Recently I’ve started using PyMC3 for Bayesian modelling, and it’s an amazing piece of software! The API only exposes as much of heavy machinery of MCMC as you need - by which I mean, just the pm.sample() method (a.k.a., as Thomas Wiecki puts it, the Magic Inference Button™). This really frees up your mind to think about your data and model, which is really the heart and soul of data science!

That being said however, I quickly realized that the water gets very deep very fast: I explored my data set, specified a hierarchical model that made sense to me, hit the Magic Inference Button™, and… uh, what now? I blinked at the angry red warnings the sampler spat out.

So began by long, rewarding and ongoing exploration of Bayesian modelling. This is a compilation of notes, tips, tricks and recipes that I’ve collected from everywhere: papers, documentation, peppering my more experienced colleagues with questions. It’s still very much a work in progress, but hopefully somebody else finds it useful!

For the Uninitiated

  • First of all, welcome! It’s a brave new world out there - where statistics is cool, Bayesian and (if you’re lucky) even easy. Dive in!

Bayesian Modelling

  • If you don’t know any probability, I’d recommend Michael Betancourt’s crash-course in practical probability theory.

  • For an introduction to general Bayesian methods and modelling, I really liked Cam Davidson Pilon’s Bayesian Methods for Hackers: it really made the whole “thinking like a Bayesian” thing click for me.

  • If you’re willing to spend some money, I’ve heard that Doing Bayesian Data Analysis by Kruschke (a.k.a. “the puppy book”) is for the bucket list.

  • Here we come to a fork in the road. The central problem in Bayesian modelling is this: given data and a probabilistic model that we think models this data, how do we find the posterior distribution of the model’s parameters? There are currently two good solutions to this problem. One is Markov chain Monte Carlo sampling (a.k.a. MCMC sampling), and the other is variational inference (a.k.a. VI). Both methods are mathematical Death Stars: extremely powerful but incredibly complicated. Nevertheless, I think it’s important to get at least a hand-wavy understanding of what these methods are. If you’re new to all this, my personal recommendation is to invest your time in learning MCMC: it’s been around longer, we know that there are sufficiently robust tools to help you, and there’s a lot more support/documentation out there.

Markov Chain Monte Carlo

Variational Inference

  • VI has been around for a while, but it was only in 2017 (2 years ago, at the time of writing) that automatic differentiation variational inference was invented. As such, variational inference is undergoing a renaissance and is currently an active area of statistical research. Since it’s such a nascent field, most resources on it are very theoretical and academic in nature.

  • Chapter 10 (on approximate inference) in Bishop’s Pattern Recognition and Machine Learning and this tutorial by David Blei are excellent, if a bit mathematically-intensive, resources.

  • The most hands-on explanation of variational inference I’ve seen is the docs for Pyro, a probabilistic programming language developed by Uber that specializes in variational inference.

Model Formulation

  • Try thinking about how your data would be generated: what kind of machine has your data as outputs? This will help you both explore your data, as well as help you arrive at a reasonable model formulation.

  • Try to avoid correlated variables. Some of the more robust samplers (cough NUTS cough cough) can cope with a posteriori correlated random variables, but sampling is much easier for everyone involved if the variables are uncorrelated. By the way, the bar is pretty low here: if the jointplot/scattergram of the two variables looks like an ellipse, thats usually okay. It’s when the ellipse starts looking like a line that you should be alarmed.

  • Try to avoid discrete latent variables, and discrete parameters in general. There is no good method to sample them in a smart way (since discrete parameters have no gradients); and with “naïve” samplers (i.e. those that do not take advantage of the gradient), the number of samples one needs to make good inferences generally scales exponentially in the number of parameters. For an instance of this, see this example on marginal Gaussian mixtures.

  • The Stan GitHub wiki has some excellent recommendations on how to choose good priors. Once you get a good handle on the basics of using PyMC3, I 100% recommend reading this wiki from start to end: the Stan community has fantastic resources on Bayesian statistics, and even though their APIs are quite different, the mathematical theory all translates over.

