quatorch 0.1.3rc0

Quaternion operations in pure PyTorch

Quaternions in PyTorch

QuaTorch

QuaTorch is a lightweight python package providing Quaternion, a torch.Tensor subclass that represents a Quaternion. It implements common special operations for quaternions such as multiplication, conjugation, inversion, normalization, log, exp, etc. It also supports conversion to/from rotation matrix and axis-angle representation. Convenient utilities are provided together, such as spherical linear interpolation (slerp) and 3D vector rotation.

Highlights

• Quaternion type: quatorch.Quaternion (subclass of torch.Tensor).
• Element-wise and algebraic ops implemented: +, -, * (quaternion product and scalar mul), abs (norm), conjugate, inverse, normalize, to_rotation_matrix, and more.
• Utilities: from_rotation_matrix, from_axis_angle, to_axis_angle, rotate_vector, slerp, log, exp, and pow.

Installation

This project targets Python 3.10+ and requires PyTorch. Install via pip (recommended):

pip install quatorch

Or install editable/development mode:

git clone 
cd QuaTorch
pip install -e .

Quick start

Basic usage examples using PyTorch tensors and Quaternion:

import torch
from quatorch import Quaternion

# Create a quaternion from four scalars (W, X, Y, Z)
q = Quaternion(1.0, 0.0, 0.0, 0.0)

# Or from a tensor of shape (..., 4)
q2 = Quaternion(torch.tensor([0.9239, 0.3827, 0.0, 0.0]))  # 45° around X

# Normalize
q2 = q2.normalize()

# Quaternion multiplication (rotation composition)
q3 = q * q2

# Rotate a vector
v = torch.tensor([1.0, 0.0, 0.0])
v_rot = q2.rotate_vector(v)

# Convert to rotation matrix
R = q2.to_rotation_matrix()

# Slerp between quaternions
t = 0.5
q_mid = q.slerp(q2, t)

API notes

• Order definition:

  • The quaternion $q=w + x\mathbf{i} + y\mathbf{j} + z\mathbf{k}$ is represented by an ordered tuple $(w, x, y, z)$ and this is the expected order for a quaternion in the whole library (i.e., watch out for XYZW-ordered incoming data).

• Construction:

  • Quaternion(data: torch.Tensor) where data has .shape[-1] == 4. An arbitrary leading shape is supported in all operations.
  • Quaternion(w, x, y, z) accepts scalars or tensors broadcastable to the same shape.

• Interoperability:

  • The class implements several torch.* functions via a small dispatcher so many PyTorch APIs behave sensibly with Quaternion objects.

Running tests

This repository includes unit tests using pytest under test/unit_tests.

From the project root, run:

uv run --with=.[cu128] pytest 

or

uv run --with=.[cpu] pytest 

Contributing

Contributions are welcome! In particular:

• Bug reports and feature requests
• Optimizing performance
• Helping improving documentation