#!/usr/bin/env python3 """ 绘制 2loss 实验的 Loss 曲线对比图。 对比: DeCLIP vs Integrated_2loss """ import json import os import numpy as np import matplotlib.pyplot as plt from pathlib import Path # 配置 DATA_DIR = "/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/decoupling_analysis/2loss/results/data" OUTPUT_DIR = "/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/decoupling_analysis/2loss/results" # 绘图配置 SMOOTHING_WINDOW = 15 COLORS = { 'DeCLIP': '#2E86AB', # 蓝色 'Integrated_2loss': '#E94F37', # 红色 } def setup_plot_style(): """设置顶会论文级别的绘图风格。""" plt.style.use('seaborn-v0_8-whitegrid') plt.rcParams.update({ 'font.family': 'serif', 'font.serif': ['Times New Roman', 'DejaVu Serif'], 'font.size': 11, 'axes.labelsize': 13, 'axes.titlesize': 14, 'axes.titleweight': 'bold', 'xtick.labelsize': 11, 'ytick.labelsize': 11, 'legend.fontsize': 11, 'figure.dpi': 150, 'savefig.dpi': 300, 'savefig.bbox': 'tight', 'savefig.pad_inches': 0.1, 'axes.linewidth': 1.2, 'grid.linewidth': 0.8, 'grid.alpha': 0.3, 'lines.linewidth': 2.0, }) def load_jsonl(filepath: str) -> list: """加载JSONL文件。""" data = [] with open(filepath, 'r', encoding='utf-8') as f: for line in f: data.append(json.loads(line.strip())) return data def smooth_data(data: np.ndarray, window: int) -> np.ndarray: """使用滑动平均平滑数据。""" if window <= 1: return data kernel = np.ones(window) / window smoothed = np.convolve(data, kernel, mode='same') for i in range(window // 2): smoothed[i] = np.mean(data[:i + window // 2 + 1]) smoothed[-(i + 1)] = np.mean(data[-(i + window // 2 + 1):]) return smoothed def plot_paper_figure( iters: np.ndarray, declip_data: dict, integrated_data: dict, save_path_png: str, save_path_pdf: str, smoothing_window: int = SMOOTHING_WINDOW ): """绘制1x3论文用图。""" fig, axes = plt.subplots(1, 3, figsize=(15, 4.5)) plot_configs = [ (axes[0], 'loss_total', '(a) Total Loss'), (axes[1], 'loss_content', '(b) Content Loss'), (axes[2], 'loss_context', '(c) Context Loss'), ] for ax, key, title in plot_configs: for model_name, data, color in [ ('DeCLIP', declip_data, COLORS['DeCLIP']), ('Integrated (2loss)', integrated_data, COLORS['Integrated_2loss']) ]: y_data = np.array(data[key]) ax.plot(iters, y_data, color=color, alpha=0.2, linewidth=1.0) y_smooth = smooth_data(y_data, smoothing_window) ax.plot(iters, y_smooth, color=color, linewidth=2.5, label=model_name) ax.set_xlabel('Iteration') ax.set_ylabel('Loss') ax.set_title(title, fontweight='bold') ax.legend(loc='upper right', framealpha=0.9) ax.grid(True, alpha=0.3) for spine in ax.spines.values(): spine.set_linewidth(1.2) plt.tight_layout() plt.savefig(save_path_png) plt.savefig(save_path_pdf) plt.close() print(f"已保存: {save_path_png}") print(f"已保存: {save_path_pdf}") def compute_statistics(declip_data: dict, integrated_data: dict) -> dict: """计算统计数据。""" stats = {} for model_name, data in [('DeCLIP', declip_data), ('Integrated_2loss', integrated_data)]: stats[model_name] = {} for key in ['loss_total', 'loss_content', 'loss_context']: values = np.array(data[key]) stats[model_name][key] = { 'mean': float(np.mean(values)), 'std': float(np.std(values)), 'min': float(np.min(values)), 'max': float(np.max(values)), 'final': float(values[-1]), 'first': float(values[0]), } return stats def main(): setup_plot_style() os.makedirs(OUTPUT_DIR, exist_ok=True) print("=" * 60) print("Loss曲线绘图工具 (2loss 版本)") print("=" * 60) # 加载数据 declip_path = os.path.join(DATA_DIR, "declip_training.jsonl") integrated_path = os.path.join(DATA_DIR, "integrated_2loss_training.jsonl") if not os.path.exists(declip_path) or not os.path.exists(integrated_path): print("错误: 数据文件不存在，请先运行 extract_training_data.py") return print("加载数据...") declip_records = load_jsonl(declip_path) integrated_records = load_jsonl(integrated_path) print(f" DeCLIP: {len(declip_records)} 条记录") print(f" Integrated_2loss: {len(integrated_records)} 条记录") # 截断到相同长度 min_len = min(len(declip_records), len(integrated_records)) declip_records = declip_records[:min_len] integrated_records = integrated_records[:min_len] print(f" 截断后: {min_len} 条记录") # 转换为numpy数组格式 iters = np.arange(min_len) declip_data = { 'loss_total': [r['loss_total'] for r in declip_records], 'loss_content': [r['loss_content'] for r in declip_records], 'loss_context': [r['loss_context'] for r in declip_records], } integrated_data = { 'loss_total': [r['loss_total'] for r in integrated_records], 'loss_content': [r['loss_content'] for r in integrated_records], 'loss_context': [r['loss_context'] for r in integrated_records], } print() print("=" * 60) print("生成图表") print("=" * 60) # 论文用图 (1x3) plot_paper_figure( iters, declip_data, integrated_data, os.path.join(OUTPUT_DIR, 'loss_comparison_2loss_paper.png'), os.path.join(OUTPUT_DIR, 'loss_comparison_2loss_paper.pdf') ) # 保存统计数据 stats = compute_statistics(declip_data, integrated_data) stats_path = os.path.join(OUTPUT_DIR, 'statistics.json') with open(stats_path, 'w', encoding='utf-8') as f: json.dump(stats, f, indent=2, ensure_ascii=False) print(f"已保存: {stats_path}") print() print("=" * 60) print("统计摘要") print("=" * 60) for model in ['DeCLIP', 'Integrated_2loss']: print(f"\n【{model}】") for loss_type in ['loss_total', 'loss_content', 'loss_context']: s = stats[model][loss_type] print(f" {loss_type}: first={s['first']:.4f}, final={s['final']:.4f}, " f"mean={s['mean']:.4f}, min={s['min']:.4f}") print() print("=" * 60) print("绘图完成!") print("=" * 60) if __name__ == "__main__": main()