The operator works on 32 and 64 bit floating point data types and is implemented both for CPU and GPU with custom kernels. Implementations include forward and backward steps.

The full code for our work Deep Iterative Surface Normal Estimation can be found here: https://github.com/nnaisense/deep-iterative-surface-normal-estimation.

Ensure that at least PyTorch >= 1.0.0 is installed, checkout repository and run:

python setup.py install

python setup.py test

Requires pytest and numpy.

import torch
from torch_sym3eig import Sym3Eig

matrices = torch.rand(100, 3, 3)
# Matrices need to be symmetric
matrices = (matrices + torch.transpose(matrices, 1, 2))/2.0

eig_vals, eig_vecs = Sym3Eig.apply(matrices)

Runtimes for 100k matrices on an i7-7700K @ 4.20GHz and a GTX Titan X, respectively:

Running benchmarks:

python test/benchmark.py 100000

• x (Tensor) - Input matrices (symmetric 3x3) (number_of_matrices x 3 x 3).

• eig_vals (Tensor) - Eigenvalues of the input matrices in no pre-defined order(number_of_matrices x 3).
• eig_vecs (Tensor) - Corresponding eigenvectors in columns(number_of_matrices x 3 x 3).

The module was initially created for and used in our Deep Iterative Surface Normal Estimation paper, which can be cited as reference:

@misc{lenssen2019,
    title={Deep Iterative Surface Normal Estimation},
    author={Jan Eric Lenssen and Christian Osendorfer and Jonathan Masci},
    year={2019},
    eprint={1904.07172},
    archivePrefix={arXiv},
}

The implementation follows the derivations from:

• Mike B. Giles. Collected matrix derivative results for forward and reverse mode algorithmic differentiation. In Advances in Automatic Differentiation, pages 35–44. Springer Berlin Heidelberg, 2008, and
• https://en.wikipedia.org/wiki/Eigenvalue_algorithm