Skip to content

litmus.fitting_methods

Contains fitting procedures to be executed by the litmus class object

HM 24

fitting_procedure(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=True, **fit_params)

Bases: logger

Generic class for lag fitting procedures. Contains parent methods for setting properties

fitting_procedures extend the logger class and inherit all init args. See the logger documentation for details

The only required argument is the stat_model to perform fitting for. All other fitting parameters (listed below) can be passed as keyword arguments at init, or via .set_config(), or reset to default values with .reset().

Parameters:

Name Type Description Default
stat_model stats_model

Statistics model to fit for

 required
Source code in litmus/fitting_methods.py 
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout,
             err_stream=sys.stderr,
             verbose=True,
             debug=True,
             **fit_params):
    """
    fitting_procedures extend the logger class and inherit all __init__ args. See the logger documentation for details

    The only required argument is the stat_model to perform fitting for. All other fitting parameters (listed below) can be passed as keyword arguments at init, or via .set_config(), or reset to default values with .reset().

    :param stats_model stat_model: Statistics model to fit for
    """

    logger.__init__(self,
                    out_stream=out_stream,
                    err_stream=err_stream,
                    verbose=verbose,
                    debug=debug,
                    )

    if not hasattr(self, "_default_params"):
        self._default_params = {}

    # --------------------
    # Attributes
    self.stat_model: litmus.models.stats_model = stat_model
    """Stats model to do fitting for"""
    self.name = "Base Fitting Procedure"
    """Name for string printing"""
    self.is_ready = False
    """Flag to see if pre-pre-fitting has been done"""
    self.has_prefit = False
    """Flag to see if pre-fitting has been done"""
    self.has_run = False
    """Flag to see if fitting procedure has run to completion"""
    self.fitting_params = {} | self._default_params
    """A keyed dict of tuning parameters for the fitting method"""
    self.seed = _utils.randint() if "seed" not in fit_params.keys() else fit_params['seed']
    """Int seed for any randomized elements in the fitting method"""
    # --------------------

    self.set_config(**(self._default_params | fit_params))
    self._tempseed = self.seed
    self._data = None
stat_model: litmus.models.stats_model = stat_model instance-attribute

Stats model to do fitting for

name = 'Base Fitting Procedure' instance-attribute

Name for string printing

is_ready = False instance-attribute

Flag to see if pre-pre-fitting has been done

has_prefit = False instance-attribute

Flag to see if pre-fitting has been done

has_run = False instance-attribute

Flag to see if fitting procedure has run to completion

fitting_params = {} | self._default_params instance-attribute

A keyed dict of tuning parameters for the fitting method

seed = _utils.randint() if 'seed' not in fit_params.keys() else fit_params['seed'] instance-attribute

Int seed for any randomized elements in the fitting method

reset() -> None

Clears all memory and resets params to defaults

Source code in litmus/fitting_methods.py 
142
143
144
145
146
147
148
149
150
def reset(self) -> None:
    """
    Clears all memory and resets params to defaults
    """
    self.set_config(**self._default_params)

    self.has_run, self.is_ready = False, False

    return
set_config(**fit_params) -> None

Configure fitting parameters for fitting_method() object Accepts any parameters present with a name in fitting_method.fitting_params Unlisted parameters will be ignored.

Source code in litmus/fitting_methods.py 
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
def set_config(self, **fit_params) -> None:
    """
    Configure fitting parameters for fitting_method() object
    Accepts any parameters present with a name in fitting_method.fitting_params
    Unlisted parameters will be ignored.
    """

    if self.debug: print("Doing config with keys", fit_params.keys())

    badkeys = [key for key in fit_params.keys() if key not in self._default_params.keys()]

    for key, val in zip(fit_params.keys(), fit_params.values()):
        if key in badkeys: continue

        # If something's changed, flag as having not run
        currval = self.__getattribute__(key)
        if self.has_run and val != currval: self.has_run = False

        self.__setattr__(key, val)
        # self.fitting_params |= {key: val}
        if self.debug: print("\t set attr", key, file=self.out_stream)

    if len(badkeys) > 0:
        self.msg_err("Tried to configure bad keys:", *badkeys, delim='\t')
    return
readyup() -> None

Performs pre-fit preparation calcs. Should only be called if not self.is_ready()

Source code in litmus/fitting_methods.py 
178
179
180
181
182
def readyup(self) -> None:
    """
    Performs pre-fit preparation calcs. Should only be called if not self.is_ready()
    """
    self.is_ready = True
prefit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None) -> None

Performs any tasks required after setup but prior to actual fitting

Parameters:

Name Type Description Default
lc_1 lightcurve

Lightcurve 1 (Main)

 required
lc_2 lightcurve

Lightcurve 2 (Response)

 required
seed int

A random seed for feeding to the fitting process. If none, will select randomly

 None
Source code in litmus/fitting_methods.py 
187
188
189
190
191
192
193
194
195
def prefit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None) -> None:
    """
    Performs any tasks required after setup but prior to actual fitting
    :param lc_1: Lightcurve 1 (Main)
    :param lc_2: Lightcurve 2 (Response)
    :param seed: A random seed for feeding to the fitting process. If none, will select randomly
    """

    self.has_prefit = True
fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None) -> None

Performs the lag recovery method for this fitting procedure. If not prefit, will run prefit()

Parameters:

Name Type Description Default
lc_1 lightcurve

Lightcurve 1 (Main)

 required
lc_2 lightcurve

Lightcurve 2 (Response)

 required
seed int

A random seed for feeding to the fitting process. If none, will select randomly

 None
Source code in litmus/fitting_methods.py 
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None) -> None:
    """
    Performs the lag recovery method for this fitting procedure. If not prefit, will run prefit()
    :param lc_1: Lightcurve 1 (Main)
    :param lc_2: Lightcurve 2 (Response)
    :param seed: A random seed for feeding to the fitting process. If none, will select randomly
    """

    # Sanity checks inherited by all subclasses
    if not self.is_ready: self.readyup()
    if isinstance(seed, int):
        self._tempseed = seed
        self._tempseed = _utils.randint()
        self._tempseed = _utils.randint()
    seed = self._tempseed

    data = self.stat_model.lc_to_data(lc_1, lc_2)
    self._data = data

    # An error message raised if this fitting procedure doesn't have .fit()
    if self.__class__.fit == fitting_procedure.fit:
        self.msg_err("Fitting \"%s\" method does not have method .fit() implemented" % self.name)

    return
get_samples(N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}

Returns MCMC-like posterior samples if fit() has been run

Parameters:

Name Type Description Default
N int

Number of samples to return. If None, return all

 None
seed int

Random seed for any stochastic elements

 None
importance_sampling bool

If true, will weight the results by

 False

Returns:

Type Description
{str: [float]}

keyed dict of samples in the constrained domain
Source code in litmus/fitting_methods.py 
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
def get_samples(self, N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}:
    """
    Returns MCMC-like posterior samples if fit() has been run
    :param N: Number of samples to return. If None, return all
    :param seed: Random seed for any stochastic elements
    :param importance_sampling: If true, will weight the results by
    :return: keyed dict of samples in the constrained domain
    """

    if not self.is_ready: self.readyup()
    if isinstance(seed, int):
        self._tempseed = seed
        self._tempseed = _utils.randint()
    seed = self._tempseed

    if self.__class__.fit == fitting_procedure.fit:
        self.msg_err("Fitting \"%s\" method does not have method .get_samples() implemented" % self.name)
get_evidence(seed: int = None, return_type='linear') -> [float, float, float]

Returns the estimated evidence for the fit model.

Parameters:

Name Type Description Default
seed int

int seed for random number generation

 None
return_type

if 'linear', returns as array-like [Z, -dZ-, dZ+]. If 'log', returns as array-like [logZ, -dlogZ-, dlogZ+]

 'linear'

Returns:

Type Description
[float, float, float]

len 3 array of type [mu, -E-, E+]
Source code in litmus/fitting_methods.py 
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
def get_evidence(self, seed: int = None, return_type='linear') -> [float, float, float]:
    """
    Returns the estimated evidence for the fit model.
    :param seed: int seed for random number generation
    :param return_type: if 'linear', returns as array-like [Z, -dZ-, dZ+]. If 'log', returns as array-like [logZ, -dlogZ-, dlogZ+]
    :return: len 3 array of type [mu, -E-, E+]
    """

    assert return_type in ['linear', 'log'], "Return type must be 'linear' or 'log'"

    if not self.is_ready: self.readyup()
    if not self.has_run: self.msg_err("Warning! Tried to call get_evidence without running first!")

    if isinstance(seed, int):
        self._tempseed = seed
        self._tempseed = _utils.randint()
    seed = self._tempseed

    if self.__class__.get_evidence == fitting_procedure.get_evidence:
        self.msg_err("Fitting \"%s\" method does not have method .get_evidence() implemented" % self.name)
        return np.array([0.0, 0.0, 0.0])
get_information(seed: int = None) -> [float, float, float]

Returns an estimate of the information (KL divergence relative to prior). Returns as array-like [I,dI-,dI+]

Parameters:

Name Type Description Default
seed int

int seed for random number generation

 None

Returns:

Type Description
[float, float, float]

len 3 array of type [mu, -E-, E+]
Source code in litmus/fitting_methods.py 
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
def get_information(self, seed: int = None) -> [float, float, float]:
    """
    Returns an estimate of the information (KL divergence relative to prior). Returns as array-like [I,dI-,dI+]
    :param seed: int seed for random number generation
    :return: len 3 array of type [mu, -E-, E+]
    """

    if not self.is_ready: self.readyup()
    if isinstance(seed, int):
        self._tempseed = seed
        self._tempseed = _utils.randint()
    seed = self._tempseed

    if self.__class__.get_information == fitting_procedure.get_information:
        self.msg_err("Fitting \"%s\" method does not have method .get_information() implemented" % self.name)

        return np.array([0.0, 0.0, 0.0])
get_peaks(seed=None)

Returns the maximum posterior position in parameter space

Parameters:

