stoianmihail / netzwerk

A collection of state-of-the-art contraction ordering algorithms. https://arxiv.org/abs/2209.12332

Netzwerk

A collection of state-of-the-art contraction ordering algorithms. Plug-in for opt_einsum and cotengra.

• Optimal and near-optimal algorithms for several classes of tensor networks
• Based on the latest research in contraction ordering and database join ordering
• Careful C++ implementations

Description

$\large \textcolor{#0065bd}{\texttt{Netzwerk}}$ hosts optimal and near-optimal algorithms for tensor network contraction ordering:

• $\large \texttt{TensorIKKBZ}$: Optimal linear contraction orders for tree tensor networks
• $\large \texttt{LinDP}$: Optimal general contraction trees given the linear contraction orders of $\large \texttt{TensorIKKBZ}$
• Coming soon..

Setup

Build the shared library, which requires CMake.

cd src/netzwerk
mkdir -p build
cd build
cmake ..
make

The following initializes the package, assuming you are still in build:

cd ../../
versioneer install
pip3 install -e .

Usage

Make sure that opt_einsum and cotengra are imported before netzwerk, as follows:

import opt_einsum as oe
import cotengra as ctg
import netzwerk

Data Generation

Example: generate FTPS of up to 100 tensors.

python3 gen.py -t ftps -s 100

The files will be stored in data/ftps.

Run

The script bench.sh runs the experiments. For instance, if you want to benchmark on the networks you have just generated:

./bench.sh opt_einsum ftps 100

The results will be stored in results. For the Sycamore circuit, simply run

./bench.sh cotengra sycamore

Plot

Use the following notebooks to plot the results:

• bench-plot.ipynb: for networks generated via gen.py
• circuit-plot.ipynb: for quantum circuits, e.g., Sycamore.