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.