fa2 1.1.1

The fastest ForceAtlas2 algorithm for Python (and NetworkX)

ForceAtlas2 for Python

The fastest Python implementation of the ForceAtlas2 graph layout algorithm, with Cython optimization for 10-100x speedup. Supports NetworkX, igraph, and raw adjacency matrices.

ForceAtlas2 is a force-directed layout algorithm designed for network visualization. It spatializes weighted undirected graphs in 2D, 3D, or higher dimensions, where edge weights define connection strength. It scales well to large graphs (>10,000 nodes) using Barnes-Hut approximation (O(n log n) complexity).

Documentation · PyPI · Paper

500-node stochastic block model (7 communities) laid out with ForceAtlas2 LinLog mode

Random geometric graph (400 nodes) laid out with ForceAtlas2

1000-node stochastic block model (8 communities) laid out in 3D with ForceAtlas2 LinLog mode

Installation

pip install fa2

For maximum performance, install with Cython (recommended):

pip install cython
pip install fa2 --no-binary fa2

To build from source:

git clone https://github.com/bhargavchippada/forceatlas2.git
cd forceatlas2
pip install cython numpy
pip install -e ".[dev]" --no-build-isolation

Dependencies

Package	Required	Purpose
numpy	Yes	Adjacency matrix handling
scipy	Yes	Sparse matrix support
tqdm	Yes	Progress bar
cython	No (recommended)	10-100x speedup
networkx	No	NetworkX graph wrapper
igraph	No	igraph graph wrapper
matplotlib	No	Visualization (pip install fa2[viz])

Python: >= 3.9 (tested on 3.9 through 3.14)

Quick Start

Simplest — No Numpy Required

from fa2.easy import layout, visualize

# Edge list in → positions out
positions = layout([("A", "B"), ("B", "C"), ("A", "C")], mode="community")

# One call to render
visualize([("A", "B"), ("B", "C"), ("A", "C")], output="png", path="graph.png")

CLI

# Layout from JSON edge list
python -m fa2 layout edges.json --mode community -o layout.json

# Render to image
python -m fa2 render edges.csv -o graph.png

# Compute quality metrics
echo '[["A","B"],["B","C"]]' | python -m fa2 metrics

With NetworkX

import networkx as nx
import matplotlib.pyplot as plt
from fa2 import ForceAtlas2

G = nx.random_geometric_graph(400, 0.2)

forceatlas2 = ForceAtlas2(
    outboundAttractionDistribution=True,  # Dissuade hubs
    edgeWeightInfluence=1.0,
    jitterTolerance=1.0,
    barnesHutOptimize=True,
    barnesHutTheta=1.2,
    scalingRatio=2.0,
    strongGravityMode=False,
    gravity=1.0,
    verbose=True,
)

positions = forceatlas2.forceatlas2_networkx_layout(G, pos=None, iterations=2000)

nx.draw_networkx_nodes(G, positions, node_size=20, node_color="blue", alpha=0.4)
nx.draw_networkx_edges(G, positions, edge_color="green", alpha=0.05)
plt.axis("off")
plt.show()

With Raw Adjacency Matrix

import numpy as np
from fa2 import ForceAtlas2

# Create a symmetric adjacency matrix
G = np.array([
    [0, 1, 1, 0],
    [1, 0, 1, 1],
    [1, 1, 0, 0],
    [0, 1, 0, 0],
], dtype=float)

forceatlas2 = ForceAtlas2(verbose=False, seed=42)
positions = forceatlas2.forceatlas2(G, iterations=1000)
# Returns: [(x1, y1), (x2, y2), ...]

With Scipy Sparse Matrix

import scipy.sparse
from fa2 import ForceAtlas2

# For large graphs, sparse matrices are more memory-efficient
G_sparse = scipy.sparse.csr_matrix(adjacency_matrix)
forceatlas2 = ForceAtlas2(verbose=False)
positions = forceatlas2.forceatlas2(G_sparse, iterations=1000)

With igraph

import igraph
from fa2 import ForceAtlas2

G = igraph.Graph.Famous("Petersen")
forceatlas2 = ForceAtlas2(verbose=False)
layout = forceatlas2.forceatlas2_igraph_layout(G, iterations=1000)
igraph.plot(G, layout=layout)

API Reference

ForceAtlas2(**kwargs)

