Glasnost

Minimalist maximum likelihood fitter in (mostly) NumPy, with extensibility for Bayesian MCMC fits. Direct access to parameters hopefully simplifies simultaneous fits where parameters are shared, and additional constraints are implemented directly into the likelihood.

Model specification

Model specification is done within name_scopes (à la TensorFlow), such that each model component and parameter is named according to the scope in which it resides, but this doesn't have to be manually constructed by the user.

with gl.name_scope("massFit"):
    with gl.name_scope("coreGaussian"):
        m = gl.Parameter(name = 'mean') # Full name is 'massFit/coreGaussian/mean'
        s = gl.Parameter(name = 'sigma') # Full name is 'massFit/coreGaussian/sigma'

        gCore = gl.Gaussian(m, s) # Full name is 'massFit/coreGaussian'

Composite models can be specifed using relative normalisations

def doubleGaussianFracModel(mean1, width1, frac, mean2, width2, nEvents):

    with gl.name_scope('doubleGaussianFracModel'):

        with gl.name_scope('gauss1'):

            m1 = gl.Parameter(mean1, name = 'mean', minVal = 4200, maxVal = 5700)
            s1 = gl.Parameter(width1, name = 'sigma', minVal = 0, maxVal = width1 * 5)

            gauss1 = gl.Gaussian({'mean' : m1, 'sigma' : s1})

        with gl.name_scope('gauss2'):

            m2 = gl.Parameter(mean2, name = 'mean', minVal = 4200, maxVal = 5700)
            s2 = gl.Parameter(width2, name = 'sigma', minVal = 0, maxVal = width2 * 5)

            gauss2 = gl.Gaussian({'mean' : m2, 'sigma' : s2})

        gaussFrac = gl.Parameter(frac, name = 'gaussFrac', minVal = 0.0, maxVal = 1.0)
        totalYield = gl.Parameter(nEvents, name = 'totalYield',
                                  minVal = 0.8 * nEvents,
                                  maxVal = 1.2 * nEvents)

    fitComponents = {gauss1.name : gauss1, gauss2.name : gauss2}
    doubleGaussian = gl.Model(initialFitFracs = {gauss1.name : gaussFrac},
                              initialFitComponents = fitComponents,
                              minVal = 5280, maxVal = 5700)

    model = gl.Model(initialFitYields = {doubleGaussian.name : totalYield},
                                         initialFitComponents = {doubleGaussian.name : doubleGaussian},
                                         minVal = 5280, maxVal = 5700)

    return model

or with explicit yields

with gl.name_scope('doubleGaussianYieldsModel'):

...

    with gl.name_scope('gauss1'):
    
        ...

        gauss1Yield = gl.Parameter(nEvents1, name = 'gauss1Yield',
                                   minVal = 0.8 * nEvents1, maxVal = 1.2 * nEvents1)

    with gl.name_scope('gauss2'):

        ...

        gauss2Yield = gl.Parameter(nEvents2, name = 'gauss2Yield',
                                   minVal = 0.8 * nEvents2, maxVal = 1.2 * nEvents2)

fitYields = {gauss1.name : gauss1Yield, gauss2.name : gauss2Yield}
fitComponents = {gauss1.name : gauss1, gauss2.name : gauss2}

model = gl.Model(initialFitYields = fitYields,
                 initialFitComponents = fitComponents,
                 minVal = 5300, maxVal = 5700)

Simultaneous models

Simultaneous fits to multiple (one-dimensional) spectra can be performed using SimultaneousModel and passing a dictionary of models and yield parameters.

fitYields1 = {gauss1.name : gauss1Yield}
fitComponents1 = {gauss1.name : gauss1}

model1 = gl.Model(name = 's1',
                  initialFitYields = fitYields1,
                  initialFitComponents = fitComponents1,
                  minVal = 5000, maxVal = 5600)

fitYields2 = {gauss2.name : gauss2Yield}
fitComponents2 = {gauss2.name : gauss2}

model2 = gl.Model(name = 's2',
                  initialFitYields = fitYields2,
                  initialFitComponents = fitComponents2,
                  minVal = 5000, maxVal = 5600)

model = gl.SimultaneousModel(name = 's',
                             initialFitComponents = {model1.name : model1, model2.name : model2})

Constraints

Constraints on model parameters can either be external, by providing a prior distribution on the parameter that reduces to a float passed at construction time

m = gl.Parameter(mean, name = 'mean', minVal = 4200, maxVal = 5700)
meanConstraint = gl.Gaussian({'mean' : 5400., 'sigma' : 0.01})
m.priorDistribution = meanConstraint # Constrain the parameter according to a Gaussian distribution

They can also be defined relative to another model parameter, by passing an arithmetic function string that describes a transformation of the other parameter, and the definition of the parameter, to the constructor

with gl.name_scope('gauss1'):

    m1 = gl.Parameter('m', name = 'mean1')
    s1 = gl.Parameter(width, name = 'sigma1', minVal = 0, maxVal = width * 5)

    gauss1 = gl.Gaussian({'mean' : m1, 'sigma' : s1})

with gl.name_scope('gauss2'):

    m2 = gl.Parameter('m1 + 100.', name = 'mean2', m1 = m1) # m2 is defined to be m1 + 100
    s2 = gl.Parameter('s1 / 2.', name = 'sigma2', s1 = s1) # s2 is defined to be s1 / 2

    gauss2 = gl.Gaussian({'mean' : m2, 'sigma' : s2})

Inference

Model parameters can be inferred from the data using Fitter, which takes backend arguments of minuit for maximum likelihood fitting with Minuit, or emcee for Bayesian MCMC with Emcee.

fitterML = gl.Fitter(model, backend = 'minuit') # Maximum likelihood
fitterBayes = gl.Fitter(model, backend = 'emcee') # MCMC

resML = fitterML.fit(data)
resBayes = fitterBayes.fit(data, nIterations = 1000, nWalkers = 10)

By default, the maxium likelihood value or the maximum of the posterior distribution for each model parameter can be accessed using param.value and the 68% interval around the maximum likelihood value (or the maxium of the posterior) can be accessed using param.error. For Bayesian MCMC, the MCMC chains can be accessed using res.chain

samples = res.chain[:, 200:, :].reshape((-1, model.getNFloatingParameters()))

Plotting

Plotting functionality is provided by the Plotter class, which plots normalised models on data histograms

plotter = gl.Plotter(model, data)
plotter.plotDataModel(nDataBins = 30)

Supported distributions

Gaussian
Uniform
Crystal-Ball
Exponential
Student's t
Beta

dpohanlon / Glasnost