Hierarchical Models

  • First of all, hierarchical models are amazing! The PyMC3 docs opine on this at length, so let’s not waste any digital ink.

  • The poster child of a Bayesian hierarchical model looks something like this (equations taken from Wikipedia):

    This hierarchy has 3 levels (some would say it has 2 levels, since there are only 2 levels of parameters to infer, but honestly whatever: by my count there are 3). 3 levels is fine, but add any more levels, and it becomes harder for to sample. Try out a taller hierarchy to see if it works, but err on the side of 3-level hierarchies.

  • If your hierarchy is too tall, you can truncate it by introducing a deterministic function of your parameters somewhere (this usually turns out to just be a sum). For example, instead of modelling your observations are drawn from a 4-level hierarchy, maybe your observations can be modeled as the sum of three parameters, where these parameters are drawn from a 3-level hierarchy.

  • More in-depth treatment here in (Betancourt and Girolami, 2013). tl;dr: hierarchical models all but require you use to use Hamiltonian Monte Carlo; also included are some practical tips and goodies on how to do that stuff in the real world.

Model Implementation

  • At the risk of over-generalizing, there are only two things that can go wrong in Bayesian modelling: either your data is wrong, or your model is wrong. And it is a hell of a lot easier to debug your data than it is to debug your model. So before you even try implementing your model, plot histograms of your data, count the number of data points, drop any NaNs, etc. etc.

  • PyMC3 has one quirky piece of syntax, which I tripped up on for a while. It’s described quite well in this comment on Thomas Wiecki’s blog. Basically, suppose you have several groups, and want to initialize several variables per group, but you want to initialize different numbers of variables for each group. Then you need to use the quirky variables[index] notation. I suggest using scikit-learn’s LabelEncoder to easily create the index. For example, to make normally distributed heights for the iris dataset:

    # Different numbers of examples for each species
    species = (48 * ['setosa'] + 52 * ['virginica'] + 63 * ['versicolor'])
    num_species = len(list(set(species)))  # 3
    # One variable per group 
    heights_per_species = pm.Normal('heights_per_species',
                                    mu=0, sd=1, shape=num_species)
    idx = sklearn.preprocessing.LabelEncoder().fit_transform(species)
    heights = heights_per_species[idx]
    
  • You might find yourself in a situation in which you want to use a centered parameterization for a portion of your data set, but a non-centered parameterization for the rest of your data set (see below for what these parameterizations are). There’s a useful idiom for you here:

    num_xs = 5
    use_centered = np.array([0, 1, 1, 0, 1])  # len(use_centered) = num_xs
    x_sd = pm.HalfCauchy('x_sd', sd=1)
    x_raw = pm.Normal('x_raw', mu=0, sd=mu_x_sd**use_centered, shape=num_xs)
    x = pm.Deterministic('x', x_sd**(1 - use_centered) * x_raw)
    

    You could even experiment with allowing use_centered to be between 0 and 1, instead of being either 0 or 1!

  • I prefer to use the pm.Deterministic function instead of simply using normal arithmetic operations (e.g. I’d prefer to write x = pm.Deterministic('x', y + z) instead of x = y + z). This means that you can index the trace object later on with just trace['x'], instead of having to compute it yourself with trace['y'] + trace['z'].

MCMC Initialization and Sampling

  • Have faith in PyMC3’s default initialization and sampling settings: someone much more experienced than us took the time to choose them! NUTS is the most efficient MCMC sampler known to man, and jitter+adapt_diag… well, you get the point.

  • However, if you’re truly grasping at straws, the more powerful initialization setting would be advi or advi+adapt_diag, which uses variational inference to initialize the sampler. An even better option would be to use advi+adapt_diag_grad, which is (at the time of writing) an experimental feature in beta.