Create a ForceAtlas2 layout engine with the following parameters:

Behavior

Parameter	Type	Default	Description
outboundAttractionDistribution	bool	False	Dissuade hubs — distributes attraction along outbound edges so hubs are pushed to borders
linLogMode	bool	False	Use Noack's LinLog model: F = log(1 + distance) instead of F = distance. Produces tighter community clusters
adjustSizes	bool	False	Prevent node overlap using anti-collision forces (Gephi parity). Pass sizes or size_attr to set node radii
edgeWeightInfluence	float	1.0	How much edge weights matter. 0 = all edges equal, 1 = normal, other values apply weight^influence
normalizeEdgeWeights	bool	False	Min-max normalize edge weights to [0, 1]. Applied after inversion
invertedEdgeWeightsMode	bool	False	Invert edge weights (w = 1/w). Applied before normalization

Performance

Parameter	Type	Default	Description
barnesHutOptimize	bool	True	Use Barnes-Hut tree approximation for repulsion. Reduces O(n^2) to O(n log n)
barnesHutTheta	float	1.2	Barnes-Hut accuracy/speed tradeoff. Lower = more accurate but slower
jitterTolerance	float	1.0	How much oscillation is tolerated during convergence. Higher = faster but less precise
backend	str	"auto"	"auto": Cython if compiled, else vectorized. "cython" / "loop": force loop-based (Cython or pure Python). "vectorized": NumPy (no BH, O(n²))

Tuning

Parameter	Type	Default	Description
scalingRatio	float	2.0	Repulsion strength. Higher = more spread out graph. Must be > 0
strongGravityMode	bool	False	Distance-independent gravity: constant pull regardless of distance from center
gravity	float	1.0	Center attraction strength. Prevents disconnected components from drifting. Must be >= 0

Layout & Other

Parameter	Type	Default	Description
dim	int	2	Number of layout dimensions. Use 3 for 3D layouts, etc.
seed	int/None	None	Random seed for reproducible layouts
verbose	bool	True	Show progress bar (tqdm) and timing breakdown

Class Methods

ForceAtlas2.inferSettings(G, **overrides)

Auto-tune parameters based on graph characteristics. Returns a configured ForceAtlas2 instance.

• G: Any supported graph type (ndarray, sparse, networkx.Graph, igraph.Graph)
• **overrides: Override any inferred parameter
• Returns: ForceAtlas2 instance

fa = ForceAtlas2.inferSettings(G, verbose=False, seed=42)
pos = fa.forceatlas2(G, iterations=100)

Methods

forceatlas2(G, pos=None, iterations=100, callbacks=None, sizes=None)

Compute layout from an adjacency matrix.

• G: numpy.ndarray or scipy.sparse matrix (must be symmetric)
• pos: Initial positions as (N, dim) array, or None for random
• iterations: Number of layout iterations (must be >= 1)
• callbacks: List of callback(iteration, nodes) functions
• sizes: Node radii as (N,) array (for adjustSizes=True)
• Returns: List of tuples with dim elements per node

forceatlas2_networkx_layout(G, pos=None, iterations=100, weight_attr=None, callbacks=None, size_attr=None, store_pos_as=None)

Compute layout for a NetworkX graph. Supports NetworkX 2.7+ and 3.x.

• G: networkx.Graph (undirected)
• pos: Initial positions as {node: tuple} dict
• weight_attr: Edge attribute name for weights
• callbacks: List of callback(iteration, nodes) functions
• size_attr: Node attribute name for sizes (used with adjustSizes)
• store_pos_as: If set, saves positions as node attributes under this key
• Returns: Dict of {node: tuple}

forceatlas2_igraph_layout(G, pos=None, iterations=100, weight_attr=None, callbacks=None, size_attr=None, store_pos_as=None)

Compute layout for an igraph graph.

