Star-Mapper/layout_engine.py

293 lines
10 KiB
Python

"""
Layout Engine - Efficient graph layout precomputation for large graphs.
Uses igraph (C-based) for maximum performance, with networkx fallback.
Supports multiple layout algorithms optimized for different graph sizes and structures.
"""
import numpy as np
import time
import logging
from collections import defaultdict
logger = logging.getLogger(__name__)
try:
import igraph as ig
HAS_IGRAPH = True
except ImportError:
HAS_IGRAPH = False
logger.warning("igraph not available, falling back to networkx layouts")
import networkx as nx
def compute_layout(nodes_dict, edges, algorithm='auto', spacing=1.0, iterations=300):
"""
Precompute node positions using efficient graph layout algorithms.
Strategy based on graph size:
- Small (<300 nodes): High-quality force-directed with many iterations
- Medium (300-3000): Force-directed with optimized parameters
- Large (3000-20000): Community-based layout (Louvain + per-community layout)
- Very large (>20000): DrL or spectral layout
Args:
spacing: Multiplier for target_range (1.0 = default, >1 = more spread)
iterations: Number of layout iterations (higher = better quality, slower)
Returns dict of {node_id: {'x': float, 'y': float}}
"""
if not nodes_dict:
return {}
start_time = time.time()
node_ids = list(nodes_dict.keys())
n = len(node_ids)
if n == 0:
return {}
if n == 1:
return {node_ids[0]: {'x': 0.0, 'y': 0.0}}
# Build adjacency
node_index = {nid: i for i, nid in enumerate(node_ids)}
edge_list = []
for edge in edges:
src = edge.get('source')
tgt = edge.get('target')
if src in node_index and tgt in node_index:
si, ti = node_index[src], node_index[tgt]
if si != ti: # skip self-loops for layout
edge_list.append((si, ti))
# Choose algorithm
if algorithm == 'auto':
if n > 20000:
algorithm = 'drl' if HAS_IGRAPH else 'spectral'
elif n > 3000:
algorithm = 'community'
elif n > 300:
algorithm = 'force_directed'
else:
algorithm = 'force_directed_hq'
logger.info(f"Computing layout for {n} nodes, {len(edge_list)} edges using '{algorithm}' (spacing={spacing}, iterations={iterations})")
if HAS_IGRAPH and algorithm != 'spectral':
positions = _layout_igraph(node_ids, edge_list, n, algorithm, iterations)
else:
positions = _layout_networkx(node_ids, edge_list, n, algorithm, iterations)
# Apply spacing multiplier by re-normalizing with scaled range
target_range = 2000 * spacing
elapsed = time.time() - start_time
logger.info(f"Layout computed in {elapsed:.2f}s")
# Re-normalize if spacing != 1.0
if abs(spacing - 1.0) > 0.01:
coords_list = [[positions[nid]['x'], positions[nid]['y']] for nid in node_ids]
coords = np.array(coords_list)
return _normalize_and_scale(node_ids, coords, target_range=target_range)
return positions
def _layout_igraph(node_ids, edge_list, n, algorithm, iterations=300):
"""Use igraph's C-based layout algorithms for maximum speed."""
g = ig.Graph(n=n, edges=edge_list, directed=False)
iters = max(50, int(iterations))
if algorithm == 'drl':
# DrL: Distributed Recursive Layout - excellent for very large graphs
layout = g.layout_drl()
elif algorithm == 'community':
# Community-based: detect communities, then layout
layout = _community_layout_igraph(g, iterations=iters)
elif algorithm == 'force_directed_hq':
# High quality Fruchterman-Reingold with more iterations
layout = g.layout_fruchterman_reingold(niter=max(iters, 500))
elif algorithm == 'fruchterman_reingold':
layout = g.layout_fruchterman_reingold(niter=iters)
elif algorithm == 'kamada_kawai':
if n < 1000:
layout = g.layout_kamada_kawai()
else:
layout = g.layout_fruchterman_reingold(niter=iters)
elif algorithm == 'circle':
layout = g.layout_circle()
else:
# Default: Fruchterman-Reingold
layout = g.layout_fruchterman_reingold(niter=iters)
coords = np.array(layout.coords)
return _normalize_and_scale(node_ids, coords)
def _community_layout_igraph(g, iterations=200):
"""
Community-based layout: detect communities with Louvain,
arrange communities in a circle, then layout nodes within each community.
