Model fitting using MCMC - The basic framework

In this tutorial we show how Bayesian model fitting using Markov Chain Monte Carlo can be done in Scalismo. To be able to focus on the main components of the framework instead of technical details, we start in this tutorial with a simple toy example, namely 1D Bayesian linear regression. The application to 3D shape modelling is discussed in depth in the next tutorial.

Related resources

Week 2 of our online course on shape model fitting may provide helpful context for this tutorial.

To run the code from this tutorial, download the following Scala file:

• Tutorial14.scala

Problem setting

In a Bayesian linear regression an outcome variable $y$ is modelled a linear function of the explanatory variable $x$. The normal linear model assumes that the distribution of $y$ is a normal distribution with a mean $a \cdot x + b$ and variance $\sigma^2$.

$$y \sim N(a \cdot x + b, \sigma^2 ).$$

In the following we will denote the unknown parameters $a$, $b$ and $\sigma^2$ by $\theta$; I.e. $\theta = (a, b, \sigma^2)$. The inference problem is to estimate the parameters $\theta$, given observations $X=(x_1, \ldots, x_n)$ and $Y=(y_1, \ldots, y_n)$. This is done by computing the posterior distribution:

$$p(\theta | Y, X) = \frac{p(Y | \theta, X)p(\theta)}{\int P(Y | \theta, X)p(\theta) \, d\theta}$$

The likelihood term $p(Y | \theta, X)$ is given by the normal distribution $N(a \cdot x + b,\sigma^2)$ define above. Hence the likelihood of observing the data $X, Y$ is

$$\prod_{i=1}^n p(y_i | \theta, x_i) = \prod_{i=1}^n N(y_i | a \cdot x_i + b, \sigma^2)$$

As prior distribution $p(\theta)$ we define

$$a \sim N(0, 5) \\ b \sim N(0, 10) \\ \sigma^2 \sim logNormal(0, 0.25)$$

Metropolis Hastings Algorithm

The way we approach such an inference problem in Scalismo is by using the Metropolis-Hastings algorithm. The Metropolis-Hastings algorithm allows us to draw samples from any distribution, given that the unnormalized distribution can be evaluated point-wise. This requirement is easy to fulfill for all shape modelling applications.

For setting up the Metropolis-Hastings algorithm, we need two things:

1. The (unnormalized) target distribution, from which we want to sample. In our case this is the posterior distribution $p(\theta \mid Y, X))$. In Scalismo the corresponding class is called the DistributionEvaluator.
2. A proposal distribution $Q(\theta' \mid \theta)$, which generates for a given sample $\theta$ a new sample $\theta'$.

The Metropolis Hastings algorithm introduces an ingenious scheme for accepting and rejecting the samples from this proposal distribution, based on their probability under the target density, such that the resulting sequence of samples is guaranteed to be distributed according to the target distribution. In practice, the algorithm works as follows: It uses the proposal generator to perturb a given sample $\theta$ to obtain a new sample $\theta'$. Then it checks, using the evaluator, which of the two samples, $\theta$ or $\theta'$ is more likely and uses this ratio as a basis for rejecting or accepting the new sample.

Implementation in Scalismo

Preparation

As in the previous tutorials, we start by importing some commonly used objects and initializing the system.

import scalismo.sampling.MHSample
import scalismo.sampling.MHDistributionEvaluator
import scalismo.sampling.evaluators.ProductEvaluator
import scalismo.sampling.MHProposalGenerator
import scalismo.sampling.proposals.GaussianRandomWalkProposal
import scalismo.sampling.proposals.MHProductProposal
import scalismo.sampling.ParameterConversion
import scalismo.sampling.algorithms.MetropolisHastings
import scalismo.sampling.loggers.MHSampleLogger
import scalismo.sampling.proposals.MHMixtureProposal
import scalismo.sampling.proposals.MHIdentityProposal
import breeze.stats.meanAndVariance

import java.io.File

import breeze.stats.distributions.Rand.FixedSeed.randBasis
import scalismo.utils.Random.FixedSeed.randBasis

  scalismo.initialize()

To make the setup simple, we generate artificial data, which follows exactly our assumptions. In this way we will be able to see how well we estimated the parameters.

