PySR searches for symbolic expressions which optimize a particular objective.
pysr_animation.mp4
| Docs | Forums | Paper | colab demo |
|---|---|---|---|
 |  |  |  |
| pip | conda | Stats |
|---|---|---|
 |  | pip:  conda:  |
If you find PySR useful, please cite the paper arXiv:2305.01582. If you've finished a project with PySR, please submit a PR to showcase your work on the research showcase page!
Contents:
PySR is an open-source tool for Symbolic Regression: a machine learning task where the goal is to find an interpretable symbolic expression that optimizes some objective.
Over a period of several years, PySR has been engineered from the ground up to be (1) as high-performance as possible, (2) as configurable as possible, and (3) easy to use. PySR is developed alongside the Julia library SymbolicRegression.jl, which forms the powerful search engine of PySR. The details of these algorithms are described in the PySR paper.
Symbolic regression works best on low-dimensional datasets, but one can also extend these approaches to higher-dimensional spaces by using "Symbolic Distillation" of Neural Networks, as explained in 2006.11287, where we apply it to N-body problems. Here, one essentially uses symbolic regression to convert a neural net to an analytic equation. Thus, these tools simultaneously present an explicit and powerful way to interpret deep neural networks.
You can install PySR with pip:
pip install pysrJulia dependencies will be installed at first import.
Similarly, with conda:
conda install -c conda-forge pysrYou can also use the Dockerfile to install PySR in a docker container
- Clone this repo.
- Within the repo's directory, build the docker container:
docker build -t pysr .- You can then start the container with an IPython execution with:
docker run -it --rm pysr ipythonFor more details, see the docker section.
If you are using PySR on a cluster where you do not have root access,
you can use Apptainer to build a container
instead of Docker. The Apptainer.def file is analogous to the Dockerfile,
and can be built with:
apptainer build --notest pysr.sif Apptainer.defand launched with
apptainer run pysr.sifOne issue you might run into can result in a hard crash at import with
a message like "GLIBCXX_... not found". This is due to another one of the Python dependencies
loading an incorrect libstdc++ library. To fix this, you should modify your
LD_LIBRARY_PATH variable to reference the Julia libraries. For example, if the Julia
version of libstdc++.so is located in $HOME/.julia/juliaup/julia-1.10.0+0.x64.linux.gnu/lib/julia/
(which likely differs on your system!), you could add:
export LD_LIBRARY_PATH=$HOME/.julia/juliaup/julia-1.10.0+0.x64.linux.gnu/lib/julia/:$LD_LIBRARY_PATH
to your .bashrc or .zshrc file.
You might wish to try the interactive tutorial here, which uses the notebook in examples/pysr_demo.ipynb.
In practice, I highly recommend using IPython rather than Jupyter, as the printing is much nicer. Below is a quick demo here which you can paste into a Python runtime. First, let's import numpy to generate some test data:
importnumpyasnpX=2*np.random.randn(100, 5)
y=2.5382*np.cos(X[:, 3]) +X[:, 0] **2-0.5We have created a dataset with 100 datapoints, with 5 features each.
The relation we wish to model is
Now, let's create a PySR model and train it. PySR's main interface is in the style of scikit-learn:
frompysrimportPySRRegressormodel=PySRRegressor(
niterations=40, # < Increase me for better resultsbinary_operators=["+", "*"],
unary_operators=[
"cos",
"exp",
"sin",
"inv(x) = 1/x",
# ^ Custom operator (julia syntax)
],
extra_sympy_mappings={"inv": lambdax: 1/x},
# ^ Define operator for SymPy as wellelementwise_loss="loss(prediction, target) = (prediction - target)^2",
# ^ Custom loss function (julia syntax)
)This will set up the model for 40 iterations of the search code, which contains hundreds of thousands of mutations and equation evaluations.
Let's train this model on our dataset:
model.fit(X, y)Internally, this launches a Julia process which will do a multithreaded search for equations to fit the dataset.
Equations will be printed during training, and once you are satisfied, you may quit early by hitting 'q' and then <enter>.
After the model has been fit, you can run model.predict(X)
to see the predictions on a given dataset using the automatically-selected expression,
or, for example, model.predict(X, 3) to see the predictions of the 3rd equation.
You may run:
print(model)to print the learned equations:
PySRRegressor.equations_= [
pickscoreequationlosscomplexity00.0000004.432479442.354317111.255691 (x0*x0) 3.437307320.011629 ((x0*x0) +-0.28087974) 3.358285530.897855 ((x0*x0) +cos(x3)) 1.368308640.857018 ((x0*x0) + (cos(x3) *2.4566472)) 0.24648385>>>>inf (((cos(x3) +-0.19699033) *2.5382123) + (x0*... 0.00000010
]This arrow in the pick column indicates which equation is currently selected by your
model_selection strategy for prediction.
