🐶 DOG BREED CLASSIFICATION

Deep Learning for Accurate Multi-Class Dog Breed Identification

Built with the tools and technologies:

Python | TensorFlow 2.x | TensorFlow Hub | Keras | Pandas | NumPy | Matplotlib | scikit-learn | Jupyter Notebook

---

Table of Contents

• Overview
• Getting Started
  • Prerequisites
  • Installation
  • Usage
• Data Exploration & Preprocessing
• Modeling & Evaluation
• Model Deployment
• License
• Contact

---

Overview

This project presents an end-to-end deep learning pipeline for multi-class dog breed classification, leveraging transfer learning with TensorFlow and TensorFlow Hub. The notebook demonstrates:

• Data loading, inspection, and preprocessing for image and label data
• Exploratory data analysis and visualization of class distributions
• One-hot encoding of 120 dog breed labels
• Efficient image preprocessing and batching using TensorFlow's tf.data API
• Model construction using a pre-trained MobileNetV2 as a feature extractor
• Training with callbacks (TensorBoard, EarlyStopping) and validation monitoring
• Model evaluation, prediction visualization, and interpretability
• Saving, loading, and deploying trained models
• Preparing Kaggle-compatible CSV submissions and making predictions on custom images

Project Highlights

• Dataset: Kaggle Dog Breed Identification (10,222 training images, 120 breeds)
• Transfer Learning: MobileNetV2 from TensorFlow Hub for robust feature extraction
• Data Augmentation: Techniques to improve generalization and prevent overfitting
• Evaluation Metric: Log Loss, with probabilistic outputs for each class
• Visualization: Batch image grids, prediction confidence plots, and TensorBoard logs
• Reproducibility: Modular code for easy adaptation and deployment

---

Getting Started

Prerequisites

To run this notebook, you will need the following libraries installed:

• tensorflow (2.x)
• tensorflow_hub
• tf_keras
• pandas
• numpy
• matplotlib
• scikit-learn
• os
• IPython

Installation

You can install all necessary libraries using pip:

pip install tensorflow tensorflow_hub tf_keras pandas numpy matplotlib scikit-learn

Usage

1. Clone the repository:
   git clone https://github.com/brej-29/dog-breed-classification.git
2. Navigate to the project directory:
   cd dog-breed-classification
3. Open the Jupyter Notebook:
   jupyter notebook DogBreedClassification.ipynb
4. Follow the notebook cells:
   • Unzip and inspect the Kaggle dataset.
   • Import core libraries (TensorFlow, TensorFlow Hub, Keras, Pandas, NumPy, Matplotlib, scikit-learn).
   • Check for GPU availability for faster training.
   • Load and explore the dataset, visualize class distributions, and inspect images.
   • Prepare labels (one-hot encoding for 120 breeds).
   • Split data into training and validation sets.
   • Preprocess images (resize, normalize) and create efficient data batches using TensorFlow's tf.data API.
   • Visualize batched data for sanity checks.
   • Build the model using transfer learning (MobileNetV2 from TensorFlow Hub).
   • Set up callbacks (TensorBoard, EarlyStopping) for monitoring and regularization.
   • Train the model on a subset, then on the full dataset.
   • Evaluate and visualize predictions, including confidence plots and batch visualizations.
   • Save and reload trained models for reproducibility.
   • Make predictions on test data and prepare a Kaggle-compatible CSV for submission.
   • Run inference on your own custom images and visualize results.

---

Data Exploration & Preprocessing

• Load and inspect the Kaggle Dog Breed Identification dataset.
• Visualize class imbalance and breed frequencies.
• One-hot encode breed labels for multi-class classification.
• Preprocess images: resize to 224x224, normalize, and convert to tensors.
• Create efficient data pipelines with batching and shuffling for training, validation, and test sets.

---

Modeling & Evaluation

• Build a transfer learning model using MobileNetV2 as a feature extractor and a dense softmax output layer.
• Compile with categorical cross-entropy loss and Adam optimizer.
• Use TensorBoard and EarlyStopping callbacks.
• Train and validate the model, visualize predictions and confidence scores.
• Prepare predictions for Kaggle submission and evaluate on custom images.

---

Model Deployment

• Save trained models in HDF5 format for future inference or retraining.
• Functions provided for loading and evaluating saved models.
• Example code for making predictions on new, custom images.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Contact

If you have any questions or feedback, feel free to reach out via my LinkedIn Profile