KAGGLE_SETUP.md

YOLOv8-MPEB Kaggle Training Guide

Quick Start for Kaggle

1. Upload Files to Kaggle Dataset

Create a new Kaggle dataset and upload these files:

• yolov8_mpeb.yaml - Model architecture
• yolov8_mpeb_modules.py - Custom modules
• dataset_example.yaml - Dataset configuration
• train_kaggle.py - Kaggle training script

2. Create a New Kaggle Notebook

1. Go to Kaggle Notebooks
2. Create a new notebook
3. Add your dataset as input (e.g., yolo-mpeb-training-code)
4. Enable GPU (Settings → Accelerator → GPU P100)

3. Run Training in Kaggle Notebook

# Cell 1: Copy training script to working directory
import shutil
from pathlib import Path

CODE_DIR = Path('/kaggle/input/yolo-mpeb-training-code/code')
shutil.copy(CODE_DIR / 'train_kaggle.py', '/kaggle/working/train_kaggle.py')
print("✓ Training script copied to working directory")

# Cell 2: Install Ultralytics (if needed)
!pip install ultralytics -q

# Cell 3: Run training
!python /kaggle/working/train_kaggle.py

Important Notes

Kaggle File System Structure

• /kaggle/input/ - READ-ONLY directory containing your input datasets
• /kaggle/working/ - WRITABLE directory for outputs, models, and temporary files
• /kaggle/temp/ - WRITABLE temporary directory

Path Configuration

The dataset_example.yaml has been configured to use /kaggle/working/VisDrone as the dataset root. This ensures:

• Dataset downloads go to a writable location
• Training outputs are saved correctly
• No "Read-only file system" errors

Dataset Download

The VisDrone dataset will be automatically downloaded to /kaggle/working/VisDrone on first run. This is approximately 2.3 GB and may take a few minutes.

Training Duration

• Estimated time: 6-8 hours on Tesla P100
• Epochs: 200
• Batch size: 32
• Image size: 640x640

Output Files

After training completes, you'll find:

• Best weights: /kaggle/working/runs/train/yolov8_mpeb/weights/best.pt
• Last weights: /kaggle/working/runs/train/yolov8_mpeb/weights/last.pt
• Training plots: /kaggle/working/runs/train/yolov8_mpeb/
• Validation results: In the training output

Saving Your Results

Since Kaggle notebooks reset after session ends, make sure to:

1. Save output - Click "Save Version" to preserve your notebook with outputs
2. Download weights - Download the .pt files before closing
3. Commit notebook - Commit your notebook to save training logs

Troubleshooting

Error: "Read-only file system"

Solution: Make sure dataset_example.yaml uses /kaggle/working/VisDrone as the path, not a relative path.

Error: "Module not found"

Solution: Ensure all files are in your Kaggle dataset and the path in train_kaggle.py matches your dataset name.

Error: "CUDA out of memory"

Solution: Reduce batch size in train_kaggle.py:

'batch': 16,  # Reduced from 32

Dataset not downloading

Solution: Check your internet connection in Kaggle. The dataset downloads from Ultralytics servers.

Model Specifications

Based on the paper: "YOLOv8-MPEB small target detection algorithm based on UAV images"

• Model: YOLOv8s-MPEB
• Parameters: 7.39M
• Model Size: 14.5 MB
• GFLOPs: 27.4
• Target mAP50: 91.9%

Custom Architecture Components

1. MobileNetV3 Backbone - Lightweight feature extraction
2. EMA Attention - Efficient Multi-scale Attention in C2f modules
3. BiFPN Fusion - Bidirectional Feature Pyramid Network
4. P2 Detection Head - Enhanced small object detection

After Training

Validate Your Model

from ultralytics import YOLO

model = YOLO('/kaggle/working/runs/train/yolov8_mpeb/weights/best.pt')
results = model.val(data='/kaggle/working/code/dataset_example.yaml')

print(f"mAP50: {results.box.map50:.4f}")
print(f"mAP50-95: {results.box.map:.4f}")

Run Inference

from ultralytics import YOLO

model = YOLO('/kaggle/working/runs/train/yolov8_mpeb/weights/best.pt')
results = model.predict('path/to/image.jpg', save=True, conf=0.25)

Support

For issues or questions:

1. Check the error message carefully
2. Verify all paths are correct
3. Ensure GPU is enabled in Kaggle settings
4. Check that all required files are in your dataset

License

This implementation is based on the YOLOv8-MPEB paper and uses the Ultralytics framework (AGPL-3.0 License).