Skip to content

Latest commit

 

History

History
373 lines (314 loc) · 19 KB

README.md

File metadata and controls

373 lines (314 loc) · 19 KB

SwissFit

Python License: GPL v2 Documentation Status

SwissFit is a general-purpose library for fitting models to data with Gaussian-distributed noise. The design of this library is inspired by Peter Lepage's lsqfit and operates in a similar manner. As such, it builds on top of the GVar library and extensively utilizes the powerful numerical tools of SciPy and Vegas to extract model parameters and their associated statistical uncertainties from maximum a posteriori estimation (i.e., traditional least squares fitting with Gaussian priors) or Markov chain Monte Carlo sampling.

The original intent of SwissFit was to provide a library for fitting simple datasets with either feedforard or radial basis function neural networks; however, it as grown out of its initial purpose and now serves as an independent scientific library focused on fitting general models to data. As such, SwissFit supports fitting with both feedfoward and radial basis function neural networks. Generic fits using maximum a posteriori estimation (via minimization of an augmented $\chi^2$) can use any of SciPy's local optimization algorithms from "least squares", SciPy's local optimization algorthms from "minimize", and SciPy's implementation of the basin hopping global optimization algorithm. SwissFit also ships with basic tools for one-dimensional empirical Bayes; however, multi-dimensional empirical Bayes via Bayesian optimization is in the works. Also in the works is support for generic Bayesian model averaging.

The current version of SwissFit is in beta status; as such, please consider it an "early release". If you have any suggestions for bug fixes or additional features, I would love to hear them, though I can only guarantee that bug fixes will be implemented.

If you are here because you looked at "Constrained curve fitting for semi-parametric models with radial basis function networks" by Curtis Taylor Peterson and Anna Hasenfratz (arXiv:2402.04175), I have provided Jupyter notebooks that reproduce our results from that paper under the examples folder. These examples will work with v0.1 of SwissFit, which is downloadable under the "releases" tab.

Documentation

Documentation is currently in the works, but accessible via ReadTheDocs

Acknowledgement

If you use SwissFit, please consider citing this repository (see "cite the repository" on the right) or arXiv:2402.04175. If you use the VegasLepage module to perform your fits using Markov chain Monte Carlo estimation, please acknowledge Vegas.

Additionally, if you use our XY model data under examples/v0p1/example_data/clockinf for your research, please acknowledge USQCD resources by adding the following statement to your acknowledgements.

"The data used in this work was generated using the computing and long-term storage facilities of the USQCD Collaboration, which are funded by the Office of Science of the U.S. Department of Energy."

If you use any other spin model data in this repository (2- & 3-state Potts, along with 4-state clock), please acknowledge SpinMonteCarlo.jl.

Features

SwissFit currently supports the following.

  • lsqfit-style least squares fitting, including priors. Priors can be transformed to represent constraints. Quality of fit and model selection criteria directly available from fit. See the example below and under examples/map_fit.py.
  • Fully integrated with GVar, which allows fit parameters to be propagated into a secondary analysis with full automatic error propagation
  • Bayesian model averaging in the flavor of PRD103(2021)114502 and PRD109(2024)014510. All model-averaged parameters returned by SwissFit are correlated! See the example below and under examples/model_average_correlation_function.py.
  • Markov Chain Monte Carlo (MCMC-based) parameter estimation via Peter Lepage's Vegas library. Similar to the MCMC estimation already available in lsqfit. See the example below and under examples/mcmc_fit.py.
  • Support for integrating radial basis function networks and feedforward neural networks in least-squares model function
  • Optimization with SciPy's least_squares optimization methods (trust region reflective, Levenberg-Marquardt, dogbox), SciPy's "minimize" local optimization methods (BFGS, Nelder-Mead, conjugate gradient, etc.), and SciPy's basin hopping global optimization algorithm
  • Basic support for surrogate-based empirical Bayes (arXiv:2402.04175)

The following are planned or already in the works for SwissFit

  • Proper documentation
  • Optimization with stochastic gradient descent, specifically Adam and its Nesterov-accelerated counterpart
  • Options for other Markov Chain Monte Carlo algorithms, such as Hamiltonian Monte Carlo
  • Empirical Bayes via Scikit-learn's Bayesian optimization module

SwissFit is currently in beta. Help us get to a v1.0.0 release by providing feedback and letting me know if you run into problems! Thank you for considering to use SwissFit for whatever problem that you are trying to tackle!

