process/generate_train_examples.py

import networkx as nx
import itertools
import os
import json
from tqdm import tqdm
import random
def build_graph(graph_dict):
    G = nx.Graph()
    for node, neighbors in graph_dict.items():
        for neighbor in neighbors:
            G.add_edge(node, neighbor)
    return G

# 图转字典形式
def graph_to_dict(G):
    return {node: sorted(list(G.neighbors(node))) for node in sorted(G.nodes)}

def split_graph(graph):
    # 确保图是连通的
    if nx.is_connected(graph):
        # 获取所有节点和需要的节点数
        nodes = list(graph.nodes())
        total_nodes = len(nodes)
        half_size = total_nodes // 2

        # 生成所有可能的节点组合，遍历分割不同的节点数量
        for size in range(1, half_size+1):
            all_combinations = itertools.combinations(nodes, half_size+1-size)

            # 遍历所有组合，检查是否可以分割成两个连通的子图
            for combination in all_combinations:
                part1 = set(combination)
                part2 = set(nodes) - part1

                # 获取子图
                subgraph1 = graph.subgraph(part1).copy()
                subgraph2 = graph.subgraph(part2).copy()

                # 检查子图是否连通
                if nx.is_connected(subgraph1) and nx.is_connected(subgraph2):
                    return subgraph1, subgraph2

        # 如果没有找到符合条件的组合，抛出异常
        raise ValueError("Unable to split graph into two connected subgraphs.")

    else:
        raise ValueError("Graph is not connected!")


# 生成连通子图（带去重 + 可恢复节点）
def find_connected_subgraphs(G):
    results = []
    seen = set()  # 用于去重
    nodes = list(G.nodes)
    max_remove = len(nodes) - 1
    for k in range(1, max_remove + 1):
        for to_remove in itertools.combinations(nodes, k):
            G_copy = G.copy()
            G_copy.remove_nodes_from(to_remove)
            if not nx.is_connected(G_copy):
                continue

            # 生成唯一 key（节点+边）用于去重
            key = (
                frozenset(G_copy.nodes),
                frozenset((min(a, b), max(a, b)) for a, b in G_copy.edges)
            )
            if key in seen:
                continue
            seen.add(key)

            # 找出可以恢复的节点
            restorable = []
            for node in to_remove:
                temp_G = G_copy.copy()
                temp_G.add_node(node)
                for neighbor in graph_dict.get(node, []):
                    if neighbor in temp_G.nodes:
                        temp_G.add_edge(node, neighbor)
                if nx.is_connected(temp_G):
                    restorable.append(node)
            results.append({
                "deleted_nodes": list(to_remove),
                "restorable_nodes": restorable,
                "graph_dict": graph_to_dict(G_copy)
            })

    return results
i = 0
# 执行
cad_classes = ['building', 'chair', 'fan', 'lamp', 'table', 'tools', 'vehicle']
for cad_class in cad_classes:
    folder_name = f"./dataset/assemblies_15/{cad_class}"
    file_list = os.listdir(folder_name)
    file_list = sorted(file_list, key=lambda x: int(x.split('_')[1]))
    folder_path = './dataset/parts_15'
    parts_list = os.listdir(folder_path)
    for file_name in tqdm(file_list):
        graph_path = os.path.join(folder_name, file_name)
        step_list = os.listdir(graph_path)
        if 'new_graph.json' not in step_list:
            continue
        with open(f'{graph_path}/new_graph.json', 'r') as f:
            graph_dict = json.load(f)
        negative_parts_list = [part for part in parts_list if part not in graph_dict]
        G = build_graph(graph_dict)
        subgraph1, subgraph2 = split_graph(G)
        results_1 = find_connected_subgraphs(subgraph1)
        results_2 = find_connected_subgraphs(subgraph2)
        i = i + 1
        # 输出结果
        examples = []
        for result in results_1:
            for restorable_node in result['restorable_nodes']:
                negative_parts = random.sample(negative_parts_list, 1)
                example = {"partial assembly": result['graph_dict'], "label": restorable_node, 'negative': [negative_part.split('.')[0] for negative_part in negative_parts]}
                examples.append(example)
        for result in results_2:
            for restorable_node in result['restorable_nodes']:
                negative_parts = random.sample(negative_parts_list, 1)
                example = {"partial assembly": result['graph_dict'], "label": restorable_node, 'negative': [negative_part.split('.')[0] for negative_part in negative_parts]}
                examples.append(example)
        with open(f'{graph_path}/train_examples.json', 'w') as f:
            json.dump(examples, f, indent=4)