|
1
|
|
|
# Licensed under a 3-clause BSD style license - see LICENSE.rst |
|
2
|
|
|
"""iminuit fitting functions.""" |
|
3
|
|
|
import logging |
|
4
|
|
|
import numpy as np |
|
5
|
|
|
from scipy.stats import chi2, norm |
|
6
|
|
|
from .likelihood import Likelihood |
|
7
|
|
|
|
|
8
|
|
|
__all__ = ["optimize_iminuit", "covariance_iminuit", "confidence_iminuit", "contour_iminuit"] |
|
9
|
|
|
|
|
10
|
|
|
log = logging.getLogger(__name__) |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
class MinuitLikelihood(Likelihood): |
|
14
|
|
|
"""Likelihood function interface for iminuit.""" |
|
15
|
|
|
|
|
16
|
|
|
def fcn(self, *factors): |
|
17
|
|
|
self.parameters.set_parameter_factors(factors) |
|
18
|
|
|
|
|
19
|
|
|
total_stat = self.function() |
|
20
|
|
|
|
|
21
|
|
|
if self.store_trace: |
|
22
|
|
|
self.store_trace_iteration(total_stat) |
|
23
|
|
|
|
|
24
|
|
|
return total_stat |
|
25
|
|
|
|
|
26
|
|
|
|
|
27
|
|
|
def setup_iminuit(parameters, function, store_trace=False, **kwargs): |
|
28
|
|
|
from iminuit import Minuit |
|
29
|
|
|
|
|
30
|
|
|
minuit_func = MinuitLikelihood(function, parameters, store_trace=store_trace) |
|
31
|
|
|
|
|
32
|
|
|
pars, errors, limits = make_minuit_par_kwargs(parameters) |
|
33
|
|
|
|
|
34
|
|
|
minuit = Minuit(minuit_func.fcn, name=list(pars.keys()), **pars) |
|
35
|
|
|
minuit.tol = kwargs.pop("tol", 0.1) |
|
36
|
|
|
minuit.errordef = kwargs.pop("errordef", 1) |
|
37
|
|
|
minuit.print_level = kwargs.pop("print_level", 0) |
|
38
|
|
|
minuit.strategy = kwargs.pop("strategy", 1) |
|
39
|
|
|
|
|
40
|
|
|
for name, error in errors.items(): |
|
41
|
|
|
minuit.errors[name] = error |
|
42
|
|
|
|
|
43
|
|
|
for name, limit in limits.items(): |
|
44
|
|
|
minuit.limits[name] = limit |
|
45
|
|
|
|
|
46
|
|
|
return minuit, minuit_func |
|
47
|
|
|
|
|
48
|
|
|
|
|
49
|
|
|
def optimize_iminuit(parameters, function, store_trace=False, **kwargs): |
|
50
|
|
|
"""iminuit optimization |
|
51
|
|
|
|
|
52
|
|
|
Parameters |
|
53
|
|
|
---------- |
|
54
|
|
|
parameters : `~gammapy.modeling.Parameters` |
|
55
|
|
|
Parameters with starting values |
|
56
|
|
|
function : callable |
|
57
|
|
|
Likelihood function |
|
58
|
|
|
store_trace : bool |
|
59
|
|
|
Store trace of the fit |
|
60
|
|
|
**kwargs : dict |
|
61
|
|
|
Options passed to `iminuit.Minuit` constructor. If there is an entry 'migrad_opts', those options |
|
62
|
|
|
will be passed to `iminuit.Minuit.migrad()`. |
|
63
|
|
|
|
|
64
|
|
|
Returns |
|
65
|
|
|
------- |
|
66
|
|
|
result : (factors, info, optimizer) |
|
67
|
|
|
Tuple containing the best fit factors, some info and the optimizer instance. |
|
68
|
|
|
""" |
|
69
|
|
|
migrad_opts = kwargs.pop("migrad_opts", {}) |
|
70
|
|
|
|
|
71
|
|
|
minuit, minuit_func = setup_iminuit( |
|
72
|
|
|
parameters=parameters, function=function, store_trace=store_trace, **kwargs |
