xiaomoguhzz's picture
Add files using upload-large-folder tool
c9aee57 verified
# 解耦蒸馏 vs 集成蒸馏 消融实验设计
> 最后更新: 2026-01-11
---
## 📋 审稿人要求
> The advantages of the decoupled learning framework—especially compared to integrated or joint distillation strategies—require clearer justification through more focused ablation studies.
>
> DeCLIP claims decoupling avoids optimization conflicts, but lacks:
> (1) direct comparison with integrated distillation (e.g., joint content+context distillation without decoupling) in terms of mIoU/mAP and training stability
---
## 🎯 核心对比定义
### 1. 解耦蒸馏 (Decoupled Distillation) - DeCLIP
```
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ DeCLIP 解耦蒸馏架构 │
│ │
│ Input Image │
│ ↓ │
│ ┌─────────────────────────────────────────────────────────────────┐ │
│ │ EVA-CLIP / OpenAI-CLIP ViT │ │
│ │ ┌─────────────────────────────────────────────────────────┐ │ │
│ │ │ Block 0 → Block 1 → ... → Block (N-2) │ │ │
│ │ │ (完整的 Attention + Residual + MLP) │ │ │
│ │ └─────────────────────────────────────────────────────────┘ │ │
│ │ ↓ │ │
│ │ ┌─────────────────────────────────────────────────────────┐ │ │
│ │ │ Block (N-1) - 最后一个 Block │ │ │
│ │ │ ┌─────────────────────────────────────────────────┐ │ │ │
│ │ │ │ LayerNorm → Q, K, V Projection │ │ │ │
│ │ │ │ │ │ │ │
│ │ │ │ ┌────────┐ ┌────────┐ │ │ │ │
│ │ │ │ │ Q │ │ V │ │ │ │ │
│ │ │ │ └────┬───┘ └────┬───┘ │ │ │ │
│ │ │ │ │ │ │ │ │ │
│ │ │ │ Q @ Q^T V @ V^T │ │ │ │
│ │ │ │ │ │ │ │ │ │
│ │ │ │ softmax softmax │ │ │ │
│ │ │ │ ↓ ↓ │ │ │ │
│ │ │ │ ┌──────────────┐ ┌──────────────┐ │ │ │ │
│ │ │ │ │ Q Self-Sim │ │ V Self-Sim │ │ │ │ │
│ │ │ │ │ (Content) │ │ (Context) │ │ │ │ │
│ │ │ │ └──────┬───────┘ └──────┬───────┘ │ │ │ │
│ │ │ │ │ │ │ │ │ │
│ │ │ │ ❌ 无残差连接 ❌ 无 MLP │ │ │ │
│ │ │ └─────────────────────────────────────────────────┘ │ │ │
│ │ └─────────────────────────────────────────────────────────┘ │ │
│ └─────────────────────────────────────────────────────────────────┘ │
│ ↓ ↓ │
│ ┌──────────────────┐ ┌──────────────────┐ │
│ │ L_content Loss │ │ L_context Loss │ │
│ │ (Q features) │ │ (V self-sim) │ │
│ └──────────────────┘ └──────────────────┘ │
│ ↓ ↓ │
│ Teacher: CLIP crops Teacher: DINOv2 corr │
└─────────────────────────────────────────────────────────────────────────────────────┘
```
**关键特点**
- 在最后一个 Block 使用 `forward_without_rcffn` 方法
- **去掉残差连接**:避免特征混合
- **去掉 MLP**:保持 Q 和 V 的纯净表示
- Q 的自相似性 → Content Loss(与 CLIP Teacher 对齐)
- V 的自相似性 → Context Loss(与 DINOv2 相关性对齐)
### 2. 集成蒸馏 (Integrated Distillation) - Baseline
```
