AutoGPT/classic/benchmark/agbenchmark/utils/dependencies/graphs.py

454 lines
14 KiB
Python

import json
import logging
import math
from pathlib import Path
from typing import Any, Dict, List, Tuple
import matplotlib.patches as patches
import matplotlib.pyplot as plt
import networkx as nx
import numpy as np
from pyvis.network import Network
from agbenchmark.generate_test import DATA_CATEGORY
from agbenchmark.utils.utils import write_pretty_json
logger = logging.getLogger(__name__)
def bezier_curve(
src: np.ndarray, ctrl: List[float], dst: np.ndarray
) -> List[np.ndarray]:
"""
Generate Bézier curve points.
Args:
- src (np.ndarray): The source point.
- ctrl (List[float]): The control point.
- dst (np.ndarray): The destination point.
Returns:
- List[np.ndarray]: The Bézier curve points.
"""
curve = []
for t in np.linspace(0, 1, num=100):
curve_point = (
np.outer((1 - t) ** 2, src)
+ 2 * np.outer((1 - t) * t, ctrl)
+ np.outer(t**2, dst)
)
curve.append(curve_point[0])
return curve
def curved_edges(
G: nx.Graph, pos: Dict[Any, Tuple[float, float]], dist: float = 0.2
) -> None:
"""
Draw curved edges for nodes on the same level.
Args:
- G (Any): The graph object.
- pos (Dict[Any, Tuple[float, float]]): Dictionary with node positions.
- dist (float, optional): Distance for curvature. Defaults to 0.2.
Returns:
- None
"""
ax = plt.gca()
for u, v, data in G.edges(data=True):
_src = pos[u]
_dst = pos[v]
src = np.array(_src)
dst = np.array(_dst)
same_level = abs(src[1] - dst[1]) < 0.01
if same_level:
control = [(src[0] + dst[0]) / 2, src[1] + dist]
curve = bezier_curve(src, control, dst)
arrow = patches.FancyArrowPatch(
posA=curve[0], # type: ignore
posB=curve[-1], # type: ignore
connectionstyle="arc3,rad=0.2",
color="gray",
arrowstyle="-|>",
mutation_scale=15.0,
lw=1,
shrinkA=10,
shrinkB=10,
)
ax.add_patch(arrow)
else:
ax.annotate(
"",
xy=_dst,
xytext=_src,
arrowprops=dict(
arrowstyle="-|>", color="gray", lw=1, shrinkA=10, shrinkB=10
),
)
def tree_layout(graph: nx.DiGraph, root_node: Any) -> Dict[Any, Tuple[float, float]]:
"""Compute positions as a tree layout centered on the root
with alternating vertical shifts."""
bfs_tree = nx.bfs_tree(graph, source=root_node)
levels = {
node: depth
for node, depth in nx.single_source_shortest_path_length(
bfs_tree, root_node
).items()
}
pos = {}
max_depth = max(levels.values())
level_positions = {i: 0 for i in range(max_depth + 1)} # type: ignore
# Count the number of nodes per level to compute the width
level_count: Any = {}
for node, level in levels.items():
level_count[level] = level_count.get(level, 0) + 1
vertical_offset = (
0.07 # The amount of vertical shift per node within the same level
)
# Assign positions
for node, level in sorted(levels.items(), key=lambda x: x[1]):
total_nodes_in_level = level_count[level]
horizontal_spacing = 1.0 / (total_nodes_in_level + 1)
pos_x = (
0.5
- (total_nodes_in_level - 1) * horizontal_spacing / 2
+ level_positions[level] * horizontal_spacing
)
# Alternately shift nodes up and down within the same level
pos_y = (
-level
+ (level_positions[level] % 2) * vertical_offset
- ((level_positions[level] + 1) % 2) * vertical_offset
)
pos[node] = (pos_x, pos_y)
level_positions[level] += 1
return pos
def graph_spring_layout(
dag: nx.DiGraph, labels: Dict[Any, str], tree: bool = True
) -> None:
num_nodes = len(list(dag.nodes()))
# Setting up the figure and axis
fig, ax = plt.subplots()
ax.axis("off") # Turn off the axis
base = 3.0
if num_nodes > 10:
base /= 1 + math.log(num_nodes)
font_size = base * 10