|
73
|
|
|
) |
|
74
|
|
|
|
|
75
|
|
|
minuit.migrad(**migrad_opts) |
|
76
|
|
|
|
|
77
|
|
|
factors = minuit.values |
|
78
|
|
|
info = { |
|
79
|
|
|
"success": minuit.valid, |
|
80
|
|
|
"nfev": minuit.nfcn, |
|
81
|
|
|
"message": _get_message(minuit, parameters), |
|
82
|
|
|
"trace": minuit_func.trace, |
|
83
|
|
|
} |
|
84
|
|
|
optimizer = minuit |
|
85
|
|
|
|
|
86
|
|
|
return factors, info, optimizer |
|
87
|
|
|
|
|
88
|
|
|
|
|
89
|
|
|
def covariance_iminuit(parameters, function, **kwargs): |
|
90
|
|
|
minuit = kwargs["minuit"] |
|
91
|
|
|
|
|
92
|
|
|
if minuit is None: |
|
93
|
|
|
minuit, _ = setup_iminuit( |
|
94
|
|
|
parameters=parameters, function=function, store_trace=False, **kwargs |
|
95
|
|
|
) |
|
96
|
|
|
minuit.hesse() |
|
97
|
|
|
|
|
98
|
|
|
message, success = "Hesse terminated successfully.", True |
|
99
|
|
|
|
|
100
|
|
|
try: |
|
101
|
|
|
covariance_factors = np.array(minuit.covariance) |
|
102
|
|
|
except (TypeError, RuntimeError): |
|
103
|
|
|
N = len(minuit.values) |
|
104
|
|
|
covariance_factors = np.nan * np.ones((N, N)) |
|
105
|
|
|
message, success = "Hesse failed", False |
|
106
|
|
|
|
|
107
|
|
|
return covariance_factors, {"success": success, "message": message} |
|
108
|
|
|
|
|
109
|
|
|
|
|
110
|
|
|
def confidence_iminuit(parameters, function, parameter, reoptimize, sigma, **kwargs): |
|
111
|
|
|
# TODO: this is ugly - design something better for translating to MINUIT parameter names. |
|
112
|
|
|
if not reoptimize: |
|
113
|
|
|
log.warning("Reoptimize = False ignored for iminuit backend") |
|
114
|
|
|
|
|
115
|
|
|
minuit, minuit_func = setup_iminuit( |
|
116
|
|
|
parameters=parameters, function=function, store_trace=False, **kwargs |
|
117
|
|
|
) |
|
118
|
|
|
migrad_opts = kwargs.get("migrad_opts", {}) |
|
119
|
|
|
minuit.migrad(**migrad_opts) |
|
120
|
|
|
|
|
121
|
|
|
# Maybe a wrapper class MinuitParameters? |
|
122
|
|
|
parameter = parameters[parameter] |
|
123
|
|
|
idx = parameters.free_parameters.index(parameter) |
|
124
|
|
|
var = _make_parname(idx, parameter) |
|
125
|
|
|
|
|
126
|
|
|
message = "Minos terminated successfully." |
|
127
|
|
|
cl = 2 * norm.cdf(sigma) - 1 |
|
128
|
|
|
|
|
129
|
|
|
try: |
|
130
|
|
|
minuit.minos(var, cl=cl, ncall=None) |
|
131
|
|
|
info = minuit.merrors[var] |
|
132
|
|
|
except (AttributeError, RuntimeError) as error: |
|
133
|
|
|
return { |
|
134
|
|
|
"success": False, |
|
135
|
|
|
"message": str(error), |
|
136
|
|
|
"errp": np.nan, |
|
137
|
|
|
"errn": np.nan, |
|
138
|
|
|
"nfev": 0, |
|
139
|
|
|
} |
|
140
|
|
|
|
|
141
|
|
|
return { |
|
142
|
|
|
"success": info.is_valid, |
|
143
|
|
|
"message": message, |
|
144
|
|
|
"errp": info.upper, |
|
145
|
|
|
"errn": -info.lower, |
|
146
|
|
|
"nfev": info.nfcn, |
|
147
|
|
|
} |
|
148
|
|
|
|
|
149
|
|
|
|
|
150
|
|
|