• G: igraph.Graph (must be undirected)
• pos: Initial positions as list or (N, dim) numpy array
• weight_attr: Edge attribute name for weights
• callbacks: List of callback(iteration, nodes) functions
• size_attr: Vertex attribute name for sizes
• store_pos_as: If set, saves positions as vertex attributes
• Returns: igraph.Layout

Advanced Usage

Reproducible Layouts

Use the seed parameter for deterministic results:

fa = ForceAtlas2(seed=42, verbose=False)
pos1 = fa.forceatlas2_networkx_layout(G, iterations=1000)

fa2 = ForceAtlas2(seed=42, verbose=False)
pos2 = fa2.forceatlas2_networkx_layout(G, iterations=1000)
# pos1 == pos2 guaranteed

LinLog Mode (Community Detection)

LinLog mode replaces the linear attraction force F = distance with a logarithmic one F = log(1 + distance) (Noack's LinLog energy model). This produces layouts where communities form tighter, more clearly separated clusters:

fa = ForceAtlas2(linLogMode=True, verbose=False)
positions = fa.forceatlas2_networkx_layout(G, iterations=2000)

The attraction grows only logarithmically with distance, so distant connected nodes are pulled less strongly relative to repulsion, naturally emphasizing community structure.

Dissuade Hubs

Push high-degree nodes to the periphery by distributing attraction force across outbound edges:

fa = ForceAtlas2(outboundAttractionDistribution=True, verbose=False)
positions = fa.forceatlas2_networkx_layout(G, iterations=2000)

Each edge's attraction is divided by the source node's mass (degree + 1), so hub nodes with many connections experience less total attraction pull. An outboundAttCompensation factor (mean node mass) is applied to maintain overall force balance.

Iteration Callbacks (Animation / History)

Track positions over time for animation or convergence analysis:

history = []

def record_positions(iteration, nodes):
    if iteration % 100 == 0:
        history.append([(n.x, n.y) for n in nodes])

fa = ForceAtlas2(verbose=False, seed=42)
final_pos = fa.forceatlas2(G, iterations=1000, callbacks=[record_positions])
# history contains snapshots every 100 iterations

Custom Edge Weights

import networkx as nx

G = nx.Graph()
G.add_edge("A", "B", strength=5.0)
G.add_edge("B", "C", strength=1.0)
G.add_edge("A", "C", strength=0.5)

fa = ForceAtlas2(edgeWeightInfluence=1.0, verbose=False)
pos = fa.forceatlas2_networkx_layout(G, weight_attr="strength", iterations=1000)

3D Layout

fa = ForceAtlas2(dim=3, verbose=False, seed=42)
pos_3d = fa.forceatlas2_networkx_layout(G, iterations=1000)
# pos_3d = {node: (x, y, z), ...}

Prevent Node Overlap (adjustSizes)

import networkx as nx

G = nx.karate_club_graph()
for n in G.nodes():
    G.nodes[n]["size"] = G.degree(n) * 0.5  # Size proportional to degree

fa = ForceAtlas2(adjustSizes=True, verbose=False, seed=42)
pos = fa.forceatlas2_networkx_layout(G, iterations=1000, size_attr="size")

Auto-Tuning (inferSettings)

fa = ForceAtlas2.inferSettings(G, verbose=False, seed=42)
pos = fa.forceatlas2_networkx_layout(G, iterations=1000)

Edge Weight Processing

# Invert weights (strong connections → weak attraction)
fa = ForceAtlas2(invertedEdgeWeightsMode=True, verbose=False)

# Normalize weights to [0, 1]
fa = ForceAtlas2(normalizeEdgeWeights=True, verbose=False)

# Both combined
fa = ForceAtlas2(invertedEdgeWeightsMode=True, normalizeEdgeWeights=True, verbose=False)

Store Positions as Node Attributes

fa = ForceAtlas2(verbose=False, seed=42)
pos = fa.forceatlas2_networkx_layout(G, iterations=1000, store_pos_as="fa2_pos")
# Now G.nodes[n]["fa2_pos"] == pos[n] for all nodes

Tuning Tips

Goal	Settings
Spread out	Increase scalingRatio (e.g., 10.0)
Compact	Decrease scalingRatio (e.g., 0.5), increase gravity
Community clusters	Enable linLogMode=True
Prevent hub dominance	Enable outboundAttractionDistribution=True
Faster convergence	Increase jitterTolerance (e.g., 5.0)
Higher quality	More iterations, lower jitterTolerance
Large graphs (>5000)	Keep barnesHutOptimize=True (default)
Strong gravity	Set strongGravityMode=True for constant-magnitude pull
Prevent overlap	adjustSizes=True with node sizes via size_attr
3D layout	dim=3 (or any integer >= 2)
Auto-tune	ForceAtlas2.inferSettings(G)
No Cython available	backend="vectorized" (auto-detected by default)

