1

This question is similar to Fit a non-linear function to data/observations with pyMCMC/pyMC, in that I'm trying to do nonlinear regression using PyMC.

However, I was wondering if anyone knew how to make my observed variable follow a non-normal distribution (i.e., T distribution) using PyMC. I know they include T distributions, but I wasn't sure how to include these as my observed variable.

Here's a quick demonstration using some faked-up data of where I'm running into issues: I'd like to use an output distribution that protects against some of the clearly outlier data points.

import numpy as np
import pymc as pm
import matplotlib.pyplot as plt

# For reproducibility
np.random.seed(1234)

x = np.linspace(0, 10*np.pi, num=150)

# Set real parameters for the sinusoid
true_freq = 0.9
true_logamp = 1.2
true_decay = 0.12
true_phase = np.pi/4


# Simulate the true trajectory
y_real = (np.exp(true_logamp - true_decay*x) *
          np.cos(true_freq*x + true_phase))

# Add some noise
y_err = y_real + 0.05*y_real.max()*np.random.randn(len(x))


# Add some outliers
num_outliers = 10
outlier_locs = np.random.randint(0, len(x), num_outliers)
y_err[outlier_locs] += (10 * y_real.max() *
                        (np.random.rand(num_outliers)))



# Bayesian Regression

def model(x, y, p0):

    log_amp = pm.Normal('log_amp', mu=np.log(p0['amplitude']),
                        tau=1/(np.log(p0['amplitude'])))

    decay = pm.Normal('decay', mu=p0['decay'],
                        tau=1/(p0['decay']))

    period = pm.TruncatedNormal('period', mu=p0['period'],
                                tau=1/(p0['period']),
                                a=1/(0.5/(np.median(np.diff(x)))),
                                b=x.max() - x.min())

    phase = pm.VonMises('phase', mu=p0['phase'], kappa=1.)

    obs_tau = pm.Gamma('obs_tau', 0.1, 0.1)

    @pm.deterministic(plot=False)
    def decaying_sinusoid(x=x, log_amp=log_amp, decay=decay,
                          period=period, phase=phase):

        return (np.exp(log_amp - decay*x) *
                np.cos((2*np.pi/period)*x + phase))

    obs = pm.Normal('obs', mu=decaying_sinusoid, tau=obs_tau, value=y,
                    observed=True)

    return locals()

p0 = {
    'amplitude' : 2.30185,
    'decay'     : 0.06697,
    'period'    : 7.11672,
    'phase'     : 0.93055,
}


MDL = pm.MCMC(model(x, y_err, p0))
MDL.sample(20000, 10000, 1)

# Plot fit
y_min = MDL.stats()['decaying_sinusoid']['quantiles'][2.5]
y_max = MDL.stats()['decaying_sinusoid']['quantiles'][97.5]
y_fit = MDL.stats()['decaying_sinusoid']['mean']
plt.plot(x, y_err, '.', label='Data')
plt.plot(x, y_fit, label='Fit')
plt.plot(x, y_real, label='True')

plt.fill_between(x, y_min, y_max, color='0.5', alpha=0.5)
plt.legend()

non-robust fit :(

Thanks!!

Community
  • 1
  • 1
pstjohn
  • 521
  • 5
  • 13

1 Answers1

2

PyMC2 has a t distribution built in, pm.T, but it is centered at zero, so you can't use it directly in this application. Instead, you can use the pm.t_like(x, nu) function, which calculates the log-likelihood from a value x and a dof parameter nu, to define a custom observed stochastic. To make such a custom distribution for your observed variable is simple: change lines 59-60 to be:

@pm.observed
def obs(mu=decaying_sinusoid, tau=obs_tau, value=y):
    return pm.t_like(value-mu, tau)
Abraham D Flaxman
  • 2,969
  • 21
  • 39
  • Thanks! This runs, but sadly takes forever. The original script takes takes 8.2 to run the 20,000 steps, while replacing it with a custom log-likelihood raises that to 6284.6 seconds. I need to just try to remove the outliers ahead of time. – pstjohn Oct 27 '14 at 19:25
  • For me it only increases run time by 30%, which is well worth it for a 10x increase in accuracy. [Here is a notebook where they race.](http://nbviewer.ipython.org/gist/aflaxman/b6bbdf036b1a0cad76ec) – Abraham D Flaxman Oct 27 '14 at 20:21