def contour_iminuit(parameters, function, x, y, numpoints, sigma, **kwargs): |
|
151
|
|
|
minuit, minuit_func = setup_iminuit( |
|
152
|
|
|
parameters=parameters, function=function, store_trace=False, **kwargs |
|
153
|
|
|
) |
|
154
|
|
|
minuit.migrad() |
|
155
|
|
|
|
|
156
|
|
|
par_x = parameters[x] |
|
157
|
|
|
idx_x = parameters.free_parameters.index(par_x) |
|
158
|
|
|
x = _make_parname(idx_x, par_x) |
|
159
|
|
|
|
|
160
|
|
|
par_y = parameters[y] |
|
161
|
|
|
idx_y = parameters.free_parameters.index(par_y) |
|
162
|
|
|
y = _make_parname(idx_y, par_y) |
|
163
|
|
|
|
|
164
|
|
|
cl = chi2(2).cdf(sigma ** 2) |
|
165
|
|
|
contour = minuit.mncontour(x=x, y=y, size=numpoints, cl=cl) |
|
166
|
|
|
# TODO: add try and except to get the success |
|
167
|
|
|
return { |
|
168
|
|
|
"success": True, |
|
169
|
|
|
"x": contour[:, 0], |
|
170
|
|
|
"y": contour[:, 1], |
|
171
|
|
|
} |
|
172
|
|
|
|
|
173
|
|
|
|
|
174
|
|
|
# this code is copied from https://github.com/iminuit/iminuit/blob/master/iminuit/_minimize.py#L95 |
|
175
|
|
|
def _get_message(m, parameters): |
|
176
|
|
|
message = "Optimization terminated successfully." |
|
177
|
|
|
success = m.accurate |
|
178
|
|
|
success &= np.all(np.isfinite([par.value for par in parameters])) |
|
179
|
|
|
if not success: |
|
180
|
|
|
message = "Optimization failed." |
|
181
|
|
|
fmin = m.fmin |
|
182
|
|
|
if fmin.has_reached_call_limit: |
|
183
|
|
|
message += " Call limit was reached." |
|
184
|
|
|
if fmin.is_above_max_edm: |
|
185
|
|
|
message += " Estimated distance to minimum too large." |
|
186
|
|
|
return message |
|
187
|
|
|
|
|
188
|
|
|
|
|
189
|
|
|
def _make_parnames(parameters): |
|
190
|
|
|
return [_make_parname(idx, par) for idx, par in enumerate(parameters)] |
|
191
|
|
|
|
|
192
|
|
|
|
|
193
|
|
|
def _make_parname(idx, par): |
|
194
|
|
|
return f"par_{idx:03d}_{par.name}" |
|
195
|
|
|
|
|
196
|
|
|
|
|
197
|
|
|
def make_minuit_par_kwargs(parameters): |
|
198
|
|
|
"""Create *Parameter Keyword Arguments* for the `Minuit` constructor. |
|
199
|
|
|
|
|
200
|
|
|
See: http://iminuit.readthedocs.io/en/latest/api.html#iminuit.Minuit |
|
201
|
|
|
""" |
|
202
|
|
|
names = _make_parnames(parameters.free_parameters) |
|
203
|
|
|
pars, errors, limits = {}, {}, {} |
|
204
|
|
|
|
|
205
|
|
|
for name, par in zip(names, parameters.free_parameters): |
|
206
|
|
|
pars[name] = par.factor |
|
207
|
|
|
|
|
208
|
|
|
min_ = None if np.isnan(par.factor_min) else par.factor_min |
|
209
|
|
|
max_ = None if np.isnan(par.factor_max) else par.factor_max |
|
210
|
|
|
limits[name] = (min_, max_) |
|
211
|
|
|
|
|
212
|
|
|
if par.error == 0 or np.isnan(par.error): |
|
213
|
|
|
error = 1 |
|
214
|
|
|
else: |
|
215
|
|
|
error = par.error / par.scale |
|
216
|
|
|
errors[name] = error |
|
217
|
|
|
|
|
218
|
|
|
return pars, errors, limits |
|
219
|
|
|
|
gammapy /
gammapy
