Intel Image Classification

This project is an end-to-end image classification service built using transfer learning with Xception, trained on the Intel Image Classification dataset, and deployed as a containerized inference API on Fly.io. The service accepts an image file and returns the predicted class along with confidence scores.

---

🚀 Project Overview

• Model: Xception (ImageNet pretrained, frozen base)
• Framework: TensorFlow / Keras
• Deployment: Docker + Gunicorn + Fly.io
• Inference API: Flask
• Input: Image file (.jpg, .png)
• Output: Predicted class + confidence

---

🗂 Dataset

Intel Image Classification Dataset

Classes:

• buildings
• forest
• glacier
• mountain
• sea
• street

Images are resized to 299 × 299, matching Xception's expected input.

---

🧠 Model Architecture

The model uses transfer learning with the following architecture:

• Base: Xception (pretrained on ImageNet, frozen)
• Global Average Pooling: Reduces spatial dimensions
• Dense Layer: 100 units with ReLU activation
• Dropout: 0.2 rate for regularization
• Output Layer: 6 units (one per class)

Loss Function:

• CategoricalCrossentropy(from_logits=True)

Optimizer:

• Adam with learning rate 1e-3

---

🏋️ Training Summary

Training configuration from train.py:

• Input size: 299 × 299
• Batch size: 16
• Epochs: 20
• Learning rate: 0.001
• Dropout rate: 0.2
• Inner layer size: 100 units
• Data augmentation: Shear, zoom, horizontal flip
• Validation split: 20%
• Checkpoint: Best model saved based on val_loss

Model saved as: models/model.keras

Training was done on GPU. Inference runs on CPU.

---

📦 Project Structure

intel-image-classification/
├── train.py                     # Training script
├── predict.py                   # Flask inference API
├── Dockerfile                   # Container definition
├── fly.toml                     # Fly.io configuration
├── requirements.txt             # Python dependencies
├── notebook.ipynb              # Exploratory notebook
├── .gitignore                   # Git ignore rules
├── 30.jpg                       # Test image
├── data/
│   └── intel-image-classification/
│       └── seg_train/seg_train/ # Training data
├── models/
│   └── model.keras              # Trained model
└── README.md                    # This file

---

🔌 API Specification

Endpoint

POST /predict

Request

• Content-Type: multipart/form-data
• Field name: file
• Value: Image file (.jpg, .png)

Response (JSON)

{
  "class": "forest",
  "confidence": 0.94
}

---

🧪 Local Testing (without Docker)

1. Install dependencies

pip install -r requirements.txt

2. Run the Flask server

python predict.py

The server runs on port 9696.

3. Test the endpoint

curl -X POST \
  -F "file=@30.jpg" \
  http://localhost:9696/predict

Expected output:

{
  "class": "mountain",
  "confidence": 0.97
}

---

🐳 Docker Usage

Build the image

docker build -t intel-image-classifier .

Run the container

docker run -p 9696:9696 intel-image-classifier

Test the container

curl -X POST \
  -F "file=@30.jpg" \
  http://localhost:9696/predict

---

☁️ Cloud Deployment (Fly.io)

The model is deployed on Fly.io and exposed as a REST API.

🔗 API Endpoint

🚀 Example Request

curl -X POST \
  -F "file=@30.jpg" \
  https://intel-image-classification.fly.dev/predict

⚙️ Requirements

Key dependencies:

tensorflow==2.15.0
flask==3.0.0
gunicorn==21.2.0
pillow==10.1.0
numpy==1.26.2

All dependencies are listed in requirements.txt.

---

🎓 Training Your Own Model

To retrain the model:

1. Download the dataset

Download the Intel Image Classification dataset and place it in:

data/intel-image-classification/seg_train/seg_train/

2. Run training

python train.py

Training features:

• Uses ModelCheckpoint to save best model
• Data augmentation (shear, zoom, flip)
• Reproducible (fixed random seed: 42)
• GPU memory growth enabled
• 20% validation split

Output:

• Best model saved to models/model.keras
• Training history logged to console

---

Screenshot:

📌 Key Learnings

• Transfer learning with frozen Xception base dramatically reduces training time
• Correct image preprocessing (preprocess_input) is critical for accuracy
• Data augmentation helps prevent overfitting on small datasets
• Containerization with Docker simplifies deployment and dependency management
• GPU is only needed for training, not inference
• Gunicorn provides production-ready WSGI server
• Fly.io deployment issues are usually DevOps, not ML

---

🐛 Common Issues

Docker build fails

• Ensure models/model.keras exists
• Check Docker daemon is running

Model loading error

• Verify TensorFlow version matches training version
• Check model file path

Port already in use

• Change port in predict.py or use different port mapping

---

🧾 License

This project is for educational purposes as part of ML Zoomcamp.

---

✨ Acknowledgements

• Intel Image Classification Dataset
• TensorFlow & Keras
• Fly.io
• ML Zoomcamp

---