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