visualization/arxiv_mc.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import argparse, json, re
from pathlib import Path
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

MONTHS = ['2407','2408','2409','2410','2411','2412','2501','2502','2503','2504','2505','2506']

TASK_PREFIX = "arxiv_mc_"

def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("root", type=str, help="Directory containing 12 month subfolders (e.g., 2407 .. 2506)")
    p.add_argument("--out-dir", type=str, default=None, help="Output directory (default: ROOT)")
    p.add_argument("--task-prefix", type=str, default=TASK_PREFIX, help="Prefix before month in task names")
    p.add_argument("--matrix-csv", type=str, default="arxiv_mc_matrix.csv")
    p.add_argument("--rowmeans-csv", type=str, default="arxiv_mc_row_means.csv")
    p.add_argument("--heatmap-png", type=str, default="arxiv_mc_heatmap.png")
    return p.parse_args()

def load_month_values_json(root: Path, month: str):
    # Accept either <root>/<month>/values.json or <root>/<month>_full/values.json etc.
    candidates = []
    # Common patterns
    candidates.append(root / month / "values.json")
    candidates.append(root / f"{month}_full" / "values.json")
    candidates.append(root / f"{month}_lr4e-5" / "values.json")
    # Fallback: any folder beginning with month
    for path in sorted(root.glob(f"{month}*/values.json")):
        if path not in candidates:
            candidates.append(path)
    for c in candidates:
        if c.exists():
            return c
    return None

def build_matrix(root: Path, task_prefix: str, metric: str = "acc_norm"):
    # rows: checkpoint month; cols: eval month
    mat = np.full((len(MONTHS), len(MONTHS)), np.nan, dtype=float)

    for i, ckpt_m in enumerate(MONTHS):
        vpath = load_month_values_json(root, ckpt_m)
        if vpath is None:
            print(f"[WARN] Missing values.json for checkpoint month {ckpt_m}")
            continue
        try:
            data = json.loads(vpath.read_text(encoding="utf-8"))
        except Exception as e:
            print(f"[WARN] Failed to parse {vpath}: {e}")
            continue
        items = data.get("tasks", [])
        # Build dict: (task, metric) -> value
        d = {}
        for it in items:
            t = it.get("task")
            m = it.get("metric")
            val = it.get("value")
            if t is None or m is None:
                continue
            d[(t, m)] = val
        # Fill row i
        for j, eval_m in enumerate(MONTHS):
            task_name = f"{task_prefix}{eval_m}"
            val = d.get((task_name, metric))
            if val is None:
                continue
            mat[i, j] = float(val)
    return mat



def plot_heatmap(mat: np.ndarray, out_png: Path, title: str = "Accuracy (rows: ckpt, cols: eval month)"):
    fig, ax = plt.subplots(figsize=(10, 8))
    # Heatmap
    im = ax.imshow(mat, aspect="auto")
    # Ticks / labels
    ax.set_xticks(range(len(MONTHS)))
    ax.set_yticks(range(len(MONTHS)))
    ax.set_xticklabels(MONTHS, rotation=45, ha="right")
    ax.set_yticklabels(MONTHS)
    ax.set_xlabel("Eval month")
    ax.set_ylabel("Checkpoint month")
    ax.set_title(title)
    # Annotate with values if not NaN
    for i in range(mat.shape[0]):
        for j in range(mat.shape[1]):
            val = mat[i, j]
            if not np.isnan(val):
                ax.text(j, i, f"{val:.2f}", ha="center", va="center", fontsize=8)
    fig.colorbar(im, ax=ax, fraction=0.046, pad=0.04, label="acc")
    fig.tight_layout()
    fig.savefig(out_png, dpi=200)
    plt.close(fig)



def main():
    args = parse_args()
    root = Path(args.root).expanduser().resolve()
    out_dir = Path(args.out_dir).expanduser().resolve() if args.out_dir else root
    out_dir.mkdir(parents=True, exist_ok=True)

    # 分别构建 acc_norm 与 acc 的矩阵
    mat_acc_norm = build_matrix(root, args.task_prefix, metric="acc_norm")
    mat_acc = build_matrix(root, args.task_prefix, metric="acc")

    # 各画一张图（文件名沿用你的参数名，或你也可以固定命名）
    plot_heatmap(mat_acc_norm, out_dir / args.heatmap_png, title="Accuracy (acc_norm)")
    # 给 acc 单独起个文件名（在不改 argparse 的前提下，直接在文件名上加后缀）
    acc_png = out_dir / (Path(args.heatmap_png).with_suffix("").as_posix() + "_acc.png")
    # 注意上面用到 with_suffix("") + 手动拼接，避免双后缀；也可以更简洁：
    # acc_png = out_dir / ("acc_heatmap.png")

    plot_heatmap(mat_acc, acc_png, title="Accuracy (acc)")

    print("Saved:")
    print("  ", out_dir / args.heatmap_png)
    print("  ", acc_png)

if __name__ == "__main__":
    main()