Performance

The Cython-compiled version provides 10-100x speedup over pure Python:

Backend Comparison (small graphs)

Graph Size	Edges	Iterations	Pure Python	Vectorized	Cython	Speedup
50 nodes	~225	100	~178ms	~11ms	~3ms	~60x
200 nodes	~377	50	~982ms	~61ms	~12ms	~82x
500 nodes	~415	20	~1,045ms	~157ms	~16ms	~65x

Large Graph Scaling (Cython, 2D)

Nodes	Edges	Iterations	Time
1,000	~10,000	50	0.08s
5,000	~52,000	10	0.19s
10,000	~105,000	5	0.28s
50,000	~525,000	1	0.87s
100,000	~1,050,000	1	1.84s
500,000	~5,250,000	1	10.9s

Dimensional Scaling (Cython, 10k nodes, 5 iterations)

Dim	Time	Overhead vs 2D
2D	0.28s	—
3D	1.06s	~3.8x
5D	4.88s	~17x

Higher dimensions use list-based NodeND (slower than scalar Node2D). The 2D path uses direct C struct fields for maximum performance.

Three backends are available via backend=:

• "auto" (default): Uses Cython if compiled, otherwise NumPy vectorized
• "vectorized": NumPy-vectorized (no Barnes-Hut, O(n²) — best for small-medium graphs without Cython)
• "loop": Pure Python loops (slowest, always available)

Benchmarks on Ubuntu Linux, Python 3.13, Cython 3.2. Barnes-Hut enabled for Cython/loop backends. Sparse random graphs with ~20 edges/node.

To verify Cython is active:

import fa2.fa2util
print(fa2.fa2util.__file__)  # Should end in .so (Linux/Mac) or .pyd (Windows), not .py

Algorithm

Based on the paper:

Jacomy M, Venturini T, Heymann S, Bastian M (2014) ForceAtlas2, a Continuous Graph Layout Algorithm for Handy Network Visualization Designed for the Gephi Software. PLoS ONE 9(6): e98679. https://doi.org/10.1371/journal.pone.0098679

The implementation follows the Gephi Java source and has been verified against both the paper and the reference code.

Force Model

ForceAtlas2 uses a "(1, 1)" energy model — inverse-distance repulsion and linear attraction:

Force	Formula	Description
Repulsion	F = k_r * m1 * m2 / d	All node pairs repel. Mass = degree + 1. Barnes-Hut quadtree approximation reduces O(n^2) to O(n log n)
Linear Attraction	F = -c * w * d	Connected nodes attract proportionally to distance and edge weight
Log Attraction	F = -c * w * log(1 + d)	LinLog mode: sub-linear attraction for community emphasis
Gravity	F = m * g / d	Pull toward center, weakens with distance (standard mode)
Strong Gravity	F = c * m * g	Distance-independent pull toward center (constant magnitude)

Adaptive Speed

Each iteration measures swinging (erratic oscillation) and traction (useful movement) across all nodes. Global speed is set proportional to traction / swinging, with per-node damping for oscillating nodes. This allows fast convergence while preventing instability.

Barnes-Hut Approximation

A 2^dim spatial tree recursively partitions the space. For distant node groups, repulsion is computed against the group's center of mass instead of individual nodes. The barnesHutTheta parameter (default 1.2) controls the distance/size threshold — higher values are faster but less accurate.

Visualization & Export

Requires: pip install fa2[viz]

from fa2.viz import plot_layout, export_layout

# Render to matplotlib figure
fig = plot_layout(G, positions, color_by_degree=True, title="My Graph")

# Export to various formats
export_layout(G, positions, fmt="json", path="graph.json")   # D3.js/Sigma.js compatible
export_layout(G, positions, fmt="png", path="graph.png")     # PNG image
export_layout(G, positions, fmt="gexf", path="graph.gexf")   # Gephi format

Layout Quality Metrics

from fa2.metrics import stress, edge_crossing_count, neighborhood_preservation

s = stress(G, positions)                           # Lower is better
crossings = edge_crossing_count(G, positions)      # 2D only
np_score = neighborhood_preservation(G, positions)  # 0-1, higher is better

MCP Server (AI Agents)

ForceAtlas2 can be used as an MCP tool by AI agents:

{
    "mcpServers": {
        "fa2": {"command": "python", "args": ["-m", "fa2.mcp_server"]}
    }
}

Requires: pip install fa2[mcp]

Tools: layout_graph, layout_and_render, evaluate_layout

Migration Guide

All versions are backwards compatible — existing code continues to work unchanged.

