pathsig 0.2.0

GPU-accelerated library for scalable and efficient differentiable path signature computations.

pathsig

pathsig is a GPU-accelerated library for differentiable signature and log-signature computations. It offers greater flexibility through projections and windowing, and achieves ~4–10× speedups for training (forward + backward) and ~10–30× speedups for forward-only evaluation in our benchmarks.

Documentation can be found at https://pathsig.readthedocs.io, and the accompanying paper is available on arXiv .

Installation

pip install pathsig

Quickstart

The example below computes truncated, windowed, and projected signatures on a CUDA tensor. Log-signatures are computed analogously, typically with a suitable projection (see the documentation).

import torch
import pathsig

x = torch.randn(32, 128, 8, device="cuda", dtype=torch.float32)

# Truncated signature
sig = pathsig.Signature(depth=4)
y = sig(x)  # (B, D)

# Windowed signature
windows = torch.tensor([[0, 32], [32, 64]], device="cuda")  # (W, 2)
sig = pathsig.Signature(depth=4, windows=windows)
y = sig(x)  # (B, W, D)

# Signature with a word projection
proj = pathsig.projections.words(
    words=[(0, 1), (2, 2, 3)],
    depth=4,
    path_dim=8,
    full_levels=(1,),
)
sig = pathsig.Signature(depth=4, projection=proj)
y = sig(x)  # (B, proj.sig_size)

License and citation

This project is licensed under the MIT License. See LICENSE for details.