File size: 2,665 Bytes
a10ba7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# PHASE 5: FINAL GOLD STANDARD IMPLEMENTATION (16x16)

## 1. Executive Summary
This phase implements the **definitive production configuration** of LIPE V2. We maintain the **16x16 input resolution** as our established baseline and focus on maximizing accuracy through architectural refinement of the fusion layer and an exhaustive training strategy.

## 2. Definitive Architecture (Gold Standard)

| Component | Final Specification | Rationale |
| :--- | :--- | :--- |
| **Input Resolution** | **16x16 pixels** | **Consistent with baseline.** High enough for detail, low enough for CPU. |
| **Feature Fusion** | **Concatenation + MLP** | **UPGRADE.** Replaces simple addition with a non-linear compression layer to better model eye-face spatial correlations. |
| **Feature Aggregation** | **Adaptive Avg Pooling** | Ensures structural stability and allows the model to remain resolution-agnostic. |
| **Loss Function** | **Adaptive Wing Loss** | Mathematically optimized for deep convergence in high-precision regression. |
| **Distillation** | **Hybrid KD (L2CS-Net)** | Transfers "expert" knowledge from ResNet-50 Teacher to Student. |

## 3. The Final Training Regimen ("Slow and Steady")
To guarantee every subject (p00-p14) reaches its absolute global minimum error:

*   **Batch Size:** **32** (Maximum gradient sensitivity for deep convergence).
*   **Early Stopping:** **Patience 50** (Extended search for the best possible weights).
*   **LR Scheduler:** **ReduceLROnPlateau** (Factor: 0.5, Patience: 15).
*   **Optimizer:** **Adam** (Initial LR: 1e-4).
*   **Hardware:** **12 CPU Workers** (Home workstation optimization).

## 4. Complexity & Resource Audit
*   **Parameters:** ~580,404 (Fits comfortably within CPU cache and < 2MB size).
*   **Computation:** ~21.25 MFLOPs (Efficiency verified: only 17% of total power budget).
*   **Latency:** ~1.2 ms/frame (~830 FPS on CPU).

## 5. Execution Roadmap

### Step 1: Baseline Archiving (Immediate after p14 current run)
- [ ] Move `report/lopo_results_v16.csv` and `checkpoints/best_student_p*.pt` to `report/baseline_v16/`.
- [ ] This archive represents the **16x16 Addition Baseline (~4.98°)**.

### Step 2: Architecture Upgrade (Fusion Layer)
- [ ] Modify `src/models/student.py`: Replace `+` with `torch.cat([app, geo], dim=1)` followed by a `nn.Linear(512, 256)` layer.

### Step 3: Final LOPO Execution
- [ ] Execute training with the "Slow and Steady" parameters.
- [ ] Expected total run time: ~30-36 hours.

### Step 4: Final Thesis Data Compilation
- [ ] Compare **16x16 Addition (Baseline)** vs. **16x16 Concatenation (Gold Standard)**.
- [ ] Document final Mean MAE and performance metrics.