From v1.0.x to v1.1.0

v1.1.0 adds new modules without changing any existing API. No code changes required.

What's new	Module	Install
Simple API — layout(), visualize() from edge lists	fa2.easy	included
CLI — python -m fa2 layout/render/metrics	fa2.__main__	included
Visualization — plot_layout(), export_layout()	fa2.viz	pip install fa2[viz]
Quality metrics — stress(), edge_crossing_count(), neighborhood_preservation()	fa2.metrics	included
MCP server — AI agent tools	fa2.mcp_server	pip install fa2[mcp]
Mode presets — "community", "hub-dissuade", "compact"	fa2.easy	included

Before (v1.0.x):

import numpy as np
from fa2 import ForceAtlas2

G = np.array([[0, 1, 1], [1, 0, 1], [1, 1, 0]], dtype=float)
fa = ForceAtlas2(linLogMode=True, verbose=False, seed=42)
pos = fa.forceatlas2(G, iterations=100)

After (v1.1.0) — same code works, plus simpler alternative:

from fa2.easy import layout

pos = layout([("A", "B"), ("B", "C"), ("A", "C")], mode="community")

From v0.3.x to v1.x

Change	v0.3.x	v1.0.0+
Python support	2.7, 3.x	3.9+ only
NetworkX	2.x only	2.7+ and 3.x
Cython	0.29.x	3.x
linLogMode	Not implemented	Implemented (correct log(1+d) formula)
seed parameter	Not available	New — for reproducibility
callbacks	Not available	New — for animation/monitoring
dim parameter	N/A	New — 3D+ layouts
adjustSizes	Silent no-op	Implemented (Gephi anti-collision parity)
inferSettings()	N/A	New — auto-tuning from graph characteristics
normalizeEdgeWeights	N/A	New — min-max normalize to [0,1]
invertedEdgeWeightsMode	N/A	New — w = 1/w inversion
backend parameter	N/A	New — "auto", "cython", "vectorized", "loop"
igraph support	Fragile	Robust (handles weighted, edgeless, directed-rejection)
Error handling	assert statements	Proper ValueError/TypeError with messages
Input validation	Minimal	Symmetry, pos/sizes shape, param ranges, self-loop warning
Barnes-Hut	Double-counting leaf repulsion	Correct one-sided repulsion (matches Gephi)
multiThreaded	Silent no-op	Raises NotImplementedError

Breaking changes (v0.3.x → v1.x)

• Python 2 dropped: Python 2.x is no longer supported.
• multiThreaded=True now raises NotImplementedError instead of being silently ignored.
• Invalid parameter values (negative scalingRatio, etc.) now raise ValueError.

Development

# Clone and install
git clone https://github.com/bhargavchippada/forceatlas2.git
cd forceatlas2
pip install cython numpy
pip install -e ".[dev]" --no-build-isolation

# Run tests (372 total)
pytest tests/ -v

# Run tests with coverage
pytest tests/ --cov=fa2 --cov-report=term-missing

# Run benchmarks only
pytest tests/test_benchmark.py --benchmark-only -s

# Lint
ruff check fa2/ tests/

# Regenerate C file after modifying fa2util.pyx
cython fa2/fa2util.pyx -3 -o fa2/fa2util.c

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Write tests for new functionality
4. Ensure all tests pass (100% coverage on forceatlas2.py)
5. Submit a pull request

Areas needing help

• multiThreaded: Parallel force computation. Gephi parallelizes repulsion + gravity (not attraction) with thread pooling. Python's GIL limits benefit, but Cython nogil or multiprocessing could help.

License

Copyright (C) 2017 Bhargav Chippada [e-mail]
Licensed under the GNU GPLv3.

Based on the Gephi ForceAtlas2 plugin:

Copyright 2008-2011 Gephi
Authors: Mathieu Jacomy <[e-mail]>
Licensed under GPL v3 / CDDL

And Max Shinn's Python port:

Copyright 2016 Max Shinn <[e-mail]>
Available under the GPLv3

Also thanks to Eugene Bosiakov (https://github.com/bosiakov/fa2l).

2D Grid graph (25x25) laid out with ForceAtlas2