Name Type Description Default
seed

int seed for random number generation

 None
Source code in litmus/fitting_methods.py 
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
def get_peaks(self, seed=None):
    """
    Returns the maximum posterior position in parameter space
    :param seed: int seed for random number generation
    """

    if not self.is_ready: self.readyup()
    if isinstance(seed, int):
        self._tempseed = seed
        self._tempseed = _utils.randint()
    seed = self._tempseed

    if self.__class__.get_peaks == fitting_procedure.get_peaks:
        self.msg_err("Fitting \"%s\" method does not have method .get_peaks() implemented" % self.name)

        return {}, np.array([])
diagnostics(plot=True) -> _types.Figure

Generates some plots to check if the solver has converged

Parameters:

Name Type Description Default
plot

If True, run plt.show()

 True

Returns:

Type Description
Figure

generated matplotlib figure
Source code in litmus/fitting_methods.py 
297
298
299
300
301
302
303
304
305
def diagnostics(self, plot=True) -> _types.Figure:
    """
    Generates some plots to check if the solver has converged
    :param plot: If True, run plt.show()
    :return: generated matplotlib figure
    """

    if self.__class__.diagnostics == fitting_procedure.diagnostics:
        self.msg_err("Fitting \"%s\" method does not have method .diagnostics() implemented" % self.name)
diagnostic_lagplot(plot=True) -> _types.Figure

Generates diagnostic plots specifically for lags

Parameters:

Name Type Description Default
plot

If True, run plt.show()

 True

Returns:

Type Description
Figure

generated matplotlib figure
Source code in litmus/fitting_methods.py 
307
308
309
310
311
312
313
314
315
def diagnostic_lagplot(self, plot=True) -> _types.Figure:
    """
    Generates diagnostic plots specifically for lags
    :param plot: If True, run plt.show()
    :return: generated matplotlib figure
    """

    if self.__class__.diagnostic_lagplot == fitting_procedure.diagnostic_lagplot:
        self.msg_err("Fitting \"%s\" method does not have method .diagnostic_lagplot() implemented" % self.name)

ICCF(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=False, **fit_params)

Bases: fitting_procedure

Fit lags using interpolated cross correlation function in the style of pyCCF. Note that this is not a Bayesian fitter and gives only approximate measures of the lag

todo - Add p value, false positive and evidence estimates (?)

Parameters:

Name Type Description Default
stat_model stats_model

Statistics model to fit for

 required
Nboot int

Number of bootstraps in ICCF

 required
Nterp int

Number of points to interpolate with in ICCF correlation evals

 required
Nlags int

Number of lags to test with.

 required
Source code in litmus/fitting_methods.py 
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout, err_stream=sys.stderr,
             verbose=True, debug=False, **fit_params):

    """
    :param stats_model stat_model: Statistics model to fit for
    :param int Nboot: Number of bootstraps in ICCF
    :param int Nterp: Number of points to interpolate with in ICCF correlation evals
    :param int Nlags: Number of lags to test with.
    """

    args_in = {**locals(), **fit_params}
    del args_in['self']
    del args_in['__class__']
    del args_in['fit_params']

    if not hasattr(self, '_default_params'):
        self._default_params = {}

    self._default_params |= {
        'Nboot': 512,
        'Nterp': 2014,
        'Nlags': 512,
    }

    super().__init__(**args_in)

    # --------------------
    # Attributes
    self.name = "ICCF Fitting Procedure"

    self.lags = np.zeros(self.Nterp)
    """Array of lags to test at"""
    self.samples = np.zeros(self.Nboot)
    """The bootstrapped samples of the best fit lag"""
    self.correls = np.zeros(self.Nterp)
    """Un-bootstrapped correlation function"""
    self.lag_mean = 0.0
    """Mean of bootstrapped best fit lags"""
    self.lag_err = 0.0
    """Std err of bootstrapped bestfit lags"""
name = 'ICCF Fitting Procedure' instance-attribute
lags = np.zeros(self.Nterp) instance-attribute

Array of lags to test at

samples = np.zeros(self.Nboot) instance-attribute

The bootstrapped samples of the best fit lag

correls = np.zeros(self.Nterp) instance-attribute

Un-bootstrapped correlation function

lag_mean = 0.0 instance-attribute

Mean of bootstrapped best fit lags

lag_err = 0.0 instance-attribute

Std err of bootstrapped bestfit lags

set_config(**fit_params)
Source code in litmus/fitting_methods.py 
374
375
def set_config(self, **fit_params):
    super().set_config(**fit_params)
readyup()
Source code in litmus/fitting_methods.py 
377
378
379
380
381
382
383
384
385
def readyup(self):
    super().readyup()
    """
    # self.lags = jnp.linspace(*self.stat_model.prior_ranges['lag'], self.Nlags)
    self.lags = np.random.randn(self.Nlags) * self.stat_model.prior_ranges['lag'].ptp() + \
                self.stat_model.prior_ranges['lag'][0]
    """
    self.is_ready = True
    self.has_prefit = False
fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # -------------------
    fitting_procedure.fit(**locals())
    seed = self._tempseed
    # -------------------

    # Unpack lightcurve
    X1, Y1, E1 = lc_1.T, lc_1.Y, lc_1.E
    X2, Y2, E2 = lc_2.T, lc_2.Y, lc_2.E

    # Get interpolated correlation for all un-bootstrapped data
    self.correls = iccf.correlfunc_jax_vmapped(self.lags, X1, Y1, X2, Y2, self.Nterp)

    # Do bootstrap fitting
    lagrange = jnp.linspace(*self.stat_model.prior_ranges['lag'], self.Nlags)
    jax_samples = iccf.correl_func_boot_jax_wrapper_nomap(lagrange, X1, Y1, X2, Y2, E1, E2,
                                                          Nterp=self.Nterp,
                                                          Nboot=self.Nboot)

    # Store Results
    self.samples = jax_samples
    self.lag_mean, self.lag_err = jax_samples.mean(), jax_samples.std()

    self.has_run = True
get_samples(N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}
Source code in litmus/fitting_methods.py 
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
def get_samples(self, N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}:
    # -------------------
    fitting_procedure.get_samples(**locals())
    seed = self._tempseed

    if importance_sampling:
        self.msg_err("Warning! Cannot use important sampling with ICCF. Try implementing manually")
        return
    # -------------------

    # Return entire sample chain or sub-set of samples
    if N is None:
        return ({'lag': self.samples})
    else:
        if N > self.Nboot:
            self.msg_err(
                "Warning, tried to get %i sub-samples from %i boot-strap iterations in ICCF" % (N, self.Nboot),
            )
        return ({'lag': np.random.choice(a=self.samples, size=N, replace=True)})
get_peaks(seed: int = None) -> ({float: [float]}, [float])
Source code in litmus/fitting_methods.py 
433
434
435
436
437
438
439
def get_peaks(self, seed: int = None) -> ({float: [float]}, [float]):
    # -------------------
    fitting_procedure.get_peaks(**locals())
    seed = self._tempseed
    # --------------
    out = self.lags[np.argmax(self.correls)]
    return {'lag': np.array([out])}

prior_sampling(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=False, **fit_params)

Bases: fitting_procedure

Randomly samples from the prior and weights with importance sampling. The crudest available sampler. For test purposes only, not suggested for actual use.

Parameters:

Name Type Description Default
stat_model stats_model

Statistics model to fit for

 required
Nsamples int

Number of samples to draw from the prior. Defaults to 4096.

 required
Source code in litmus/fitting_methods.py 
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout, err_stream=sys.stderr,
             verbose=True, debug=False, **fit_params):

    """
    :param stats_model stat_model: Statistics model to fit for
    :param stats_model stat_model: Statistics model to fit for
    :param int Nsamples: Number of samples to draw from the prior. Defaults to 4096.
    """

    # ------------------------------------
    args_in = {**locals(), **fit_params}
    del args_in['self']
    del args_in['__class__']
    del args_in['fit_params']

    if not hasattr(self, '_default_params'):
        self._default_params = {}
    self._default_params |= {
        'Nsamples': 4096
    }

    super().__init__(**args_in)
    # --------------------
    # Attributes

    self.name = "Prior Sampling Fitting Procedure"

    self.samples = {key: np.zeros(self.Nsamples) for key in self.stat_model.paramnames()}
    """Samples drawn from the prior"""
    self.log_prior = np.zeros(self.Nsamples)
    """Log prior densities of the samples"""
    self.log_likes = np.zeros(self.Nsamples)
    """Log likelihood of the samples"""
    self.log_density = np.zeros(self.Nsamples)
    """Log joint density of the samples"""
    self.weights = np.zeros(self.Nsamples)
    """Normalized weights for drawing the samples"""
name = 'Prior Sampling Fitting Procedure' instance-attribute
samples = {key: np.zeros(self.Nsamples)for key in self.stat_model.paramnames()} instance-attribute

Samples drawn from the prior

log_prior = np.zeros(self.Nsamples) instance-attribute

Log prior densities of the samples

log_likes = np.zeros(self.Nsamples) instance-attribute

Log likelihood of the samples

log_density = np.zeros(self.Nsamples) instance-attribute

Log joint density of the samples

weights = np.zeros(self.Nsamples) instance-attribute

Normalized weights for drawing the samples

fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # -------------------
    fitting_procedure.fit(**locals())
    seed = self._tempseed
    # -------------------

    # Generate samples & calculate likelihoods todo - These currently give posterior densities and not log likes
    data = self.stat_model.lc_to_data(lc_1, lc_2)
    samples = self.stat_model.prior_sample(num_samples=self.Nsamples, seed=seed)
    log_density = self.stat_model.log_density(data=data, params=samples)
    log_prior = self.stat_model.log_prior(params=samples)
    log_likes = log_density - log_prior
    likes = np.exp(log_likes)

    # Store results
    self.log_prior = log_prior
    self.log_likes = log_likes
    self.log_density = log_density
    self.samples = samples
    self.weights = likes / likes.sum()

