'): k = k[:-5]+'.type' |
---|
| 152 | if k not in self._pars: |
---|
| 153 | formatted_pars = ", ".join(sorted(self._pars.keys())) |
---|
| 154 | raise KeyError("invalid parameter %r for %s--use one of: %s" |
---|
| 155 | %(k, self.model, formatted_pars)) |
---|
| 156 | if '.' in k and not k.endswith('.width'): |
---|
| 157 | self.model.setParam(k, v) |
---|
| 158 | elif isinstance(v, parameter.BaseParameter): |
---|
| 159 | self._pars[k] = v |
---|
| 160 | elif isinstance(v, (tuple,list)): |
---|
| 161 | low, high = v |
---|
| 162 | self._pars[k].value = (low+high)/2 |
---|
| 163 | self._pars[k].range(low,high) |
---|
[95d58d3] | 164 | else: |
---|
[e3efa6b3] | 165 | self._pars[k].value = v |
---|
| 166 | |
---|
| 167 | def set_fitted(self, param_list): |
---|
[6fe5100] | 168 | """ |
---|
[e3efa6b3] | 169 | Flag a set of parameters as fitted parameters. |
---|
[6fe5100] | 170 | """ |
---|
[e3efa6b3] | 171 | for k,p in self._pars.items(): |
---|
[4e9f227] | 172 | p.fixed = (k not in param_list or k in self.constraints) |
---|
[4a0dc427] | 173 | self.fitted_par_names = [k for k in param_list if k not in self.constraints] |
---|
[bf5e985] | 174 | self.computed_par_names = [k for k in param_list if k in self.constraints] |
---|
| 175 | self.fitted_pars = [self._pars[k] for k in self.fitted_par_names] |
---|
| 176 | self.computed_pars = [self._pars[k] for k in self.computed_par_names] |
---|
[6fe5100] | 177 | |
---|
[e3efa6b3] | 178 | # ===== Fitness interface ==== |
---|
| 179 | def parameters(self): |
---|
| 180 | return self._pars |
---|
[6fe5100] | 181 | |
---|
[e3efa6b3] | 182 | def update(self): |
---|
| 183 | for k,v in self._pars.items(): |
---|
[4e9f227] | 184 | #print "updating",k,v,v.value |
---|
[e3efa6b3] | 185 | self.model.setParam(k,v.value) |
---|
| 186 | self._dirty = True |
---|
[6fe5100] | 187 | |
---|
[e3efa6b3] | 188 | def _recalculate(self): |
---|
| 189 | if self._dirty: |
---|
[9f7fbd9] | 190 | self._residuals, self._theory \ |
---|
| 191 | = self.data.residuals(self.model.evalDistribution) |
---|
[e3efa6b3] | 192 | self._dirty = False |
---|
[6fe5100] | 193 | |
---|
[e3efa6b3] | 194 | def numpoints(self): |
---|
| 195 | return numpy.sum(self.data.idx) # number of fitted points |
---|
[6fe5100] | 196 | |
---|
[e3efa6b3] | 197 | def nllf(self): |
---|
| 198 | return 0.5*numpy.sum(self.residuals()**2) |
---|
| 199 | |
---|
| 200 | def theory(self): |
---|
| 201 | self._recalculate() |
---|
| 202 | return self._theory |
---|
| 203 | |
---|
| 204 | def residuals(self): |
---|
| 205 | self._recalculate() |
---|
| 206 | return self._residuals |
---|
| 207 | |
---|
| 208 | # Not implementing the data methods for now: |
---|
| 209 | # |
---|
| 210 | # resynth_data/restore_data/save/plot |
---|
[6fe5100] | 211 | |
---|
[191c648] | 212 | class ParameterExpressions(object): |
---|
| 213 | def __init__(self, models): |
---|
| 214 | self.models = models |
---|
| 215 | self._setup() |
---|
| 216 | |
---|
| 217 | def _setup(self): |
---|
| 218 | exprs = {} |
---|
| 219 | for M in self.models: |
---|
| 220 | exprs.update((".".join((M.name, k)), v) for k, v in M.constraints.items()) |
---|
| 221 | if exprs: |
---|
| 222 | symtab = dict((".".join((M.name, k)), p) |
---|
| 223 | for M in self.models |
---|
| 224 | for k,p in M.parameters().items()) |
---|
| 225 | self.update = compile_constraints(symtab, exprs) |
---|
