1 | """ |
---|
2 | BumpsFitting module runs the bumps optimizer. |
---|
3 | """ |
---|
4 | import time |
---|
5 | |
---|
6 | import numpy |
---|
7 | |
---|
8 | from bumps import fitters |
---|
9 | from bumps.mapper import SerialMapper |
---|
10 | |
---|
11 | from sans.fit.AbstractFitEngine import FitEngine |
---|
12 | from sans.fit.AbstractFitEngine import FResult |
---|
13 | |
---|
14 | class BumpsMonitor(object): |
---|
15 | def __init__(self, handler, max_step=0): |
---|
16 | self.handler = handler |
---|
17 | self.max_step = max_step |
---|
18 | |
---|
19 | def config_history(self, history): |
---|
20 | history.requires(time=1, value=2, point=1, step=1) |
---|
21 | |
---|
22 | def __call__(self, history): |
---|
23 | self.handler.progress(history.step[0], self.max_step) |
---|
24 | if len(history.step)>1 and history.step[1] > history.step[0]: |
---|
25 | self.handler.improvement() |
---|
26 | self.handler.update_fit() |
---|
27 | |
---|
28 | class ConvergenceMonitor(object): |
---|
29 | """ |
---|
30 | ConvergenceMonitor contains population summary statistics to show progress |
---|
31 | of the fit. This is a list [ (best, 0%, 25%, 50%, 75%, 100%) ] or |
---|
32 | just a list [ (best, ) ] if population size is 1. |
---|
33 | """ |
---|
34 | def __init__(self): |
---|
35 | self.convergence = [] |
---|
36 | |
---|
37 | def config_history(self, history): |
---|
38 | history.requires(value=1, population_values=1) |
---|
39 | |
---|
40 | def __call__(self, history): |
---|
41 | best = history.value[0] |
---|
42 | try: |
---|
43 | p = history.population_values[0] |
---|
44 | n,p = len(p), numpy.sort(p) |
---|
45 | QI,Qmid, = int(0.2*n),int(0.5*n) |
---|
46 | self.convergence.append((best, p[0],p[QI],p[Qmid],p[-1-QI],p[-1])) |
---|
47 | except: |
---|
48 | self.convergence.append((best, )) |
---|
49 | |
---|
50 | class SasProblem(object): |
---|
51 | """ |
---|
52 | Wrap the SAS model in a form that can be understood by bumps. |
---|
53 | """ |
---|
54 | def __init__(self, param_list, model=None, data=None, fitresult=None, |
---|
55 | handler=None, curr_thread=None, msg_q=None): |
---|
56 | """ |
---|
57 | :param Model: the model wrapper fro sans -model |
---|
58 | :param Data: the data wrapper for sans data |
---|
59 | """ |
---|
60 | self.model = model |
---|
61 | self.data = data |
---|
62 | self.param_list = param_list |
---|
63 | self.res = None |
---|
64 | self.theory = None |
---|
65 | |
---|
66 | @property |
---|
67 | def name(self): |
---|
68 | return self.model.name |
---|
69 | |
---|
70 | @property |
---|
71 | def dof(self): |
---|
72 | return self.data.num_points - len(self.param_list) |
---|
73 | |
---|
74 | def summarize(self): |
---|
75 | """ |
---|
76 | Return a stylized list of parameter names and values with range bars |
---|
77 | suitable for printing. |
---|
78 | """ |
---|
79 | output = [] |
---|
80 | bounds = self.bounds() |
---|
81 | for i,p in enumerate(self.getp()): |
---|
82 | name = self.param_list[i] |
---|
83 | low,high = bounds[:,i] |
---|
84 | range = ",".join((("[%g"%low if numpy.isfinite(low) else "(-inf"), |
---|
85 | ("%g]"%high if numpy.isfinite(high) else "inf)"))) |
---|
86 | if not numpy.isfinite(p): |
---|
87 | bar = "*invalid* " |
---|
88 | else: |
---|
89 | bar = ['.']*10 |
---|
90 | if numpy.isfinite(high-low): |
---|
91 | position = int(9.999999999 * float(p-low)/float(high-low)) |
---|
92 | if position < 0: bar[0] = '<' |
---|
93 | elif position > 9: bar[9] = '>' |
---|
94 | else: bar[position] = '|' |
---|
95 | bar = "".join(bar) |
---|
96 | output.append("%40s %s %10g in %s"%(name,bar,p,range)) |
---|
