visualization/ppl_heat_map.py

import json
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import argparse
from pathlib import Path

# 固定的月份顺序
months = [2407, 2408, 2409, 2410, 2411, 2412,
          2501, 2502, 2503, 2504, 2505, 2506]

# ---------- 数据加载 ----------
def load_results(path: Path):
    results = []
    if path.is_file():
        files = [path]
    elif path.is_dir():
        files = sorted(path.rglob("ppl_metrics.jsonl"))
        if not files:
            raise FileNotFoundError(f"目录 {path} 下没有找到 ppl_metrics.jsonl")
    else:
        raise FileNotFoundError(f"{path} 不存在")

    for file in files:
        # ⬇️ 从文件夹名提取 ckpt 月份，只取前4位数字
        ckpt_str = ''.join(filter(str.isdigit, file.parent.name[:4]))

        with open(file, "r", encoding="utf-8") as f:
            for line in f:
                if line.strip():
                    row = json.loads(line)
                    # 加上 ckpt_month 字段（训练月）
                    row["ckpt_month"] = ckpt_str
                    # 原来的 month 就是 val_month（验证月）
                    row["val_month"] = str(row["month"])
                    results.append(row)
    return results


# ---------- 构造矩阵 ----------
def build_matrix(results, metric="val_ppl"):
    n = len(months)
    mat = np.full((n, n), np.nan)

    # ckpt_month -> val_month -> list
    data = {}
    for row in results:
        ckpt = str(row["ckpt_month"])
        val = str(row["val_month"])
        value = row.get(metric, None)
        if value is None:
            continue
        data.setdefault(ckpt, {}).setdefault(val, []).append(value)

    # 对齐 months 顺序，填矩阵
    for i, ckpt in enumerate(months):
        for j, val in enumerate(months):
            if str(ckpt) in data and str(val) in data[str(ckpt)]:
                mat[i, j] = np.mean(data[str(ckpt)][str(val)])

    return mat


# ---------- 绘制热图 ----------
def plot_heatmap(matrix, metric="val_ppl", save_path="heatmap.png"):
    plt.figure(figsize=(10, 8))
    sns.heatmap(matrix, xticklabels=months, yticklabels=months,
                cmap="viridis", annot=True, fmt=".2f")
    plt.xlabel("Validation month (j)")
    plt.ylabel("Checkpoint month (i)")
    plt.title(f"{metric} Heatmap (12x12)")
    plt.tight_layout()
    plt.savefig(save_path, dpi=300)
    print(f"[Saved] {save_path}")

# ---------- 参数 ----------
def parse_args():
    parser = argparse.ArgumentParser(description="Build ppl/loss heatmap from jsonl results.")
    parser.add_argument("input", type=str,
                        help="输入的 ppl_metrics.jsonl 文件 或 包含多个子文件夹的目录")
    parser.add_argument("--metric", type=str, default="val_ppl",
                        choices=["val_ppl", "val_loss"],
                        help="选择绘制的指标 (默认: val_ppl)")
    parser.add_argument("--out", type=str, default=None,
                        help="输出图片路径 (默认: 输入路径下 heatmap.png)")
    return parser.parse_args()

# ---------- 主程序 ----------
if __name__ == "__main__":
    args = parse_args()
    results = load_results(Path(args.input))
    mat = build_matrix(results, metric=args.metric)

    if args.out is None:
        out_path = Path(args.input)
        if out_path.is_dir():
            out_path = out_path / f"{args.metric}_heatmap.png"
        else:
            out_path = out_path.with_suffix(f"_{args.metric}_heatmap.png")
    else:
        out_path = Path(args.out)

    plot_heatmap(mat, metric=args.metric, save_path=str(out_path))