| 226 | else: |
---|
| 227 | self.update = lambda: 0 |
---|
| 228 | |
---|
| 229 | def __call__(self): |
---|
| 230 | self.update() |
---|
| 231 | |
---|
| 232 | def __getstate__(self): |
---|
| 233 | return self.models |
---|
| 234 | |
---|
| 235 | def __setstate__(self, state): |
---|
| 236 | self.models = state |
---|
| 237 | self._setup() |
---|
| 238 | |
---|
[6fe5100] | 239 | class BumpsFit(FitEngine): |
---|
| 240 | """ |
---|
| 241 | Fit a model using bumps. |
---|
| 242 | """ |
---|
| 243 | def __init__(self): |
---|
| 244 | """ |
---|
| 245 | Creates a dictionary (self.fit_arrange_dict={})of FitArrange elements |
---|
| 246 | with Uid as keys |
---|
| 247 | """ |
---|
| 248 | FitEngine.__init__(self) |
---|
| 249 | self.curr_thread = None |
---|
| 250 | |
---|
| 251 | def fit(self, msg_q=None, |
---|
| 252 | q=None, handler=None, curr_thread=None, |
---|
| 253 | ftol=1.49012e-8, reset_flag=False): |
---|
[e3efa6b3] | 254 | # Build collection of bumps fitness calculators |
---|
[bf5e985] | 255 | models = [SasFitness(model=M.get_model(), |
---|
| 256 | data=M.get_data(), |
---|
| 257 | constraints=M.constraints, |
---|
[5044543] | 258 | fitted=M.pars, |
---|
| 259 | initial_values=M.vals if reset_flag else None) |
---|
[bf5e985] | 260 | for M in self.fit_arrange_dict.values() |
---|
| 261 | if M.get_to_fit()] |
---|
[233c121] | 262 | if len(models) == 0: |
---|
| 263 | raise RuntimeError("Nothing to fit") |
---|
[e3efa6b3] | 264 | problem = FitProblem(models) |
---|
| 265 | |
---|
[191c648] | 266 | # TODO: need better handling of parameter expressions and bounds constraints |
---|
| 267 | # so that they are applied during polydispersity calculations. This |
---|
| 268 | # will remove the immediate need for the setp_hook in bumps, though |
---|
| 269 | # bumps may still need something similar, such as a sane class structure |
---|
| 270 | # which allows a subclass to override setp. |
---|
| 271 | problem.setp_hook = ParameterExpressions(models) |
---|
[4e9f227] | 272 | |
---|
[e3efa6b3] | 273 | # Run the fit |
---|
| 274 | result = run_bumps(problem, handler, curr_thread) |
---|
[6fe5100] | 275 | if handler is not None: |
---|
| 276 | handler.update_fit(last=True) |
---|
[e3efa6b3] | 277 | |
---|
[eff93b8] | 278 | # TODO: shouldn't reference internal parameters of fit problem |
---|
[e3efa6b3] | 279 | varying = problem._parameters |
---|
| 280 | # collect the results |
---|
| 281 | all_results = [] |
---|
| 282 | for M in problem.models: |
---|
| 283 | fitness = M.fitness |
---|
| 284 | fitted_index = [varying.index(p) for p in fitness.fitted_pars] |
---|
[e1442d4] | 285 | param_list = fitness.fitted_par_names + fitness.computed_par_names |
---|
[e3efa6b3] | 286 | R = FResult(model=fitness.model, data=fitness.data, |
---|
[e1442d4] | 287 | param_list=param_list) |
---|
[e3efa6b3] | 288 | R.theory = fitness.theory() |
---|
| 289 | R.residuals = fitness.residuals() |
---|
[5044543] | 290 | R.index = fitness.data.idx |
---|
[e3efa6b3] | 291 | R.fitter_id = self.fitter_id |
---|
[eff93b8] | 292 | # TODO: should scale stderr by sqrt(chisq/DOF) if dy is unknown |
---|
[e3efa6b3] | 293 | R.success = result['success'] |
---|
[e1442d4] | 294 | if R.success: |
---|
| 295 | R.stderr = numpy.hstack((result['stderr'][fitted_index], |
---|
| 296 | numpy.NaN*numpy.ones(len(fitness.computed_pars)))) |
---|