97 | return "\n".join(output) |
---|
98 | |
---|
99 | def nllf(self, p=None): |
---|
100 | residuals = self.residuals(p) |
---|
101 | return 0.5*numpy.sum(residuals**2) |
---|
102 | |
---|
103 | def setp(self, p): |
---|
104 | for k,v in zip(self.param_list, p): |
---|
105 | self.model.setParam(k,v) |
---|
106 | #self.model.set_params(self.param_list, params) |
---|
107 | |
---|
108 | def getp(self): |
---|
109 | return numpy.array([self.model.getParam(k) for k in self.param_list]) |
---|
110 | #return numpy.asarray(self.model.get_params(self.param_list)) |
---|
111 | |
---|
112 | def bounds(self): |
---|
113 | return numpy.array([self._getrange(p) for p in self.param_list]).T |
---|
114 | |
---|
115 | def labels(self): |
---|
116 | return self.param_list |
---|
117 | |
---|
118 | def _getrange(self, p): |
---|
119 | """ |
---|
120 | Override _getrange of park parameter |
---|
121 | return the range of parameter |
---|
122 | """ |
---|
123 | lo, hi = self.model.details[p][1:3] |
---|
124 | if lo is None: lo = -numpy.inf |
---|
125 | if hi is None: hi = numpy.inf |
---|
126 | return lo, hi |
---|
127 | |
---|
128 | def randomize(self, n): |
---|
129 | p = self.getp() |
---|
130 | # since randn is symmetric and random, doesn't matter |
---|
131 | # point value is negative. |
---|
132 | # TODO: throw in bounds checking! |
---|
133 | return numpy.random.randn(n, len(self.param_list))*p + p |
---|
134 | |
---|
135 | def chisq(self): |
---|
136 | """ |
---|
137 | Calculates chi^2 |
---|
138 | |
---|
139 | :param params: list of parameter values |
---|
140 | |
---|
141 | :return: chi^2 |
---|
142 | |
---|
143 | """ |
---|
144 | return numpy.sum(self.res**2)/self.dof |
---|
145 | |
---|
146 | def residuals(self, params=None): |
---|
147 | """ |
---|
148 | Compute residuals |
---|
149 | :param params: value of parameters to fit |
---|
150 | """ |
---|
151 | if params is not None: self.setp(params) |
---|
152 | #import thread |
---|
153 | #print "params", params |
---|
154 | self.res, self.theory = self.data.residuals(self.model.evalDistribution) |
---|
155 | return self.res |
---|
156 | |
---|
157 | BOUNDS_PENALTY = 1e6 # cost for going out of bounds on unbounded fitters |
---|
158 | class MonitoredSasProblem(SasProblem): |
---|
159 | """ |
---|
160 | SAS problem definition for optimizers which do not have monitoring or bounds. |
---|
161 | """ |
---|
162 | def __init__(self, param_list, model=None, data=None, fitresult=None, |
---|
163 | handler=None, curr_thread=None, msg_q=None, update_rate=1): |
---|
164 | """ |
---|
165 | :param Model: the model wrapper fro sans -model |
---|
166 | :param Data: the data wrapper for sans data |
---|
167 | """ |
---|
168 | SasProblem.__init__(self, param_list, model, data) |
---|
169 | self.msg_q = msg_q |
---|
170 | self.curr_thread = curr_thread |
---|
171 | self.handler = handler |
---|
172 | self.fitresult = fitresult |
---|
173 | #self.last_update = time.time() |
---|
174 | #self.func_name = "Functor" |
---|
175 | #self.name = "Fill in proper name!" |
---|
176 | |
---|
177 | def residuals(self, p): |
---|
178 | """ |
---|
179 | Cost function for scipy.optimize.leastsq, which does not have a monitor |
---|
180 | built into the algorithm, and instead relies on a monitor built into |
---|
181 | the cost function. |
---|
182 | """ |
---|
183 | # Note: technically, unbounded fitters and unmonitored fitters are |
---|
184 | self.setp(p) |
---|
185 | |
---|
186 | # Compute penalty for being out of bounds which increases the farther |
---|
187 | # you get out of bounds. This allows derivative following algorithms |
---|
