ml-zoomcamp-capstone

Image Classification

Deep Learning Deployment with Streamlit & TensorFlow

---

Project Overview

This capstone project is an image classification model that identifies six categories:
[ Buildings, Forest, Glacier, Mountain, Sea, and Street ].

---

Problem Description

This project aims to classify images into one of six scene categories using a Deep Learning model. The goal is to:

• Train a robust image classifier using transfer learning.
• Deploy the model as a web application using Streamlit.
• Package the deployment using Docker for easy reproducibility.

---

Dataset

The dataset used for this project is the Intel Image Classification Dataset, which consists of labeled images for six different scene categories. The dataset is preprocessed and split into training, validation, and test sets.

Kaggel Link: Intel Image Classification on Kaggle

---

Exploratory Data Analysis (EDA)

To better understand the dataset, we performed an extensive EDA, which includes:

• Data Structure: Verified the organization of training, validation, and test sets.
• Image/Label Shape: Checked image consistency and label correctness.
• Class Distribution: Ensured balanced distribution across categories.
• Visualize Pictures: Displayed sample images to confirm dataset integrity.

For dataset analysis, refer to the Jupyter Notebook in the repository.

---

Model Training & Implementation

• Architecture: Pretrained Xception model with fine-tuning.
• Loss Function: Categorical Cross-Entropy.
• Optimizer: Adam Optimizer.
• Evaluation Metrics: Accuracy, Validation Loss.
• Data Augmentation: Applied techniques like flipping, rotation, and brightness adjustment.

Model Callbacks

• ModelCheckpoint: Saves the best-performing model.
• EarlyStopping: Stops training if no improvement is observed.
• ReduceLROnPlateau: Reduces learning rate on validation loss plateau.

---

Deployment

This model is deployed using Streamlit and Docker.

How It Works

1. Upload an image (JPG, PNG, JPEG).
2. Select image processing method:
   • Resize with padding (Maintain Aspect Ratio).
   • Central crop (Remove edges).
3. Prediction: The model classifies the image and outputs:
   • Predicted Class.
   • Confidence Score.

Hosting & Infrastructure

• Streamlit Cloud for web deployment. [Live Demo: Image Classifier]
• Google Drive (gdown) for downloading model weights.
• Docker for containerized application deployment.

---

Installation & Setup

1. Clone the Repository

git clone https://github.com/Pei-Tong/ml-zoomcamp-capstone.git
cd ml-zoomcamp-capstone

2. Install Dependencies

pip install -r requirements.txt

3. Train the model from scratch

python train.py

4-1. Run Streamlit App

streamlit run app.py

4-2. Run with Docker

docker build -t image-classifier .
docker run -p 5000:5000 image-classifier

4-3. Run on Streamlit Cloud

Live Demo: Image Classifier

---

Project Files

File	Description
app.py	Streamlit UI for the image classification app.
train.py	Model training script using TensorFlow.
predict.py	Model inference script.
service.py	BentoML service file for model serving.
Dockerfile	Docker setup for deployment.
requirements.txt	Required dependencies for the project.

---

Model Performance

The final model achieves high accuracy on the test dataset. The evaluation metrics are as follows:

Metric	Score
Training Accuracy	88.05%
Validation Accuracy	89.68%
Test Accuracy	89.27%

---

Web Interface

The following is a screenshot of the Customer Purchase Prediction app interface:

---

Evaluation Criteria Mapping

Criteria	Status
Problem Description	✅ Described in README
EDA	✅ Included dataset analysis in Jupyter Notebook
Model Training	✅ Used transfer learning, tuning, and augmentation
Exporting Notebook to Script	✅ train.py provided for training
Reproducibility	✅ Dataset link provided, setup instructions included
Model Deployment	✅ Streamlit & Docker used for deployment
Dependency Management	✅ requirements.txt included
Containerization	✅ Dockerfile provided
Cloud Deployment	✅ Hosted on Streamlit Cloud

---

Future Improvements

• Fine-tuning the model with larger datasets.
• Implementing real-time image classification.
• Deploying on cloud platforms like AWS/GCP.