  
  val a = 0.2
  val b = 3
  val sigma = 0.5
  val errorDist = breeze.stats.distributions.Gaussian(0, sigma)
  val data = for (x <- 0 until 100) yield {
    (x.toDouble, a * x + b + errorDist.draw())
  }

Before we discuss the two main components, the Evaluator and Proposal generator in detail, we first define a class for representing the parameters $\theta = (a, b, \sigma^2)$:

  case class Parameters(a : Double, b:  Double, sigma : Double)

To be able to make use of the proposal generators that Scalismo provides, we will also need to define a conversion object, which tells Scalismo how our parameters can be converted to a tuple and back.

  given tuple3ParameterConversion : ParameterConversion[Tuple3[Double, Double, Double], Parameters] with
      def from(p: Parameters): Tuple3[Double, Double, Double] = (p.a, p.b, p.sigma)
      def to(t: Tuple3[Double, Double, Double]): Parameters = Parameters(a = t._1, b = t._2, sigma = t._3)  

Evaluators: Modelling the target density

In Scalismo, the target density is represented by classes, which we will refer to as Evaluators. Any Evaluator is a subclass of the class DistributionEvalutor, which is defined in Scalismo as follows:

trait DistributionEvaluator[A]:
  /** log probability/density of sample */
  def logValue(sample: A): Double

Note: This trait is already defined in Scalismo, don't paste it into your code.

We see that the only thing we need to define is the log probability of a sample.

In our case, we will define separate evaluators for the prior distribution $p(\theta)$ and the likelihood $p(Y | \theta, X)$.

The likelihood function, defined above, can be implemented as follows:

  case class LikelihoodEvaluator(data: Seq[(Double, Double)])
      extends MHDistributionEvaluator[Parameters]:

    override def logValue(theta: MHSample[Parameters]): Double = 

      val likelihoods = for ((x, y) <- data) yield 
        val likelihood = breeze.stats.distributions.Gaussian(
          theta.parameters.a * x + theta.parameters.b,
          theta.parameters.sigma
        )

        likelihood.logPdf(y)      
      likelihoods.sum

Notice that we work in Scalismo with log probabilities, and hence the product in above formula becomes a sum.

In a similar way, we encode the prior distribution:

  object PriorEvaluator extends MHDistributionEvaluator[Parameters]:

    val priorDistA = breeze.stats.distributions.Gaussian(0, 1)
    val priorDistB = breeze.stats.distributions.Gaussian(0, 10)
    val priorDistSigma = breeze.stats.distributions.LogNormal(0, 0.25)

    override def logValue(theta: MHSample[Parameters]): Double = 
      priorDistA.logPdf(theta.parameters.a)
        + priorDistB.logPdf(theta.parameters.b)
        + priorDistSigma.logPdf(theta.parameters.sigma)    

The target density (i.e. the posterior distribution) can be computed by taking the product of the prior and the likelihood.

  val posteriorEvaluator = ProductEvaluator(PriorEvaluator, LikelihoodEvaluator(data))

Note that the posteriorEvaluator represents the unnormalized posterior, as we did not normalize by the probability of the data $p(y)$.

The proposal generator

Now that the evaluators are in place, our next task is to set up the proposal distributions. In Scalismo, we can define a proposal distribution by implementing concrete subclasses, of the following class

abstract class MHProposalGenerator[A]:
   def propose(current: A): A
   def logTransitionProbability(from: A, to: A): Double

The type A refers to the type of the parameters that we are using. The propose method takes the current parameters and, based on their values, proposes a new one. The method logTransitionProbability yields the logProbability of transitioning from the state represented by the parameter values from to the state represented by the parameter values in to.

Fortunately, we usually do not have to implement these methods by ourselves. Scalismo already provides some proposal generators, which can be flexibly combined to build up more powerful generators.

The most generic one is the GaussianRandomWalkProposal, which takes the given parameters and perturbs them by adding an increment from a zero-mean Gaussian with given standard deviation. The following codes defines a proposal for each of our parameter vectors.

  val genA = GaussianRandomWalkProposal(0.01, "rw-a-0.1").forType[Double]
  val genB = GaussianRandomWalkProposal(0.05, "rw-b-0.5").forType[Double]
  val genSigma = GaussianRandomWalkProposal(0.01, "rw-sigma-0.01").forType[Double]

As we expect the distribution to have more variability in the value of $b$ than $a$, we choose the values for the step size (standard deviation) accordingly. We also provide a tag when defining a proposal generator. This is helpful for debugging and optimizing the chain. Finally, note also that we explicitly specified a type (here Double) of the specified sample.

