acaster

autocast

Automatic type conversion for Python.

autocast is a lightweight library that automatically converts objects from one type to another using a graph of registered transformation functions. If a direct conversion isn't available, autocast finds the shortest path through intermediate types to get you the result you need.

It comes with built-in support for NumPy, PyTorch, Open3D, and OmegaConf, but it is designed to be easily extensible for your own classes.

📦 Installation

pip install acaster

To install with support for all optional backends (PyTorch, Open3D, etc.):

pip install "acaster[all]"

🚀 Quick Start

Basic Usage

The core function is acast(object, target_type).

import torch
import numpy as np
from autocast import acast

# Create a Torch tensor
tensor = torch.ones((3, 3))

# Automatically convert to NumPy array
array = acast(tensor, np.ndarray)

print(type(array)) 
# <class 'numpy.ndarray'>

Advanced Usage (Open3D)

Build an Open3D Point Cloud directly from PyTorch tensors or Numpy arrays without manually converting them to Open3D Vector3dVector:

import open3d as o3d
import torch
from autocast import acast

H, W = 448, 448
rgb_image = torch.rand(H, W, 3) # H x W x 3
predicted_pts = torch.rand(H * W, 3) # H * W x 3 

# Cast a tuple of (points, colors) directly to a PointCloud
pcd = acast((predicted_pts, rgb_image), o3d.geometry.PointCloud)

print(type(pcd))
# <class 'open3d.cpu.pybind.geometry.PointCloud'> 

# o3d.io.write_point_cloud("output.ply", pcd)

Recursive Collection Handling

autocast automatically handles nested lists and dictionaries.

data = {
    "points": torch.rand(100, 3),
    "colors": torch.rand(100, 3),
    "meta": [torch.tensor(1.0), torch.tensor(2.0)]
}

# Convert everything inside the dict to NumPy arrays
numpy_data = acast(data, np.ndarray)

print(type(numpy_data['points'])) 
# <class 'numpy.ndarray'>

🛠 Features

• Graph-Based Casting: Uses Breadth-First Search (BFS) to find conversion paths. If you have A -> B and B -> C, autocast can do A -> C automatically.
• Zero-Overhead Imports: Backend modules (like torch or open3d) are only imported if they are installed in your environment.
• Extensible: Register your own types and conversion functions easily.

🧩 Extending autocast

You can register your own casts using the add_acast function or the @acaster decorator.

Option 1: The Decorator

from autocast import acaster, acast

class Cat:
    def speak(self): return "Meow"

class Dog:
    def speak(self): return "Woof"

# Register a function that converts Cat -> Dog
@acaster(Cat, Dog)
def cat_to_dog(cat_instance):
    return Dog()

my_cat = Cat()
my_dog = acast(my_cat, Dog)

Option 2: Manual Registration

from autocast import add_acast

def str_to_int(s):
    return int(s)

add_acast(str, int, str_to_int)

print(acast("123", int)) # 123

🤝 Contributing

We welcome contributions! If you want to add support for a new library (e.g., Pandas, Polars, JAX), please submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/new-cast)
3. Commit your changes
4. Push to the branch
5. Open a Pull Request

To run the tests, install pytests and run python -m pytest.

Publishing (Maintainers Only)

To release a new version of acaster:

1. Bump Version: Update version = "x.y.z" in pyproject.toml.
2. Tag Release:

git tag v0.x.y
git push origin --tags

3. Build & Upload:

rm -rf dist/
python -m build
twine upload dist/*

📄 License

Distributed under the MIT License. See LICENSE for more information.

Acknowledgements

The idea for this library was inspired by erezsh/autocast.