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ Integrated 集成蒸馏架构 │
│ │
│ Input Image │
│ ↓ │
│ ┌─────────────────────────────────────────────────────────────────┐ │
│ │ EVA-CLIP / OpenAI-CLIP ViT │ │
│ │ ┌─────────────────────────────────────────────────────────┐ │ │
│ │ │ Block 0 → Block 1 → ... → Block (N-1) │ │ │
│ │ │ 完整的 Attention + Residual + MLP(所有层) │ │ │
│ │ └─────────────────────────────────────────────────────────┘ │ │
│ │ ↓ │ │
│ │ ┌─────────────────────────────────────────────────────────┐ │ │
│ │ │ Final Processing │ │ │
│ │ │ Head (Linear) → LayerNorm │ │ │
│ │ └─────────────────────────────────────────────────────────┘ │ │
│ └─────────────────────────────────────────────────────────────────┘ │
│ ↓ │
│ ┌──────────────────┐ │
│ │ Final Features │ │
│ │ (融合的特征) │ │
│ └────────┬─────────┘ │
│ │ │
│ ┌─────────────────┼─────────────────┐ │
│ ↓ ↓ ↓ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ L_content │ │ L_context │ │ L_region │ │
│ │ (ROI池化) │ │ (self-sim) │ │ (ROI sim) │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
│ ↓ ↓ ↓ │
│ Teacher: CLIP Teacher: DINOv2 Teacher: DINOv2 │
│ │
│ L_total = L_content + L_context + L_region │
│ (所有 Loss 作用于同一个融合特征!) │
└─────────────────────────────────────────────────────────────────────────────────────┘
```
**关键特点**
- 使用完整的 ViT forward(包含所有残差连接和 MLP)
- 最终输出是 **融合后的特征**
- Content 和 Context Loss 都作用于 **同一个融合特征**
- 可能导致 **梯度冲突**:两个任务的梯度方向可能相反
---
## 📊 实验方案
### 实验 1:训练动态曲线(Training Dynamics)
**目的**:证明解耦蒸馏训练更稳定
**记录指标**(每个 iteration):
- `L_content` vs Iteration
- `L_context` vs Iteration
- `L_total` vs Iteration
**预期结果**
- DeCLIP:曲线平稳下降,收敛快
- Integrated:曲线震荡,收敛慢,可能出现 Loss 反弹
### 实验 2:梯度冲突分析(Gradient Conflict Analysis)
**目的**:数学证明两个任务存在优化冲突
**方法**(参考 PCGrad 论文):
```python
# 在 Integrated baseline 训练时:
# 1. 分别计算 L_content 和 L_context 对最后 N 层的梯度
grad_content = torch.autograd.grad(L_content, params, retain_graph=True)
grad_context = torch.autograd.grad(L_context, params, retain_graph=True)
# 2. 计算梯度余弦相似度
cos_sim = cosine_similarity(grad_content, grad_context)
# 3. 逐层统计
# cos_sim < 0: 梯度冲突
# cos_sim ≈ 0: 梯度正交
# cos_sim > 0: 梯度一致
```
**分析范围**:逐层分析每个 ViT Block
**预期结果**
- 梯度余弦相似度主要分布在 **负值区间**(冲突)或 **接近 0**(正交)
- 说明两个任务的梯度方向经常相反或不相关
### 实验 3:PCA 可视化
**目的**:直观展示解耦效果
**对比内容**
- DeCLIP 解耦蒸馏的最终输出特征
- Integrated 集成蒸馏的最终输出特征
**方法**
- 选取挑战性图片(多物体、复杂背景)
- 提取特征的前 3 主成分,映射为 RGB
**预期结果**
- Integrated:特征"模糊"或"混乱"
- DeCLIP:边界清晰,语义准确
### 实验 4:CKA 分析
**目的**:分析特征相似性
**对比内容**
1. DeCLIP 的 Q 特征 vs V 特征
2. Integrated 输出特征 vs DeCLIP Q 特征
3. Integrated 输出特征 vs DeCLIP V 特征
