File size: 10,092 Bytes
c9aee57 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 | #!/usr/bin/env python3
"""
论文用简化图表 - 2张图讲清楚结论
图1: Total Loss 对比 (DeCLIP vs Integrated)
图2: 梯度冲突分析 (1×2: 分布 + 逐层趋势)
"""
import json
import os
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
from collections import defaultdict
# 配置
BASE_DIR = "/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private"
DECLIP_LOG = f"{BASE_DIR}/logs/DeCLIP_EVA-B_DINOv2-B_560/out.log"
INTEGRATED_2LOSS_LOG = f"{BASE_DIR}/logs/Integrated_EVA-B_DINOv2-B_560_2loss/out.log"
GRADIENT_2LOSS_PATH = f"{BASE_DIR}/logs/Integrated_EVA-B_DINOv2-B_560_grad_analysis_2loss/gradient_analysis.jsonl"
OUTPUT_DIR = f"{BASE_DIR}/decoupling_analysis/2loss/results/paper_figures"
SMOOTHING_WINDOW = 20
# 颜色
COLOR_DECLIP = '#2E86AB' # 蓝色
COLOR_INTEGRATED = '#E94F37' # 红色
COLOR_CONFLICT = '#E94F37' # 红色
COLOR_ALIGNED = '#2E86AB' # 蓝色
def setup_plot_style():
"""设置论文级别绘图风格"""
plt.rcParams.update({
'font.family': 'serif',
'font.serif': ['Times New Roman', 'DejaVu Serif'],
'font.size': 12,
'axes.labelsize': 14,
'axes.titlesize': 15,
'axes.titleweight': 'bold',
'xtick.labelsize': 12,
'ytick.labelsize': 12,
'legend.fontsize': 12,
'figure.dpi': 150,
'savefig.dpi': 300,
'savefig.bbox': 'tight',
'savefig.pad_inches': 0.1,
'axes.linewidth': 1.5,
'axes.spines.top': False,
'axes.spines.right': False,
'lines.linewidth': 2.5,
})
def parse_training_log(log_path: str) -> list:
"""解析训练日志"""
import re
records = []
with open(log_path, 'r', encoding='utf-8') as f:
for line in f:
if "Train Epoch:" not in line:
continue
try:
total_match = re.search(r'Loss:\s*([\d.e+-]+)\s*\(', line)
if total_match:
records.append(float(total_match.group(1)))
except:
continue
return records
def load_gradient_data(filepath: str) -> dict:
"""加载梯度分析数据"""
layer_data = defaultdict(list)
with open(filepath, 'r', encoding='utf-8') as f:
for line in f:
record = json.loads(line.strip())
for layer_name, cos_sim in record.get('layer_cos_sims', {}).items():
layer_data[layer_name].append(cos_sim)
return dict(layer_data)
def smooth_data(data: np.ndarray, window: int) -> np.ndarray:
"""滑动平均"""
if window <= 1 or len(data) < window:
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_loss_comparison(declip_loss, integrated_loss, output_path):
"""
图1: Total Loss 对比
简洁版:只画 Total Loss,突出差距
"""
min_len = min(len(declip_loss), len(integrated_loss))
declip_loss = np.array(declip_loss[:min_len])
integrated_loss = np.array(integrated_loss[:min_len])
iters = np.arange(min_len)
fig, ax = plt.subplots(figsize=(8, 5))
# 绘制曲线
ax.plot(iters, declip_loss, color=COLOR_DECLIP, alpha=0.15, linewidth=1.0)
ax.plot(iters, smooth_data(declip_loss, SMOOTHING_WINDOW),
color=COLOR_DECLIP, linewidth=3, label='DeCLIP (Decoupled)')
ax.plot(iters, integrated_loss, color=COLOR_INTEGRATED, alpha=0.15, linewidth=1.0)
ax.plot(iters, smooth_data(integrated_loss, SMOOTHING_WINDOW),
color=COLOR_INTEGRATED, linewidth=3, label='Integrated')
# 标注最终值
final_declip = declip_loss[-1]
final_integrated = integrated_loss[-1]
# 添加水平虚线和数值标注
ax.axhline(y=final_declip, color=COLOR_DECLIP, linestyle='--', alpha=0.5, linewidth=1.5)
ax.axhline(y=final_integrated, color=COLOR_INTEGRATED, linestyle='--', alpha=0.5, linewidth=1.5)
ax.text(min_len * 1.02, final_declip, f'{final_declip:.2f}',
color=COLOR_DECLIP, fontsize=14, fontweight='bold', va='center')
ax.text(min_len * 1.02, final_integrated, f'{final_integrated:.2f}',
color=COLOR_INTEGRATED, fontsize=14, fontweight='bold', va='center')
# 添加差距标注
ratio = final_integrated / final_declip
mid_y = (final_declip + final_integrated) / 2
ax.annotate('', xy=(min_len * 0.95, final_declip), xytext=(min_len * 0.95, final_integrated),
arrowprops=dict(arrowstyle='<->', color='gray', lw=2))
ax.text(min_len * 0.92, mid_y, f'{ratio:.1f}×', fontsize=13,
color='gray', fontweight='bold', ha='right', va='center')
ax.set_xlabel('Iteration')
ax.set_ylabel('Total Loss')
ax.set_title('Decoupled Distillation Converges Better')
ax.legend(loc='upper right', framealpha=0.95)