    # Mark as having completed a run
    self.has_run = True
get_samples(N: int = None, seed: int = None, importance_sampling: bool = True) -> {str: [float]}
Source code in litmus/fitting_methods.py 
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
def get_samples(self, N: int = None, seed: int = None, importance_sampling: bool = True) -> {str: [float]}:
    # -------------------
    fitting_procedure.get_samples(**locals())
    seed = self._tempseed
    # -------------------

    if N is None:
        N = self.Nsamples
    else:
        if N > self.Nsamples:
            self.msg_err("Warning, tried to get %i sub-samples from %i samples" % (N, self.Nsamples))

    if importance_sampling:
        weights = self.weights / self.weights.sum()
    else:
        weights = None

    I = np.random.choice(a=np.arange(self.Nsamples), size=N, replace=True,
                         p=weights)
    return ({
        key: val[I] for key, val in zip(self.samples.keys(), self.samples.values())
    })
get_evidence(seed=None, return_type='linear') -> [float, float, float]
Source code in litmus/fitting_methods.py 
540
541
542
543
544
545
546
547
548
549
550
551
552
553
def get_evidence(self, seed=None, return_type='linear') -> [float, float, float]:
    # -------------------
    fitting_procedure.get_evidence(**locals())
    seed = self._tempseed
    # -------------------
    density = np.exp(self.log_density - self.log_density.max())

    Z = density.mean() * self.stat_model.prior_volume * np.exp(self.log_density.max())
    uncert = density.std() / np.sqrt(self.Nsamples) * self.stat_model.prior_volume

    if return_type == 'linear':
        np.array([Z, -uncert, uncert])
    elif return_type == 'log':
        np.array([np.log(Z), np.log(1 - uncert / Z), np.log(1 + uncert / Z)])
get_information(seed: int = None) -> [float, float, float]
Source code in litmus/fitting_methods.py 
555
556
557
558
559
560
561
562
563
564
565
def get_information(self, seed: int = None) -> [float, float, float]:
    # -------------------
    fitting_procedure.get_information(**locals())
    seed = self._tempseed
    # -------------------
    info_partial = np.random.choice(self.log_density - self.log_prior, self.Nsamples,
                                    p=self.weights)
    info = info_partial.mean() * self.stat_model.prior_volume
    uncert = info_partial.std() / np.sqrt(self.Nsamples) * self.stat_model.prior_volume

    return np.array([info, -uncert, uncert])

nested_sampling(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=False, **fit_params)

Bases: fitting_procedure

Fits the Bayesian model with Nested Sampling by using JAXNS. Highly accurate evidence / posterior distributions, but can be slow for models with more than a few parameters. Use only if hessian_scan and SVI_scan fail.

Parameters:

Name Type Description Default
stat_model stats_model

Statistics model to fit for

 required
num_live_points int

Number of live points to use in nested sampling fitting. Defaults to 500.

 required
max_samples int

Maximum samples before terminating the run. Defaults to 10_000.

 required
num_parallel_samplers int

Number of parallel samplers to fit with. Defaults to 1.

 required
evidence_uncert float

Termination condition for evidence uncertainty. Default to 1E-3.

 required
live_evidence_frac float

Termination condition for live fraction of evidence remaining. Defaults to log(1 + 1e-3).

 required
Source code in litmus/fitting_methods.py 
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout, err_stream=sys.stderr,
             verbose=True, debug=False, **fit_params):
    """
    :param stats_model stat_model: Statistics model to fit for
    :param int num_live_points: Number of live points to use in nested sampling fitting. Defaults to 500.
    :param int max_samples: Maximum samples before terminating the run. Defaults to 10_000.
    :param int num_parallel_samplers: Number of parallel samplers to fit with. Defaults to 1.
    :param float evidence_uncert: Termination condition for evidence uncertainty. Default to 1E-3.
    :param float live_evidence_frac: Termination condition for live fraction of evidence remaining. Defaults to log(1 + 1e-3).
    """

    args_in = {**locals(), **fit_params}
    del args_in['self']
    del args_in['__class__']
    del args_in['fit_params']

    if not hasattr(self, '_default_params'):
        self._default_params = {}
    self._default_params |= {
        'num_live_points': 500,
        'max_samples': 10_000,
        'num_parallel_samplers': 1,
        'evidence_uncert': 1E-3,
        'live_evidence_frac': np.log(1 + 1e-3),
    }

    super().__init__(**args_in)

    self._jaxnsmodel = None
    self._jaxnsresults = None
    self._jaxnstermination = None
    self._jaxnsresults = None

    # --------------------
    # Attributes
    self.name = "Nested Sampling Fitting Procedure"
    self.sampler = None
    """The JAXNS nested sampler object"""
    self.logevidence = jnp.zeros(3)
    """JAXNS log evidence"""
    self.priorvolume = 0.0
    """Prior volume for correcting from unit cube to actual density"""
name = 'Nested Sampling Fitting Procedure' instance-attribute
sampler = None instance-attribute

The JAXNS nested sampler object

logevidence = jnp.zeros(3) instance-attribute

JAXNS log evidence

priorvolume = 0.0 instance-attribute

Prior volume for correcting from unit cube to actual density

prefit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
def prefit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # ---------------------
    fitting_procedure.prefit(**locals())
    if seed is None: seed = _utils.randint()
    # ---------------------

    # Get uniform prior bounds
    bounds = np.array([self.stat_model.prior_ranges[key] for key in self.stat_model.paramnames()])
    lo, hi = jnp.array(bounds[:, 0]), jnp.array(bounds[:, 1])

    data = self.stat_model.lc_to_data(lc_1, lc_2)

    # Construct jaxns friendly prior & likelihood
    def prior_model():
        x = yield jaxns.Prior(tfpd.Uniform(low=lo, high=hi), name='x')
        return x

    def log_likelihood(x):
        params = _utils.dict_unpack(x, keys=self.stat_model.paramnames())
        with numpyro.handlers.block(hide=self.stat_model.paramnames()):
            LL = self.stat_model._log_likelihood(params, data=data)
        return LL

    # jaxns object setup
    self._jaxnsmodel = jaxns.Model(prior_model=prior_model,
                                   log_likelihood=log_likelihood,
                                   )

    self._jaxnstermination = jaxns.TerminationCondition(
        dlogZ=self.evidence_uncert,
        max_samples=self.max_samples,
    )

    # Build jaxns Nested Sampler
    self.sampler = jaxns.NestedSampler(
        model=self._jaxnsmodel,
        max_samples=self.max_samples,
        verbose=self.debug,
        num_live_points=self.num_live_points,
        num_parallel_workers=self.num_parallel_samplers,
        difficult_model=True,
    )

    self.has_prefit = True
fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # ---------------------
    fitting_procedure.fit(**locals())
    if seed is None: seed = _utils.randint()
    # ---------------------
    if not self.has_prefit:
        self.prefit(lc_1, lc_2, seed)
    self.readyup()
    # ---------------------

    # -----------------
    # Run the sampler!
    termination_reason, state = self.sampler(jax.random.PRNGKey(seed),
                                             self._jaxnstermination)

    # -----------------
    # Extract & save results
    self._jaxnsresults = self.sampler.to_results(
        termination_reason=termination_reason,
        state=state
    )

    self.has_run = True
get_samples(N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}
Source code in litmus/fitting_methods.py 
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
def get_samples(self, N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}:
    if seed is None: seed = _utils.randint()

    samples, logweights = self._jaxnsresults.samples['x'], self._jaxnsresults.log_dp_mean

    if N is None:
        if importance_sampling:
            out = {key: samples.T[i] for i, key in enumerate(self.stat_model.paramnames())}
            return out
        else:
            N = samples.shape[0]

    if importance_sampling:
        logweights = jnp.zeros_like(logweights)

    samples = jaxns.resample(
        key=jax.random.PRNGKey(seed),
        samples=samples,
        log_weights=logweights,
        S=N
    )

    out = {key: samples.T[i] for i, key in enumerate(self.stat_model.paramnames())}

    return (out)
get_evidence(seed: int = None, return_type='linear') -> [float, float, float]

Returns the -1, 0 and +1 sigma values for model evidence from nested sampling. This represents an estimate of numerical uncertainty

Source code in litmus/fitting_methods.py 
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
def get_evidence(self, seed: int = None, return_type='linear') -> [float, float, float]:
    """
    Returns the -1, 0 and +1 sigma values for model evidence from nested sampling.
    This represents an estimate of numerical uncertainty
    """

    if seed is None: seed = _utils.randint()

    l, l_e = self._jaxnsresults.log_Z_mean, self._jaxnsresults.log_Z_uncert

    if return_type == 'linear':

        out = np.exp([
            l,
            l - l_e,
            l + l_e
        ])

        out -= np.array([0, out[0], out[0]])
    elif return_type == 'log':
        out = np.array([l, -l_e, l_e])
    else:
        self.msg_err(
            "Warning! Tried to call get_evidence in %s with bad return_type. Should be 'log' or 'linear'" % self.name)
        out = [0.0, 0.0, 0.0]

    return out
get_information(seed: int = None) -> [float, float, float]

Use the Nested Sampling shells to estimate the model information relative to prior

Source code in litmus/fitting_methods.py 
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
def get_information(self, seed: int = None) -> [float, float, float]:
    """
    Use the Nested Sampling shells to estimate the model information relative to prior
    """
    # todo - this is outmoded

    if seed is None: seed = _utils.randint()

    samples, logweights = self._jaxnsresults.samples, self._jaxnsresults.log_dp_mean

    weights = np.exp(logweights)
    weights /= weights.sum()

    log_density = self._jaxnsresults.log_posterior_density
    prior_values = self.stat_model.log_prior(samples)

    info = np.sum((log_density - prior_values) * weights)

    partial_info = np.random.choice((log_density - prior_values), len(log_density), p=weights)
    uncert = partial_info.std() / np.sqrt(len(log_density))

    return np.array(info, uncert, uncert)
get_peaks(seed: int = None) -> ({str: [float]}, float)
Source code in litmus/fitting_methods.py 
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
def get_peaks(self, seed: int = None) -> ({str: [float]}, float):

    # todo - this is outmoded

    # ---------------------
    if seed is None: seed = _utils.randint()
    # ---------------------

    self.msg_err("get_peaks currently placeholder.")
    return ({key: np.array([]) for key in self.stat_model.paramnames()}, np.array([]))

    # ---------------------

    NS = self.sampler
    samples = self.get_samples()
    log_densities = NS._results.log_posterior_density

    # Find clusters
    indices = clustering.clusterfind_1D(samples['lag'])

    # Break samples and log-densities up into clusters
    sorted_samples = clustering.sort_by_cluster(samples, indices)
    sort_logdens = clustering.sort_by_cluster(log_densities, indices)

    Nclusters = len(sorted_samples)

    # Make an empty dictionary to store positions in
    peak_locations = {key: np.zeros([Nclusters]) for key in samples.keys()}
    peaklikes = np.zeros([Nclusters])

    for i, group, lds in enumerate(sorted_samples, sort_logdens):
        j = np.argmax(lds)
        for key in samples.keys():
            peak_locations[key][i] = group[key][j]
        peaklikes[i] = lds[j]

    return (peak_locations, peaklikes)
diagnostics(show=True) -> _types.Figure
Source code in litmus/fitting_methods.py 
807
808
809
810
811
def diagnostics(self, show=True) -> _types.Figure:

    # todo fix this to make the show work properly
    jaxns.plot_diagnostics(self._jaxnsresults)
    return plt.gcf()

hessian_scan(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=False, **fit_params)

Bases: fitting_procedure

Litmus's main hessian scan fitting procedure.