  • Never initialize the sampler with the MAP estimate! In low dimensional problems the MAP estimate (a.k.a. the mode of the posterior) is often quite a reasonable point. But in high dimensions, the MAP becomes very strange. Check out Ferenc Huszár’s blog post on high-dimensional Gaussians to see why. Besides, at the MAP all the derivatives of the posterior are zero, and that isn’t great for derivative-based samplers.

MCMC Trace Diagnostics

  • You’ve hit the Magic Inference Button™, and you have a trace object. Now what? First of all, make sure that your sampler didn’t barf itself, and that your chains are safe for consumption (i.e., analysis).
  1. Run the chain for as long as you have the patience or resources for. Make sure that the tune parameter increases commensurately with the draws parameter.

  2. Check for divergences. PyMC3’s sampler will spit out a warning if there are diverging chains, but the following code snippet may make things easier:

    # Display the total number and percentage of divergent chains
    diverging = trace['diverging']
    print('Number of Divergent Chains: {}'.format(diverging.nonzero()[0].size))
    diverging_perc = divergent.nonzero()[0].size / len(trace) * 100
    print('Percentage of Divergent Chains: {:.1f}'.format(diverging_perc))
    
  3. Check the traceplot (pm.traceplot(trace)). You’re looking for traceplots that look like “fuzzy caterpillars”. If the trace moves into some region and stays there for a long time (a.k.a. there are some “sticky regions”), that’s cause for concern! That indicates that once the sampler moves into some region of parameter space, it gets stuck there (probably due to high curvature or other bad topological properties).

  4. In addition to the traceplot, there are a ton of other plots you can make with your trace:

    • pm.plot_posterior(trace): check if your posteriors look reasonable.
    • pm.forestplot(trace): check if your variables have reasonable credible intervals, and Gelman-Rubin scores close to 1.
    • pm.autocorrplot(trace): check if your chains are impaired by high autocorrelation. Also remember that thinning your chains is a waste of time at best, and deluding yourself at worst. See Chris Fonnesbeck’s comment on this GitHub issue and Junpeng Lao’s reply to Michael Betancourt’s tweet
    • pm.energyplot(trace): ideally the energy and marginal energy distributions should look very similar. Long tails in the distribution of energy levels indicates deteriorated sampler efficiency.
    • pm.densityplot(trace): a souped-up version of pm.plot_posterior. It doesn’t seem to be wildly useful unless you’re plotting posteriors from multiple models.
  5. PyMC3 has a nice helper function to pretty-print a summary table of the trace: pm.summary(trace) (I usually tack on a .round(2) for my sanity). Look out for:
    • the values (a.k.a. the Gelman-Rubin statistic, a.k.a. the potential scale reduction factor, a.k.a. the PSRF): are they all close to 1? If not, something is horribly wrong. Consider respecifying or reparameterizing your model. You can also inspect these in the forest plot.
    • the sign and magnitude of the inferred values: do they make sense, or are they unexpected and unreasonable? This could indicate a poorly specified model. (E.g. parameters of the unexpected sign that have low uncertainties might indicate that your model needs interaction terms.)
  6. As a drastic debugging measure, try to pm.sample with draws=1, tune=500, and discard_tuned_samples=False, and inspect the traceplot. During the tuning phase, we don’t expect to see friendly fuzzy caterpillars, but we do expect to see good (if noisy) exploration of parameter space. So if the sampler is getting stuck during the tuning phase, that might explain why the trace looks horrible.

  7. If you get scary errors that describe mathematical problems (e.g. ValueError: Mass matrix contains zeros on the diagonal. Some derivatives might always be zero.), then you’re shit out of luck exceptionally unlucky: those kinds of errors are notoriously hard to debug. I can only point to the Folk Theorem of Statistical Computing:

    If you’re having computational problems, probably your model is wrong.

Fixing Divergences

There were N divergences after tuning. Increase 'target_accept' or reparameterize.

