Learning representations of algorithms is an emerging area of machine learning, seeking to bridge concepts from neural networks with classical algorithms. The CLRS Algorithmic Reasoning Benchmark (CLRS) consolidates and extends previous work toward evaluation algorithmic reasoning by providing a suite of implementations of classical algorithms. These algorithms have been selected from the third edition of the standard Introduction to Algorithms by Cormen, Leiserson, Rivest and Stein.
The CLRS Algorithmic Reasoning Benchmark can be installed with pip, either from PyPI:
pip install dm-clrs
or directly from GitHub (updated more frequently):
pip install git+https://github.com/google-deepmind/clrs.git
You may prefer to install it in a virtual environment if any requirements clash with your Python installation:
python3 -m venv clrs_env
source clrs_env/bin/activate
pip install git+https://github.com/google-deepmind/clrs.git
Once installed you can run our example baseline model:
python3 -m clrs.examples.run
If this is the first run of the example, the dataset will be downloaded and
stored in --dataset_path
(default '/tmp/CLRS30').
Alternatively, you can also download and extract https://storage.googleapis.com/dm-clrs/CLRS30_v1.0.0.tar.gz
CLRS implements the selected algorithms in an idiomatic way, which aligns as closely as possible to the original CLRS 3ed pseudocode. By controlling the input data distribution to conform to the preconditions we are able to automatically generate input/output pairs. We additionally provide trajectories of "hints" that expose the internal state of each algorithm, to both optionally simplify the learning challenge and to distinguish between different algorithms that solve the same overall task (e.g. sorting).
In the most generic sense, algorithms can be seen as manipulating sets of objects, along with any relations between them (which can themselves be decomposed into binary relations). Accordingly, we study all of the algorithms in this benchmark using a graph representation. In the event that objects obey a more strict ordered structure (e.g. arrays or rooted trees), we impose this ordering through inclusion of predecessor links.
For each algorithm, we provide a canonical set of train, eval and test trajectories for benchmarking out-of-distribution generalization.
Trajectories | Problem Size | |
---|---|---|
Train | 1000 | 16 |
Eval | 32 x multiplier | 16 |
Test | 32 x multiplier | 64 |
Here, "problem size" refers to e.g. the length of an array or number of nodes in a graph, depending on the algorithm. "multiplier" is an algorithm-specific factor that increases the number of available eval and test trajectories to compensate for paucity of evaluation signals. "multiplier" is 1 for all algorithms except:
- Maximum subarray (Kadane), for which "multiplier" is 32.
- Quick select, minimum, binary search, string matchers (both naive and KMP), and segment intersection, for which "multiplier" is 64.
The trajectories can be used like so:
train_ds, num_samples, spec = clrs.create_dataset(
folder='/tmp/CLRS30', algorithm='bfs',
split='train', batch_size=32)
for i, feedback in enumerate(train_ds.as_numpy_iterator()):
if i == 0:
model.init(feedback.features, initial_seed)
loss = model.feedback(rng_key, feedback)
Here, feedback
is a namedtuple
with the following structure:
Feedback = collections.namedtuple('Feedback', ['features', 'outputs'])
Features = collections.namedtuple('Features', ['inputs', 'hints', 'lengths'])
where the content of Features
can be used for training and outputs
is
reserved for evaluation. Each field of the tuple is an ndarray
with a leading
batch dimension. Because hints
are provided for the full algorithm trajectory,
these contain an additional time dimension padded up to the maximum length
max(T)
of any trajectory within the dataset. The lengths
field specifies the
true length t <= max(T)
for each trajectory, which can be used e.g. for loss
masking.
The examples
directory contains a full working Graph Neural Network (GNN)
example using JAX and the DeepMind JAX Ecosystem of libraries. It allows
training of multiple algorithms on a single processor, as described in
"A Generalist Neural Algorithmic Learner".
Our initial CLRS-30 benchmark includes the following 30 algorithms. We aim to support more algorithms in the future.