188 | # to point back toward the feasible region. |
---|
189 | penalty = self.bounds_penalty() |
---|
190 | if penalty > 0: |
---|
191 | self.theory = numpy.ones(self.data.num_points) |
---|
192 | self.res = self.theory*(penalty/self.data.num_points) + BOUNDS_PENALTY |
---|
193 | return self.res |
---|
194 | |
---|
195 | # If no penalty, then we are not out of bounds and we can use the |
---|
196 | # normal residual calculation |
---|
197 | SasProblem.residuals(self, p) |
---|
198 | |
---|
199 | # send update to the application |
---|
200 | if True: |
---|
201 | #self.fitresult.set_model(model=self.model) |
---|
202 | # copy residuals into fit results |
---|
203 | self.fitresult.residuals = self.res+0 |
---|
204 | self.fitresult.iterations += 1 |
---|
205 | self.fitresult.theory = self.theory+0 |
---|
206 | |
---|
207 | self.fitresult.p = numpy.array(p) # force copy, and coversion to array |
---|
208 | self.fitresult.set_fitness(fitness=self.chisq()) |
---|
209 | if self.msg_q is not None: |
---|
210 | self.msg_q.put(self.fitresult) |
---|
211 | |
---|
212 | if self.handler is not None: |
---|
213 | self.handler.set_result(result=self.fitresult) |
---|
214 | self.handler.update_fit() |
---|
215 | |
---|
216 | if self.curr_thread != None: |
---|
217 | try: |
---|
218 | self.curr_thread.isquit() |
---|
219 | except: |
---|
220 | #msg = "Fitting: Terminated... Note: Forcing to stop " |
---|
221 | #msg += "fitting may cause a 'Functor error message' " |
---|
222 | #msg += "being recorded in the log file....." |
---|
223 | #self.handler.stop(msg) |
---|
224 | raise |
---|
225 | |
---|
226 | return self.res |
---|
227 | |
---|
228 | def bounds_penalty(self): |
---|
229 | from numpy import sum, where |
---|
230 | p, bounds = self.getp(), self.bounds() |
---|
231 | return (sum(where(p<bounds[:,0], bounds[:,0]-p, 0)**2) |
---|
232 | + sum(where(p>bounds[:,1], bounds[:,1]-p, 0)**2) ) |
---|
233 | |
---|
234 | class BumpsFit(FitEngine): |
---|
235 | """ |
---|
236 | Fit a model using bumps. |
---|
237 | """ |
---|
238 | def __init__(self): |
---|
239 | """ |
---|
240 | Creates a dictionary (self.fit_arrange_dict={})of FitArrange elements |
---|
241 | with Uid as keys |
---|
242 | """ |
---|
243 | FitEngine.__init__(self) |
---|
244 | self.curr_thread = None |
---|
245 | |
---|
246 | def fit(self, msg_q=None, |
---|
247 | q=None, handler=None, curr_thread=None, |
---|
248 | ftol=1.49012e-8, reset_flag=False): |
---|
249 | """ |
---|
250 | """ |
---|
251 | fitproblem = [] |
---|
252 | for fproblem in self.fit_arrange_dict.itervalues(): |
---|
253 | if fproblem.get_to_fit() == 1: |
---|
254 | fitproblem.append(fproblem) |
---|
255 | if len(fitproblem) > 1 : |
---|
256 | msg = "Bumps can't fit more than a single fit problem at a time." |
---|
257 | raise RuntimeError, msg |
---|
258 | elif len(fitproblem) == 0 : |
---|
259 | raise RuntimeError, "No problem scheduled for fitting." |
---|
260 | model = fitproblem[0].get_model() |
---|
261 | if reset_flag: |
---|
262 | # reset the initial value; useful for batch |
---|
263 | for name in fitproblem[0].pars: |
---|
264 | ind = fitproblem[0].pars.index(name) |
---|
265 | model.setParam(name, fitproblem[0].vals[ind]) |
---|
266 | data = fitproblem[0].get_data() |
---|
267 | |
---|
268 | self.curr_thread = curr_thread |
---|
269 | |
---|
270 | result = FResult(model=model, data=data, param_list=self.param_list) |
---|
271 | result.pars = fitproblem[0].pars |
---|
272 | result.fitter_id = self.fitter_id |
---|
273 | result.index = data.idx |
---|
274 | if handler is not None: |
---|
275 | handler.set_result(result=result) |
---|
276 | |