Requirements

All versions of the above libraries should at least be compatible with Python>=3.10. Library dependencies are automatically installed.

Installation

SwissFit will be uploaded to PyPI for simple installation sometime in the near future. For now, install SwissFit as follows. First, clone this repository into whatever folder that you wish. Then cd into your cloned directory for SwissFit and install by running setup.py as

# Update pip - optional, but recommended
pip3 install --upgrade pip

# Install SwissFit
pip3 install swissfit

That's all. The setup.py script will install SwissFit for you, along with all of SwissFit's dependences; namely, Numpy, SciPy, Scikit-learn, GVar, Vegas, and Matplotlib.

Basic example usage

Let's get familiar with SwissFit by fitting a simple sine function. The full example code can be found under examples/simple_fit.py or examples/simple_fit.ipynb. Choose the sine function to be $$f(x) = a\sin(bx),$$ with $a=2.0$ and $b=0.5$. First, let's import everything that we'll need.

""" SwissFit imports """
from swissfit import fit # SwissFit fitter
from swissfit.optimizers import scipy_least_squares # SciPy's least squares methods

""" Other imports """
import gvar as gvar # Peter Lepage's GVar library
import numpy as np # NumPy

To extract the parameters of the sine function from data, we need to define a fit function; let's do so:

def sin(x, p):
    return p['c'][0] * gvar.sin(p['c'][-1] * x)

SwissFit operates around Python dictionaries. Therefore, you'll see that the fit parameters are encoded by a Python dictionary in our fit function. Now we need data. Let's create a function that generates an artificial dataset for us to fit to.

def create_dataset(a, b, error):
    # Actual parameters of the sine function
    real_fit_parameters = {'c': [a, b]}

    # Real dataset
    np.random.seed(0) # Seed random number generator
    data = {} # Dictionary to hold data

    # Input data
    data['x'] = np.linspace(0., 2. * np.pi / b, 20)

    # Output data
    data['y'] = [
        gvar.gvar(
            np.random.normal(sin(xx, real_fit_parameters), error), # Random mean
            error # Error on mean
        )
        for xx in data['x']
    ]

    # Return dataset
    return data

This function takes in the values for $a$, $b$ and the error that we want our artificial dataset to possess. It returns a dictionary with inputs data['x'] in $[0,2\pi/b]$ and outputs data['y'] that are uncorrelated GVar variables. Note that SwissFit is fully capable of handling correlated GVar variables. This dictionary of inputs is what we will feed into SwissFit. Before we create our SwissFit object, let's generate our artificial dataset and define our priors.

# Artificial dataset
data = create_dataset(
  2.0, # a
  0.5, # b
  0.1  # error
)
    
# Create priors
prior = {'c': [gvar.gvar('1.5(1.5)'), gvar.gvar('0.75(0.75)')]}

Again, SwissFit operates around Python dictionaries. Therefore, you see that both our dataset and priors are defined as Python dictionaries. We're now ready to create our SwissFit object.

fitter = fit.SwissFit(
    data = data,
    prior = prior,
    fit_fcn = sin,
)

To fit to data, we also need to create an optimizer object:

optimizer = scipy_least_squares.SciPyLeastSquares()

Now we are ready to fit. It is as simple as passing the SwissFit optimizer object through the call method of the SwissFit object

fitter(optimizer)

Now that we have done our fit, we can print the output and save our (correlated) fit parameters.

print(optimizer)
fit_parameters = fitter.p

The output of print is:

SwissFit: 🧀
   chi2/dof [dof] = 1.04 [20]   Q = 0.41   (Bayes) 
   chi2/dof [dof] = 1.15 [18]   Q = 0.3   (freq.) 
   AIC [k] = 24.85 [2]   logML = 7.511*

Parameters*:
     c
             1                  2.007(33)   [1.5(1.5)]
             2                 0.4990(21)   [0.75(75)]

