tensor-network-visualization 1.3.3

Minimal Matplotlib visualizations for TensorKrowch, TensorNetwork, Quimb, TeNPy, and traced PyTorch/NumPy einsum tensor networks.

Tensor-Network-Visualization

Minimal Matplotlib visualizations for TensorKrowch, TensorNetwork, Quimb, TeNPy, and traced PyTorch/NumPy einsum tensor networks.

Repository: https://github.com/DOKOS-TAYOS/Tensor-Network-Visualization

What It Does

• renders tensor networks in 2d or 3d
• works with multiple tensor-network ecosystems through one shared API
• returns standard Matplotlib Figure and Axes objects for further customization
• can also reconstruct and render tensor networks from traced binary einsum contractions

There is no custom UI layer. The package builds a normalized graph and draws it with Matplotlib.

Supported Engines

• tensorkrowch
• tensornetwork
• quimb
• tenpy
• einsum

Documentation

• Extended guide: docs/guide.md
• Examples catalog: examples/README.md
• Third-party licenses: THIRD_PARTY_LICENSES.md

Installation

As a dependency

The base package includes the shared rendering core. Install the extra that matches the backend you want to visualize:

pip install "tensor-network-visualization[tensorkrowch]"
pip install "tensor-network-visualization[tensornetwork]"
pip install "tensor-network-visualization[quimb]"
pip install "tensor-network-visualization[tenpy]"
pip install "tensor-network-visualization[einsum]"

[einsum] installs PyTorch for executing traced contractions. If you already have NumPy and only need to render an existing trace, the base package is enough.

Local development

Inside the project virtual environment:

.\.venv\Scripts\python -m pip install -e ".[dev]"

Runtime-only editable install:

.\.venv\Scripts\python -m pip install -e .

The repository also keeps:

• requirements.txt -> -e .
• requirements.dev.txt -> -e ".[dev]"

These are thin wrappers around the editable installs above.

Quick Start

from tensor_network_viz import PlotConfig, show_tensor_network

config = PlotConfig(figsize=(8, 6))

fig, ax = show_tensor_network(
    network,
    engine="tensorkrowch",
    view="2d",
    config=config,
    show=False,
)

ax.set_title("My tensor network")
fig.savefig("network.png", bbox_inches="tight")

show_tensor_network(...) is the main dispatcher. You provide:

• network: the backend-native object or node/tensor collection
• engine: which backend adapter to use
• view: 2d or 3d
• config: optional PlotConfig
• show=False if you want to save or modify the figure before displaying it

The function returns (fig, ax).

Public API

From the root package:

from tensor_network_viz import (
    EinsumTrace,
    PlotConfig,
    einsum,
    pair_tensor,
    show_tensor_network,
)

Engine-specific helpers are also available:

from tensor_network_viz.tensorkrowch import plot_tensorkrowch_network_2d
from tensor_network_viz.tensornetwork import plot_tensornetwork_network_3d
from tensor_network_viz.quimb import plot_quimb_network_2d
from tensor_network_viz.tenpy import plot_tenpy_network_3d
from tensor_network_viz.einsum_module import plot_einsum_network_2d

Accepted Inputs at a Glance

• tensorkrowch: a network object with nodes or leaf_nodes, or an iterable of nodes
• tensornetwork: an iterable of tensornetwork.Node
• quimb: a TensorNetwork or an iterable of Tensor
• tenpy: finite, segment, or infinite MPS, and finite or infinite MPO
• einsum: an EinsumTrace or an ordered iterable of pair_tensor

See the extended guide for backend-specific details and caveats.

Plot Configuration

PlotConfig controls figure size, colors, line widths, label visibility, scale parameters, custom positions, and layout iterations.

from tensor_network_viz import PlotConfig

config = PlotConfig(
    figsize=(10, 6),
    show_tensor_labels=True,
    show_index_labels=True,
    layout_iterations=300,
)

Important options:

• positions: custom node positions keyed by node id
• validate_positions=True: warn about unknown ids or wrong coordinate dimensions
• layout_iterations: tune the force-directed fallback layout

Automatic layout now prefers structural embeddings before falling back to forces:

• linear backbones are drawn as straight chains
• regular 2D meshes are kept on regular lattices
• trees use a deterministic hierarchical layout
• 3D views start from a principal plane and only lift nodes when the planar embedding becomes ambiguous
• dangling legs and other free exits in 3D use deterministic orthogonal directions instead of radial spreading

Important Backend Notes

• Quimb hyper-indices shared by more than two tensors are rendered through internal virtual hubs.
• Infinite TeNPy MPS and MPO objects are drawn as one periodic unit cell.
• The einsum backend reconstructs the network of fundamental tensors rather than plotting intermediate contraction results.
• If you pass only a subset of nodes/tensors, connections to outside objects appear as dangling legs.
• Disconnected components are supported.

Examples

The repository includes runnable scripts for every backend plus an extra TensorKrowch TSP example. See examples/README.md for commands and a short explanation of each script.

Development

.\.venv\Scripts\python -m ruff check .
.\.venv\Scripts\python -m pyright
.\.venv\Scripts\python -m pytest