- The Magic Inference Button™

  • Divergences in HMC occur when the sampler finds itself in regions of extremely high curvature (such as the opening of the a hierarchical funnel). Broadly speaking, the sampler is prone to malfunction in such regions, causing the sampler to fly off towards to infinity. The ruins the chains by heavily biasing the samples.

  • Remember: if you have even one diverging chain, you should be worried.

  • Increase target_accept: usually 0.9 is a good number (currently the default in PyMC3 is 0.8). This will help get rid of false positives from the test for divergences. However, divergences that don’t go away are cause for alarm.

  • Increasing tune can sometimes help as well: this gives the sampler more time to 1) find the typical set and 2) find good values for step sizes, scaling factors, etc. If you’re running into divergences, it’s always possible that the sampler just hasn’t started the mixing phase and is still trying to find the typical set.

  • Consider a non-centered parameterization. This is an amazing trick: it all boils down to the familiar equation from STAT 101, but it honestly works wonders. See Thomas Wiecki’s blog post on it, and this page from the PyMC3 documentation.

  • If that doesn’t work, there may be something wrong with the way you’re thinking about your data: consider reparameterizing your model, or respecifying it entirely.

  • Now, here’s a little secret: a small number of divergences is acceptable, and even to be expected but only on particularly long and treacherous traces. That’s not to say that you can ignore all your divergences if you’re taking draws=10000! Put it this way: if you have a single-digit number of traces, you should be worried, and checking your posteriors, traces, correlation matrix, s etc., and if you have a double-digit number of traces you should be alarmed… and doing exactly the same thing.

Other Common Warnings

  • It’s worth noting that far and away the worst warning to get is the one about divergences. While a divergent chain indicates that your inference may be flat-out invalid, the rest of these warnings indicate that your inference is merely (lol, “merely”) inefficient.

  • The number of effective samples is smaller than XYZ for some parameters.
    • Quoting Junpeng Lao on discourse.pymc3.io: “A low number of effective samples is usually an indication of strong autocorrelation in the chain.”
    • Make sure you’re using an efficient sampler like NUTS. (And not, for instance, Metropolis-Hastings. (I mean seriously, it’s the 21st century, why would you ever want Metropolis-Hastings?))
    • Tweak the acceptance probability (target_accept) - it should be large enough to ensure good exploration, but small enough to not reject all proposals and get stuck.
  • The gelman-rubin statistic is larger than XYZ for some parameters. This indicates slight problems during sampling.
    • When PyMC3 samples, it runs several chains in parallel. Loosely speaking, the Gelman-Rubin statistic measures how similar these chains are. Ideally it should be close to 1.
    • Increasing the tune parameter may help, for the same reasons as described in the Fixing Divergences section.
  • The chain reached the maximum tree depth. Increase max_treedepth, increase target_accept or reparameterize.
    • NUTS puts a cap on the depth of the trees that it evaluates during each iteration, which is controlled through the max_treedepth. Reaching the maximum allowable tree depth indicates that NUTS is prematurely pulling the plug to avoid excessive compute time.
    • Yeah, what the Magic Inference Button™ says: try increasing max_treedepth or target_accept.

Model Reparameterization

  • Countless warnings have told you to engage in this strange activity of “reparameterization”. What even is that? Luckily, the Stan User Manual (specifically the Reparameterization and Change of Variables section) has an excellent explanation of reparameterization, and even some practical tips to help you do it (although your mileage may vary on how useful those tips will be to you).

  • Asides from meekly pointing to other resources, there’s not much I can do to help: this stuff really comes from a combination of intuition, statistical knowledge and good ol’ experience. I can, however, cite some examples to give you a better idea.

    • The non-centered parameterization is a classic example. If you have a parameter whose mean and variance you are also modelling, the non-centered parameterization decouples the sampling of mean and variance from the sampling of the parameter, so that they are now independent. In this way, we avoid “funnels”.
    • The horseshoe distribution is known to be a good shrinkage prior, as it is very spikey near zero, and has very long tails. However, modelling it using one parameter can give multimodal posteriors - an exceptionally bad result. The trick is to reparameterize and model it as the product of two parameters: one to create spikiness at zero, and one to create long tails (which makes sense: to sample from the horseshoe, take the product of samples from a normal and a half-Cauchy).