Parameters:

Name Type Description Default
stat_model stats_model

Statistics model to fit for

 required
Nlags' int

64

 required
opt_tol' float

1E-2

 required
opt_tol_init' float

1E-4

 required
step_size' float

0.001

 required
constrained_domain' bool

False

 required
max_opt_eval' int

1_000

 required
max_opt_eval_init' int

5_000

 required
LL_threshold' float

100.0

 required
init_samples' int

5_000

 required
grid_bunching' float

Amount to bunch up points about peaks in grid smoothing, with 0.0 being even spacing and 1.0 being MCMC_like sample spacing. Defaults to 0.5.

 required
grid_depth'

None

 required
grid_Nterp'

None

 required
grid_relaxation' float

0.1, # deprecated, remove

 required
grid_firstdepth' float

2.0

 required
reverse' bool

Defaults to True.

 required
split_lags' bool

Defaults to True.

 required
optimizer_args_init' dict

{}

 required
optimizer_args' dict

{}

 required
seed_params' dict

{}

 required
precondition' str

Type of preconditioning to use in scan. Defaults to 'diag'.

 required
interp_scale' str

Scale to peform grid smoothing interpolation on . Defaults to 'log'.

 required
Source code in litmus/fitting_methods.py 
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout, err_stream=sys.stderr,
             verbose=True, debug=False, **fit_params):

    """
    :param stats_model stat_model: Statistics model to fit for
    :param int Nlags': 64
    :param float opt_tol': 1E-2
    :param float opt_tol_init': 1E-4
    :param float step_size': 0.001
    :param bool constrained_domain': False
    :param int max_opt_eval': 1_000
    :param int max_opt_eval_init': 5_000
    :param float LL_threshold': 100.0
    :param int init_samples': 5_000
    :param float grid_bunching': Amount to bunch up points about peaks in grid smoothing, with 0.0 being even spacing and 1.0 being MCMC_like sample spacing. Defaults to 0.5.
    :param grid_depth': None
    :param grid_Nterp': None
    :param float grid_relaxation': 0.1,  # deprecated, remove
    :param float grid_firstdepth': 2.0
    :param bool reverse': Defaults to True.
    :param bool split_lags': Defaults to True.
    :param dict optimizer_args_init': {}
    :param dict optimizer_args': {}
    :param dict seed_params': {}
    :param str precondition': Type of preconditioning to use in scan. Defaults to 'diag'.
    :param str interp_scale': Scale to peform grid smoothing interpolation on . Defaults to 'log'.
    """
    args_in = {**locals(), **fit_params}
    del args_in['self']
    del args_in['__class__']
    del args_in['fit_params']

    if not hasattr(self, '_default_params'):
        self._default_params = {}

    self._default_params |= {
        'Nlags': 64,
        'opt_tol': 1E-2,
        'opt_tol_init': 1E-4,
        'step_size': 0.001,
        'constrained_domain': False,
        'max_opt_eval': 1_000,
        'max_opt_eval_init': 5_000,
        'LL_threshold': 100.0,
        'init_samples': 5_000,
        'grid_bunching': 0.5,
        'grid_depth': None,
        'grid_Nterp': None,
        'grid_relaxation': 0.1,  # deprecated, remove
        'grid_firstdepth': 2.0,
        'reverse': True,
        'split_lags': True,
        'optimizer_args_init': {},
        'optimizer_args': {},
        'seed_params': {},
        'precondition': 'diag',
        'interp_scale': 'log',
    }

    self._allowable_interpscales = ['linear', 'log']

    super().__init__(**args_in)

    # -----------------------------------

    self.name = "Hessian Scan Fitting Procedure"

    self.lags: [float] = np.zeros(self.Nlags)
    self.converged: np.ndarray[bool] = np.zeros_like(self.lags, dtype=bool)

    self.scan_peaks: dict = {}
    self.log_evidences: list = []
    self.log_evidences_uncert: list = []

    self.diagnostic_hessians: list = []
    self.diagnostic_densities: list = []
    self.diagnostic_grads: list = []
    self.diagnostic_ints: list = []
    self.diagnostic_tgrads: list = []

    self.params_toscan = self.stat_model.free_params()
    if 'lag' in self.params_toscan: self.params_toscan.remove('lag')

    self.precon_matrix: np.ndarray[np.float64] = np.eye(len(self.params_toscan), dtype=np.float64)
    self.solver: jaxopt.BFGS = None

    self.estmap_params = {}
name = 'Hessian Scan Fitting Procedure' instance-attribute
lags: [float] = np.zeros(self.Nlags) instance-attribute
converged: np.ndarray[bool] = np.zeros_like(self.lags, dtype=bool) instance-attribute
scan_peaks: dict = {} instance-attribute
log_evidences: list = [] instance-attribute
log_evidences_uncert: list = [] instance-attribute
diagnostic_hessians: list = [] instance-attribute
diagnostic_densities: list = [] instance-attribute
diagnostic_grads: list = [] instance-attribute
diagnostic_ints: list = [] instance-attribute
diagnostic_tgrads: list = [] instance-attribute
params_toscan = self.stat_model.free_params() instance-attribute
precon_matrix: np.ndarray[np.float64] = np.eye(len(self.params_toscan), dtype=np.float64) instance-attribute
solver: jaxopt.BFGS = None instance-attribute
estmap_params = {} instance-attribute
readyup()
Source code in litmus/fitting_methods.py 
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
def readyup(self):

    # Get grid properties
    if self.grid_depth is None:
        self.grid_depth = int(1 / (1 - self.grid_relaxation) * 5)
    if self.grid_Nterp is None:
        self.grid_Nterp = self.Nlags * 10

    # Make list of lags for scanning
    self.lags = np.linspace(*self.stat_model.prior_ranges['lag'], self.Nlags + 1, endpoint=False)[1:]
    self.converged = np.zeros_like(self.lags, dtype=bool)

    free_dims = len(self.stat_model.free_params())
    self.scan_peaks = {key: np.array([]) for key in self.stat_model.paramnames()}
    self.diagnostic_hessians = []
    self.diagnostic_grads = []
    self.diagnostic_densities = []
    self.log_evidences_uncert = []

    self.params_toscan = [key for key in self.stat_model.paramnames() if
                          key not in ['lag'] and key in self.stat_model.free_params()
                          ]

    self.is_ready = True
estimate_MAP(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)

Parameters:

Name Type Description Default
lc_1 lightcurve
required
lc_2 lightcurve
required
seed int
None

Returns:

Type Description
Source code in litmus/fitting_methods.py 
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
def estimate_MAP(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    """
    :param lc_1:
    :param lc_2:
    :param seed:
    :return:
    """

    data = self.stat_model.lc_to_data(lc_1, lc_2)

    # ----------------------------------
    # Find seed for optimization if not supplies
    if self.stat_model.free_params() != self.seed_params.keys():
        seed_params, ll_start = self.stat_model.find_seed(data, guesses=self.init_samples, fixed=self.seed_params)

        self.msg_run("Beginning scan at constrained-space position:")
        for it in seed_params.items():
            self.msg_run('\t %s: \t %.2f' % (it[0], it[1]))
        self.msg_run(
            "Log-Density for this is: %.2f" % ll_start)

    else:
        seed_params = self.seed_params
        ll_start = self.stat_model.log_density(seed_params,
                                               data=data
                                               )

    # ----------------------------------
    # SCANNING FOR OPT

    self.msg_run("Moving non-lag params to new location...")
    estmap_params = self.stat_model.scan(start_params=seed_params,
                                         optim_params=[key for key in self.stat_model.free_params() if
                                                       key != 'lag'],
                                         data=data,
                                         optim_kwargs=self.optimizer_args_init,
                                         precondition=self.precondition
                                         )
    ll_firstscan = self.stat_model.log_density(estmap_params, data)
    if 'lag' in self.stat_model.free_params():
        self.msg_run("Optimizer settled at new fit:")
        for it in estmap_params.items():
            self.msg_run('\t %s: \t %.2f' % (it[0], it[1]))
        self.msg_run(
            "Log-Density for this is: %.2f" % ll_firstscan
        )

        self.msg_run("Finding a good lag...")
        test_lags = self.stat_model.prior_sample(self.init_samples)['lag']
        test_samples = _utils.dict_extend(estmap_params, {'lag': test_lags})
        ll_test = self.stat_model.log_density(test_samples, data)
        bestlag = test_lags[ll_test.argmax()]

        self.msg_run("Grid finds good lag at %.2f:" % bestlag)
        self.msg_run(
            "Log-Density for this is: %.2f" % ll_test.max()
        )

        if ll_test.max() > ll_firstscan:
            bestlag = bestlag
        else:
            bestlag = estmap_params['lag']

        estmap_params = self.stat_model.scan(start_params=estmap_params | {'lag': bestlag},
                                             optim_params=['lag'],
                                             data=data,
                                             optim_kwargs=self.optimizer_args_init,
                                             precondition=self.precondition
                                             )

        self.msg_run("Lag-only opt settled at new lag %.2f..." % estmap_params['lag'])

    ll_end = self.stat_model.log_density(estmap_params,
                                         data=data
                                         )

    # ----------------------------------
    # CHECKING OUTPUTS

    self.msg_run("Optimizer settled at new fit:")
    for it in estmap_params.items():
        self.msg_run('\t %s: \t %.2f' % (it[0], it[1]))
    self.msg_run(
        "Log-Density for this is: %.2f" % ll_end
    )

    # ----------------------------------
    # CHECKING OUTPUTS

    if ll_end < ll_start:
        self.msg_err("Warning! Optimization seems to have diverged. Defaulting to seed params. \n"
                     "Please consider running with different optim_init inputs")
        estmap_params = seed_params
    return estmap_params
make_grid(data, seed_params=None, interp_scale='log') -> _types.ArrayN

Generates a grid of test lags for use in the hessian scan via the grid smoothing algorithm listed in the paper

Parameters:

Name Type Description Default
data

data to condition the model on

 required
seed_params

An initial guess for the seed parameters formed by common sense. If None or incomplete, is filled using the statmodels find_seed method.