ax.set_xlim(0, min_len * 1.1)
ax.set_ylim(0, max(integrated_loss) * 1.1)
ax.grid(True, alpha=0.3, linestyle='-', linewidth=0.5)
plt.tight_layout()
plt.savefig(output_path.replace('.png', '.png'))
plt.savefig(output_path.replace('.png', '.pdf'))
plt.close()
print(f"已保存: {output_path}")
def plot_gradient_analysis(layer_data, output_path):
"""
图2: 梯度冲突分析 (1×2)
左: 分布直方图
右: 逐层趋势
"""
# 收集所有数据
all_values = []
for cos_sims in layer_data.values():
all_values.extend(cos_sims)
all_array = np.array(all_values)
# 计算统计
conflict_ratio = np.mean(all_array < 0) * 100
mean_cos = np.mean(all_array)
# 逐层统计
layer_names = sorted(layer_data.keys(), key=lambda x: int(x.split('_')[1]))
layer_means = [np.mean(layer_data[name]) for name in layer_names]
layer_stds = [np.std(layer_data[name]) for name in layer_names]
layer_indices = [int(name.split('_')[1]) for name in layer_names]
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
# ===== 左图: 分布直方图 =====
ax1 = axes[0]
bins = np.linspace(-0.5, 0.5, 31)
n, bins_edges, patches = ax1.hist(all_array, bins=bins, edgecolor='white',
linewidth=0.8, alpha=0.9)
# 根据值着色
for i, (patch, left_edge) in enumerate(zip(patches, bins_edges[:-1])):
right_edge = bins_edges[i + 1]
center = (left_edge + right_edge) / 2
if center < 0:
patch.set_facecolor(COLOR_CONFLICT)
else:
patch.set_facecolor(COLOR_ALIGNED)
# 添加零线
ax1.axvline(x=0, color='black', linestyle='-', linewidth=2)
# 添加统计标注框
textstr = f'Conflict: {conflict_ratio:.0f}%\nMean: {mean_cos:.3f}'
props = dict(boxstyle='round,pad=0.5', facecolor='white', alpha=0.95, edgecolor='gray')
ax1.text(0.05, 0.95, textstr, transform=ax1.transAxes, fontsize=13,
verticalalignment='top', bbox=props, fontweight='bold')
# 图例
legend_elements = [
mpatches.Patch(facecolor=COLOR_CONFLICT, label='Conflict (cos < 0)'),
mpatches.Patch(facecolor=COLOR_ALIGNED, label='Aligned (cos > 0)')
]
ax1.legend(handles=legend_elements, loc='upper right', framealpha=0.95)
ax1.set_xlabel('Gradient Cosine Similarity')
ax1.set_ylabel('Frequency')
ax1.set_title('(a) Distribution of Gradient Similarity')
ax1.set_xlim(-0.55, 0.55)
# ===== 右图: 逐层趋势 =====
ax2 = axes[1]
# 绘制柱状图
colors = [COLOR_CONFLICT if m < 0 else COLOR_ALIGNED for m in layer_means]
bars = ax2.bar(layer_indices, layer_means, color=colors, alpha=0.85,
edgecolor='white', linewidth=1)
ax2.errorbar(layer_indices, layer_means, yerr=layer_stds, fmt='none',
color='black', capsize=4, capthick=1.5, elinewidth=1.5)
# 零线
ax2.axhline(y=0, color='black', linestyle='-', linewidth=1.5)
# 添加趋势箭头和标注
ax2.annotate('', xy=(11, layer_means[-1] - 0.05), xytext=(0, layer_means[0] + 0.05),
arrowprops=dict(arrowstyle='->', color='gray', lw=2,
connectionstyle='arc3,rad=-0.2'))
ax2.text(5.5, -0.38, 'Deeper = More Conflict', fontsize=11,
color='gray', ha='center', style='italic')
ax2.set_xlabel('Layer Index')
ax2.set_ylabel('Mean Cosine Similarity')
ax2.set_title('(b) Conflict Increases in Deeper Layers')
ax2.set_xticks(layer_indices)
ax2.set_ylim(-0.45, 0.15)
ax2.grid(True, alpha=0.3, axis='y', linestyle='-', linewidth=0.5)
plt.tight_layout()
plt.savefig(output_path)
plt.savefig(output_path.replace('.png', '.pdf'))
plt.close()
print(f"已保存: {output_path}")
def main():
setup_plot_style()
os.makedirs(OUTPUT_DIR, exist_ok=True)
print("=" * 60)
print("生成论文用图表")
print("=" * 60)
# 加载数据
print("\n加载数据...")
declip_loss = parse_training_log(DECLIP_LOG)
integrated_loss = parse_training_log(INTEGRATED_2LOSS_LOG)
layer_data = load_gradient_data(GRADIENT_2LOSS_PATH)
print(f" DeCLIP: {len(declip_loss)} 条记录")
print(f" Integrated: {len(integrated_loss)} 条记录")
print(f" 梯度数据: {len(layer_data)} 层")
# 生成图表
print("\n生成图表...")
# 图1: Loss 对比
plot_loss_comparison(
declip_loss, integrated_loss,
os.path.join(OUTPUT_DIR, 'fig1_loss_comparison.png')
)
# 图2: 梯度冲突分析
plot_gradient_analysis(
layer_data,
os.path.join(OUTPUT_DIR, 'fig2_gradient_analysis.png')
)
print("\n" + "=" * 60)
print("完成!图表保存在:")
print(f" {OUTPUT_DIR}/")
print("=" * 60)
if __name__ == "__main__":
main()
|