CmdStanRb

Ruby interface to Stan, a library a for performing high performance statistical computations, i.e. full Bayesian Inference.

Installation

gem install cmd_stan_rb

Installation for IRuby notebooks

gem install cmd_stan_rb

… and then re-register your Ruby kernel by

iruby register --force

Before you can start…

This project makes use of CmdStan to compile models with the Stan compiler and incorporating additional tooling from the Stan ecosystem. There are battle-tested similar projects for other Programming languages are i.e.

• CmdStanPy
• CmdStanR
• stannis

Because of its dependency to CmdStan, you need to get CmdStan on board to make CmdStanRb work. In the future I'd like to happen this automatically, but for now you need to do this manually.

The easiest way is cloning the repository to a location of your choice:

git clone https://github.com/stan-dev/cmdstan

Then, remember the path to that directory, as we need CmdStanRb to point to that directory by setting a config variable

Sometimes you have to cd into your cmdstan directory and run

git submodule update --init --recursive

Usage

To see some examples with detailed context, you may want to consult one of the example jupyter notebooks. I.e, here's how one performs linear regression with CmdStandRb. If you like monkeys and pirates, check the example of an A/B test by fitting models based on exponential distribution.

# Tell CmdStanRb where your CmdStan repository is located.
# If you skip this step, the default value is "vendor/cmdstan"
CmdStanRb.configuration.cmdstan_dir = "~/path/to/cmdstan"

# Optional: Tell CmdStanRb where to store the models
CmdStanRb.configuration.model_dir = "/Users/robin/cmd_stan_rb-models"

# Prepare Stan compiler and assistent tooling
CmdStanRb.build_binaries

# Now we're ready to fit some models 👨‍🔬
# First, define your model. The first argument is a name,
# which is used to identify your model for re-use later.
model =
  Stan::Model.new("bernoulli-test") do
    # Stan code goes here as a string
    %q{
      data {
        int<lower=0> N;
        int<lower=0,upper=1> y[N];
      }

      parameters {
        real<lower=0,upper=1> theta;
      }

      model {
        theta ~ beta(1,1);  // uniform prior on interval 0,1
        y ~ bernoulli(theta);
      }
    }
  end

# Translate model
model.compile

# Specify data you've observed
model.data = {
  "N" => 4,
  "y" => [0,1,0,0]
}

# Run simulation to obtain samples from posterior distribution
result = model.fit(warmup: 500, samples: 2000) # Arguments are optional

# Print result
puts model.show

The latter command shows the simulation result:

Inference for Stan model: b439_model
1 chains: each with iter=(1000); warmup=(0); thin=(1); 1000 iterations saved.

Warmup took (0.0062) seconds, 0.0062 seconds total
Sampling took (0.013) seconds, 0.013 seconds total

                Mean     MCSE  StdDev     5%   50%   95%    N_Eff  N_Eff/s    R_hat
lp__            -4.4  3.5e-02    0.77   -5.9  -4.1  -3.8  4.9e+02  3.7e+04  1.0e+00
accept_stat__   0.90  4.4e-03    0.15   0.53  0.97   1.0  1.2e+03  9.2e+04  1.0e+00
stepsize__       1.1      nan    0.00    1.1   1.1   1.1      nan      nan      nan
treedepth__      1.4  1.5e-02    0.48    1.0   1.0   2.0  1.0e+03  7.6e+04  1.0e+00
n_leapfrog__     2.3  3.2e-02    0.97    1.0   3.0   3.0  9.2e+02  6.8e+04  1.0e+00
divergent__     0.00      nan    0.00   0.00  0.00  0.00      nan      nan      nan
energy__         4.9  5.2e-02     1.1    3.9   4.6   6.8  4.3e+02  3.2e+04  1.0e+00
theta           0.34  9.3e-03    0.18  0.071  0.32  0.64  3.7e+02  2.7e+04  1.0e+00

Samples were drawn using hmc with nuts.
For each parameter, N_Eff is a crude measure of effective sample size,
and R_hat is the potential scale reduction factor on split chains (at
convergence, R_hat=1).

Diving deeper into the model output

The return value of model.fit, called posterior below, holds all (sampled) posterior distributions for each parameter, like theta. For each parameter there is a method generated with the same name. The method returns a Ruby Hash reflecting the sampled posterior distribution of that parameter.

Loading previously compiled models

CmdStanRb stores models in a diretory you can define by

CmdStanRb.configuration.model_dir = "/Users/robin/cmd_stan_rb-models"

If you skip this, CmdStanRb will create a folder in your home directory named cmd_stan_rb-models and store the model files and the compiled model binaries there.

Having compiled a model with name my-model-01, you can load it again by

model = Stan::Model.load("my-model-01")

Note that once loaded, the model's data has been wiped out, so model.data is blank and needs to get set again to fit again.