We can now combine these individual proposal generators to a proposal generator for our Parameter class as follows:

  val parameterGenerator = MHProductProposal(genA, genB, genSigma).forType[Parameters]

It might also be a good idea to sometimes only vary the noise genSigma but not the other parameters. To achieve this, we introduce another proposal, the MHIdentityProposal. As the name suggests, it does not do anything, but simply returns the same parameters it gets passed. While this does not sound very useful by itself, by combining it with other proposals we can achieve our goal:

  val identProposal = MHIdentityProposal.forType[Double]
  val noiseOnlyGenerator = MHProductProposal(identProposal, identProposal, genSigma).forType[Parameters]

We have now two different generators, which generate new parameters given a current set of parameter values. A good strategy is to sometimes vary all the parameters, and sometimes only the noise. This can be done using a MHMixtureProposal:

  val mixtureGenerator = MHMixtureProposal((0.1, noiseOnlyGenerator), (0.9, parameterGenerator))

In this case we use the noiseOnlyGenerator 10% of the times and the parameterGenerator 90% of the times.

Building the Markov Chain

Now that we have all the components set up, we can assemble the Markov Chain.

  val chain = MetropolisHastings(mixtureGenerator, posteriorEvaluator)

To be able to diagnose the chain, in case of problems, we also instantiate a logger, which logs all the accepted and rejected samples.

  val logger = MHSampleLogger[Parameters]()

The Markov chain has to start somewhere. We define the starting point by defining an initial sample.

  val initialSample = MHSample(Parameters(0.0, 0.0, 1.0), generatedBy="initial")

To drive the sampling generation, we define an interator, which we initialize with the initial sample. We also provide the logger as an argument.

  val mhIterator = chain.iterator(initialSample, logger)

Our initial parameters might be far away from a high-probability area of our target density. Therefore it might take a few hundred or even a few thousand iterations before the produced samples start to follow the required distribution. We therefore have to drop the samples in this burn-in phase, before we use the samples:

  val samples = mhIterator.drop(1000).take(5000).toIndexedSeq

As we have generated synthetic data, we can check if the expected value, computed from this samples, really corresponds to the parameters from which we sampled our data:

  val meanAndVarianceA = meanAndVariance(samples.map(_.parameters.a))
  println(s"Estimates for parameter a: mean = ${meanAndVarianceA.mean}, var = ${meanAndVarianceA.variance}")
  val meanAndVarianceB = meanAndVariance(samples.map(_.parameters.b))
  println(s"Estimates for parameter b: mean = ${meanAndVarianceB.mean}, var = ${meanAndVarianceB.variance}")
  val meanAndVarianceSigma = meanAndVariance(samples.map(_.parameters.sigma))
  println(s"Estimates for parameter sigma: mean = ${meanAndVarianceSigma.mean}, var = ${meanAndVarianceSigma.variance}")

In the next tutorial, we see an example of how the exact same mechanism can be used for fitting shape models. Before we discuss this, we should, however, spend some time to discuss how the chain can be debugged in case something goes wrong. You can safely skip this section and come back to it later if you first want to see a practical example.

Debugging the Markov chain

The logger that we defined for the chain stored all the accepted and rejected samples. We can use this to get interesting diagnostics.

For example, we can check how often the individual samples got accepted.

  println("Acceptance ratios: " +logger.samples.acceptanceRatios)

When running this code we see that the acceptance ratio of the proposal where we vary all the parameters, is around 0.12. The proposal, which only changes the noise value has, as expected, a much higher acceptance ratio of aroun 0.75.

Sometimes it happens that a chain is efficient in the early stages (the burn-in phase), but many samples get rejected in later stages. To diagnose such situations, we can compute the acceptance ratios also only for the last $n$ samples:

  println("acceptance ratios over the last 100 samples: " +logger.samples.takeLast(100).acceptanceRatios)

Such diagnostics helps us to spot when a proposal is not effective and gives us an indication how to tune our sampler to achieve optimal performance.