• Stars
    star
    467
  • Rank 90,730 (Top 2 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created over 1 year ago
  • Updated 3 months ago

Reviews

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

Repository Details

Tracr: TRAnsformer Compiler for RASP.

Tracr is a compiler for converting RASP programs (Weiss et al. 2021) into transformer weights. Please see our tech report for a detailed description of the compiler.

Directory structure:

  • rasp contains an implementation of RASP embedded in Python.
  • compiler contains the compiler itself.
  • transformer contains the implementation of the transformer.
  • craft contains the intermediate representation used by the compiler: essentially a small linear algebra-based library with named dimensions.

This is not an officially supported Google product.

Installation

Just clone and pip install:

git clone https://github.com/deepmind/tracr
cd tracr
pip3 install .

Usage example: RASP reverse program

Consider the RASP reverse program:

opp_index = length - indices - 1;
flip = select(indices, opp_index, ==);
reverse = aggregate(flip, tokens);

To compile this with Tracr, we would first implement the program using Tracr's RASP library:

from tracr.rasp import rasp

length = make_length()  # `length` is not a primitive in our implementation.
opp_index = length - rasp.indices - 1
flip = rasp.Select(rasp.indices, opp_index, rasp.Comparison.EQ)
reverse = rasp.Aggregate(flip, rasp.tokens)

Where:

def make_length():
  all_true_selector = rasp.Select(rasp.tokens, rasp.tokens, rasp.Comparison.TRUE)
  return rasp.SelectorWidth(all_true_selector)

We can then compile the RASP program to a transformer with:

from tracr.compiler import compiling

bos = "BOS"
model = compiling.compile_rasp_to_model(
    reverse,
    vocab={1, 2, 3},
    max_seq_len=5,
    compiler_bos=bos,
)

This yields a transformer as a Haiku model. This model isn't intended to provide everything you might need, but rather serves as a kind of "documentation-in-code" for the semantics of the generated parameters. The expectation is that the user can then write or contribute an adapter that converts parameters from this reference model to another transformer implementation.

Using this model we can perform a forward pass:

>>> out = model.apply([bos, 1, 2, 3])
>>> out.decoded
["BOS", 3, 2, 1]

Success! We have a transformer that reverses its input tokens.

Note: compiled models always expect a BOS token in order to support selectors which don't attend to any of the input tokens. This is necessary to preserve intuitive RASP semantics; the alternative would have been to treat all-False selector rows as equivalent to all-True (which is what softmax in an attention layer would naturally do). For more details, see our paper.

You can also inspect some of the intermediate activations of the model, using out.residuals, out.layer_outputs, and out.attn_logits.

For more examples of RASP programs we can compile, check out compiler/lib.py.

For an interactive example of compiling a model and visualizing its computation, check out the notebook at examples/Visualize_Tracr_Models.ipynb.

Developer README

If you'd like to extend Tracr to fit your purposes, here's some information on how Tracr works under the hood.

How Tracr works conceptually

To compile a program, Tracr does the following.

  1. Trace RASP program into a graph representation. This involves creating a graph node for each RASP expression and inferring dependencies between these graph nodes.

  2. Infer bases. Tracr is designed to have each node output to a separate subspace of the residual stream. To do this, we first infer the set of all possible token values that each node can take, then using that information, decide on a subspace for each node, and augment each node in the graph with the basis vectors for that node's subspace.

  3. Convert nodes to Craft components. Craft is the name of our internal intermediate representation that does linear algebra on named subspaces. In this stage, each expression node is converted to a Craft component that actually performs the linear algebra operations necessary to implement the expression. This includes converting sequence operators to MLP weights, and selectors to weights of attention heads. (We compute the appropriate weights directly using the theory of universal approximation for MLPs - no gradient descent required!)

  4. Convert Craft graph to Craft model. In this stage, we convert from a graph representation to a layout that looks more like an actual transformer. At this stage, we essentially have a working model, but with the linear algebra done using Craft rather than JAX + Haiku.

  5. Convert Craft model to Haiku model. Finally, we convert our intermediate representation of the model to a full Haiku model.

Two details worth expanding on here are subspaces and corresponding bases. Each node writes to a separate subspace of the residual stream, where each subspace is simply a unique chunk of the residual stream vector. For example, the first node might write to the first 5 components of the residual stream; the second node the next 5; and so on. In terms of what the embeddings actually associated with each node, Tracr employs two different kinds of bases:

  • Categorical representation - in which each unique token value is represented as a unique one-hot vector in that node's subspace. This is the representation used by default.
  • Numerical representation - in which each unique token value is mapped to a unique scalar value. This is necessary for some uses of the aggregate operation - essentially, ones which involve taking a mean - and some other operations are represented more efficiently with this representation.