(You may change model_selection after .fit(X, y) as well.)
model.equations_ is a pandas DataFrame containing all equations, including callable format
(lambda_format),
SymPy format (sympy_format - which you can also get with model.sympy()), and even JAX and PyTorch format
(both of which are differentiable - which you can get with model.jax() and model.pytorch()).
Note that PySRRegressor stores the state of the last search, and will restart from where you left off the next time you call .fit(), assuming you have set warm_start=True.
This will cause problems if significant changes are made to the search parameters (like changing the operators). You can run model.reset() to reset the state.
You will notice that PySR will save two files:
hall_of_fame...csv and hall_of_fame...pkl.
The csv file is a list of equations and their losses, and the pkl file is a saved state of the model.
You may load the model from the pkl file with:
model=PySRRegressor.from_file("hall_of_fame.2022-08-10_100832.281.pkl")There are several other useful features such as denoising (e.g., denoise=True),
feature selection (e.g., select_k_features=3).
For examples of these and other features, see the examples page.
For a detailed look at more options, see the options page.
You can also see the full API at this page.
There are also tips for tuning PySR on this page.
The following code makes use of as many PySR features as possible. Note that is just a demonstration of features and you should not use this example as-is. For details on what each parameter does, check out the API page.
model=PySRRegressor(
procs=4,
populations=8,
# ^ 2 populations per core, so one is always running.population_size=50,
# ^ Slightly larger populations, for greater diversity.ncycles_per_iteration=500,
# ^ Generations between migrations.niterations=10000000, # Run foreverearly_stop_condition=(
"stop_if(loss, complexity) = loss < 1e-6 && complexity < 10"# Stop early if we find a good and simple equation
),
timeout_in_seconds=60*60*24,
# ^ Alternatively, stop after 24 hours have passed.maxsize=50,
# ^ Allow greater complexity.maxdepth=10,
# ^ But, avoid deep nesting.binary_operators=["*", "+", "-", "/"],
unary_operators=["square", "cube", "exp", "cos2(x)=cos(x)^2"],
constraints={
"/": (-1, 9),
"square": 9,
"cube": 9,
"exp": 9,
},
# ^ Limit the complexity within each argument.# "inv": (-1, 9) states that the numerator has no constraint,# but the denominator has a max complexity of 9.# "exp": 9 simply states that `exp` can only have# an expression of complexity 9 as input.nested_constraints={
"square": {"square": 1, "cube": 1, "exp": 0},
"cube": {"square": 1, "cube": 1, "exp": 0},
"exp": {"square": 1, "cube": 1, "exp": 0},
},
# ^ Nesting constraints on operators. For example,# "square(exp(x))" is not allowed, since "square": {"exp": 0}.complexity_of_operators={"/": 2, "exp": 3},
# ^ Custom complexity of particular operators.complexity_of_constants=2,
# ^ Punish constants more than variablesselect_k_features=4,
# ^ Train on only the 4 most important featuresprogress=True,
# ^ Can set to false if printing to a file.weight_randomize=0.1,
# ^ Randomize the tree much more frequentlycluster_manager=None,
# ^ Can be set to, e.g., "slurm", to run a slurm# cluster. Just launch one script from the head node.precision=64,
# ^ Higher precision calculations.warm_start=True,
# ^ Start from where left off.turbo=True,
# ^ Faster evaluation (experimental)extra_sympy_mappings={"cos2": lambdax: sympy.cos(x)**2},
# extra_torch_mappings={sympy.cos: torch.cos},# ^ Not needed as cos already defined, but this# is how you define custom torch operators.# extra_jax_mappings={sympy.cos: "jnp.cos"},# ^ For JAX, one passes a string.
)You can also test out PySR in Docker, without installing it locally, by running the following command in the root directory of this repo:
docker build -t pysr .This builds an image called pysr for your system's architecture,
which also contains IPython. You can select a specific version
of Python and Julia with:
docker build -t pysr --build-arg JLVERSION=1.10.0 --build-arg PYVERSION=3.11.6 .You can then run with this dockerfile using:
docker run -it --rm -v "$PWD:/data" pysr ipythonwhich will link the current directory to the container's /data directory
and then launch ipython.
If you have issues building for your system's architecture,
you can emulate another architecture by including --platform linux/amd64,
before the build and run commands.
We are eager to welcome new contributors! Check out our contributors guide for tips 🚀. If you have an idea for a new feature, don't hesitate to share it on the issues or discussions page.