"""在 moleculestm 环境中运行：把数据集里所有唯一 SMILES 编码成 MoleculeSTM(Graph 版) 表征缓存。

用法（工作目录 = lnp_ml 根）:
    conda activate moleculestm
    export HF_HUB_OFFLINE=1                      # 只用本地权重
    python scripts/precompute_moleculestm.py \
        --stm-repo /path/to/MoleculeSTM \
        --ckpt MoleculeSTM_ckpt/pretrained_MoleculeSTM/<某个Graph变体>/molecule_model.pth \
        --out data/processed/moleculestm_embeddings.npz

产物 .npz 含 "smiles" (N,) 与 "embeddings" (N, D=300)，供主项目查表使用。
"""
import argparse
import sys
from pathlib import Path

import numpy as np
import pandas as pd
import torch

DATA_FILES = [
    "data/interim/internal.csv",
    "data/external/all_data_LiON.csv",
]
GLOB_FILES = [
    "data/external/all_amine_split_for_LiON/cv_*/train.csv",
    "data/external/all_amine_split_for_LiON/cv_*/test.csv",
]


def collect_smiles(root: Path) -> list:
    seen: set = set()
    files = [root / f for f in DATA_FILES]
    for pattern in GLOB_FILES:
        files += sorted(root.glob(pattern))
    for f in files:
        if not f.exists():
            print(f"跳过不存在的文件: {f}")
            continue
        df = pd.read_csv(f, low_memory=False)
        if "smiles" in df.columns:
            seen.update(df["smiles"].dropna().astype(str).tolist())
    return sorted(seen)


def build_encoder(stm_repo: str, ckpt: str, device: str):
    """加载 MoleculeSTM Graph 分子编码器（GNN_graphpred 包装的 GIN）。"""
    sys.path.insert(0, stm_repo)
    from MoleculeSTM.models import GNN, GNN_graphpred  # 官方模型定义

    node = GNN(num_layer=5, emb_dim=300, JK="last", drop_ratio=0.0, gnn_type="gin")
    model = GNN_graphpred(
        num_layer=5, emb_dim=300, num_tasks=1, JK="last",
        graph_pooling="mean", molecule_node_model=node,
    )
    state = torch.load(ckpt, map_location="cpu")
    model.load_state_dict(state)   # 与官方一致，key 完全匹配
    model.eval().to(device)
    return model


def smiles_to_data(smiles: str, stm_repo: str):
    """SMILES -> torch_geometric Data（用官方 mol_to_graph_data_obj_simple）。"""
    from rdkit import Chem
    # 官方图特征化函数（不同版本路径可能是 datasets.utils 或 datasets.molecule_datasets）
    from MoleculeSTM.datasets.utils import mol_to_graph_data_obj_simple

    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        return None
    return mol_to_graph_data_obj_simple(mol)


@torch.no_grad()
def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--stm-repo", required=True, help="官方 MoleculeSTM 仓库路径")
    ap.add_argument("--ckpt", required=True, help="molecule_model.pth 路径")
    ap.add_argument("--out", default="data/processed/moleculestm_embeddings.npz")
    ap.add_argument("--root", default=".", help="lnp_ml 仓库根目录")
    ap.add_argument("--batch-size", type=int, default=256)
    ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
    args = ap.parse_args()

    from torch_geometric.data import DataLoader
    from torch_geometric.nn import global_mean_pool

    model = build_encoder(args.stm_repo, args.ckpt, args.device)
    smiles = collect_smiles(Path(args.root))
    print(f"收集到 {len(smiles)} 个唯一 SMILES，开始编码...")

    # 图特征化（记录成功的 SMILES，跳过 RDKit 解析失败的）
    data_list, kept = [], []
    for s in smiles:
        d = smiles_to_data(s, args.stm_repo)
        if d is not None:
            data_list.append(d)
            kept.append(s)
    if len(kept) < len(smiles):
        print(f"警告: {len(smiles) - len(kept)} 个 SMILES 无法图特征化，已跳过")

    loader = DataLoader(data_list, batch_size=args.batch_size, shuffle=False)
    chunks = []
    for batch in loader:
        batch = batch.to(args.device)
        graph_repr, _ = model(batch.x, batch.edge_index, batch.edge_attr, batch.batch)  # [B, 300]
        chunks.append(graph_repr.cpu().numpy().astype(np.float32))
        print(f"  已编码 {sum(c.shape[0] for c in chunks)}/{len(kept)}")
    embeddings = np.vstack(chunks)

    out = Path(args.out)
    out.parent.mkdir(parents=True, exist_ok=True)
    np.savez_compressed(out, smiles=np.array(kept), embeddings=embeddings)
    print(f"已保存 {embeddings.shape} 到 {out}")


if __name__ == "__main__":
    main()