**预期结果**
- DeCLIP 的 Q 和 V 有一定差异(证明确实解耦了)
- Integrated 输出与 DeCLIP Q/V 的关系模糊
### 实验 5:定量表格对比(暂不实现,用户自行评估)
**指标**
- Detection: mAP
- Segmentation: mIoU
---
## 🔧 训练配置
### 统一配置
| 配置项 | 值 |
|-------|---|
| VFM Teacher | DINOv2 |
| Epochs | 6 |
| Learning Rate | 1e-5 |
| Batch Size | 16 (total) |
| Optimizer | AdamW |
| Scheduler | Cosine |
| Dataset | COCO |
### 模型配置
| Student Model | VFM Size |
|--------------|----------|
| EVA02-CLIP-B-16 | DINOv2-B |
| EVA02-CLIP-L-14-336 | DINOv2-L |
| OpenAI-CLIP-B-16 | DINOv2-B |
| OpenAI-CLIP-L-14-336 | DINOv2-L |
### Loss 配置
**DeCLIP(解耦)**
- Content Loss 权重: 1.0
- Context Loss 权重: 0.25
- Region Loss 权重: 0.05
**Integrated(集成)**
- Content Loss 权重: 1.0
- Context Loss 权重: 1.0(直接相加,不用系数)
---
## 📁 代码组织
```
decoupling_analysis/
├── DESIGN.md # 本设计文档
├── PROGRESS.md # 进度记录
├── gradient_analysis.py # 梯度冲突分析
├── pca_visualization.py # PCA 可视化
└── cka_analysis.py # CKA 分析
scripts/decoupling_ablation/
├── dist_integrated_eva_vitb16_coco.sh # EVA-B/16 多卡训练
├── dist_integrated_eva_vitL14_336_coco.sh # EVA-L/14-336 多卡训练
├── dist_integrated_openai_vitb16_coco.sh # OpenAI-B/16 多卡训练
├── dist_integrated_openai_vitL14_coco.sh # OpenAI-L/14 多卡训练
├── debug_integrated_eva_vitb16_coco.sh # EVA-B/16 单卡调试
├── debug_integrated_eva_vitL14_336_coco.sh # EVA-L/14-336 单卡调试
├── debug_integrated_openai_vitb16_coco.sh # OpenAI-B/16 单卡调试
└── debug_integrated_openai_vitL14_coco.sh # OpenAI-L/14 单卡调试
src/training/
├── integrated_distill.py # 集成蒸馏实现 (IntegratedDistillation 类)
├── main.py # (已修改) 添加 integrated 版本支持
└── params.py # (已修改) 添加 --version integrated 选项
```
---
## 📝 实现步骤
### Step 1: 实现 Integrated Distillation ✅
- [x] 创建 `src/training/integrated_distill.py`
- [x] 使用完整的 ViT forward(`mode="vanilla"`,不调用 `forward_without_rcffn`
- [x] 在最终输出特征上同时计算 Content + Context Loss
### Step 2: 实现 Gradient Analysis ✅
- [x] 创建 `gradient_analysis.py`
- [x] 在训练过程中记录梯度余弦相似度
- [x] 逐层分析
### Step 3: 实现 PCA Visualization ✅
- [x] 创建 `pca_visualization.py`
- [x] 对比 DeCLIP vs Integrated 的特征
### Step 4: 实现 CKA Analysis ✅
- [x] 创建 `cka_analysis.py`
- [x] 分析 Q, V, 输出特征的相似性
### Step 5: 创建训练脚本 ✅
- [x] EVA-CLIP B/L
- [x] OpenAI CLIP B/L
### Step 6: 代码审阅与修复 ✅ (2026-01-11)
- [x] 修复 `extract_roi_features_integrated` 缺少 `.transpose(-2,-1)` 的维度问题
- [x] 修复 `region_scd_loss` 中不必要的 transpose
- [x] 移除未使用的导入 (`autocast`, `math`)
---
## ⚠️ 注意事项
1. **PCGrad 仅用于分析**:不使用它来解决冲突,只用来证明冲突存在
2. **Loss 直接相加**:Integrated 的 Content 和 Context Loss 直接相加,不调权重
3. **训练配置一致**:除了解耦 vs 集成的差异,其他配置完全一致
4. **评估用户自行**:训练完毕后用户自己进行 mAP/mIoU 评估