font_size = max(10, base * 10)
node_size = max(300, base * 1000)
if tree:
root_node = [node for node, degree in dag.in_degree() if degree == 0][0]
pos = tree_layout(dag, root_node)
else:
# Adjust k for the spring layout based on node count
k_value = 3 / math.sqrt(num_nodes)
pos = nx.spring_layout(dag, k=k_value, iterations=50)
# Draw nodes and labels
nx.draw_networkx_nodes(dag, pos, node_color="skyblue", node_size=int(node_size))
nx.draw_networkx_labels(dag, pos, labels=labels, font_size=int(font_size))
# Draw curved edges
curved_edges(dag, pos) # type: ignore
plt.tight_layout()
plt.show()
def rgb_to_hex(rgb: Tuple[float, float, float]) -> str:
return "#{:02x}{:02x}{:02x}".format(
int(rgb[0] * 255), int(rgb[1] * 255), int(rgb[2] * 255)
)
def get_category_colors(categories: Dict[Any, str]) -> Dict[str, str]:
unique_categories = set(categories.values())
colormap = plt.cm.get_cmap("tab10", len(unique_categories)) # type: ignore
return {
category: rgb_to_hex(colormap(i)[:3])
for i, category in enumerate(unique_categories)
}
def graph_interactive_network(
dag: nx.DiGraph,
labels: Dict[Any, Dict[str, Any]],
html_graph_path: str = "",
) -> None:
nt = Network(notebook=True, width="100%", height="800px", directed=True)
category_colors = get_category_colors(DATA_CATEGORY)
# Add nodes and edges to the pyvis network
for node, json_data in labels.items():
label = json_data.get("name", "")
# remove the first 4 letters of label
label_without_test = label[4:]
node_id_str = node.nodeid
# Get the category for this label
category = DATA_CATEGORY.get(
label, "unknown"
) # Default to 'unknown' if label not found
# Get the color for this category
color = category_colors.get(category, "grey")
nt.add_node(
node_id_str,
label=label_without_test,
color=color,
data=json_data,
)
# Add edges to the pyvis network
for edge in dag.edges():
source_id_str = edge[0].nodeid
target_id_str = edge[1].nodeid
edge_id_str = (
f"{source_id_str}_to_{target_id_str}" # Construct a unique edge id
)
if not (source_id_str in nt.get_nodes() and target_id_str in nt.get_nodes()):
logger.warning(
f"Skipping edge {source_id_str} -> {target_id_str} due to missing nodes"
)
continue
nt.add_edge(source_id_str, target_id_str, id=edge_id_str)
# Configure physics for hierarchical layout
hierarchical_options = {
"enabled": True,
"levelSeparation": 200, # Increased vertical spacing between levels
"nodeSpacing": 250, # Increased spacing between nodes on the same level
"treeSpacing": 250, # Increased spacing between different trees (for forest)
"blockShifting": True,
"edgeMinimization": True,
"parentCentralization": True,
"direction": "UD",
"sortMethod": "directed",
}
physics_options = {
"stabilization": {
"enabled": True,
"iterations": 1000, # Default is often around 100
},
"hierarchicalRepulsion": {
"centralGravity": 0.0,
"springLength": 200, # Increased edge length
"springConstant": 0.01,
"nodeDistance": 250, # Increased minimum distance between nodes
"damping": 0.09,
},
"solver": "hierarchicalRepulsion",
"timestep": 0.5,
}
nt.options = {
"nodes": {
"font": {
"size": 20, # Increased font size for labels
"color": "black", # Set a readable font color
},
"shapeProperties": {"useBorderWithImage": True},
},
"edges": {
"length": 250, # Increased edge length
},
"physics": physics_options,
"layout": {"hierarchical": hierarchical_options},
}
# Serialize the graph to JSON and save in appropriate locations
graph_data = {"nodes": nt.nodes, "edges": nt.edges}
logger.debug(f"Generated graph data:\n{json.dumps(graph_data, indent=4)}")