 None
interp_scale

What scale to perform interpolation between the test lags at, 'log' or 'linear'

 'log'

Returns:

Type Description
ArrayN

Array of lags of len self.Nlags
Source code in litmus/fitting_methods.py 
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
def make_grid(self, data, seed_params=None, interp_scale='log') -> _types.ArrayN:
    """
    Generates a grid of test lags for use in the hessian scan via the grid smoothing algorithm listed in the paper
    :param data: data to condition the model on
    :param seed_params: An initial guess for the seed parameters formed by common sense. If None or incomplete, is filled using the statmodels find_seed method.
    :param interp_scale: What scale to perform interpolation between the test lags at, 'log' or 'linear'
    :return: Array of lags of len self.Nlags
    """

    assert interp_scale in ['log', 'linear'], "Interp scale was %s, must be in 'log' or 'linear'" % interp_scale

    if not self.is_ready: self.readyup()

    if 'lag' in self.stat_model.fixed_params():
        lags = np.array([np.mean(self.stat_model.prior_ranges['lag'])])
        self.Nlags = 1
        self.readyup()
        return lags

    # If no seed parameters specified, use stored
    if seed_params is None:
        seed_params = self.estmap_params

    # If these params are incomplete, use find_seed to complete them
    if seed_params.keys() != self.stat_model.paramnames():
        seed_params, llstart = self.stat_model.find_seed(data, guesses=self.init_samples, fixed=seed_params)

    lags = np.linspace(*self.stat_model.prior_ranges['lag'], int(self.Nlags * self.grid_firstdepth) + 1,
                       endpoint=False)[1:]
    lag_terp = np.linspace(*self.stat_model.prior_ranges['lag'], self.grid_Nterp)

    log_density_all, lags_all = np.empty(shape=(1,)), np.empty(shape=(1,))
    for i in range(self.grid_depth):
        params = _utils.dict_extend(seed_params, {'lag': lags})
        log_density_all = np.concatenate([log_density_all, self.stat_model.log_density(params, data)])
        lags_all = np.concatenate([lags_all, lags])
        I = lags_all.argsort()
        log_density_all, lags_all = log_density_all[I], lags_all[I]

        if interp_scale == 'linear':

            density = np.exp(log_density_all - log_density_all.max())

            # Linearly interpolate the density profile
            density_terp = np.interp(lag_terp, lags_all, density, left=0, right=0)
            density_terp /= density_terp.sum()


        elif interp_scale == 'log':

            density = np.exp(log_density_all - log_density_all.max())

            # Linearly interpolate the density profile
            log_density_terp = np.interp(lag_terp, log_density_all - log_density_all.max(), density,
                                         left=log_density_all[0], right=log_density_all[-1])
            density_terp = np.exp(log_density_terp)
            density_terp /= density_terp.sum()

        gets = np.linspace(0, 1, self.grid_Nterp)
        percentiles = np.cumsum(density_terp) * self.grid_bunching + gets * (1 - self.grid_bunching)
        percentiles /= percentiles.max()

        lags = np.interp(np.linspace(0, 1, self.Nlags), percentiles, lag_terp,
                         left=lag_terp.min(),
                         right=lag_terp.max()
                         )

    return lags
prefit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
def prefit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # -------------------
    fitting_procedure.prefit(**locals())
    seed = self._tempseed
    # -------------------

    data = self.stat_model.lc_to_data(lc_1, lc_2)

    # ----------------------------------
    # Estimate the MAP

    self.estmap_params = self.estimate_MAP(lc_1, lc_2, seed)
    self.estmap_tol = self.stat_model.opt_tol(self.estmap_params, data,
                                              integrate_axes=self.stat_model.free_params())
    self.msg_run("Estimated to be within ±%.2eσ of local optimum" % self.estmap_tol)
    # ----------------------------------

    # Make a grid

    lags = self.make_grid(data, seed_params=self.estmap_params)
    if self.split_lags:
        split_index = abs(lags - self.estmap_params['lag']).argmin()
        lags_left, lags_right = lags[:split_index], lags[split_index:]
        lags = np.concatenate([lags_right, lags_left[::-1]])
        if self.reverse: lags = lags = np.concatenate([lags_left[::-1], lags_right])
    elif self.reverse:
        lags = lags[::-1]
    self.lags = lags

    self.has_prefit = True
fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # -------------------
    fitting_procedure.fit(**locals())
    seed = self._tempseed
    # -------------------
    # Setup + prefit if not run
    self.msg_run("Starting Hessian Scan")

    data = self.stat_model.lc_to_data(lc_1, lc_2)

    if not self.has_prefit:
        self.prefit(lc_1, lc_2, seed)
    best_params = self.estmap_params.copy()

    # ----------------------------------
    # Create scanner and perform setup
    params_toscan = self.params_toscan
    lags_forscan = self.lags.copy()

    solver, runsolver, [converter, deconverter, optfunc, runsolver_jit] = self.stat_model._scanner(data,
                                                                                                   optim_params=params_toscan,
                                                                                                   optim_kwargs=self.optimizer_args,
                                                                                                   return_aux=True
                                                                                                   )
    self.solver = solver
    x0, y0 = converter(best_params)
    state = solver.init_state(x0, y0, data)

    # ----------------------------------
    # Sweep over lags
    scanned_optima, grads, Hs = [], [], []
    tols, Zs, Ints, tgrads = [], [], [], []
    for i, lag in enumerate(lags_forscan):
        self.msg_run(":" * 23)
        self.msg_run("Scanning at lag=%.2f ..." % lag)

        # Get current param site in packed-function friendly terms
        opt_params, aux_data, state = runsolver_jit(solver, best_params | {'lag': lag}, state)

        # --------------
        # Check if the optimization has succeeded or broken

        l_1 = self.stat_model.log_density(best_params | {'lag': lag}, data)
        l_2 = self.stat_model.log_density(opt_params | {'lag': lag}, data)
        bigdrop = l_2 - l_1 < -self.LL_threshold
        diverged = np.any(np.isinf(np.array([x for x in self.stat_model.to_uncon(opt_params).values()])))

        self.msg_run("Change of %.2f against %.2f" % (l_2 - l_1, self.LL_threshold))

        if not bigdrop and not diverged:
            self.converged[i] = True

            is_good = [True, True, True]

            # ======
            # Check position & Grad
            try:
                uncon_params = self.stat_model.to_uncon(opt_params)
                log_height = self.stat_model.log_density_uncon(uncon_params, data)
            except:
                self.msg_err("Something wrong!")
                is_good[0] = False

            # ======
            # Check tolerances & hessians
            try:
                H = self.stat_model.log_density_uncon_hess(uncon_params, data, keys=params_toscan)
                assert np.linalg.det(H), "Error in H calc"
                tol = self.stat_model.opt_tol(opt_params, data, integrate_axes=params_toscan)
            except:
                self.msg_err("Something wrong in Hessian / Tolerance!:")
                is_good[1] = False

            # ======
            # Get evidence
            try:
                laplace_int = self.stat_model.laplace_log_evidence(opt_params, data,
                                                                   integrate_axes=params_toscan,
                                                                   constrained=self.constrained_domain)
                tgrad = self.stat_model.uncon_grad_lag(opt_params) if not self.constrained_domain else 0
                Z = laplace_int + tgrad
                assert not np.isnan(Z), "Error in Z calc"
            except:
                self.msg_err("Something wrong in Evidence!:")
                is_good[2] = False

            # Check and save if good
            if np.all(is_good):
                self.msg_run(
                    "Seems to have converged at iteration %i / %i with tolerance %.2e" % (i, self.Nlags, tol))

                if tol < 1.0:
                    best_params = opt_params
                else:
                    self.msg_run("Possibly stuck in a furrow. Resetting start params")
                    best_params = self.estmap_params.copy()

                scanned_optima.append(opt_params.copy())
                tols.append(tol)

                grads.append(aux_data['grad'])
                Hs.append(H)
                Ints.append(laplace_int)
                tgrads.append(tgrad)
                Zs.append(Z)
            else:
                self.msg_run("Seems to have severely diverged at iteration %i / %i" % (i, self.Nlags))
                reason = ["Eval", "hessian/tol", "evidence"]
                for a, b in zip(reason, is_good):
                    self.msg_run("%s:\t%r" % (a, b))

        else:
            self.converged[i] = False
            self.msg_run("Unable to converge at iteration %i / %i" % (i, self.Nlags),
                         "\nLarge Drop?:\t", bigdrop,
                         "\nOptimizer Diverged:\t", diverged)

    if sum(self.converged) == 0:
        self.msg_err("All slices catastrophically diverged! Try different starting conditions and/or grid spacing")

    self.msg_run("Scanning Complete. Calculating laplace integrals...")

