/distrax

Primary LanguagePythonApache License 2.0Apache-2.0

Distrax

CI status pypi

Distrax is a lightweight library of probability distributions and bijectors. It acts as a JAX-native reimplementation of a subset of TensorFlow Probability (TFP), with some new features and emphasis on extensibility.

Installation

You can install the latest released version of Distrax from PyPI via:

pip install distrax

or you can install the latest development version from GitHub:

pip install git+https://github.com/deepmind/distrax.git

To run the tests or examples you will need to install additional requirements.

Design Principles

The general design principles for the DeepMind JAX Ecosystem are addressed in this blog. Additionally, Distrax places emphasis on the following:

  1. Readability. Distrax implementations are intended to be self-contained and read as close to the underlying math as possible.
  2. Extensibility. We have made it as simple as possible for users to define their own distribution or bijector. This is useful for example in reinforcement learning, where users may wish to define custom behavior for probabilistic agent policies.
  3. Compatibility. Distrax is not intended as a replacement for TFP, and TFP contains many advanced features that we do not intend to replicate. To this end, we have made the APIs for distributions and bijectors as cross-compatible as possible, and provide utilities for transforming between equivalent Distrax and TFP classes.

Features

Distributions

Distributions in Distrax are simple to define and use, particularly if you're used to TFP. Let's compare the two side-by-side:

import distrax
import jax
import jax.numpy as jnp

from tensorflow_probability.substrates import jax as tfp
tfd = tfp.distributions

key = jax.random.PRNGKey(1234)
mu = jnp.array([-1., 0., 1.])
sigma = jnp.array([0.1, 0.2, 0.3])

dist_distrax = distrax.MultivariateNormalDiag(mu, sigma)
dist_tfp = tfd.MultivariateNormalDiag(mu, sigma)

samples = dist_distrax.sample(seed=key)

# Both print 1.775
print(dist_distrax.log_prob(samples))
print(dist_tfp.log_prob(samples))

In addition to behaving consistently, Distrax distributions and TFP distributions are cross-compatible. For example:

mu_0 = jnp.array([-1., 0., 1.])
sigma_0 = jnp.array([0.1, 0.2, 0.3])
dist_distrax = distrax.MultivariateNormalDiag(mu_0, sigma_0)

mu_1 = jnp.array([1., 2., 3.])
sigma_1 = jnp.array([0.2, 0.3, 0.4])
dist_tfp = tfd.MultivariateNormalDiag(mu_1, sigma_1)

# Both print 85.237
print(dist_distrax.kl_divergence(dist_tfp))
print(tfd.kl_divergence(dist_distrax, dist_tfp))

Distrax distributions implement the method sample_and_log_prob, which provides samples and their log-probability in one line. For some distributions, this is more efficient than calling separately sample and log_prob:

mu = jnp.array([-1., 0., 1.])
sigma = jnp.array([0.1, 0.2, 0.3])
dist_distrax = distrax.MultivariateNormalDiag(mu, sigma)

samples = dist_distrax.sample(seed=key, sample_shape=())
log_prob = dist_distrax.log_prob(samples)

# A one-line equivalent of the above is:
samples, log_prob = dist_distrax.sample_and_log_prob(seed=key, sample_shape=())

TFP distributions can be passed to Distrax meta-distributions as inputs. For example:

key = jax.random.PRNGKey(1234)

mu = jnp.array([-1., 0., 1.])
sigma = jnp.array([0.2, 0.3, 0.4])
dist_tfp = tfd.Normal(mu, sigma)

metadist_distrax = distrax.Independent(dist_tfp, reinterpreted_batch_ndims=1)
samples = metadist_distrax.sample(seed=key)
print(metadist_distrax.log_prob(samples))  # Prints 0.38871175

To use Distrax distributions in TFP meta-distributions, Distrax provides the wrapper to_tfp. A wrapped Distrax distribution can be directly used in TFP:

key = jax.random.PRNGKey(1234)

distrax_dist = distrax.Normal(0., 1.)
wrapped_dist = distrax.to_tfp(distrax_dist)
metadist_tfp = tfd.Sample(wrapped_dist, sample_shape=[3])

samples = metadist_tfp.sample(seed=key)
print(metadist_tfp.log_prob(samples))  # Prints -3.3409896

