---

Transfer Learning for Alpaca Classification

This project demonstrates the use of transfer learning with the MobileNet architecture to classify images of alpacas using a small, specialised dataset. The Alpaca Dataset (Small), sourced from Kaggle, provides diverse images of alpacas in varying poses, lighting conditions, and backgrounds. The aim of this project is to fine-tune MobileNet, a pre-trained convolutional neural network, to achieve high performance in a niche image classification task with limited data.

---

Project Overview

This project explores:

1. The principles of transfer learning.

2. Modifications to MobileNet for task-specific fine-tuning.

3. Data preprocessing techniques for small datasets.

4. Training, evaluation, and deployment of a lightweight model for image classification.

---

Dataset

The Alpaca Dataset (Small) contains high-quality images of alpacas. It includes a variety of poses and backgrounds to challenge the model’s ability to generalise. The dataset was curated to mimic real-world scenarios where data availability is limited.

Source: Kaggle Alpaca Dataset (Small)

---

Project Structure

• dataset/: Contains the Alpaca Dataset.

• notebooks/: Contains the main Jupyter Notebook where the analysis, training, and results are documented.

• models/: Saved weights and configurations of the trained MobileNet model.

• src/: Includes the test_utils.py file, which provides helper functions or classes used in the notebook.

• results/: Outputs such as evaluation metrics and test results.

• README.md: Project documentation (this file).

---

Workflow

The project follows these steps:

1. Data Collection: Import and review the Alpaca Dataset.

2. Data Preprocessing: Resize, normalise, and augment the dataset for training.

3. Model Preparation: Load the pre-trained MobileNet model and adjust its layers for classification.

4. Training: Fine-tune the model on the Alpaca Dataset.

5. Evaluation: Assess the model using accuracy and loss metrics.

6. Deployment: Test the model on unseen data to validate its generalisation capability.

---

Requirements

To run the project, you need the following:

• Python 3.8 or later

• TensorFlow 2.x

• NumPy

• Matplotlib

• Pandas

• Jupyter Notebook

Install dependencies using:

pip install -r requirements.txt

---

Results

The fine-tuned MobileNet achieved high accuracy on the Alpaca Dataset, demonstrating the power of transfer learning in adapting pre-trained models for small, task-specific datasets. Evaluation metrics and sample predictions are available in the results/ folder.

---

Key Findings

• Transfer learning enables efficient training on small datasets.

• Data augmentation helps prevent overfitting.

• MobileNet’s lightweight architecture is ideal for resource-constrained tasks.

---

Future Work

• Explore additional datasets to test model generalisation further.

• Implement real-time alpaca detection using the trained model.

• Experiment with alternative architectures like ResNet or EfficientNet.

---

Acknowledgements

• Deeplearning.ai for the MobileNet tutorial.

• Kaggle for providing the Alpaca Dataset.

---

License

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