• Stars
    star
    527
  • Rank 84,091 (Top 2 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created over 3 years ago
  • Updated 5 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

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 DeepMind JAX Ecosystem citation.

More Repositories

1

deepmind-research

This repository contains implementations and illustrative code to accompany DeepMind publications
Jupyter Notebook
13,132
star
2

alphafold

Open source code for AlphaFold.
Python
12,602
star
3

sonnet

TensorFlow-based neural network library
Python
9,769
star
4

mujoco

Multi-Joint dynamics with Contact. A general purpose physics simulator.
Jupyter Notebook
8,113
star
5

pysc2

StarCraft II Learning Environment
Python
8,001
star
6

lab

A customisable 3D platform for agent-based AI research
C
7,101
star
7

graph_nets

Build Graph Nets in Tensorflow
Python
5,352
star
8

graphcast

Python
4,517
star
9

open_spiel

OpenSpiel is a collection of environments and algorithms for research in general reinforcement learning and search/planning in games.
C++
4,231
star
10

alphageometry

Python
4,079
star
11

learning-to-learn

Learning to Learn in TensorFlow
Python
4,064
star
12

dm_control

Google DeepMind's software stack for physics-based simulation and Reinforcement Learning environments, using MuJoCo.
Python
3,793
star
13

acme

A library of reinforcement learning components and agents
Python
3,466
star
14

trfl

TensorFlow Reinforcement Learning
Python
3,136
star
15

dm-haiku

JAX-based neural network library
Python
2,848
star
16

alphatensor

Python
2,670
star
17

dnc

A TensorFlow implementation of the Differentiable Neural Computer.
Python
2,478
star
18

gemma

Open weights LLM from Google DeepMind.
Python
2,421
star
19

mctx

Monte Carlo tree search in JAX
Python
2,313
star
20

code_contests

C++
2,064
star
21

optax

Optax is a gradient processing and optimization library for JAX.
Python
1,670
star
22

kinetics-i3d

Convolutional neural network model for video classification trained on the Kinetics dataset.
Python
1,639
star
23

penzai

A JAX research toolkit for building, editing, and visualizing neural networks.
Python
1,639
star
24

mathematics_dataset

This dataset code generates mathematical question and answer pairs, from a range of question types at roughly school-level difficulty.
Python
1,621
star
25

bsuite

bsuite is a collection of carefully-designed experiments that investigate core capabilities of a reinforcement learning (RL) agent
Python
1,497
star
26

educational

Jupyter Notebook
1,398
star
27

jraph

A Graph Neural Network Library in Jax
Python
1,349
star
28

rc-data

Question answering dataset featured in "Teaching Machines to Read and Comprehend
Python
1,285
star
29

mujoco_menagerie

A collection of high-quality models for the MuJoCo physics engine, curated by Google DeepMind.
Jupyter Notebook
1,278
star
30

tapnet

Tracking Any Point (TAP)
Jupyter Notebook
1,266
star
31

rlax

Python
1,223
star
32

scalable_agent

A TensorFlow implementation of Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures.
Python
981
star
33

android_env

RL research on Android devices.
Python
977
star
34

neural-processes

This repository contains notebook implementations of the following Neural Process variants: Conditional Neural Processes (CNPs), Neural Processes (NPs), Attentive Neural Processes (ANPs).
Jupyter Notebook
969
star
35

mujoco_mpc

Real-time behaviour synthesis with MuJoCo, using Predictive Control
C++
959
star
36

dramatron

Dramatron uses large language models to generate coherent scripts and screenplays.
Jupyter Notebook
947
star
37

tree

tree is a library for working with nested data structures
Python
925
star
38

materials_discovery

Jupyter Notebook
866
star
39

xmanager

A platform for managing machine learning experiments
Python
815
star
40

open_x_embodiment

Jupyter Notebook
785
star
41

chex

Python
751
star
42

ferminet

An implementation of the Fermionic Neural Network for ab-initio electronic structure calculations
Python
707
star
43

reverb

Reverb is an efficient and easy-to-use data storage and transport system designed for machine learning research
C++
700
star
44

funsearch

Jupyter Notebook
699
star
45

alphadev

Python
688
star
46

pycolab

A highly-customisable gridworld game engine with some batteries included. Make your own gridworld games to test reinforcement learning agents!
Python
659
star
47

concordia

A library for generative social simulation
Python
634
star
48

hanabi-learning-environment

hanabi_learning_environment is a research platform for Hanabi experiments.
Python
614
star
49

recurrentgemma

Open weights language model from Google DeepMind, based on Griffin.
Python
603
star
50

ai-safety-gridworlds

This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents.
Python
577
star
51

meltingpot

A suite of test scenarios for multi-agent reinforcement learning.
Python
576
star
52

ithaca

Restoring and attributing ancient texts using deep neural networks
Jupyter Notebook
547
star
53