Bijectors

A "bijector" in Distrax is an invertible function that knows how to compute its Jacobian determinant. Bijectors can be used to create complex distributions by transforming simpler ones. Distrax bijectors are functionally similar to TFP bijectors, with a few API differences. Here is an example comparing the two:

import distrax
import jax.numpy as jnp

from tensorflow_probability.substrates import jax as tfp
tfb = tfp.bijectors
tfd = tfp.distributions

# Same distribution.
distrax.Transformed(distrax.Normal(loc=0., scale=1.), distrax.Tanh())
tfd.TransformedDistribution(tfd.Normal(loc=0., scale=1.), tfb.Tanh())

Additionally, Distrax bijectors can be composed and inverted:

bij_distrax = distrax.Tanh()
bij_tfp = tfb.Tanh()

# Same bijector.
inv_bij_distrax = distrax.Inverse(bij_distrax)
inv_bij_tfp = tfb.Invert(bij_tfp)

# These are both the identity bijector.
distrax.Chain([bij_distrax, inv_bij_distrax])
tfb.Chain([bij_tfp, inv_bij_tfp])

All TFP bijectors can be passed to Distrax, and can be freely composed with Distrax bijectors. For example, all of the following will work:

distrax.Inverse(tfb.Tanh())

distrax.Chain([tfb.Tanh(), distrax.Tanh()])

distrax.Transformed(tfd.Normal(loc=0., scale=1.), tfb.Tanh())

Distrax bijectors can also be passed to TFP, but first they must be transformed with to_tfp:

bij_distrax = distrax.to_tfp(distrax.Tanh())

tfb.Invert(bij_distrax)

tfb.Chain([tfb.Tanh(), bij_distrax])

tfd.TransformedDistribution(tfd.Normal(loc=0., scale=1.), bij_distrax)

Distrax also comes with Lambda, a convenient wrapper for turning simple JAX functions into bijectors. Here are a few Lambda examples with their TFP equivalents:

distrax.Lambda(lambda x: x)
# tfb.Identity()

distrax.Lambda(lambda x: 2*x + 3)
# tfb.Chain([tfb.Shift(3), tfb.Scale(2)])

distrax.Lambda(jnp.sinh)
# tfb.Sinh()

distrax.Lambda(lambda x: jnp.sinh(2*x + 3))
# tfb.Chain([tfb.Sinh(), tfb.Shift(3), tfb.Scale(2)])

Unlike TFP, bijectors in Distrax do not take event_ndims as an argument when they compute the Jacobian determinant. Instead, Distrax assumes that the number of event dimensions is statically known to every bijector, and uses Block to lift bijectors to a different number of dimensions. For example:

x = jnp.zeros([2, 3, 4])

# In TFP, `event_ndims` can be passed to the bijector.
bij_tfp = tfb.Tanh()
ld_1 = bij_tfp.forward_log_det_jacobian(x, event_ndims=0)  # Shape = [2, 3, 4]

# Distrax assumes `Tanh` is a scalar bijector by default.
bij_distrax = distrax.Tanh()
ld_2 = bij_distrax.forward_log_det_jacobian(x)  # ld_1 == ld_2

# With `event_ndims=2`, TFP sums the last 2 dimensions of the log det.
ld_3 = bij_tfp.forward_log_det_jacobian(x, event_ndims=2)  # Shape = [2]

# Distrax treats the number of dimensions statically.
bij_distrax = distrax.Block(bij_distrax, ndims=2)
ld_4 = bij_distrax.forward_log_det_jacobian(x)  # ld_3 == ld_4

Distrax bijectors implement the method forward_and_log_det (some bijectors additionally implement inverse_and_log_det), which allows to obtain the forward mapping and its log Jacobian determinant in one line. For some bijectors, this is more efficient than calling separately forward and forward_log_det_jacobian. (Analogously, when available, inverse_and_log_det can be more efficient than inverse and inverse_log_det_jacobian.)

x = jnp.zeros([2, 3, 4])
bij_distrax = distrax.Tanh()

y = bij_distrax.forward(x)
ld = bij_distrax.forward_log_det_jacobian(x)