Estimator:
   algorithm = SciPy least squares
   fun = 10.427412170606441
   optimality = 0.0017098526837675543
   nfev = 16
   njev = 14
   status = 2
   message = `ftol` termination condition is satisfied.
   success = True

We can also grab many quality of fit & information criteria directly from fitter as follows.

print(
    'chi2_data:', fitter.chi2_data,
    '\nchi2_prior:', fitter.chi2_prior,
    '\nchi2:', fitter.chi2,
    '\ndof (Bayes):', fitter.dof,
    '\ndof (freq.):', fitter.frequentist_dof,
    '\np-value:', fitter.Q,
    '\nmarginal likelihood:', fitter.logml,
    '\nAkaike information criterion:', fitter.aic
)

The output of the above print statement is

chi2_data: 20.628697369539452 
chi2_prior: 0.22612697167343043 
chi2: 20.854824341212883 
dof (Bayes): 20 
dof (freq.): 18 
p-value: 0.40572144469143007 
marginal likelihood: 7.511209597426163 
Akaike information criterion: 24.628697369539452

Because the output of fitter.p are correlated GVar variables, we can pass these parameters through any function that we want and get an output with Gaussian errors fully propagated through. For example, we could calculate f(0.5) and f(1.0), along with their covariance

# Calculate f(0.5, f(1.0)
fa = sin(0.5, fit_parameters)
fb = sin(1.0, fit_parameters)

# Print f(0.5) & f(1.0)
print('f(0.5) f(1.0):', fa, fb)
    
# Print covariance matrix of (fa, fb)
print('covariance of f(0.5) & f(1.0):\n', gvar.evalcov([fa, fb]))

We could do the same thing for any other derived quantity. That's the power of automatic error propagation by automatic differentiation! The output of the above block of code is:

f(0.5) f(1.0): 0.4955(85) 0.960(16)
covariance of f(0.5) & f(1.0):
 [[7.29612481e-05 1.40652271e-04]
 [1.40652271e-04 2.71200285e-04]]

Okay, that's all fine an dandy, but how to we visualize the result of our fit? This is no longer a exercise in using SwissFit - we now simply manipulate the GVar variables that we get from our fit. To produce the plot above, we use Matplotlib.

# Import Matplotlib
import matplotlib.pyplot as plt

# Plot fit data
plt.errorbar(
    data['x'], 
    gvar.mean(data['y']), 
    gvar.sdev(data['y']), 
    color = 'k', markerfacecolor = 'none',
    markeredgecolor = 'k',
    capsize = 6., fmt = 'o'
)

# Get result of fit function
x = np.linspace(data['x'][0], data['x'][-1], 100)
y = sin(x, fit_parameters)

# Plot error of fit function from fit as a colored band
plt.fill_between(
    x,
    gvar.mean(y) - gvar.sdev(y),
    gvar.mean(y) + gvar.sdev(y),
    color = 'maroon', alpha = 0.5
)

# x/y label
plt.xlabel('x', fontsize = 20.)
plt.ylabel('$a\\sin(bx)$', fontsize = 20.)

# Show fit parameters
plt.text(
    7.25, 0.75,
    '$a=' + str(fit_parameters['c'][0]) + '$, \n $b=' + str(fit_parameters['c'][-1]) + '$',
    fontsize = 15.
)

# Grid
plt.grid('on')

This produces the following figure.

Markov Chain Monte Carlo estimation

The example above estimates the mean and covariance of fit parameters via maximum a posteriori estimation (MAP). The MAP estimate for the mean can be poor because what we are calculating is really the posterior mode. Alternatively, we can estimate the mean of the fit parameters by sampling directly from the posterior. SwissFit has budding support for this kind of parameter estimation, which currently only supports sampling via the Vegas algorithm (see arXiv:2009.05112 for details). The infrastructure for sampling with Vegas is provided by Peter Lepage's Vegas library. MCMC estimation with SwissFit is simple and follows essentially the same steps as the MAP estimation example above. Simply replace

from swissfit.optimizers import scipy_least_squares

...

optimizer = scipy_least_squares.SciPyLeastSquares()
fitter(optimizer)

with

from swissfit.monte_carlo import vegas as vegas_lepage

...

estimator = vegas_lepage.VegasLepage()
fitter(estimator)

Everything else is the same, including printing out information about the fit. The printout should look like the following. Note, however, that SwissFit does not currently calculate the correlation between the fit parameters and the underlying dataset for MCMC-based estimates of the fit parameters. I hope to alleviate this deficit in the future.