# FIXME: use more reliable method to find the right location for these files.
# This will fail in all cases except if run from the root of our repo.
home_path = Path.cwd()
write_pretty_json(graph_data, home_path / "frontend" / "public" / "graph.json")
flutter_app_path = home_path.parent / "frontend" / "assets"
# Optionally, save to a file
# Sync with the flutter UI
# this literally only works in the AutoGPT repo, but this part of the code
# is not reached if BUILD_SKILL_TREE is false
write_pretty_json(graph_data, flutter_app_path / "tree_structure.json")
validate_skill_tree(graph_data, "")
# Extract node IDs with category "coding"
coding_tree = extract_subgraph_based_on_category(graph_data.copy(), "coding")
validate_skill_tree(coding_tree, "coding")
write_pretty_json(
coding_tree,
flutter_app_path / "coding_tree_structure.json",
)
data_tree = extract_subgraph_based_on_category(graph_data.copy(), "data")
# validate_skill_tree(data_tree, "data")
write_pretty_json(
data_tree,
flutter_app_path / "data_tree_structure.json",
)
general_tree = extract_subgraph_based_on_category(graph_data.copy(), "general")
validate_skill_tree(general_tree, "general")
write_pretty_json(
general_tree,
flutter_app_path / "general_tree_structure.json",
)
scrape_synthesize_tree = extract_subgraph_based_on_category(
graph_data.copy(), "scrape_synthesize"
)
validate_skill_tree(scrape_synthesize_tree, "scrape_synthesize")
write_pretty_json(
scrape_synthesize_tree,
flutter_app_path / "scrape_synthesize_tree_structure.json",
)
if html_graph_path:
file_path = str(Path(html_graph_path).resolve())
nt.write_html(file_path)
def extract_subgraph_based_on_category(graph, category):
"""
Extracts a subgraph that includes all nodes and edges required to reach all nodes
with a specified category.
:param graph: The original graph.
:param category: The target category.
:return: Subgraph with nodes and edges required to reach the nodes
with the given category.
"""
subgraph = {"nodes": [], "edges": []}
visited = set()
def reverse_dfs(node_id):
if node_id in visited:
return
visited.add(node_id)
node_data = next(node for node in graph["nodes"] if node["id"] == node_id)
# Add the node to the subgraph if it's not already present.
if node_data not in subgraph["nodes"]:
subgraph["nodes"].append(node_data)
for edge in graph["edges"]:
if edge["to"] == node_id:
if edge not in subgraph["edges"]:
subgraph["edges"].append(edge)
reverse_dfs(edge["from"])
# Identify nodes with the target category and initiate reverse DFS from them.
nodes_with_target_category = [
node["id"] for node in graph["nodes"] if category in node["data"]["category"]
]
for node_id in nodes_with_target_category:
reverse_dfs(node_id)
return subgraph
def is_circular(graph):
def dfs(node, visited, stack, parent_map):
visited.add(node)
stack.add(node)
for edge in graph["edges"]:
if edge["from"] == node:
if edge["to"] in stack:
# Detected a cycle
cycle_path = []
current = node
while current != edge["to"]:
cycle_path.append(current)
current = parent_map.get(current)
cycle_path.append(edge["to"])
cycle_path.append(node)
return cycle_path[::-1]
elif edge["to"] not in visited:
parent_map[edge["to"]] = node
cycle_path = dfs(edge["to"], visited, stack, parent_map)
if cycle_path:
return cycle_path
stack.remove(node)
return None
visited = set()
stack = set()
parent_map = {}
for node in graph["nodes"]:
node_id = node["id"]
if node_id not in visited:
cycle_path = dfs(node_id, visited, stack, parent_map)
if cycle_path:
return cycle_path
return None
def get_roots(graph):
"""
Return the roots of a graph. Roots are nodes with no incoming edges.
"""
# Create a set of all node IDs
all_nodes = {node["id"] for node in graph["nodes"]}
# Create a set of nodes with incoming edges
nodes_with_incoming_edges = {edge["to"] for edge in graph["edges"]}
# Roots are nodes that have no incoming edges
roots = all_nodes - nodes_with_incoming_edges
return list(roots)
def validate_skill_tree(graph, skill_tree_name):
"""
Validate if a given graph represents a valid skill tree
and raise appropriate exceptions if not.
:param graph: A dictionary representing the graph with 'nodes' and 'edges'.
:raises: ValueError with a description of the invalidity.
"""
# Check for circularity
cycle_path = is_circular(graph)
if cycle_path:
cycle_str = " -> ".join(cycle_path)
raise ValueError(
f"{skill_tree_name} skill tree is circular! "
f"Detected circular path: {cycle_str}."
)
# Check for multiple roots
roots = get_roots(graph)
if len(roots) > 1:
raise ValueError(f"{skill_tree_name} skill tree has multiple roots: {roots}.")
elif not roots:
raise ValueError(f"{skill_tree_name} skill tree has no roots.")