from __future__ import annotations import argparse import tempfile from dataclasses import dataclass from pathlib import Path from typing import Any, Dict, Iterable from torch.utils.data import Dataset, DataLoader import pandas as pd import yaml import ipdb # 记录必要的字段信息是否提供 field_dict = { 'dataset': {'modalities': ['meta_features', 'node_level_task_graphs', 'time_series']}, 'graphs': ['', '', '', ''], 'time_series': ['', '', '', ' ', ''] } def read_yaml(path: Path) -> dict[str, Any]: with path.open("r", encoding="utf-8") as f: return yaml.safe_load(f) def read_all_configs(dataset_yaml_path: str) -> Dict[str, Any]: """ 从dataset.yaml开始，读取所有相关配置文件 Args: dataset_yaml_path: dataset.yaml文件路径 Returns: 整合所有配置的字典 """ # 获取协议目录路径 protocol_dir = Path(dataset_yaml_path).parent # 1. 读取dataset.yaml dataset_config = read_yaml(Path(dataset_yaml_path)) print(f"已读取 dataset.yaml: {dataset_config['dataset_name']}") # 2. 读取图配置 graph_configs = {} if 'node_level_task_graphs' in dataset_config['modalities']: graph_ids = dataset_config['modalities']['node_level_task_graphs'].get('graph_ids', []) for graph_id in graph_ids: graph_yaml_path = protocol_dir / f"input_multimodal/node_level_task_graphs/{graph_id}/graph.yaml" if graph_yaml_path.exists(): graph_config = read_yaml(Path(graph_yaml_path)) graph_configs[graph_id] = graph_config print(f"已读取 graph.yaml: {graph_id}") else: print(f"警告: 未找到 graph.yaml: {graph_yaml_path}") # 3. 读取时序配置 ts_configs = {} if 'time_series' in dataset_config['modalities']: ts_ids = dataset_config['modalities']['time_series'].get('time_series_ids', []) for ts_id in ts_ids: ts_yaml_path = protocol_dir / f"input_multimodal/time_series/{ts_id}/series.yaml" if ts_yaml_path.exists(): ts_config = read_yaml(Path(ts_yaml_path)) ts_configs[ts_id] = ts_config print(f"已读取 series.yaml: {ts_id}") else: print(f"警告: 未找到 series.yaml: {ts_yaml_path}") # 4. 整合所有配置 all_configs = { 'dataset': dataset_config, 'graphs': graph_configs, 'time_series': ts_configs } return all_configs def _first_existing(candidates: Iterable[Path]) -> Path: candidate_list = [Path(p).expanduser() for p in candidates] for p in candidate_list: # 从多个可能路径里面，返回其中的一个存在的路径 if p.is_file(): return p.resolve() shown = ", ".join(str(c) for c in candidate_list) raise FileNotFoundError(f"None of the candidate paths exist: {shown}") def _load_frame(path: Path) -> pd.DataFrame: """Load a single-table file as used in protocol YAMLs (csv / parquet / h5).""" path = Path(path).expanduser().resolve() suf = path.suffix.lower() if suf == ".csv": return pd.read_csv(path) if suf in (".tsv", ".txt"): return pd.read_csv(path, sep="\t") if suf in (".parquet", ".pq"): return pd.read_parquet(path) if suf in (".h5", ".hdf5", ".hdf"): return pd.read_hdf(path) raise ValueError(f"Unsupported data file extension for {path}") def _meta_candidates(dataset_dir: Path, file_name: str) -> list[Path]: p = Path(file_name).expanduser() if p.is_absolute(): return [p] return [ dataset_dir / "input_multimodal" / file_name, ] def _graph_file_candidates(dataset_dir: Path, graph_level: str, graph_id: str, file_name: str) -> list[Path]: p = Path(file_name).expanduser() if p.is_absolute(): return [p] base = [ dataset_dir / "input_multimodal" / graph_level / graph_id / file_name, ] return base def _series_file_candidates(dataset_dir: Path, series_id: str, file_name: str) -> list[Path]: # 生成了所有潜在可能的存储路径 p = Path(file_name).expanduser() if p.is_absolute(): return [p] return [ dataset_dir / "input_multimodal" / "time_series" / series_id / file_name, dataset_dir / file_name, dataset_dir / "input_multimodal" / file_name, ] def _parse_meta(meta: pd.DataFrame) -> Dict[str, Any]: # 对meta_df进行解析, 解析df中的每一列表头信息，转化为字典形式 meta_dict = {} for col in meta.columns: meta_dict[col] = list(meta[col]) return meta_dict def _parse_graph(graph: pd.DataFrame, graph_features: list[str] | None) -> Dict[str, Any]: ''' graph字典中必要的字段信息包括： - 稀疏的邻接表 - 每一条边的权重 - 边上的特征 (可选) ''' graph_dict = {} required_fields = ['src_id', 'dst_id', 'weight'] for field in required_fields: if field not in graph.columns: raise ValueError(f"Missing required column: {field}") graph_dict['edge'] = list(zip(graph['src_id'], graph['dst_id'])) graph_dict['weight'] = graph['weight'].tolist() if all(x == 1.0 for x in list(graph_dict['weight'])): graph_dict['weighted'] = False else: graph_dict['weighted'] = True if graph_features is not None: # 如果图中有特征，则将特征添加到字典中 for feature in graph_features: graph_dict[feature] = list(graph[feature]) return