    # --------
    # Save and apply
    self.diagnostic_grads = grads
    self.diagnostic_hessians = Hs
    self.diagnostic_tgrads = np.array(tgrads).squeeze().flatten()
    self.diagnostic_ints = np.array(Ints).squeeze().flatten()

    self.scan_peaks = _utils.dict_combine(scanned_optima)
    self.diagnostic_densities = self.stat_model.log_density(self.scan_peaks, data)
    self.log_evidences = np.array(Zs).squeeze().flatten()
    self.log_evidences_uncert = np.square(tols).squeeze().flatten()

    self.msg_run("Hessian Scan Fitting complete.", "-" * 23, "-" * 23, delim='\n')
    self.has_run = True
refit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
def refit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # -------------------
    fitting_procedure.fit(**locals())
    seed = self._tempseed
    # -------------------

    data = self.stat_model.lc_to_data(lc_1, lc_2)

    peaks = _utils.dict_divide(self.scan_peaks)
    I = np.arange(len(peaks))
    select = np.argwhere(self.log_evidences_uncert > self.opt_tol).squeeze()
    if not (_utils.isiter(select)): select = np.array([select])

    peaks, I = np.array(peaks)[select], I[select]

    self.msg_run("Doing re-fitting of %i lags" % len(peaks))
    newtols = []
    for j, i, peak in zip(range(len(I)), I, peaks):

        self.msg_run(":" * 23, "Refitting lag %i/%i at lag %.2f" % (j, len(peaks), peak['lag']), delim='\n')

        ll_old = self.stat_model.log_density(peak, data)
        old_tol = self.log_evidences_uncert[i]

        new_peak = self.stat_model.scan(start_params=peak,
                                        optim_params=self.params_toscan,
                                        data=data,
                                        optim_kwargs=self.optimizer_args,
                                        precondition=self.precondition
                                        )
        ll_new = self.stat_model.log_density(new_peak, data)
        if ll_old > ll_new or np.isnan(ll_new):
            self.msg_err("New peak bad (LL from %.2e to %.2e. Trying new start location" % (ll_old, ll_new))
            new_peak = self.stat_model.scan(start_params=self.estmap_params,
                                            optim_params=self.params_toscan,
                                            data=data,
                                            optim_kwargs=self.optimizer_args,
                                            precondition=self.precondition
                                            )
            ll_new = self.stat_model.log_density(new_peak, data)

            if ll_old > ll_new:
                self.msg_err("New peak bad (LL from %.2e to %.2e. Trying new start location" % (ll_old, ll_new))
                continue

        new_peak_uncon = self.stat_model.to_uncon(new_peak)
        new_grad = _utils.dict_pack(new_grad, keys=self.params_toscan)
        new_hessian = self.stat_model.log_density_uncon_hess(new_peak_uncon, data, keys=self.params_toscan)

        try:
            int = self.stat_model.laplace_log_evidence(new_peak, data, constrained=self.constrained_domain)
            tgrad = self.stat_model.uncon_grad_lag(new_peak)
            Z = tgrad + int
            Hinv = np.linalg.inv(new_hessian)
        except:
            self.msg_run("Optimization failed on %i/%i" % (j, len(peaks)))
            continue

        tol = self.stat_model.opt_tol(new_peak, data, self.params_toscan)

        if tol < old_tol:
            self.diagnostic_tgrads[i] = tgrad
            self.diagnostic_ints[i] = int
            self.log_evidences[i] = tgrad + int

            self.diagnostic_grads[i] = new_grad
            self.diagnostic_hessians[i] = new_hessian
            self.log_evidences_uncert[i] = tol ** 2
            self.msg_run("Settled at new tol %.2e" % tol)
        else:
            self.msg_run(
                "Something went wrong at this refit! Consider changing the optimizer_args and trying again")
    self.msg_run("Refitting complete.")
diagnostics(plot=True) -> _types.Figure
Source code in litmus/fitting_methods.py 
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
def diagnostics(self, plot=True) -> _types.Figure:

    loss = self.log_evidences_uncert

    lagplot = self.scan_peaks['lag']
    I = self.scan_peaks['lag'].argsort()
    lagplot = lagplot[I]
    Y = self.estmap_tol[I]

    # ---------
    fig = plt.figure()
    plt.ylabel("Loss Norm, $ \\vert \Delta x / \sigma_x \\vert$")
    plt.xlabel("Scan Lag No.")
    plt.plot(lagplot, loss, 'o-', c='k', label="Scan Losses")
    plt.scatter(self.estmap_params['lag'], Y, c='r', marker='x', s=40, label="Initial MAP Scan Loss")
    plt.axhline(self.opt_tol, ls='--', c='k', label="Nominal Tolerance Limit")
    plt.legend(loc='best')

    fig.text(.5, .05, "How far each optimization slice is from its peak. Lower is good.", ha='center')
    plt.yscale('log')
    plt.grid()
    plt.show()

    return fig
diagnostic_lagplot(show=True) -> _types.Figure
Source code in litmus/fitting_methods.py 
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
def diagnostic_lagplot(self, show=True) -> _types.Figure:
    f, (a1, a2) = plt.subplots(2, 1, sharex=True)

    lags_forint, logZ_forint, density_forint = self._get_slices('lags', 'logZ', 'densities')

    # ---------------------
    for a in (a1, a2):
        a.scatter(lags_forint, np.exp(logZ_forint - logZ_forint.max()), label="Evidence")
        a.scatter(lags_forint, np.exp(density_forint - density_forint.max()), label="Density")
        a.grid()

    a2.set_yscale('log')
    a1.legend()

    # --------------
    # Outputs
    if show: plt.show()
    return f
get_evidence(seed: int = None, return_type='linear') -> [float, float, float]
Source code in litmus/fitting_methods.py 
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
def get_evidence(self, seed: int = None, return_type='linear') -> [float, float, float]:
    # -------------------
    fitting_procedure.get_evidence(**locals())
    seed = self._tempseed
    # -------------------

    assert self.interp_scale in self._allowable_interpscales, "Interp scale %s not recognised. Must be selection from %s" % (
        self.interp_scale, self._allowable_interpscales)

    lags_forint, logZ_forint, logZ_uncert_forint = self._get_slices('lags', 'logZ', 'dlogZ')
    minlag, maxlag = self.stat_model.prior_ranges['lag']

    if maxlag - minlag == 0:
        Z = np.exp(logZ_forint.max())
        imax = logZ_forint.argmax()
        uncert_plus, uncert_minus = logZ_uncert_forint[imax], logZ_uncert_forint[imax]


    else:

        if self.interp_scale == 'linear':
            dlag = [*np.diff(lags_forint) / 2, 0]
            dlag[1:] += np.diff(lags_forint) / 2
            dlag[0] += lags_forint.min() - minlag
            dlag[-1] += maxlag - lags_forint.max()

            dlogZ = logZ_forint + np.log(dlag)
            dZ = np.exp(dlogZ - dlogZ.max())
            Z = dZ.sum() * np.exp(dlogZ.max())

            # -------------------------------------
            # Get Uncertainties

            # todo Fix this to be generic and move outside of scope
            # Estimate uncertainty from ~dt^2 error scaling
            dlag_sub = [*np.diff(lags_forint[::2]) / 2, 0]
            dlag_sub[1:] += np.diff(lags_forint[::2]) / 2
            dlag_sub[0] += lags_forint.min() - minlag
            dlag_sub[-1] += maxlag - lags_forint.max()

            dlogZ_sub = logZ_forint[::2] + np.log(dlag_sub)
            dZ_sub = np.exp(dlogZ_sub - dlogZ_sub.max())
            Z_subsample = dZ_sub.sum() * np.exp(dlogZ_sub.max())
            uncert_numeric = abs(Z - Z_subsample) / np.sqrt(17)

            uncert_tol = np.square(dZ * logZ_uncert_forint).sum()
            uncert_tol = np.sqrt(uncert_tol)
            uncert_tol *= np.exp(dlogZ.max())


        elif self.interp_scale == 'log':
            # dZ = dXdY/dln|Y|
            dlag = np.diff(lags_forint)
            dY = np.diff(np.exp(logZ_forint - logZ_forint.max()))
            dE = np.diff(logZ_forint)
            dZ = dlag * dY / dE
            Z = np.sum(dZ) * np.exp(logZ_forint.max())

            uncert_tol = 4 * np.square(
                np.exp(logZ_forint - logZ_forint.max())[:-1] - dZ / dE
            ) * logZ_uncert_forint[:-1]
            uncert_tol += np.square(logZ_uncert_forint[-1] * np.exp(logZ_forint - logZ_forint.max())[-1])
            uncert_tol = np.sqrt(uncert_tol.sum())
            uncert_tol *= np.exp(logZ_forint.max())

            dlag_sub = np.diff(lags_forint[::2])
            dY_sub = np.diff(np.exp(logZ_forint[::2] - logZ_forint.max()))
            dE_sub = np.diff(logZ_forint[::2])
            dZ_sub = dlag_sub * dY_sub / dE_sub
            Z_subsample = np.sum(dZ_sub) * np.exp(logZ_forint.max())
            uncert_numeric = abs(Z - Z_subsample) / np.sqrt(17)

        self.msg_debug("Evidence Est: \t %.2e" % Z)
        self.msg_debug(
            "Evidence uncerts: \n Numeric: \t %.2e \n Convergence: \t %.2e" % (uncert_numeric, uncert_tol))

        uncert_plus = uncert_numeric + uncert_tol.sum()
        uncert_minus = uncert_numeric

    if return_type == 'linear':
        return np.array([Z, -uncert_minus, uncert_plus])
    elif return_type == 'log':
        return np.array([np.log(Z), np.log(1 - uncert_minus / Z), np.log(1 + uncert_plus / Z)])
get_samples(N: int = 1, seed: int = None, importance_sampling: bool = False) -> {str: [float]}
Source code in litmus/fitting_methods.py 
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
def get_samples(self, N: int = 1, seed: int = None, importance_sampling: bool = False) -> {str: [float]}:
    # -------------------
    fitting_procedure.get_samples(**locals())
    seed = self._tempseed
    # -------------------

    assert self.interp_scale in self._allowable_interpscales, "Interp scale %s not recognised. Must be selection from %s" % (
        self.interp_scale, self._allowable_interpscales)
    lags_forint, logZ_forint, peaks, covars = self._get_slices('lags', 'logZ', 'peaks', 'covars')

    if len(lags_forint) == 0:
        self.msg_err("Zero good slices in evidence integral!")
        return (0, -np.inf, np.inf)