Model Diagnostics

  • Admittedly the distinction between the previous section and this one is somewhat artificial (since problems with your chains indicate problems with your model), but I still think it’s useful to make this distinction because these checks indicate that you’re thinking about your data in the wrong way, (i.e. you made a poor modelling decision), and not that the sampler is having a hard time doing its job.
  1. Run the following snippet of code to inspect the pairplot of your variables one at a time (if you have a plate of variables, it’s fine to pick a couple at random). It’ll tell you if the two random variables are correlated, and help identify any troublesome neighborhoods in the parameter space (divergent samples will be colored differently, and will cluster near such neighborhoods).

    pm.pairplot(trace,
                sub_varnames=[variable_1, variable_2],
                divergences=True,
                color='C3',
                kwargs_divergence={'color': 'C2'})
    
  2. Look at your posteriors (either from the traceplot, density plots or posterior plots). Do they even make sense? E.g. are there outliers or long tails that you weren’t expecting? Do their uncertainties look reasonable to you? If you had a plate of variables, are their posteriors different? Did you expect them to be that way? If not, what about the data made the posteriors different? You’re the only one who knows your problem/use case, so the posteriors better look good to you!

  3. Broadly speaking, there are four kinds of bad geometries that your posterior can suffer from:
    • highly correlated posteriors: this will probably cause divergences or traces that don’t look like “fuzzy caterpillars”. Either look at the jointplots of each pair of variables, or look at the correlation matrix of all variables. Try using a centered parameterization, or reparameterize in some other way, to remove these correlations.
    • posteriors that form “funnels”: this will probably cause divergences. Try using a non-centered parameterization.
    • long tailed posteriors: this will probably raise warnings about max_treedepth being exceeded. If your data has long tails, you should model that with a long-tailed distribution. If your data doesn’t have long tails, then your model is ill-specified: perhaps a more informative prior would help.
    • multimodal posteriors: right now this is pretty much a death blow. At the time of writing, all samplers have a hard time with multimodality, and there’s not much you can do about that. Try reparameterizing to get a unimodal posterior. If that’s not possible (perhaps you’re modelling multimodality using a mixture model), you’re out of luck: just let NUTS sample for a day or so, and hopefully you’ll get a good trace.
  4. Pick a small subset of your raw data, and see what exactly your model does with that data (i.e. run the model on a specific subset of your data). I find that a lot of problems with your model can be found this way.

  5. Run posterior predictive checks (a.k.a. PPCs): sample from your posterior, plug it back in to your model, and “generate new data sets”. PyMC3 even has a nice function to do all this for you: pm.sample_ppc. But what do you do with these new data sets? That’s a question only you can answer! The point of a PPC is to see if the generated data sets reproduce patterns you care about in the observed real data set, and only you know what patterns you care about. E.g. how close are the PPC means to the observed sample mean? What about the variance?
    • For example, suppose you were modelling the levels of radon gas in different counties in a country (through a hierarchical model). Then you could sample radon gas levels from the posterior for each county, and take the maximum within each county. You’d then have a distribution of maximum radon gas levels across counties. You could then check if the actual maximum radon gas level (in your observed data set) is acceptably within that distribution. If it’s much larger than the maxima, then you would know that the actual likelihood has longer tails than you assumed (e.g. perhaps you should use a Student’s T instead of a normal?)
    • Remember that how well the posterior predictive distribution fits the data is of little consequence (e.g. the expectation that 90% of the data should fall within the 90% credible interval of the posterior). The posterior predictive distribution tells you what values for data you would expect if we were to remeasure, given that you’ve already observed the data you did. As such, it’s informed by your prior as well as your data, and it’s not its job to adequately fit your data!