SwissFit: 🧀
   chi2/dof [dof] = 1.04 [20]   Q = 0.41   (Bayes) 
   chi2/dof [dof] = 1.15 [18]   Q = 0.3   (freq.) 
   AIC [k] = 24.86 [2]   logML = 7.5247(23)

Parameters:
     c
             1                  2.006(33)   [1.5(1.5)]
             2                 0.4990(21)   [0.75(75)]

Estimator:
   algorithm = Peter Lepage's Vegas++
   nitn = 10 (adapt) 
   nitn = 10 (MCMC)

Simple Bayesian model averaging: subset selection

SwissFit supports model averaging in the form of "Bayesian model averaging". See PRD103(2021)114502 and PRD109(2024)014510 for details. Let's go through the example outlined in Section IV.A of PRD103(2021)114502 using SwissFit. The code for generating the data in this example can be found in the companion code for PRD109(2024)014510. For completeness, it is reproduced below.

import gvar as gv
import numpy as np

def create_synthetic_data(
        a0=2.0,a1=10.4,
        e0=0.8,e1=1.16,
        NT=32,
        rho=0.6,
        var=0.09,
        nsamp=1000
    ): 
    # arXiv:2008.01069 
    # jwsitison/improved_model_avg_paper/improved_model_averaging/synth_data.py
    def F(nt): return a0*np.exp(-e0*nt)+a1*np.exp(-e1*nt)
    def corr_fcn(nt,ntp): return rho**np.abs(nt-ntp)

    Fnt = np.fromfunction(F,(NT,))
    corr = np.fromfunction(corr_fcn, (NT,NT))
    eta = gv.raniter(gv.correlate([gv.gvar(0.,np.sqrt(var))]*NT,corr))
    
    dataset = [Fnt*(1.+next(eta)) for _ in range(nsamp)]
    return [*range(NT)], gv.dataset.avg_data(dataset)

xdata,ydata = create_synthetic_data()

Now that we have our synthetic dataset, let's define a model function for it.

def model(nt,p):
    return p['a0'][0]*gv.exp(-p['e0']*nt)

We want to extract a model-averaged estimate for a0 and e0 from fits to subsets of the data created by create_synthetic_data(). The following code utilizes SwissFit to extract an estimate of a0 and e0 over the subsets explored in PRD103(2021)114502. Each fit is performed as in the latter two examples and they are collected in an array called fits.

from swissfit import fit as fitter
from swissfit.optimizers import scipy_least_squares

tmin_max = 28
p0 = {'a0': [0.1], 'e0': [0.1]}
optimizer = scipy_least_squares.SciPyLeastSquares()

fits = [
  fitter.SwissFit(
    data = {'x': xdata[nt:], 'y': ydata[nt:]},
    p0 = p0,
    fit_fcn = model
  )(optimizer)
  for nt in range(0,tmin_max+1)
]

To model average, we simply create a BayesianModelAveraging object by passing the above array of SwissFit fits and the entire dataset (ydata) into its constuctor.

from swissfit.model_averaging import model_averaging
model_average = model_averaging.BayesianModelAveraging(models=fits, ydata=ydata)

To get the result of the model average, simply call model_average.p, as we do with regular SwissFit objects.

model_average_result = model_average.p
a0,e0 = model_average_result['a0'][0],model_average_result['e0'][0]
print('a0:',a0)
print('e0:',e0)

The above code should yield the following result.

a0: 2.084(39)
e0: 0.80141(78)

Note that the BayesianModelAveraging class calculates the full model-averaged covariance matrix. Hence, the fit parameters in model_average_result dictionary above are fully correlated GVar variables. For more details, see the example code under examples/model_average_correlation_function.py.