# A one-line equivalent of the above is:
y, ld = bij_distrax.forward_and_log_det(x)

Jitting Distrax

Distrax distributions and bijectors can be passed as arguments to jitted functions. User-defined distributions and bijectors get this property for free by subclassing distrax.Distribution and distrax.Bijector respectively. For example:

mu_0 = jnp.array([-1., 0., 1.])
sigma_0 = jnp.array([0.1, 0.2, 0.3])
dist_0 = distrax.MultivariateNormalDiag(mu_0, sigma_0)

mu_1 = jnp.array([1., 2., 3.])
sigma_1 = jnp.array([0.2, 0.3, 0.4])
dist_1 = distrax.MultivariateNormalDiag(mu_1, sigma_1)

jitted_kl = jax.jit(lambda d_0, d_1: d_0.kl_divergence(d_1))

# Both print 85.237
print(jitted_kl(dist_0, dist_1))
print(dist_0.kl_divergence(dist_1))
A note about vmap and pmap

The serialization logic that enables Distrax objects to be passed as arguments to jitted functions also enables functions to map over them as data using jax.vmap and jax.pmap.

However, support for this behavior is experimental and incomplete. Use caution when applying jax.vmap or jax.pmap to functions that take Distrax objects as arguments, or return Distrax objects.

Simple objects such as distrax.Categorical may behave correctly under these transformations, but more complex objects such as distrax.MultivariateNormalDiag may generate exceptions when used as inputs or outputs of a vmap-ed or pmap-ed function.

Subclassing Distributions and Bijectors

User-defined distributions can be created by subclassing distrax.Distribution. This can be achieved by implementing only a few methods:

class MyDistribution(distrax.Distribution):

  def __init__(self, ...):
    ...

  def _sample_n(self, key, n):
    samples = ...
    return samples

  def log_prob(self, value):
    log_prob = ...
    return log_prob

  def event_shape(self):
    event_shape = ...
    return event_shape

  def _sample_n_and_log_prob(self, key, n):
    # Optional. Only when more efficient implementation is possible.
    samples, log_prob = ...
    return samples, log_prob

Similarly, more complicated bijectors can be created by subclassing distrax.Bijector. This can be achieved by implementing only one or two class methods:

class MyBijector(distrax.Bijector):

  def __init__(self, ...):
    super().__init__(...)

  def forward_and_log_det(self, x):
    y = ...
    logdet = ...
    return y, logdet

  def inverse_and_log_det(self, y):
    # Optional. Can be omitted if inverse methods are not needed.
    x = ...
    logdet = ...
    return x, logdet

Examples

The examples directory contains some representative examples of full programs that use Distrax.

hmm.py demonstrates how to use distrax.HMM to combine distributions that model the initial states, transitions, and observation distributions of a Hidden Markov Model, and infer the latent rates and state transitions in a changing noisy signal.

vae.py contains an example implementation of a variational auto-encoder that is trained to model the binarized MNIST dataset as a joint distrax.Bernoulli distribution over the pixels.

flow.py illustrates a simple example of modelling MNIST data using distrax.MaskedCoupling layers to implement a normalizing flow, and training the model with gradient descent.

Acknowledgements

We greatly appreciate the ongoing support of the TensorFlow Probability authors in assisting with the design and cross-compatibility of Distrax.

Special thanks to Aleyna Kara and Kevin Murphy for contributing the code upon which the Hidden Markov Model and associated example are based.

Citing Distrax

This repository is part of the DeepMind JAX Ecosystem, to cite Distrax please use the citation:

@software{deepmind2020jax,
  title = {The {D}eep{M}ind {JAX} {E}cosystem},
  author = {DeepMind and Babuschkin, Igor and Baumli, Kate and Bell, Alison and Bhupatiraju, Surya and Bruce, Jake and Buchlovsky, Peter and Budden, David and Cai, Trevor and Clark, Aidan and Danihelka, Ivo and Dedieu, Antoine and Fantacci, Claudio and Godwin, Jonathan and Jones, Chris and Hemsley, Ross and Hennigan, Tom and Hessel, Matteo and Hou, Shaobo and Kapturowski, Steven and Keck, Thomas and Kemaev, Iurii and King, Michael and Kunesch, Markus and Martens, Lena and Merzic, Hamza and Mikulik, Vladimir and Norman, Tamara and Papamakarios, George and Quan, John and Ring, Roman and Ruiz, Francisco and Sanchez, Alvaro and Sartran, Laurent and Schneider, Rosalia and Sezener, Eren and Spencer, Stephen and Srinivasan, Srivatsan and Stanojevi\'{c}, Milo\v{s} and Stokowiec, Wojciech and Wang, Luyu and Zhou, Guangyao and Viola, Fabio},
  url = {http://github.com/deepmind},
  year = {2020},
}