- Sorting
- Insertion sort
- Bubble sort
- Heapsort (Williams, 1964)
- Quicksort (Hoare, 1962)
- Searching
- Minimum
- Binary search
- Quickselect (Hoare, 1961)
- Divide and conquer
- Maximum subarray (Kadane's variant) (Bentley, 1984)
- Greedy
- Activity selection (Gavril, 1972)
- Task scheduling (Lawler, 1985)
- Dynamic programming
- Matrix chain multiplication
- Longest common subsequence
- Optimal binary search tree (Aho et al., 1974)
- Graphs
- Depth-first search (Moore, 1959)
- Breadth-first search (Moore, 1959)
- Topological sorting (Knuth, 1973)
- Articulation points
- Bridges
- Kosaraju's strongly connected components algorithm (Aho et al., 1974)
- Kruskal's minimum spanning tree algorithm (Kruskal, 1956)
- Prim's minimum spanning tree algorithm (Prim, 1957)
- Bellman-Ford algorithm for single-source shortest paths (Bellman, 1958)
- Dijkstra's algorithm for single-source shortest paths (Dijkstra et al., 1959)
- Directed acyclic graph single-source shortest paths
- Floyd-Warshall algorithm for all-pairs shortest-paths (Floyd, 1962)
- Strings
- Naïve string matching
- Knuth-Morris-Pratt (KMP) string matcher (Knuth et al., 1977)
- Geometry
- Segment intersection
- Graham scan convex hull algorithm (Graham, 1972)
- Jarvis' march convex hull algorithm (Jarvis, 1973)
Models consist of a processor and a number of encoders and decoders. We provide JAX implementations of the following GNN baseline processors:
- Deep Sets (Zaheer et al., NIPS 2017)
- End-to-End Memory Networks (Sukhbaatar et al., NIPS 2015)
- Graph Attention Networks (Veličković et al., ICLR 2018)
- Graph Attention Networks v2 (Brody et al., ICLR 2022)
- Message-Passing Neural Networks (Gilmer et al., ICML 2017)
- Pointer Graph Networks (Veličković et al., NeurIPS 2020)
If you want to implement a new processor, the easiest way is to add
it in the processors.py
file and make it available through the
get_processor_factory
method there. A processor should have a __call__
method like this:
__call__(self,
node_fts, edge_fts, graph_fts,
adj_mat, hidden,
nb_nodes, batch_size)
where node_fts
, edge_fts
and graph_fts
will be float arrays of shape
batch_size
x nb_nodes
x H, batch_size
x nb_nodes
x nb_nodes
x H,
and batch_size
x H with encoded features for
nodes, edges and graph respectively, adj_mat
a
batch_size
x nb_nodes
x nb_nodes
boolean
array of connectivity built from hints and inputs, and hidden
a
batch_size
x nb_nodes
x H float array with the previous-step outputs
of the processor. The method should return a batch_size
x nb_nodes
x H
float array.
For more fundamentally different baselines, it is necessary to create a new
class that extends the Model API (as found within clrs/_src/model.py
).
clrs/_src/baselines.py
provides one example of how this can be done.
We provide a tensorflow_dataset
generator class in dataset.py
. This file can
be modified to generate different versions of the available algorithms, and it
can be built by using tfds build
after following the installation instructions
at https://www.tensorflow.org/datasets.
Alternatively, you can generate samples without going through tfds
by
instantiating samplers with the build_sampler
method in
clrs/_src/samplers.py
, like so:
sampler, spec = clrs.build_sampler(
name='bfs',
seed=42,
num_samples=1000,
length=16)
def _iterate_sampler(batch_size):
while True:
yield sampler.next(batch_size)
for feedback in _iterate_sampler(batch_size=32):
...
Most recently, we are offering CLRS-Text, a text-based variant of the benchmark suitable for training and evaluating the algorithmic reasoning capabilities of language models. Please see the relevant subfolder for a dedicated README file.
You may also see the companion paper on CLRS-Text.
Adding a new algorithm to the task suite requires the following steps:
- Determine the input/hint/output specification of your algorithm, and include
it within the
SPECS
dictionary ofclrs/_src/specs.py
. - Implement the desired algorithm in an abstractified form. Examples of this
can be found throughout the
clrs/_src/algorithms/
folder.
- Choose appropriate moments within the algorithm’s execution to create probes
that capture the inputs, outputs and all intermediate state (using
the
probing.push
function). - Once generated, probes must be formatted using the
probing.finalize
method, and should be returned together with the algorithm output.
- Implement an appropriate input data sampler for your algorithm,
and include it in the
SAMPLERS
dictionary withinclrs/_src/samplers.py
.
Once the algorithm has been added in this way, it can be accessed with the
build_sampler
method, and will also be incorporated to the dataset if
regenerated with the generator class in dataset.py
, as described above.
To cite the CLRS Algorithmic Reasoning Benchmark:
@article{deepmind2022clrs,
title={The CLRS Algorithmic Reasoning Benchmark},
author={Petar Veli\v{c}kovi\'{c} and Adri\`{a} Puigdom\`{e}nech Badia and
David Budden and Razvan Pascanu and Andrea Banino and Misha Dashevskiy and
Raia Hadsell and Charles Blundell},
journal={arXiv preprint arXiv:2205.15659},
year={2022}
}
To cite the CLRS-Text Algorithmic Reasoning Language Benchmark:
@article{deepmind2024clrstext,
title={The CLRS-Text Algorithmic Reasoning Language Benchmark},
author={Larisa Markeeva and Sean McLeish and Borja Ibarz and Wilfried Bounsi
and Olga Kozlova and Alex Vitvitskyi and Charles Blundell and
Tom Goldstein and Avi Schwarzschild and Petar Veli\v{c}kovi\'{c}},
journal={arXiv preprint arXiv:2406.04229},
year={2024}
}