| 297 | R.pvec = numpy.hstack((result['value'][fitted_index], |
---|
| 298 | [p.value for p in fitness.computed_pars])) |
---|
| 299 | R.fitness = numpy.sum(R.residuals**2)/(fitness.numpoints() - len(fitted_index)) |
---|
| 300 | else: |
---|
| 301 | R.stderr = numpy.NaN*numpy.ones(len(param_list)) |
---|
| 302 | R.pvec = numpy.asarray( [p.value for p in fitness.fitted_pars+fitness.computed_pars]) |
---|
| 303 | R.fitness = numpy.NaN |
---|
[e3efa6b3] | 304 | R.convergence = result['convergence'] |
---|
| 305 | if result['uncertainty'] is not None: |
---|
| 306 | R.uncertainty_state = result['uncertainty'] |
---|
| 307 | all_results.append(R) |
---|
| 308 | |
---|
[6fe5100] | 309 | if q is not None: |
---|
[e3efa6b3] | 310 | q.put(all_results) |
---|
[6fe5100] | 311 | return q |
---|
[e3efa6b3] | 312 | else: |
---|
| 313 | return all_results |
---|
[6fe5100] | 314 | |
---|
[e3efa6b3] | 315 | def run_bumps(problem, handler, curr_thread): |
---|
[85f17f6] | 316 | def abort_test(): |
---|
| 317 | if curr_thread is None: return False |
---|
| 318 | try: curr_thread.isquit() |
---|
| 319 | except KeyboardInterrupt: |
---|
| 320 | if handler is not None: |
---|
| 321 | handler.stop("Fitting: Terminated!!!") |
---|
| 322 | return True |
---|
| 323 | return False |
---|
| 324 | |
---|
[7945367] | 325 | fitclass, options = get_fitter() |
---|
| 326 | steps = options.get('steps', 0) |
---|
| 327 | if steps == 0: |
---|
| 328 | pop = options.get('pop',0)*len(problem._parameters) |
---|
| 329 | samples = options.get('samples', 0) |
---|
| 330 | steps = (samples+pop-1)/pop if pop != 0 else samples |
---|
| 331 | max_step = steps + options.get('burn', 0) |
---|
[35086c3] | 332 | pars = [p.name for p in problem._parameters] |
---|
[e1442d4] | 333 | #x0 = numpy.asarray([p.value for p in problem._parameters]) |
---|
[e3efa6b3] | 334 | options['monitors'] = [ |
---|
[35086c3] | 335 | BumpsMonitor(handler, max_step, pars, problem.dof), |
---|
[e3efa6b3] | 336 | ConvergenceMonitor(), |
---|
| 337 | ] |
---|
[95d58d3] | 338 | fitdriver = fitters.FitDriver(fitclass, problem=problem, |
---|
[042f065] | 339 | abort_test=abort_test, **options) |
---|
[249a7c6] | 340 | omp_threads = int(os.environ.get('OMP_NUM_THREADS','0')) |
---|
[e1442d4] | 341 | mapper = MPMapper if omp_threads == 1 else SerialMapper |
---|
[6fe5100] | 342 | fitdriver.mapper = mapper.start_mapper(problem, None) |
---|
[233c121] | 343 | #import time; T0 = time.time() |
---|
[6fe5100] | 344 | try: |
---|
| 345 | best, fbest = fitdriver.fit() |
---|
| 346 | except: |
---|
| 347 | import traceback; traceback.print_exc() |
---|
| 348 | raise |
---|
[95d58d3] | 349 | finally: |
---|
| 350 | mapper.stop_mapper(fitdriver.mapper) |
---|
[e3efa6b3] | 351 | |
---|
| 352 | |
---|
| 353 | convergence_list = options['monitors'][-1].convergence |
---|
| 354 | convergence = (2*numpy.asarray(convergence_list)/problem.dof |
---|
| 355 | if convergence_list else numpy.empty((0,1),'d')) |
---|
[e1442d4] | 356 | |
---|
| 357 | success = best is not None |
---|
[e3efa6b3] | 358 | return { |
---|
[e1442d4] | 359 | 'value': best if success else None, |
---|
| 360 | 'stderr': fitdriver.stderr() if success else None, |
---|
| 361 | 'success': success, |
---|
[e3efa6b3] | 362 | 'convergence': convergence, |
---|
| 363 | 'uncertainty': getattr(fitdriver.fitter, 'state', None), |
---|
| 364 | } |
---|
[6fe5100] | 365 | |
---|