    # Get weights and peaks etc
    Npeaks = len(lags_forint)
    minlag, maxlag = self.stat_model.prior_ranges['lag']

    Y = np.exp(logZ_forint - logZ_forint.max()).squeeze()

    dlag = [*np.diff(lags_forint) / 2, 0]
    dlag[1:] += np.diff(lags_forint) / 2
    dlag[0] += lags_forint.min() - minlag
    dlag[-1] += maxlag - lags_forint.max()

    if sum(dlag) == 0:
        dlag = 1.0

    weights = Y * dlag
    weights /= weights.sum()

    # Get hessians and peak locations
    if Npeaks > 1:
        I = np.random.choice(range(Npeaks), N, replace=True, p=weights)
    else:
        I = np.zeros(N)

    to_choose = [(I == i).sum() for i in range(Npeaks)]  # number of samples to draw from peak i

    # Sweep over scan peaks and add scatter
    outs = []
    for i in range(Npeaks):
        if to_choose[i] > 0:
            peak_uncon = self.stat_model.to_uncon(peaks[i])

            # Get normal dist properties in uncon space in vector form
            mu = _utils.dict_pack(peak_uncon, keys=self.params_toscan)
            cov = covars[i]

            # Generate samples
            samps = np.random.multivariate_normal(mean=mu, cov=cov, size=to_choose[i])
            samps = _utils.dict_unpack(samps.T, keys=self.params_toscan, recursive=False)
            samps = _utils.dict_extend(peak_uncon, samps)

            # Reconvert to constrained space
            samps = self.stat_model.to_con(samps)

            # -------------
            # Add linear interpolation 'smudging' to lags
            if Npeaks > 1 and 'lag' in self.stat_model.free_params():

                # Get nodes
                tnow, ynow = lags_forint[i], Y[i]
                if i == 0:
                    yprev, ynext = ynow, Y[i + 1]
                    tprev, tnext = min(self.stat_model.prior_ranges['lag']), lags_forint[i + 1]
                elif i == Npeaks - 1:
                    yprev, ynext = Y[i - 1], ynow
                    tprev, tnext = lags_forint[i - 1], max(self.stat_model.prior_ranges['lag'])
                else:
                    yprev, ynext = Y[i - 1], Y[i + 1]
                    tprev, tnext = lags_forint[i - 1], lags_forint[i + 1]
                # --

                # Perform CDF shift
                Ti, Yi = [tprev, tnow, tnext], [yprev, ynow, ynext]
                if self.interp_scale == 'linear':
                    tshift = linscatter(Ti, Yi, N=to_choose[i])
                elif self.interp_scale == 'log':
                    tshift = expscatter(Ti, Yi, N=to_choose[i])
                if np.isnan(tshift).any():
                    self.msg_err("Something wrong with the lag shift at node %i in sample generation" % i)
                else:
                    samps['lag'] += tshift
            # -------------

            if np.isnan(samps['lag']).any():
                self.msg_err("Something wrong with the lags at node %i in sample generation" % i)
            else:
                outs.append(samps)

    outs = {key: np.concatenate([out[key] for out in outs]) for key in self.stat_model.paramnames()}
    return (outs)

SVI_scan(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=False, **fit_params)

Bases: hessian_scan

An alternative to hessian_scan that fits each slice with stochastic variational inference instead of the laplace approximation. Typically slower, but more robust against numerical failure in low SNR signals and gives more accurate evidence estimates.

Inherits all fitting parameters and their default values from hessian_scan, but gains new parameters

Parameters:

Name Type Description Default
ELBO_threshold float

If a slice log-evidence decreases by this amount or more, consider it a furrow. Defaults to 100.0.

 required
ELBO_optimstep float

Size of the stochastic optimisation step in adam. Defaults to 5E-3.

 required
ELBO_particles int

Number of particles for estimating the ELBO at each optimisation step. Defaults to 128.

 required
ELBO_Nsteps int

Number of steps to take in optimisation of the ELBO for each slice. Defaults to 128.

 required
ELBO_Nsteps_init int

Number of steps to take in finding the initial slice ELBO / solution. Defaults to 1_000.

 required
ELBO_fraction int

Fraction of the ELBO run (both slice and initial) to search for minimum variance estimate. Defaults to 0.25.

 required
Source code in litmus/fitting_methods.py 
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout, err_stream=sys.stderr,
             verbose=True, debug=False, **fit_params):

    """
    Inherits all fitting parameters and their default values from hessian_scan, but gains new parameters

    :param float ELBO_threshold: If a slice log-evidence decreases by this amount or more, consider it a furrow. Defaults to 100.0.
    :param float ELBO_optimstep: Size of the stochastic optimisation step in adam. Defaults to 5E-3.
    :param int ELBO_particles: Number of particles for estimating the ELBO at each optimisation step. Defaults to 128.
    :param int ELBO_Nsteps: Number of steps to take in optimisation of the ELBO for each slice. Defaults to 128.
    :param int ELBO_Nsteps_init: Number of steps to take in finding the initial slice ELBO / solution. Defaults to 1_000.
    :param int ELBO_fraction: Fraction of the ELBO run (both slice and initial) to search for minimum variance estimate. Defaults to 0.25.
    """
    args_in = {**locals(), **fit_params}
    del args_in['self']
    del args_in['__class__']
    del args_in['fit_params']

    if not hasattr(self, '_default_params'):
        self._default_params = {}

    self._default_params |= {
        'ELBO_threshold': 100.0,
        'ELBO_optimstep': 5E-3,
        'ELBO_particles': 128,
        'ELBO_Nsteps': 128,
        'ELBO_Nsteps_init': 1_000,
        'ELBO_fraction': 0.25,
    }

    super().__init__(**args_in)

    # -----------------------------

    self.name = "SVI Scan Fitting Procedure"

    self.diagnostic_losses = []
    self.diagnostic_loss_init = []
    self.diagnostic_ints = []
name = 'SVI Scan Fitting Procedure' instance-attribute
diagnostic_losses = [] instance-attribute
diagnostic_loss_init = [] instance-attribute
diagnostic_ints = [] instance-attribute
fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # -------------------
    fitting_procedure.fit(**locals())
    seed = self._tempseed
    # -------------------

    self.msg_run("Starting SVI Scan")

    if not self.has_prefit:
        self.prefit(lc_1, lc_2, seed)

    data = self.stat_model.lc_to_data(lc_1, lc_2)

    # ----------------------------------
    # Estimate the MAP and its hessian for starting conditions

    estmap_uncon = self.stat_model.to_uncon(self.estmap_params)

    fix_param_dict_con = {key: self.estmap_params[key] for key in self.stat_model.fixed_params()}
    fix_param_dict_uncon = {key: estmap_uncon[key] for key in self.stat_model.fixed_params()}

    init_hess = -1 * self.stat_model.log_density_uncon_hess(estmap_uncon, data=data, keys=self.params_toscan)

    # Convert these into SVI friendly objects and fit an SVI at the map
    self.msg_run("Performing SVI slice at the MAP estimate")
    init_loc = _utils.dict_pack(estmap_uncon, keys=self.params_toscan)
    init_tril = jnp.linalg.cholesky(jnp.linalg.inv(init_hess))

    bad_starts = False
    if np.isnan(init_loc).any() or np.isnan(init_tril).any():
        self.msg_err("Issue with finding initial solver state for SVI. Proceeding /w numpyro defaults")
        bad_starts = True
    # ----------------------------------
    self.msg_debug("\t Constructing slice model")

    def slice_function(data, lag):
        """
        This is the conditional model that SVI will map
        """

        params = {}
        for key in self.stat_model.free_params():
            if key != 'lag':
                val = quickprior(self.stat_model, key)
                params |= {key: val}
        params |= {'lag': lag}
        params |= fix_param_dict_con

        with numpyro.handlers.block(hide=self.stat_model.paramnames()):
            LL = self.stat_model._log_likelihood(params, data)

        dilute = -np.log(self.stat_model.prior_ranges['lag'][1] - self.stat_model.prior_ranges['lag'][
            0]) if 'lag' in self.stat_model.free_params() else 0.0
        numpyro.factor('lag_prior', dilute)

    # SVI settup
    self.msg_debug("\t Constructing and running optimizer and SVI guides")
    optimizer = numpyro.optim.Adam(step_size=self.ELBO_optimstep)
    autoguide = numpyro.infer.autoguide.AutoMultivariateNormal(slice_function)
    autosvi = numpyro.infer.SVI(slice_function, autoguide, optim=optimizer,
                                loss=numpyro.infer.Trace_ELBO(self.ELBO_particles),
                                )

    self.msg_debug("\t Running SVI")
    MAP_SVI_results = autosvi.run(jax.random.PRNGKey(seed), self.ELBO_Nsteps_init,
                                  data=data, lag=self.estmap_params['lag'],
                                  init_params={'auto_loc': init_loc,
                                               'auto_scale_tril': init_tril
                                               } if not bad_starts else None
                                  )

    self.msg_debug("\t Success. Extracting solution")
    BEST_loc, BEST_tril = MAP_SVI_results.params['auto_loc'], MAP_SVI_results.params['auto_scale_tril']

    self.diagnostic_loss_init = MAP_SVI_results.losses

    # ----------------------------------
    # Main Scan

    lags_forscan = self.lags
    l_old = -np.inf

    scanned_optima = []
    ELBOS_tosave = []
    ElBOS_uncert = []
    diagnostic_hessians = []
    diagnostic_losses = []

    for i, lag in enumerate(lags_forscan):
        print(":" * 23)
        self.msg_run("Doing SVI fit at lag=%.2f ..." % lag)

        svi_loop_result = autosvi.run(jax.random.PRNGKey(seed),
                                      self.ELBO_Nsteps,
                                      data=data, lag=lag,
                                      init_params={'auto_loc': BEST_loc,
                                                   'auto_scale_tril': BEST_tril
                                                   },
                                      progress_bar=False
                                      )