---|
277 | if True: # bumps |
---|
278 | problem = SasProblem(param_list=self.param_list, |
---|
279 | model=model.model, |
---|
280 | data=data) |
---|
281 | run_bumps(problem, result, ftol, |
---|
282 | handler, curr_thread, msg_q) |
---|
283 | else: # scipy levenburg marquardt |
---|
284 | problem = SasProblem(param_list=self.param_list, |
---|
285 | model=model.model, |
---|
286 | data=data, |
---|
287 | handler=handler, |
---|
288 | fitresult=result, |
---|
289 | curr_thread=curr_thread, |
---|
290 | msg_q=msg_q) |
---|
291 | run_levenburg_marquardt(problem, result, ftol) |
---|
292 | |
---|
293 | if handler is not None: |
---|
294 | handler.update_fit(last=True) |
---|
295 | if q is not None: |
---|
296 | q.put(result) |
---|
297 | return q |
---|
298 | #if success < 1 or success > 5: |
---|
299 | # result.fitness = None |
---|
300 | return [result] |
---|
301 | |
---|
302 | def run_bumps(problem, result, ftol, handler, curr_thread, msg_q): |
---|
303 | def abort_test(): |
---|
304 | if curr_thread is None: return False |
---|
305 | try: curr_thread.isquit() |
---|
306 | except KeyboardInterrupt: |
---|
307 | if handler is not None: |
---|
308 | handler.stop("Fitting: Terminated!!!") |
---|
309 | return True |
---|
310 | return False |
---|
311 | |
---|
312 | fitopts = fitters.FIT_OPTIONS[fitters.FIT_DEFAULT] |
---|
313 | fitclass = fitopts.fitclass |
---|
314 | options = fitopts.options.copy() |
---|
315 | max_steps = fitopts.options.get('steps', 0) + fitopts.options.get('burn', 0) |
---|
316 | if 'monitors' not in options: |
---|
317 | options['monitors'] = [BumpsMonitor(handler, max_steps)] |
---|
318 | options['monitors'] += [ ConvergenceMonitor() ] |
---|
319 | options['ftol'] = ftol |
---|
320 | fitdriver = fitters.FitDriver(fitclass, problem=problem, |
---|
321 | abort_test=abort_test, **options) |
---|
322 | mapper = SerialMapper |
---|
323 | fitdriver.mapper = mapper.start_mapper(problem, None) |
---|
324 | try: |
---|
325 | best, fbest = fitdriver.fit() |
---|
326 | except: |
---|
327 | import traceback; traceback.print_exc() |
---|
328 | raise |
---|
329 | finally: |
---|
330 | mapper.stop_mapper(fitdriver.mapper) |
---|
331 | #print "best,fbest",best,fbest,problem.dof |
---|
332 | result.fitness = 2*fbest/problem.dof |
---|
333 | #print "fitness",result.fitness |
---|
334 | result.stderr = fitdriver.stderr() |
---|
335 | result.pvec = best |
---|
336 | # TODO: track success better |
---|
337 | result.success = True |
---|
338 | result.theory = problem.theory |
---|
339 | # For the convergence plot |
---|
340 | pop = numpy.asarray(options['monitors'][-1].convergence) |
---|
341 | result.convergence = 2*pop/problem.dof |
---|
342 | # Bumps uncertainty state |
---|
343 | try: result.uncertainty_state = fitdriver.fitter.state |
---|
344 | except AttributeError: pass |
---|
345 | |
---|
346 | def run_levenburg_marquardt(problem, result, ftol): |
---|
347 | # This import must be here; otherwise it will be confused when more |
---|
348 | # than one thread exist. |
---|
349 | from scipy import optimize |
---|
350 | |
---|
351 | out, cov_x, _, mesg, success = optimize.leastsq(problem.residuals, |
---|
352 | problem.getp(), |
---|
353 | ftol=ftol, |
---|
354 | full_output=1) |
---|
355 | if cov_x is not None and numpy.isfinite(cov_x).all(): |
---|
356 | stderr = numpy.sqrt(numpy.diag(cov_x)) |
---|
357 | else: |
---|
358 | stderr = [] |
---|
359 | result.fitness = problem.chisq() |
---|
360 | result.stderr = stderr |
---|
361 | result.pvec = out |
---|
362 | result.success = success |
---|
363 | result.theory = problem.theory |
---|
364 | |
---|