/numpyro

Probabilistic programming with NumPy powered by JAX for autograd and JIT compilation to GPU/TPU/CPU.

Primary LanguagePythonApache License 2.0Apache-2.0

Build Status Documentation Status Latest Version

NumPyro

Probabilistic programming with NumPy powered by JAX for autograd and JIT compilation to GPU/TPU/CPU.

Docs | Examples | Forum


What is NumPyro?

NumPyro is a small probabilistic programming library that provides a NumPy backend for Pyro. We rely on JAX for automatic differentiation and JIT compilation to GPU / CPU. This is an alpha release under active development, so beware of brittleness, bugs, and changes to the API as the design evolves.

NumPyro is designed to be lightweight and focuses on providing a flexible substrate that users can build on:

  • Pyro Primitives: NumPyro programs can contain regular Python and NumPy code, in addition to Pyro primitives like sample and param. The model code should look very similar to Pyro except for some minor differences between PyTorch and Numpy's API. See the example below.
  • Inference algorithms: NumPyro currently supports Hamiltonian Monte Carlo, including an implementation of the No U-Turn Sampler. One of the motivations for NumPyro was to speed up Hamiltonian Monte Carlo by JIT compiling the verlet integrator that includes multiple gradient computations. With JAX, we can compose jit and grad to compile the entire integration step into an XLA optimized kernel. We also eliminate Python overhead by JIT compiling the entire tree building stage in NUTS (this is possible using Iterative NUTS). There is also a basic Variational Inference implementation for reparameterized distributions.
  • Distributions: The numpyro.distributions module provides distribution classes, constraints and bijective transforms. The distribution classes wrap over samplers implemented to work with JAX's functional pseudo-random number generator. The design of the distributions module largely follows from PyTorch. A major subset of the API is implemented, and it contains most of the common distributions that exist in PyTorch. As a result, Pyro and PyTorch users can rely on the same API and batching semantics as in torch.distributions. In addition to distributions, constraints and transforms are very useful when operating on distribution classes with bounded support.
  • Effect handlers: Like Pyro, primitives like sample and param can be provided nonstandard interpretations using effect-handlers from the numpyro.handlers module, and these can be easily extended to implement custom inference algorithms and inference utilities.

A Simple Example - 8 Schools

Let us explore NumPyro using a simple example. We will use the eight schools example from Gelman et al., Bayesian Data Analysis: Sec. 5.5, 2003, which studies the effect of coaching on SAT performance in eight schools.

The data is given by:

>>> J = 8
>>> y = np.array([28.0, 8.0, -3.0, 7.0, -1.0, 1.0, 18.0, 12.0])
>>> sigma = np.array([15.0, 10.0, 16.0, 11.0, 9.0, 11.0, 10.0, 18.0])

, where y are the treatment effects and sigma the standard error. We build a hierarchical model for the study where we assume that the group-level parameters theta for each school are sampled from a Normal distribution with unknown mean mu and standard deviation tau, while the observed data are in turn generated from a Normal distribution with mean and standard deviation given by theta (true effect) and sigma, respectively. This allows us to estimate the population-level parameters mu and tau by pooling from all the observations, while still allowing for individual variation amongst the schools using the group-level theta parameters.

>>> # Eight Schools example
... def eight_schools(J, sigma, y=None):
...     mu = numpyro.sample('mu', dist.Normal(0, 5))
...     tau = numpyro.sample('tau', dist.HalfCauchy(5))
...     with numpyro.plate('J', J):
...         theta = numpyro.sample('theta', dist.Normal(mu, tau))
...         numpyro.sample('obs', dist.Normal(theta, sigma), obs=y)

Let us infer the values of the unknown parameters in our model by running MCMC using the No-U-Turn Sampler (NUTS). Note the usage of the extra_fields argument in MCMC.run. By default, we only collect samples from the target (posterior) distribution when we run inference using MCMC. However, collecting additional fields like potential energy or the acceptance probability of a sample can be easily achieved by using the extra_fields argument. For a list of possible fields that can be collected, see the HMCState object. In this example, we will additionally collect the potential_energy for each sample.

>>> nuts_kernel = NUTS(eight_schools)
>>> mcmc = MCMC(nuts_kernel, num_warmup=500, num_samples=1000)
>>> rng_key = random.PRNGKey(0)
>>> mcmc.run(rng_key, J, sigma, y=y, extra_fields=('potential_energy',))

We can print the summary of the MCMC run, and examine if we observed any divergences during inference. Additionally, since we collected the potential energy for each of the samples, we can easily compute the expected log joint density.

