Skip to content

Library for fitting models using a likelihood framework.

License

Notifications You must be signed in to change notification settings

harmslab/dataprob

 
 

Repository files navigation

dataprob

dataprob was designed to allow experimentalists to fit parameters from arbitrary models to experimental data.

  • ease of use: Users write a python function that describes their model, then load in their experimental data as a dataframe. A full analysis can be run with two python commands.
  • dataframe centric: Users use a dataframe to specify parameter bounds, guesses, fixedness, and priors. Observed data can be passed in as a dataframe or numpy vector. All outputs are simple pandas dataframes.
  • consistent experience: Users can run maximum-likelihood, bootstrap resampling, or Bayesian MCMC analyses with an identical interface and nearly identical diagnostic outputs.
  • interpretable: Provides simple diagnostic plots and runs tests assessing fit results, flagging problems with residuals and co-varying parameters.

Simple example

The following code generates noisy linear data and uses dataprob to find the maximum likelihood estimate of its slope and intercept.

import dataprob
import numpy as np

# Generate "experimental" linear data (slope = 5, intercept = 5.7) that has
# random noise on each point.
x_array = np.linspace(0,10,25)
noise = np.random.normal(loc=0,scale=0.5,size=x_array.shape)
y_obs = 5*x_array + 5.7 + noise

# 1. Define a linear model
def linear_model(m=1,b=1,x=[]):
    return m*x + b

# 2. Set up the analysis. 'method' can be "ml", "mcmc", or "bootstrap"
f = dataprob.setup(linear_model,
                   method="ml",
                   non_fit_kwargs={"x":x_array})

# 3. Fit the parameters of linear_model model to y_obs, assuming uncertainty
#    of 0.5 on each observed point.
f.fit(y_obs=y_obs,
      y_std=0.5)

# 4. Access results
print(f.fit_df)
fig = dataprob.plot_summary(f)
fig = dataprob.plot_corner(f)

The f.fit_df dataframe will look something like:

index name estimate std low_95 high_95 ... prior_std
m m 5.009 0.045 4.817 5.202 ... NaN
b b 5.644 0.274 4.465 6.822 ... NaN

The plots will be:

Installation

We recommend installing dataprob with pip:

pip install dataprob

To install from source and run tests:

git clone https://github.com/harmslab/dataprob.git
cd dataprob
pip install .

# to run test-suite
pytest --runslow

Examples

A good way to learn how to use the library is by working through examples. The following notebooks are included in the dataprob/examples/ directory. They are self-contained demonstrations in which dataprob is used to analyze various classes of experimental data. The links below launch each notebook in Google colab:

Documentation

Full documentation is on readthedocs.