graph_dict def _parse_time_series(time_series: pd.DataFrame, scfg: dict[str, Any], time2graph_map: Dict[str, Any]|None) -> Dict[str, Any]: time_series_dict = {} variables_mode = scfg.get("variables_mode") time_series_dict['variables_mode'] = variables_mode # 可以通过mode确定是单特征还是多特征 if variables_mode == "single": # 如果是单一特征，index就是timestamp，col_name就是sample_id, 特征名就是series_id time_type = str(time_series.index.dtype) time_series_dict['time_type'] = time_type time_series_dict['timestamp'] = time_series.index.tolist() time_series_dict['value'] = {} if time2graph_map: # for col in time_series.columns: time_series_dict['value'][time2graph_map.get(int(col), col)] = list(time_series[col]) else: for col in time_series.columns: time_series_dict['value'][col] = list(time_series[col]) else: # 如果是多特征，表头会是 sample_id, timestamp, feature_name1, feature_name2, ... # 需要提取timestamp, 并根据timestamp的顺序，提取对应的特征value，结构应该是feature_name*: {sample_id: [values]}的形式 index_cfg = scfg.get("index") or {} id_col = index_cfg.get("id_colunm") or index_cfg.get("id_column") time_col = index_cfg.get("time_column") if not id_col: raise ValueError("time_series.index must configure id_colunm or id_column") if not time_col: raise ValueError("time_series.index must configure time_column") required_columns = [id_col, time_col] missing_columns = [col for col in required_columns if col not in time_series.columns] if missing_columns: raise ValueError(f"Missing required time series column(s): {missing_columns}") series_features = index_cfg.get("feature_column") if not series_features: series_features = [col for col in time_series.columns if col not in required_columns] elif isinstance(series_features, str): series_features = [series_features] else: series_features = list(series_features) if not series_features: raise ValueError("No feature columns configured or inferred for time series") missing_features = [col for col in series_features if col not in time_series.columns] if missing_features: raise ValueError(f"Missing configured feature column(s): {missing_features}") sorted_time_series = time_series.sort_values([id_col, time_col]) time_series_dict['feature_names'] = series_features time_series_dict['time_type'] = str(time_series[time_col].dtype) time_series_dict['timestamp'] = sorted(time_series[time_col].drop_duplicates().tolist()) timestamp_by_sample = {} time_series_dict['value'] = {feature: {} for feature in series_features} for sample_id, sample_df in sorted_time_series.groupby(id_col, sort=False): output_sample_id = time2graph_map.get(sample_id, sample_id) if time2graph_map else sample_id timestamp_by_sample[output_sample_id] = sample_df[time_col].tolist() for feature in series_features: time_series_dict['value'][feature][output_sample_id] = sample_df[feature].tolist() if any(timestamps != time_series_dict['timestamp'] for timestamps in timestamp_by_sample.values()): time_series_dict['timestamp_by_sample'] = timestamp_by_sample return time_series_dict @dataclass class LoadedDataset: dataset_dir: Path config: dict[str, Any] meta_features: Dict[str, Any] graphs: Dict[str, Any] time_series: Dict[str, Any] # def _check_graphs(graph_config: dict[str, Any]) -> None: # check # 会将所有的yaml读成一个config class MultiModalDataset(LoadedDataset): """ 使用 ``read_all_configs`` 返回的整合配置，在 ``dataset_dir`` 下按协议解析数据路径并加载表格。 ``dataset_dir`` 通常为协议根目录（含 ``input_multimodal/`` 与 ``dataset.yaml`` 的父目录）， 或等价的数据根目录；路径解析会依次尝试 ``dataset_dir/`` 与 ``dataset_dir/input_multimodal/...`` 等候选位置。 """ def __init__(self, config: dict[str, Any], dataset_dir: Path | str) -> None: root = Path(dataset_dir).expanduser().resolve() dataset_cfg = config["dataset"] modalities = dataset_cfg.get("modalities") or {} time2graph_map = {} # 读取meta_features数据 meta_block: Dict[str, Any] = {} if isinstance(modalities.get("meta_features"), dict): mf = modalities["meta_features"] meta_file = mf.get("file_name") if mf.get("required") and not meta_file: raise ValueError("meta_features is required but file_name is not configured") if meta_file: meta_path = _first_existing(_meta_candidates(root, str(meta_file))) meta_tmp = _load_frame(meta_path) meta_features = _parse_meta(meta_tmp) # 对meta_df进行解析 sample_def = dataset_cfg.get("sample_defination") or dataset_cfg.get( "sample_definition" ) id_col_graph = None id_col_time_series = None if isinstance(sample_def, dict): id_col_graph = sample_def.get("sample_id_graph") id_col_time_series = sample_def.get("sample_id_time_series") meta_block = { "path": str(meta_path), "features": meta_features, "sample_id_graph": id_col_graph, "sample_id_time_series": id_col_time_series, } if id_col_graph is not None and id_col_time_series is not None: # 如果id_col_graph和id_col_time_series都存在，则需要构建一个id_col_graph和id_col_time_series的映射关系，将两种模态进行对齐 # 默认是将时序的id_col_time_series映射到图的id_col_graph time2graph_map = {time_id: graph_id for time_id, graph_id in zip(meta_tmp[id_col_time_series], meta_tmp[id_col_graph])} # 读取graphs数据 graphs: Dict[str, Any] = {} for gid, gcfg in (config.get("graphs") or {}).items(): if not isinstance(gcfg, dict): continue fname = gcfg.get("file_name") graph_level = gcfg.get("graph_level") # 图的级别，例如：node_level_task_graph, graph_level_task_graph graph_directed = gcfg.get("directed") # 图的类型，例如：directed, undirected edge_features_cfg = gcfg.get("edge_features") or {} graph_features = edge_features_cfg.get("feature_name") if edge_features_cfg.get("available") else None # 图中的特征，例如：edge_features if not fname: continue gpath = _first_existing(_graph_file_candidates(root, str(graph_level), str(gid), str(fname))) tmp_graph = _load_frame(gpath) # 继续解析图中的具体特征 tmp_graph_dict = _parse_graph(tmp_graph, graph_features) tmp_graph_dict['directed'] = graph_directed tmp_graph_dict['edge_features'] = graph_features graphs[str(gid)] = tmp_graph_dict # 将图数据加载到字典中， time_series: Dict[str, Any] = {} for sid, scfg in (config.get("time_series") or {}).items(): if not isinstance(scfg, dict): continue fname = scfg.get("file_name") if not fname: continue spath = _first_existing(_series_file_candidates(root, str(sid), str(fname))) tmp_time_series = _load_frame(spath) frequency = scfg.get("temporal_granularity") tmp_time_series_dict = _parse_time_series(tmp_time_series, scfg, time2graph_map) tmp_time_series_dict['frequency'] = frequency time_series[str(sid)] = tmp_time_series_dict super().__init__( dataset_dir=root, config=config, meta_features=meta_block, graphs=graphs, time_series=time_series, ) if __name__ == '__main__': protocol_dir = Path(__file__).resolve().parent dataset_yaml_path = protocol_dir / "dataset.yaml" configs = read_all_configs(str(dataset_yaml_path)) dataset = MultiModalDataset(configs, protocol_dir) ipdb.set_trace() # 在断点中检查是不是该有的信息都被加载了 # 以下是一些断言，确保dataset.config, dataset.meta_features, dataset.graphs, dataset.time_series等字段都存在且正确 assert isinstance(dataset.config, dict), "dataset.config should be a dict" assert "dataset" in dataset.config, "dataset.config should include dataset config" assert isinstance(dataset.meta_features, dict), "dataset.meta_features should be a dict" assert "features" in dataset.meta_features, "dataset.meta_features should include parsed features" assert isinstance(dataset.meta_features["features"], dict), ( "dataset.meta_features['features'] should be a dict" ) assert isinstance(dataset.graphs, dict), "dataset.graphs should be a dict" assert "turbine_proximity" in dataset.graphs, ( "dataset.graphs should include turbine_proximity" ) assert isinstance(dataset.graphs["turbine_proximity"], dict), ( "turbine_proximity should be a parsed dict" ) assert isinstance(dataset.time_series, dict), "dataset.time_series should be a dict" assert "sdwpf_scada" in dataset.time_series, ( "dataset.time_series should include sdwpf_scada" ) sdwpf_scada = dataset.time_series["sdwpf_scada"] assert isinstance(sdwpf_scada, dict), ( "sdwpf_scada should be a parsed dict" ) for field in ("variables_mode", "timestamp", "value", "frequency"): assert field in sdwpf_scada, f"sdwpf_scada should include {field}" print(f"Dataset: {dataset.config['dataset'].get('dataset_name')}") print(f"Protocol dir: {dataset.dataset_dir}") print(f"Meta features: columns={len(dataset.meta_features.get('features', {}))}") for graph_id, graph_dict in dataset.graphs.items(): print( f"Graph[{graph_id}]: edges={len(graph_dict.get('edge', []))}, " f"directed={graph_dict.get('directed')}" ) for series_id, series_data in dataset.time_series.items(): if isinstance(series_data, dict): value = series_data.get("value", {}) if series_data.get("variables_mode") == "single": feature_count = 1 sample_count = len(value) else: feature_count = len(value) sample_count = len(next(iter(value.values()), {})) print( f"TimeSeries[{series_id}]: mode={series_data.get('variables_mode')}, " f"features={feature_count}, timestamps={len(series_data.get('timestamp', []))}, " f"samples={sample_count}" ) else: print(f"TimeSeries[{series_id}]: shape={series_data.shape}")