>>> mcmc.print_summary()

                mean       std    median      5.0%     95.0%     n_eff     r_hat
        mu      3.94      2.81      3.16      0.03      9.28    114.51      1.06
       tau      3.20      2.97      2.40      0.38      7.28     24.06      1.07
  theta[0]      5.56      5.26      4.10     -1.67     13.52     63.57      1.05
  theta[1]      4.48      4.15      3.26     -2.44     11.25    148.63      1.05
  theta[2]      3.62      4.40      3.26     -3.85     10.75    445.91      1.01
  theta[3]      4.25      4.24      3.24     -2.99     10.68    366.29      1.04
  theta[4]      3.25      3.94      3.29     -3.34      9.84    311.03      1.00
  theta[5]      3.66      4.27      2.77     -2.79     11.06    344.57      1.02
  theta[6]      5.74      4.67      4.34     -1.92     13.25     58.42      1.05
  theta[7]      4.29      4.63      3.23     -2.14     12.37    342.50      1.02

Number of divergences: 139

>>> pe = mcmc.get_extra_fields()['potential_energy']
>>> print('Expected log joint density: {:.2f}'.format(np.mean(-pe)))

Expected log joint density: -51.42

The values above 1 for the split Gelman Rubin diagnostic (r_hat) indicates that the chain has not fully converged. The low value for the effective sample size (n_eff), particularly for tau, and the number of divergent transitions looks problematic. Fortunately, this is a common pathology that can be rectified by using a non-centered paramaterization for tau in our model. This is straightforward to do in NumPyro by using a TransformedDistribution instance. Let us rewrite the same model but instead of sampling theta from a Normal(mu, tau), we will instead sample it from a base Normal(0, 1) distribution that is transformed using an AffineTransform. Note that by doing so, NumPyro runs HMC by generating samples for the base Normal(0, 1) distribution instead. We see that the resulting chain does not suffer from the same pathology — the Gelman Rubin diagnostic is 1 for all the parameters and the effective sample size looks quite good!

>>> # Eight Schools example - Non-centered Reparametrization
... def eight_schools_noncentered(J, sigma, y=None):
...     mu = numpyro.sample('mu', dist.Normal(0, 5))
...     tau = numpyro.sample('tau', dist.HalfCauchy(5))
...     with numpyro.plate('J', J):
...         theta = numpyro.sample('theta', 
...                                dist.TransformedDistribution(dist.Normal(0., 1.),
...                                                             dist.transforms.AffineTransform(mu, tau)))
...         numpyro.sample('obs', dist.Normal(theta, sigma), obs=y)

>>> nuts_kernel = NUTS(eight_schools_noncentered)
>>> mcmc = MCMC(nuts_kernel, num_warmup=500, num_samples=1000)
>>> rng_key = random.PRNGKey(0)
>>> mcmc.run(rng_key, J, sigma, y=y, extra_fields=('potential_energy',))
>>> mcmc.print_summary()

                mean       std    median      5.0%     95.0%     n_eff     r_hat
        mu      4.38      3.04      4.50     -0.92      9.05    876.02      1.00
       tau      3.36      2.89      2.63      0.01      7.56    755.65      1.00
  theta[0]      5.99      5.42      5.44     -1.33     15.13    825.18      1.00
  theta[1]      4.80      4.50      4.78     -1.63     13.01   1114.97      1.00
  theta[2]      3.94      4.63      4.23     -3.41     11.06    914.68      1.00
  theta[3]      4.76      4.62      4.73     -2.31     12.11    958.40      1.00
  theta[4]      3.62      4.66      3.75     -3.87     11.17   1091.53      1.00
  theta[5]      3.92      4.43      4.06     -2.41     11.09   1179.74      1.00
  theta[6]      5.88      4.84      5.34     -1.45     13.11    881.38      1.00
  theta[7]      4.63      4.86      4.64     -3.57     11.80   1065.27      1.00

Number of divergences: 0

>>> pe = mcmc.get_extra_fields()['potential_energy']
>>> # Compare with the earlier value
>>> print('Expected log joint density: {:.2f}'.format(np.mean(-pe)))

Expected log joint density: -46.23

Now, let us assume that we have a new school for which we have not observed any test scores, but we would like to generate predictions. NumPyro provides a Predictive class for such a purpose. Note that in the absence of any observed data, we simply use the population-level parameters to generate predictions. The Predictive utility conditions the unobserved mu and tau sites to values drawn from the posterior distribution from our last MCMC run, and runs the model forward to generate predictions.

>>> # New School
... def new_school():
...     mu = numpyro.sample('mu', dist.Normal(0, 5))
...     tau = numpyro.sample('tau', dist.HalfCauchy(5))
...     return numpyro.sample('obs', dist.Normal(mu, tau))


>>> predictive = Predictive(new_school, mcmc.get_samples())
>>> samples_predictive = predictive.get_samples(random.PRNGKey(1))
>>> print(np.mean(samples_predictive['obs']))

4.419043

More Examples

For some more examples on specifying models and doing inference in NumPyro:

Pyro users will note that the API for model specification and inference is largely the same as Pyro, including the distributions API, by design. However, there are some important core differences (reflected in the internals) that users should be aware of. e.g. in NumPyro, there is no global parameter store or random state, to make it possible for us to leverage JAX's JIT compilation. Also, users may need to write their models in a more functional style that works better with JAX. Refer to FAQs for a list of differences.

Installation

Limited Windows Support: Note that NumPyro is untested on Windows, and will require building jaxlib from source. See this JAX issue for more details.