        NEW_loc, NEW_tril = svi_loop_result.params['auto_loc'], svi_loop_result.params['auto_scale_tril']

        # --------------
        # Check if the optimization has suceeded or broken

        l_old = l_old
        l_new = self._getELBO(svi_loop_result.losses)[0]
        diverged = bool(np.isinf(NEW_loc).any() + np.isinf(NEW_tril).any())
        big_drop = l_new - l_old < - self.ELBO_threshold

        self.msg_run(
            "From %.2f to %.2f, change of %.2f against %.2f" % (l_old, l_new, l_new - l_old, self.ELBO_threshold))

        if not big_drop and not diverged:
            self.msg_run("Seems to have converged at iteration %i / %i" % (i, self.Nlags))

            self.converged[i] = True
            l_old = l_new
            BEST_loc, BEST_tril = NEW_loc, NEW_tril

            uncon_params = self.stat_model.to_uncon(self.estmap_params | {'lag': lag}) | _utils.dict_unpack(NEW_loc,
                                                                                                            self.params_toscan)
            con_params = self.stat_model.to_con(uncon_params)
            scanned_optima.append(con_params)

            H = np.dot(NEW_tril, NEW_tril.T)
            H = (H + H.T) / 2
            H = jnp.linalg.inv(-H)
            diagnostic_hessians.append(H)

            diagnostic_losses.append(svi_loop_result.losses)

            ELBO, uncert = self._getELBO(svi_loop_result.losses)
            ELBOS_tosave.append(ELBO)
            ElBOS_uncert.append(uncert)


        else:
            self.msg_run("Unable to converge at iteration %i / %i" % (i, self.Nlags))
            self.msg_debug("Reason for failure: \n large ELBO drop: \t %r \n diverged: \t %r" % (
                big_drop, diverged))

    self.msg_run("Scanning Complete. Calculating ELBO integrals...")
    if sum(self.converged) == 0:
        self.msg_err("All slices catastrophically diverged! Try different starting conditions and/or grid spacing")

    self.diagnostic_ints = np.array(ELBOS_tosave)

    self.log_evidences_uncert = np.array(ElBOS_uncert)
    self.diagnostic_losses = np.array(diagnostic_losses)
    self.diagnostic_hessians = np.array(diagnostic_hessians)

    self.scan_peaks = _utils.dict_combine(scanned_optima)
    self.diagnostic_densities = self.stat_model.log_density(self.scan_peaks, data)

    # ---------------------------------------------------------------------------------
    # For each of these peaks, estimate the evidence
    # todo - vmap and parallelize

    Zs, tgrads = [], []
    for j, params in enumerate(scanned_optima):
        Z = self.diagnostic_ints[j]
        tgrad = self.stat_model.uncon_grad_lag(params) if not self.constrained_domain else 0
        tgrad = 0
        tgrads.append(tgrad)
        Zs.append(Z + tgrad)

    self.log_evidences = np.array(Zs)
    self.diagnostic_tgrads = np.array(tgrads)
    self.has_run = True

    self.msg_run("SVI Fitting complete.", "-" * 23, "-" * 23, delim='\n')
refit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
1847
1848
1849
1850
def refit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    # TODO - fill this out

    return
diagnostics(plot=True) -> _types.Figure
Source code in litmus/fitting_methods.py 
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
def diagnostics(self, plot=True) -> _types.Figure:

    f, (a2, a1) = plt.subplots(2, 1)

    for i, x in enumerate(self.diagnostic_losses):
        a1.plot(x - (self.diagnostic_ints[i]), c='k', alpha=0.25)
    a2.plot(self.diagnostic_loss_init, c='k')

    a1.axvline(int((1 - self.ELBO_fraction) * self.ELBO_Nsteps), c='k', ls='--')

    a1.set_yscale('symlog')
    a2.set_yscale('symlog')
    a1.grid(), a2.grid()

    a1.set_xlim(0, self.ELBO_Nsteps)
    a2.set_xlim(0, self.ELBO_Nsteps_init)

    a1.set_title("Scan SVIs")
    a2.set_title("Initial MAP SVI")

    f.supylabel("Loss - loss_final (log scale)")
    a1.set_xlabel("iteration Number"), a2.set_xlabel("iteration Number")

    txt = "Trace plots of ELBO convergence. All lines should be flat by the right hand side.\n" \
          "Top panel is for initial guess and need only be flat. Bottom panel should be flat within" \
          "averaging range, i.e. to the right of dotted line."
    f.supxlabel('$\begin{center}X-axis\\*\textit{\small{%s}}\end{center}$' % txt)

    f.tight_layout()

    if plot: plt.show()

    return f

JAVELIKE(stat_model: stats_model, out_stream=sys.stdout, err_stream=sys.stderr, verbose=True, debug=False, **fit_params)

Bases: fitting_procedure

A direct MCMC implementation using the AEIS in the style of JAVELIN Note that, because NumPyro fits in the unconstrained domain while JAVELIN fits in the constrained domain, the behaviour of the two will be slightly different near the prior boundaries.

Note that this is for example / comparison only, and should not be used for actual fitting as it cannot handle the multimodal distributions of seasonal lightcurves

Parameters:

Name Type Description Default
alpha float

Size of the stretch in the stretch-move. Defaults to 2.0.

 required
num_chains float

Num live points in the AEIS ensemble. Defaults to 256.

 required
num_samples float

Samples per live point in the AEIS chain. Defaults to 200_000 // 256 per chain total (i.e. 200_000 total)

 required
num_warmup float

Number of warmup samples per chain. Defaults to 5_000.

 required
Source code in litmus/fitting_methods.py 
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
def __init__(self, stat_model: stats_model,
             out_stream=sys.stdout, err_stream=sys.stderr,
             verbose=True, debug=False, **fit_params):

    """
    :param float alpha: Size of the stretch in the stretch-move. Defaults to 2.0.
    :param float num_chains: Num live points in the AEIS ensemble. Defaults to 256.
    :param float num_samples: Samples per live point in the AEIS chain. Defaults to 200_000 // 256 per chain total (i.e. 200_000 total)
    :param float num_warmup: Number of warmup samples per chain. Defaults to 5_000.
    """
    args_in = {**locals(), **fit_params}
    del args_in['self']
    del args_in['__class__']
    del args_in['fit_params']

    if not hasattr(self, '_default_params'):
        self._default_params = {}
    self._default_params |= {
        'alpha': 2.0,
        'num_chains': 256,
        'num_samples': 200_000 // 256,
        'num_warmup': 5_000,
    }

    self.sampler: numpyro.infer.MCMC = None
    """NumPyro MCMC wrapper"""
    self.kernel: numpyro.infer.AEIS = None
    """numpyro MCMC sampler kernel to use"""
    self.limited_model: Callable = None
    """The function to deploy the AEIS against"""

    super().__init__(**args_in)

    # -----------------------------

    self.name = "AEIS JAVELIN Emulator fitting Procedure"
sampler: numpyro.infer.MCMC = None instance-attribute

NumPyro MCMC wrapper

kernel: numpyro.infer.AEIS = None instance-attribute

numpyro MCMC sampler kernel to use

limited_model: Callable = None instance-attribute

The function to deploy the AEIS against

name = 'AEIS JAVELIN Emulator fitting Procedure' instance-attribute
readyup()
Source code in litmus/fitting_methods.py 
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
def readyup(self):

    fixed_vals = {key: self.stat_model.prior_ranges[key][0] for key in self.stat_model.fixed_params()}

    def limited_model(data):
        with numpyro.handlers.block(hide=self.stat_model.fixed_params()):
            params = {key: val for key, val in zip(self.stat_model.paramnames(), self.stat_model.prior())}

        params |= fixed_vals
        with numpyro.handlers.block(hide=self.stat_model.paramnames()):
            LL = self.stat_model._log_density(params, data)

        numpyro.factor('ll', LL)

    self.limited_model = limited_model

    self.kernel = numpyro.infer.AIES(self.limited_model,
                                     moves={numpyro.infer.AIES.StretchMove(a=self.alpha): 1.0}
                                     )

    self.sampler = numpyro.infer.MCMC(self.kernel,
                                      num_warmup=self.num_warmup,
                                      num_samples=self.num_samples,
                                      num_chains=self.num_chains,
                                      chain_method='vectorized',
                                      progress_bar=self.verbose)

    self.is_ready = True
prefit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
1991
1992
1993
1994
1995
1996
1997
1998
def prefit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    if seed is None: seed = self.seed
    if not self.is_ready: self.readyup()

    self.msg_run("Running warmup with %i chains and %i samples" % (self.num_chains, self.num_warmup))
    # self.sampler.warmup(jax.random.PRNGKey(seed), self.stat_model.lc_to_data(lc_1, lc_2))

    self.has_prefit = True
fit(lc_1: lightcurve, lc_2: lightcurve, seed: int = None)
Source code in litmus/fitting_methods.py 
2000
2001
2002
2003
2004
2005
2006
2007
2008
def fit(self, lc_1: lightcurve, lc_2: lightcurve, seed: int = None):
    if seed is None: seed = self.seed
    if not self.is_ready: self.readyup()
    if not self.has_prefit: self.prefit(lc_1, lc_2, seed=seed)

    self.msg_run("Running sampler with %i chains and %i samples" % (self.num_chains, self.num_samples))

    self.sampler.run(jax.random.PRNGKey(seed), self.stat_model.lc_to_data(lc_1, lc_2))
    self.has_run = True
get_samples(N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}
Source code in litmus/fitting_methods.py 
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
def get_samples(self, N: int = None, seed: int = None, importance_sampling: bool = False) -> {str: [float]}:
    if seed is None: seed = self.seed
    if not self.has_run:
        self.msg_err("Can't get samples before running!")
    if importance_sampling:
        self.msg_err("JAVELIKE Already distributed according to posterior (ideally)")
    samps = self.sampler.get_samples()

    if not (N is None):
        M = _utils.dict_dim(samps)[1]
        if M > N: self.msg_err("Tried to get %i sub-samples from chain of %i total samples." % (M, N))

        I = np.random.choice(np.arange(M), N, replace=True)
        samps = {key: samps[key][I] for key in samps.keys()}
    return (samps)