dqn

Lua/Torch implementation of DQN (Nature, 2015)
Lua
546
star
54

uncertain_ground_truth

Dermatology ddx dataset, Jax implementations of Monte Carlo conformal prediction, plausibility regions and statistical annotation aggregation from our recent work on uncertain ground truth (TMLR'23 and ArXiv pre-print).
Python
534
star
55

long-form-factuality

Benchmarking long-form factuality in large language models. Original code for our paper "Long-form factuality in large language models".
Python
526
star
56

surface-distance

Library to compute surface distance based performance metrics for segmentation tasks.
Python
526
star
57

tracr

Python
496
star
58

alphamissense

Python
494
star
59

dsprites-dataset

Dataset to assess the disentanglement properties of unsupervised learning methods
Jupyter Notebook
476
star
60

narrativeqa

This repository contains the NarrativeQA dataset. It includes the list of documents with Wikipedia summaries, links to full stories, and questions and answers.
Shell
452
star
61

clrs

Jupyter Notebook
444
star
62

lab2d

A customisable 2D platform for agent-based AI research
C++
420
star
63

dqn_zoo

DQN Zoo is a collection of reference implementations of reinforcement learning agents developed at DeepMind based on the Deep Q-Network (DQN) agent.
Python
406
star
64

alphastar

Python
403
star
65

dm_pix

PIX is an image processing library in JAX, for JAX.
Python
386
star
66

opro

official code for "Large Language Models as Optimizers"
Python
383
star
67

mathematics_conjectures

Jupyter Notebook
367
star
68

spriteworld

Spriteworld: a flexible, configurable python-based reinforcement learning environment
Python
367
star
69

torax

TORAX: Tokamak transport simulation in JAX
Python
361
star
70

dm_env

A Python interface for reinforcement learning environments
Python
343
star
71

dm_robotics

Libraries, tools and tasks created and used at DeepMind Robotics.
Python
341
star
72

spiral

We provide a pre-trained model for unconditional 19-step generation of CelebA-HQ images
C++
327
star
73

launchpad

Python
310
star
74

leo

Implementation of Meta-Learning with Latent Embedding Optimization
Python
302
star
75

enn

Python
291
star
76

streetlearn

A C++/Python implementation of the StreetLearn environment based on images from Street View, as well as a TensorFlow implementation of goal-driven navigation agents solving the task published in “Learning to Navigate in Cities Without a Map”, NeurIPS 2018
C++
285
star
77

gqn-datasets

Datasets used to train Generative Query Networks (GQNs) in the ‘Neural Scene Representation and Rendering’ paper.
Python
269
star
78

treescope

An interactive HTML pretty-printer for machine learning research in IPython notebooks.
Python
256
star
79

multi_object_datasets

Multi-object image datasets with ground-truth segmentation masks and generative factors.
Python
254
star
80

AQuA

A algebraic word problem dataset, with multiple choice questions annotated with rationales.
238
star
81

synjax

Python
238
star
82

grid-cells

Implementation of the supervised learning experiments in Vector-based navigation using grid-like representations in artificial agents, as published at https://www.nature.com/articles/s41586-018-0102-6
Python
236
star
83

card2code

A code generation dataset for generating the code that implements Hearthstone and Magic The Gathering card effects.
236
star
84

arnheim

Jupyter Notebook
235
star
85

torch-hdf5

Torch interface to HDF5 library
Lua
234
star
86

kfac-jax

Second Order Optimization and Curvature Estimation with K-FAC in JAX.
Python
231
star
87

dm_memorytasks

A set of 13 diverse machine-learning tasks that require memory to solve.
Python
221
star
88

Temporal-3D-Pose-Kinetics

Exploiting temporal context for 3D human pose estimation in the wild: 3D poses for the Kinetics dataset
Python
218
star
89

dm_alchemy

DeepMind Alchemy task environment: a meta-reinforcement learning benchmark
Python
197
star
90

neural_testbed

Jupyter Notebook
191
star
91

perception_test

Jupyter Notebook
184
star
92

jmp

JMP is a Mixed Precision library for JAX.
Python
183
star
93

neural_networks_chomsky_hierarchy

Neural Networks and the Chomsky Hierarchy
Python
183
star
94

xquad

180
star
95

nanodo

Python
180
star
96

pg19

179
star
97

spectral_inference_networks

Implementation of Spectral Inference Networks, ICLR 2019
Python
165
star
98

barkour_robot

Barkour Robot: Agile Quadruped Robots by Google DeepMind
C++
165
star
99

onetwo

Python
164
star
100

abstract-reasoning-matrices

Progressive matrices dataset, as described in: Measuring abstract reasoning in neural networks (Barrett*, Hill*, Santoro*, Morcos, Lillicrap), ICML2018
162
star