To install NumPyro with a CPU version of JAX, you can use pip:

pip install numpyro

To use NumPyro on the GPU, you will need to first install jax and jaxlib with CUDA support.

To run NumPyro on Cloud TPUs, you can use pip to install NumPyro as above and setup the TPU backend as detailed here.

You can also install NumPyro from source:

git clone https://github.com/pyro-ppl/numpyro.git
# install jax/jaxlib first for CUDA support
pip install -e .[dev]

Frequently Asked Questions

  1. Unlike in Pyro, numpyro.sample('x', dist.Normal(0, 1)) does not work. Why?

    You are most likely using a numpyro.sample statement outside an inference context. JAX does not have a global random state, and as such, distribution samplers need an explicit random number generator key (PRNGKey) to generate samples from. NumPyro's inference algorithms use the seed handler to thread in a random number generator key, behind the scenes.

    Your options are:

    • Call the distribution directly and provide a PRNGKey, e.g. dist.Normal(0, 1).sample(PRNGKey(0))

    • Provide the rng_key argument to numpyro.sample. e.g. numpyro.sample('x', dist.Normal(0, 1), rng_key=PRNGKey(0)).

    • Wrap the code in a seed handler, used either as a context manager or as a function that wraps over the original callable. e.g.

      with handlers.seed(rng_seed=0):
          x = numpyro.sample('x', dist.Beta(1, 1))  # random.PRNGKey(0) is used
          y = numpyro.sample('y', dist.Bernoulli(x))  # uses different PRNGKey split from the last one

      , or as a higher order function:

      def fn():
          x = numpyro.sample('x', dist.Beta(1, 1))
          y = numpyro.sample('y', dist.Bernoulli(x))
          return y
         
      print(handlers.seed(fn, rng_seed=0)())
  2. Can I use the same Pyro model for doing inference in NumPyro?

    As you may have noticed from the examples, NumPyro supports all Pyro primitives like sample, param, plate and module, and effect handlers. Additionally, we have ensured that the distributions API is based on torch.distributions, and the inference classes like SVI and MCMC have the same interface. This along with the similarity in the API for NumPy and PyTorch operations ensures that models containing Pyro primitive statements can be used with either backend with some minor changes. Example of some differences along with the changes needed, are noted below:

    • Any torch operation in your model will need to be written in terms of the corresponding jax.numpy operation. Additionally, not all torch operations have a numpy counterpart (and vice-versa), and sometimes there are minor differences in the API.
    • pyro.sample statements outside an inference context will need to be wrapped in a seed handler, as mentioned above.
    • There is no global parameter store, and as such using numpyro.param outside an inference context will have no effect. To retrieve the optimized parameter values from SVI, use the SVI.get_params method. Note that you can still use param statements inside a model and NumPyro will use the substitute effect handler internally to substitute values from the optimizer when running the model in SVI.
    • PyTorch neural network modules will need to rewritten as stax neural networks. See the VAE example for differences in syntax between the two backends.
    • JAX works best with functional code, particularly if we would like to leverage JIT compilation, which NumPyro does internally for many inference subroutines. As such, if your model has side-effects that are not visible to the JAX tracer, it may need to rewritten in a more functional style.

    For most small models, changes required to run inference in NumPyro should be minor. Additionally, we are working on pyro-api which allows you to write the same code and dispatch it to multiple backends, including NumPyro. This will necessarily be more restrictive, but has the advantage of being backend agnostic. See the documentation for an example, and let us know your feedback.

  3. How can I contribute to the project?

    Thanks for your interest in the project! You can take a look at beginner friendly issues that are marked with the good first issue tag on Github. Also, please feel to reach out to us on the forum.

Future / Ongoing Work

In the near term, we plan to work on the following. Please open new issues for feature requests and enhancements:

  • Improving robustness of inference on different models, profiling and performance tuning.
  • Supporting more functionality as part of the pyro-api generic modeling interface.
  • More inference algorithms, particularly those that require second order derivaties or use HMC.
  • Integration with Funsor to support inference algorithms with delayed sampling.
  • Other areas motivated by Pyro's research goals and application focus, and interest from the community.

Citing NumPyro

The motivating ideas behind NumPyro and a description of Iterative NUTS can be found in this paper that appeared in NeurIPS 2019 Program Transformations for Machine Learning Workshop.

If you use NumPyro, please consider citing:

@article{phan2019composable,
  title={Composable Effects for Flexible and Accelerated Probabilistic Programming in NumPyro},
  author={Phan, Du and Pradhan, Neeraj and Jankowiak, Martin},
  journal={arXiv preprint arXiv:1912.11554},
  year={2019}
}

as well as

@article{bingham2018pyro,
  author = {Bingham, Eli and Chen, Jonathan P. and Jankowiak, Martin and Obermeyer, Fritz and
            Pradhan, Neeraj and Karaletsos, Theofanis and Singh, Rohit and Szerlip, Paul and
            Horsfall, Paul and Goodman, Noah D.},
  title = {{Pyro: Deep Universal Probabilistic Programming}},
  journal = {arXiv preprint arXiv:1810.09538},
  year = {2018}
}