A final detail is BOS tokens. The compiler relies on beginning-of-sequence tokens to in order to implement a number of operations. This is why token sequences fed into the final model must start with a BOS token.

How Tracr works in practice

The flow of compilation execution begins in compiler/compiling.py, in the compile_rasp_to_model function. This function is fairly short and maps directly to the stages outlined above, so don't be afraid to read the source!

Running tests

We use absltest, which is unittest-compatible, and is therefore in turn pytest-compatible.

First, install test dependencies:

pip3 install absl-py pytest

Then, in the checkout directory, simply run pytest. This should take about 60 seconds.

Citing Tracr

Please use the bibtex for our tech report:

@article{lindner2023tracr,
  title = {Tracr: Compiled Transformers as a Laboratory for Interpretability},
  author = {Lindner, David and Kramár, János and Rahtz, Matthew and McGrath, Thomas and Mikulik, Vladimir},
  journal={arXiv preprint arXiv:2301.05062},
  year={2023}
}

More Repositories

1

deepmind-research

This repository contains implementations and illustrative code to accompany DeepMind publications
Jupyter Notebook
12,817
star
2

alphafold

Open source code for AlphaFold.
Python
11,700
star
3

sonnet

TensorFlow-based neural network library
Python
9,691
star
4

pysc2

StarCraft II Learning Environment
Python
7,904
star
5

mujoco

Multi-Joint dynamics with Contact. A general purpose physics simulator.
Jupyter Notebook
7,202
star
6

lab

A customisable 3D platform for agent-based AI research
C
7,012
star
7

graph_nets

Build Graph Nets in Tensorflow
Python
5,325
star
8

graphcast

Python
4,242
star
9

learning-to-learn

Learning to Learn in TensorFlow
Python
4,063
star
10

open_spiel

OpenSpiel is a collection of environments and algorithms for research in general reinforcement learning and search/planning in games.
C++
4,019
star
11

alphageometry

Python
3,580
star
12

dm_control

Google DeepMind's software stack for physics-based simulation and Reinforcement Learning environments, using MuJoCo.
Python
3,473
star
13

acme

A library of reinforcement learning components and agents
Python
3,372
star
14

trfl

TensorFlow Reinforcement Learning
Python
3,139
star
15

dm-haiku

JAX-based neural network library
Python
2,697
star
16

alphatensor

Python
2,616
star
17

dnc

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

mctx

Monte Carlo tree search in JAX
Python
2,209
star
19

gemma

Open weights LLM from Google DeepMind.
Jupyter Notebook
2,061
star
20

code_contests

C++
2,010
star
21

kinetics-i3d

Convolutional neural network model for video classification trained on the Kinetics dataset.
Python
1,639
star
22

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
23

optax

Optax is a gradient processing and optimization library for JAX.
Python
1,492
star
24

bsuite

bsuite is a collection of carefully-designed experiments that investigate core capabilities of a reinforcement learning (RL) agent
Python
1,465
star
25

penzai

A JAX research toolkit for building, editing, and visualizing neural networks.
Python
1,405
star
26

educational

Jupyter Notebook
1,382
star
27

jraph

A Graph Neural Network Library in Jax
Python
1,306
star
28

rc-data

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

rlax

Python
1,185
star
30

tapnet

Tracking Any Point (TAP)
Python
1,033
star
31

scalable_agent

A TensorFlow implementation of Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures.
Python
972
star
32

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
966
star
33

android_env

RL research on Android devices.
Python
946
star
34

mujoco_menagerie

A collection of high-quality models for the MuJoCo physics engine, curated by Google DeepMind.
Jupyter Notebook
926
star
35

dramatron

Dramatron uses large language models to generate coherent scripts and screenplays.
Jupyter Notebook
904
star
36

tree

tree is a library for working with nested data structures
Python
891
star
37

xmanager

A platform for managing machine learning experiments
Python
796
star
38

mujoco_mpc

Real-time behaviour synthesis with MuJoCo, using Predictive Control
C++
771
star
39

materials_discovery

Python
770
star
40

chex

Python
716
star
41

reverb

Reverb is an efficient and easy-to-use data storage and transport system designed for machine learning research
C++
692
star
42

alphadev

Python
662
star
43

pycolab

A highly-customisable gridworld game engine with some batteries included. Make your own gridworld games to test reinforcement learning agents!
Python
654
star
44

ferminet

An implementation of the Fermionic Neural Network for ab-initio electronic structure calculations
Python
643
star
45

hanabi-learning-environment

hanabi_learning_environment is a research platform for Hanabi experiments.
Python
614
star
46

