based-robotics / jaxadi

Transforms your CasADi functions into batchable JAX-compatible functions. By combining the power of CasADi with the flexibility of JAX, JAXADi enables the creation of efficient code that runs smoothly on CPUs, GPUs, and TPUs.
https://based-robotics.github.io/jaxadi/
MIT License
94 stars 3 forks source link
batch casadi gpu jax

CI PyPI version PyPI downloads

JAXADI Logo

JaxADi is a Python library designed to bridge the gap between casadi.Function and JAX-compatible functions. By leveraging the strengths of both CasADi and JAX, JAXADI opens up exciting opportunities for building highly efficient, batchable code that can be executed seamlessly across CPUs, GPUs, and TPUs.

JAXADI can be particularly useful in scenarios involving:

Please dive into quick tutorial to get you up to speed in no time

Installation

You can install JAXADI using pip:

pip install jaxadi

For a complete environment setup for examples, we recommend using Conda/Mamba:

mamba env create -f environment.yml

Usage

JAXADI provides a simple and intuitive API:

import casadi as cs
import numpy as np
from jaxadi import translate, convert
from jax import numpy as jnp

x = cs.SX.sym("x", 2, 2)
y = cs.SX.sym("y", 2, 2)
# Define a complex nonlinear function
z = x @ y  # Matrix multiplication
z_squared = z * z  # Element-wise squaring
z_sin = cs.sin(z)  # Element-wise sine
result = z_squared + z_sin  # Element-wise addition
# Create the CasADi function
casadi_fn = cs.Function("complex_nonlinear_func", [x, y], [result])
# Get JAX-compatible function string representation
jax_fn_string = translate(casadi_fn)
print(jax_fn_string)
# Define JAX function from CasADi one
jax_fn = convert(casadi_fn, compile=True)
# Run compiled function
input_x = jnp.array(np.random.rand(2, 2))
input_y = jnp.array(np.random.rand(2, 2))
output = jax_fn(input_x, input_y)

Note: For now translation does not support functions with very large number of operations, due to the translation implementation. Secret component of translation is work-tree expansion, which might lead to large overhead in number of symbols. We are working on finding the compromise in both speed and extensive functions support.

Examples

JAXADI comes with several examples to help you get started:

  1. Basic Translation: Learn how to translate CasADi functions to JAX.

  2. Lowering Operations: Understand the lowering process in JaxADi.

  3. Function Conversion: See how to fully convert CasADi functions to JAX.

  4. Pendulum Rollout: Batched rollout of the nonlinear passive nonlinear pendulum

  5. Pinocchio Integration: Explore how to convert Pinocchio-based CasADi functions to JAX.

  6. MJX Comparison: Compare the transformed Pinnocchio forward kinematics with one provided by Mujoco MJX

Note: To run the Pinocchio and MJX examples, ensure you have them properly installed in your environment.

Performance Benchmarks

speedup

The process of benchmarking and evaluating the performance of Jaxadi is described in the benchmarks directory.

Contributing

We welcome contributions! Please see our Contributing Guide for more details.

Citation

If you use JaxADi in your research, please cite it as follows:

@misc{jaxadi2024,
  title = {JaxADi: Bridging CasADi and JAX for Efficient Numerical Computing},
  author = {Alentev, Igor and Kozlov, Lev and Nedelchev, Simeon},
  year = {2024},
  url = {https://github.com/based-robotics/jaxadi},
  note = {Accessed: [Insert Access Date]}
}

Acknowledgements

This project draws inspiration from cusadi, with a focus on simplicity and JAX integration.

Contact

For questions, issues, or suggestions, please open an issue on our GitHub repository.

We hope JAXADI empowers your numerical computing and optimization tasks! Happy coding!