MustafaMarwat / Transfer-Learning

This project implements a deep learning-based image classifier to distinguish between alpaca and non-alpaca images. The classifier uses transfer learning with MobileNetV2, a pre-trained Convolutional Neural Network (CNN) trained on the ImageNet dataset. The goal is to create an efficient and accurate model for alpaca recognition.

Transfer Learning: Image Classifier with MobileNetV2

This project implements a deep learning-based image classifier to distinguish between alpaca and non-alpaca images. The classifier uses transfer learning with MobileNetV2, a pre-trained Convolutional Neural Network (CNN) trained on the ImageNet dataset. The goal is to create an efficient and accurate model for alpaca recognition.

Table of Contents

• Project Objectives
• Key Features
• Dataset
• Model Architecture
• Training
• Fine-tuning
• Results
• Dependencies
• Contributing

Project Objectives

1. Create Dataset:

• Establish a dataset from a directory containing alpaca and non-alpaca images.
• Implement data splitting into training and validation sets.

2. Preprocess and Augment Data:

• Apply data augmentation techniques using the Sequential API to enhance diversity in the training set.
• Utilize prefetching to prevent memory bottlenecks during data preprocessing.

3. Transfer Learning with MobileNetV2:

• Explore the architecture of MobileNetV2, focusing on depthwise separable convolutions and bottleneck layers.
• Delete the top layer of MobileNetV2, which contained ImageNet classification labels.
• Integrate a new classifier layer for binary classification (alpaca or not alpaca).

4. Model Training:

• Develop a custom model using the Functional API and MobileNetV2 base.
• Fine-tune the model's final layers to improve accuracy.
• Compile and train the model using Binary Crossentropy loss and Adam optimizer.

5. Fine-tuning:

• Unfreeze selected layers of MobileNetV2 for fine-tuning.
• Re-run the optimizer with a smaller learning rate to adapt the model to new data.
• Explore the impact of fine-tuning on model accuracy.

Key Features

• Efficient alpaca recognition using MobileNetV2.
• Transfer learning for leveraging pretrained model knowledge.
• Data augmentation for enhancing training set diversity.
• Fine-tuning for adapting the model to specific data.

Dataset

The dataset consists of alpaca and non-alpaca images stored in the dataset/ directory. It is split into training and validation sets using the image_dataset_from_directory function.

Model Architecture

The model architecture is based on MobileNetV2 with a customized classifier for binary classification. It utilizes depthwise separable convolutions and bottleneck layers to achieve efficient feature extraction.

Training

The model is trained using the train_dataset with data augmentation and preprocessing. The training process involves fine-tuning the final layers for improved accuracy.

Fine-tuning

Fine-tuning involves unfreezing selected layers of MobileNetV2 and re-running the optimizer with a smaller learning rate. This step adapts the model to new data and captures high-level details.

Results

The trained model achieves accurate alpaca recognition with improved performance through fine-tuning.

Dependencies

• TensorFlow
• Matplotlib
• NumPy

Contributing

Feel free to contribute to this project by opening issues or submitting pull requests. Your feedback and contributions are highly appreciated.