| import re
|
| import matplotlib.pyplot as plt
|
| import numpy as np
|
| from matplotlib.ticker import FuncFormatter
|
| from pathlib import Path
|
| plt.rcParams['font.sans-serif'] = ['SimHei']
|
| plt.rcParams['axes.unicode_minus'] = False
|
|
|
| def parse_loss_log(log_file, sample_interval=50, debug=False):
|
| """解析日志文件 - 修复全局步数计算"""
|
| steps = []
|
| losses = []
|
|
|
|
|
| loss_pattern = r'\[.*?\]:\s*epoch\s+\d+-step\s+(\d+)\s+loss:\s+([\d.]+)'
|
|
|
| try:
|
| with open(log_file, 'r', encoding='utf-8', errors='ignore') as f:
|
| loss_count = 0
|
| last_global_step = 0
|
|
|
| for line in f:
|
| line = line.strip()
|
| if not line:
|
| continue
|
|
|
| loss_match = re.search(loss_pattern, line)
|
| if loss_match:
|
| global_step = int(loss_match.group(1))
|
| loss = float(loss_match.group(2))
|
|
|
| last_global_step = global_step
|
|
|
| if loss_count % sample_interval == 0:
|
| steps.append(global_step)
|
| losses.append(loss)
|
|
|
| loss_count += 1
|
|
|
| except Exception as e:
|
| print(f"❌ 解析 {log_file} 错误: {e}")
|
| return None, None
|
|
|
| if debug and steps:
|
| arr = np.array(steps)
|
| print(f" 📄 匹配 {loss_count} 条 | 最后全局 Step: {last_global_step:,}")
|
| print(f" 📊 采样后: {len(steps)} 点 | 最大 Step: {arr.max():,} ({arr.max() / 1000:.1f}K)")
|
|
|
| return np.array(steps) if steps else None, np.array(losses) if losses else None
|
|
|
|
|
| def smooth_curve(data, window_size=50):
|
| if len(data) < window_size:
|
| return data
|
| kernel = np.ones(window_size) / window_size
|
| return np.convolve(data, kernel, mode='valid')
|
|
|
|
|
| def plot_loss_comparison(log_configs, save_path='loss_comparison.png',
|
| window_size=50, max_step_k=None, debug=False):
|
| if not log_configs:
|
| print("❌ 没有配置任何日志文件")
|
| return
|
|
|
| colors = ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#9467bd', '#8c564b']
|
| fig, ax = plt.subplots(figsize=(10, 6), dpi=150)
|
| all_max_step = 0
|
|
|
| for idx, (model_name, log_file) in enumerate(log_configs.items()):
|
| print(f"📁 解析: {model_name} -> {Path(log_file).name}")
|
| steps, losses = parse_loss_log(log_file, sample_interval=50, debug=debug)
|
|
|
| if steps is None or len(steps) == 0:
|
| print(f"⚠️ 跳过 {model_name}: 无有效数据")
|
| continue
|
|
|
|
|
| if len(losses) >= window_size:
|
| losses_smooth = smooth_curve(losses, window_size)
|
| steps_smooth = steps[window_size - 1:]
|
| else:
|
| losses_smooth, steps_smooth = losses, steps
|
|
|
| steps_k = steps_smooth / 100
|
| if steps_k.max() > all_max_step:
|
| all_max_step = steps_k.max()
|
|
|
|
|
| color = colors[idx % len(colors)]
|
| ax.plot(steps_k, losses_smooth, color=color, linewidth=1.8, label=model_name, alpha=0.95)
|
|
|
|
|
| if len(losses_smooth) > 0:
|
| ax.plot(steps_k[-1], losses_smooth[-1], 'o', color=color, markersize=4)
|
| ax.text(steps_k[-1] + all_max_step * 0.01, losses_smooth[-1],
|
| f'{losses_smooth[-1]:.3f}', fontsize=7, color=color, va='center')
|
|
|
| if ax.lines:
|
|
|
| x_max = max_step_k if max_step_k else all_max_step * 1.05
|
| ax.set_xlim(0, x_max)
|
| print(f"📏 X 轴范围: 0 - {x_max:.1f}K (数据最大 {all_max_step:.1f}K)")
|
|
|
| ax.xaxis.set_major_formatter(FuncFormatter(lambda x, p: f'{int(x)}'))
|
| ax.grid(True, linestyle='--', alpha=0.3, linewidth=0.5, zorder=0)
|
| ax.set_xlabel('Training Step (K)', fontsize=11, fontweight='bold')
|
| ax.set_ylabel('Loss', fontsize=11, fontweight='bold')
|
| ax.legend(loc='upper right', fontsize=9, framealpha=0.95, edgecolor='gray')
|
| ax.set_title('Training Loss Comparison', fontsize=13, fontweight='bold', pad=15)
|
|
|
| plt.tight_layout()
|
| plt.savefig(save_path, dpi=150, bbox_inches='tight')
|
| print(f"✅ 已保存: {save_path}")
|
| plt.close()
|
| else:
|
| print("❌ 没有有效数据可绘制")
|
|
|
|
|
|
|
| if __name__ == "__main__":
|
| LOG_CONFIGS = {
|
| 'ProMoE-B-全层MOE': 'training.log',
|
| }
|
|
|
| SAVE_PATH = 'loss_comparison.png'
|
| SMOOTH_WINDOW = 50
|
| MAX_STEP_K = None
|
| DEBUG = True
|
|
|
| print("🎨 生成 Loss 对比图...")
|
| plot_loss_comparison(LOG_CONFIGS, SAVE_PATH, SMOOTH_WINDOW, MAX_STEP_K, debug=DEBUG) |
