test_time_scaling/plot_reward_perplexity.py

import json
import numpy as np
import matplotlib.pyplot as plt
from sklearn.linear_model import LinearRegression
import os
import glob

# 自动查找所有子目录下的reward_perplexity_stats.json
stat_files = glob.glob(os.path.join('.', '**', 'reward_perplexity_stats.json'), recursive=True)

for stat_file in stat_files:
    with open(stat_file) as f:
        stats = json.load(f)
    rewards = np.array(stats["step_rewards"])
    rewards = rewards - np.mean(rewards)
    rewards = rewards / np.std(rewards)
    perplexities = np.array(stats["step_perplexities"])
    inv_ppl = 1 / perplexities
    # inv_ppl = perplexities
    # 计算皮尔逊相关系数R
    if len(inv_ppl) > 1 and np.std(inv_ppl) > 0 and np.std(rewards) > 0:
        R = np.corrcoef(inv_ppl, rewards)[0, 1]
    else:
        R = float('nan')
    # 画散点图
    plt.figure(figsize=(8,6))
    plt.scatter(inv_ppl, rewards, alpha=0.2)
    plt.xlabel("1 / Perplexity")
    plt.ylabel("Reward (0/1)")
    plt.title(f"Reward vs. Inverse Perplexity\n{stat_file}\nR={R:.3f}")
    # 拟合线性关系
    X = inv_ppl.reshape(-1, 1)
    y = rewards
    reg = LinearRegression().fit(X, y)
    y_pred = reg.predict(X)
    plt.plot(inv_ppl, y_pred, color='red', label='Linear fit')
    plt.legend()
    plt.tight_layout()
    plt.savefig(stat_file.replace('.json', '_reward_vs_invperplexity.png'))
    plt.close()
    print(f"{stat_file}: Linear fit coef: {reg.coef_}, intercept: {reg.intercept_}, R: {R:.3f}")