"""
try:
communities = g.community_multilevel()
membership = communities.membership
except Exception:
return g.layout_fruchterman_reingold(niter=200)
n = g.vcount()
positions = np.zeros((n, 2))
# Group nodes by community
comm_nodes = defaultdict(list)
for i, comm_id in enumerate(membership):
comm_nodes[comm_id].append(i)
num_communities = len(comm_nodes)
# Arrange communities in a circle
comm_positions = {}
radius = max(500, num_communities * 50)
for idx, comm_id in enumerate(sorted(comm_nodes.keys())):
angle = 2 * np.pi * idx / max(num_communities, 1)
comm_positions[comm_id] = (radius * np.cos(angle), radius * np.sin(angle))
# Layout nodes within each community
for comm_id, node_indices in comm_nodes.items():
cx, cy = comm_positions[comm_id]
if len(node_indices) == 1:
positions[node_indices[0]] = [cx, cy]
continue
# Create subgraph
subgraph = g.subgraph(node_indices)
# Use FR layout for the subgraph
sub_layout = subgraph.layout_fruchterman_reingold(niter=iterations)
sub_coords = np.array(sub_layout.coords)
# Scale based on community size
scale = max(30, np.sqrt(len(node_indices)) * 15)
if sub_coords.std() > 0:
sub_coords = (sub_coords - sub_coords.mean(axis=0)) / max(sub_coords.std(), 1e-6) * scale
# Offset to community position
for local_idx, global_idx in enumerate(node_indices):
positions[global_idx] = [cx + sub_coords[local_idx][0], cy + sub_coords[local_idx][1]]
return ig.Layout(positions.tolist())
def _layout_networkx(node_ids, edge_list, n, algorithm, iterations=300):
"""Fallback to networkx layouts."""
G = nx.Graph()
G.add_nodes_from(range(n))
G.add_edges_from(edge_list)
iters = max(50, int(iterations))
if algorithm == 'spectral':
try:
pos = nx.spectral_layout(G, dim=2, scale=1000)
except Exception:
pos = nx.spring_layout(G, seed=42, scale=1000, iterations=iters)
elif algorithm in ('force_directed_hq', 'force_directed'):
k = 2.0 / np.sqrt(max(n, 1))
pos = nx.spring_layout(G, k=k, iterations=iters, scale=1000, seed=42)
elif algorithm == 'community':
pos = _community_layout_networkx(G, n)
elif algorithm == 'kamada_kawai':
if n < 500:
pos = nx.kamada_kawai_layout(G, scale=1000)
else:
pos = nx.spring_layout(G, seed=42, scale=1000, iterations=iters)
elif algorithm == 'circle':
pos = nx.circular_layout(G, scale=1000)
else:
k = 2.0 / np.sqrt(max(n, 1))
pos = nx.spring_layout(G, k=k, iterations=iters, scale=1000, seed=42)
coords = np.array([pos[i] for i in range(n)])
return _normalize_and_scale(node_ids, coords)
def _community_layout_networkx(G, n):
"""Community-based layout using networkx."""
try:
from networkx.algorithms.community import greedy_modularity_communities
communities = list(greedy_modularity_communities(G))
except Exception:
return nx.spring_layout(G, seed=42, scale=1000, iterations=200)
positions = {}
num_communities = len(communities)
radius = max(500, num_communities * 50)
for idx, comm in enumerate(communities):
angle = 2 * np.pi * idx / max(num_communities, 1)
cx = radius * np.cos(angle)
cy = radius * np.sin(angle)
comm_nodes = list(comm)
if len(comm_nodes) == 1:
positions[comm_nodes[0]] = np.array([cx, cy])
continue
subG = G.subgraph(comm_nodes)
sub_pos = nx.spring_layout(subG, seed=42, scale=max(20, np.sqrt(len(comm_nodes)) * 15), iterations=200)
for node, pos in sub_pos.items():
positions[node] = pos + np.array([cx, cy])
return positions
def _normalize_and_scale(node_ids, coords, target_range=2000):
"""Normalize coordinates to a centered range for consistent rendering."""
if len(coords) == 0:
return {}
# Center
center = coords.mean(axis=0)
coords = coords - center
# Scale to target range
max_extent = max(np.abs(coords).max(), 1e-6)
coords = coords / max_extent * (target_range / 2)
# Add slight jitter to prevent perfect overlaps
jitter = np.random.RandomState(42).uniform(-1, 1, coords.shape) * 2
coords += jitter
positions = {}
for i, node_id in enumerate(node_ids):
positions[node_id] = {'x': float(coords[i][0]), 'y': float(coords[i][1])}
return positions
def get_available_algorithms():
"""Return list of available layout algorithms."""
algos = [
{'id': 'auto', 'name': 'Auto (best for size)', 'description': 'Automatically selects based on graph size'},
{'id': 'force_directed', 'name': 'Force-Directed', 'description': 'Classic spring-electric layout'},
{'id': 'force_directed_hq', 'name': 'Force-Directed (HQ)', 'description': 'Higher quality, more iterations'},
{'id': 'community', 'name': 'Community-Based', 'description': 'Groups by community detection'},
{'id': 'circle', 'name': 'Circular', 'description': 'Nodes arranged in a circle'},
]
if HAS_IGRAPH:
algos.extend([
{'id': 'drl', 'name': 'DrL (Large Graphs)', 'description': 'Distributed Recursive Layout for 10k+ nodes'},
{'id': 'kamada_kawai', 'name': 'Kamada-Kawai', 'description': 'Energy-based layout (small graphs)'},
])
else:
algos.append(
{'id': 'spectral', 'name': 'Spectral', 'description': 'Eigenvalue-based layout (fast for large graphs)'},
)
return algos