funsearch

Jupyter Notebook
611
star
47

ai-safety-gridworlds

This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents.
Python
577
star
48

dqn

Lua/Torch implementation of DQN (Nature, 2015)
Lua
546
star
49

ithaca

Restoring and attributing ancient texts using deep neural networks
Jupyter Notebook
540
star
50

meltingpot

A suite of test scenarios for multi-agent reinforcement learning.
Python
516
star
51

distrax

Python
509
star
52

recurrentgemma

Open weights language model from Google DeepMind, based on Griffin.
Python
505
star
53

surface-distance

Library to compute surface distance based performance metrics for segmentation tasks.
Python
493
star
54

dsprites-dataset

Dataset to assess the disentanglement properties of unsupervised learning methods
Jupyter Notebook
463
star
55

alphamissense

Python
455
star
56

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
432
star
57

lab2d

A customisable 2D platform for agent-based AI research
C++
415
star
58

open_x_embodiment

Jupyter Notebook
409
star
59

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
60

clrs

Python
376
star
61

spriteworld

Spriteworld: a flexible, configurable python-based reinforcement learning environment
Python
366
star
62

dm_pix

PIX is an image processing library in JAX, for JAX.
Python
363
star
63

concordia

A library for generative social simulation
Python
351
star
64

mathematics_conjectures

Jupyter Notebook
348
star
65

alphastar

Python
346
star
66

spiral

We provide a pre-trained model for unconditional 19-step generation of CelebA-HQ images
C++
327
star
67

dm_env

A Python interface for reinforcement learning environments
Python
326
star
68

dm_robotics

Libraries, tools and tasks created and used at DeepMind Robotics.
Python
315
star
69

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
315
star
70

long-form-factuality

Benchmarking long-form factuality in large language models. Original code for our paper "Long-form factuality in large language models".
Python
314
star
71

launchpad

Python
305
star
72

leo

Implementation of Meta-Learning with Latent Embedding Optimization
Python
302
star
73

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++
279
star
74

gqn-datasets

Datasets used to train Generative Query Networks (GQNs) in the ‘Neural Scene Representation and Rendering’ paper.
Python
269
star
75

enn

Python
265
star
76

multi_object_datasets

Multi-object image datasets with ground-truth segmentation masks and generative factors.
Python
247
star
77

AQuA

A algebraic word problem dataset, with multiple choice questions annotated with rationales.
238
star
78

card2code

A code generation dataset for generating the code that implements Hearthstone and Magic The Gathering card effects.
236
star
79

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
80

arnheim

Jupyter Notebook
235
star
81

synjax

Python
233
star
82

torch-hdf5

Torch interface to HDF5 library
Lua
231
star
83

dm_memorytasks

A set of 13 diverse machine-learning tasks that require memory to solve.
Python
220
star
84

Temporal-3D-Pose-Kinetics

Exploiting temporal context for 3D human pose estimation in the wild: 3D poses for the Kinetics dataset
Python
214
star
85

opro

official code for "Large Language Models as Optimizers"
Python
199
star
86

dm_alchemy

DeepMind Alchemy task environment: a meta-reinforcement learning benchmark
Python
197
star
87

neural_testbed

Jupyter Notebook
187
star
88

kfac-jax

Second Order Optimization and Curvature Estimation with K-FAC in JAX.
Python
187
star
89

pg19

179
star
90

xquad

173
star
91

jmp

JMP is a Mixed Precision library for JAX.
Python
171
star
92

spectral_inference_networks

Implementation of Spectral Inference Networks, ICLR 2019
Python
165
star
93

abstract-reasoning-matrices

Progressive matrices dataset, as described in: Measuring abstract reasoning in neural networks (Barrett*, Hill*, Santoro*, Morcos, Lillicrap), ICML2018
162
star
94

neural_networks_chomsky_hierarchy

Neural Networks and the Chomsky Hierarchy
Python
162
star
95

xitari

This is the 0.4 release of the Arcade Learning Environment (ALE), a platform designed for AI research. ALE is based on Stella, an Atari 2600 VCS emulator.
C++
159
star
96

tensor_annotations

Annotating tensor shapes using Python types
Python
158
star
97

symplectic-gradient-adjustment

A colab that implements the Symplectic Gradient Adjustment optimizer from "The mechanics of n-player differentiable games"
Jupyter Notebook
150
star
98

mc_gradients

Jupyter Notebook
149
star
99

interval-bound-propagation

This repository contains a simple implementation of Interval Bound Propagation (IBP) using TensorFlow: https://arxiv.org/abs/1810.12715
Python
148
star
100

s6

C++
146
star