🚀 YOLOv8 Scaling Study — LayerScale

Is bigger always better?
A systematic study of width vs. depth scaling in YOLOv8 on the KITTI dataset.

---

📌 Overview

This repository presents a controlled scaling experiment on YOLOv8, analyzing how architectural changes impact:

• 📊 Accuracy (mAP@50)
• ⚡ Inference Speed (FPS)
• 💾 Model Size & Parameters
• ⏱️ Training Cost

We evaluate 5 model variants across two scaling axes:

• Width Scaling → Channel counts per layer
• Depth Scaling → Number of layer repetitions

---

🧠 Key Findings

• 🏆 Best Model: depth_0.33
• ✅ Higher Accuracy than baseline (+0.2% mAP@50)
• ⚡ +50% Faster Inference
• 💾 −43% Model Size
• ⏱️ −35% Training Time

🔥 Insight: Scaling is non-monotonic — bigger models do NOT guarantee better performance.

---

📊 Results & Visualizations

1. 🏁 Overall Comparison

---

2. 📉 Scaling Curves (Width vs Depth)

• Width reduction → sharp accuracy drop
• Depth reduction → graceful degradation

---

3. ⚖️ Efficiency Frontier

• Identifies optimal trade-offs between speed vs accuracy
• depth_0.33 lies on the Pareto frontier

---

4. 🧩 Model Leaderboard

---

5. 🔍 Per-Class Analysis

• Smaller model improves minor classes
• Larger model slightly better on complex classes

---

6. 🎯 Radar — All-Rounder Model

• Shows overall balance across all metrics

---

7. 📈 Metric Correlations

• FPS vs Params → -1.0 correlation
• Recall impacts mAP more than Precision

---

8. 📊 Class Score Spread

• Smaller model = more consistent performance

---

9. ⏱️ Training Cost Analysis

• Highlights diminishing returns of larger models

---

⚙️ Experimental Setup

• Model: YOLOv8
• Dataset: KITTI (Mini-KITTI subset)
• Framework: PyTorch / Ultralytics
• Hardware: GPU (specify if needed)

Variants Tested:

Model	Width	Depth
Baseline	1.0	1.0
Width 0.5	0.5	1.0
Width 0.33	0.33	1.0
Depth 0.5	1.0	0.5
Depth 0.33	1.0	0.33

---

🧪 Reproducing the Results

# Clone repo
git clone https://github.com/Y-R-A-V-R-5/layerscale.git
cd layerscale

# Install dependencies
pip install -r requirements.txt

# Run training
python train.py --config configs/depth_033.yaml