🐾 Deep Learning Animal Image Classifier

An end-to-end multi-class image classification pipeline powered by PyTorch and MobileNetV2.

📌 Project Overview

The Animal Image Classifier is a lightweight, efficient, and highly accurate computer vision system designed to categorize images of animals into their respective classes (e.g., Bear, Bird, Cat, Dog, Elephant, Zebra).

Built to be a practical demonstration of Transfer Learning, this project leverages a pretrained convolutional neural network to achieve fast convergence with minimal training data. It takes raw image folders, processes them, trains a newly attached classification head, and evaluates the model's predictive power.

---

🏗️ End-to-End Architecture

This project follows a systematic ML engineering lifecycle, abstracting away unnecessary complexity while ensuring reproducibility.

graph TD
    A[Raw Image Data<br><code>animal_data/</code>] -->|Split via split-folders| B(Train/Test Splits<br><code>dataset_split/</code>)
    B --> C{Data Loader & Transforms}
    C -->|Batched Tensors & Labels| D[Pretrained MobileNetV2 Backbone]
    D -->|Feature Extraction| E[Custom Linear Classifier Head]
    E --> F((Training Loop))
    F -->|Loss calculation & Backprop| E
    F -->|Validation| G((Evaluation Metrics))
    E --> H[Prediction / Inference]

Loading

🧠 How It Works (The Core Mechanism)

1. Data Ingestion & Formatting: The pipeline expects a clean directory structure where each folder represents a class label. Using standard torchvision.datasets.ImageFolder, it dynamically maps folders to PyTorch Tensors.
2. Pre-processing: Raw images are normalized and resized to align perfectly with the distribution that MobileNetV2 was originally trained on (ImageNet). This ensures maximum feature extraction quality.
3. Model Configuration: We initialize the MobileNetV2 backbone with pre-trained weights. By freezing the base layers, we prevent the destruction of foundational pattern recognition capabilities (like edge detection or texture recognition) that the model learned from millions of images.
4. Transfer Learning & Optimizations: A custom Dense/Linear layer acts as our new classifier head. We utilize the AdamW optimizer paired with Cross-Entropy Loss to calculate gradients and iteratively update the network's weights.
5. Inference Function: After training, isolated helper functions allow single .jpg or .png files to be passed directly through the network, outputting a Softmax probability curve and the top-1 class prediction.

---

🔬 Model & Technology Stack

🚀 Tech Stack

• Framework: PyTorch (torch, torchvision, torchinfo)
• Data Engineering: split-folders, Pillow
• Visualization: matplotlib
• Environment: Jupyter Notebook / Python 3.x

🧬 Why MobileNetV2?

• Efficiency: Utilizes depthwise separable convolutions to drastically reduce the number of parameters.
• Speed: Extremely lightweight, meaning inference can be run entirely on standard CPUs in milliseconds.
• Accuracy: Due to ImageNet pre-training, it recognizes real-world animal textures out-of-the-box.

---

📂 Repository Structure

.
├── Animal-Classifier.ipynb       # Core training & inference pipeline
├── animal_data/                  # (User-provided) Raw image folders per class
├── dataset_split/                # (Auto-generated) Train & test distributions
├── kangaroo.jpg                  # Example prediction image
├── dolphin3.jpg                  # Example prediction image
└── README.md                     # Project documentation

Note

Note on folder naming: While this codebase currently resides in a Facial_Emotion_Recognition folder on the host machine, the actual logic, modeling, and output presented here are fully dedicated to multi-class Animal Image Classification.

---

⚙️ Setup & Installation

Follow these steps to replicate the environment and begin training on your local system:

1. Environment Configuration (Linux / macOS)

# Initialize a virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Upgrade pip
pip install --upgrade pip

# Install required dependencies
pip install torch torchvision torchinfo split-folders pillow matplotlib jupyter ipykernel

Tip

GPU Acceleration: If you have an NVIDIA GPU, ensure you install the CUDA-compatible binaries from the Official PyTorch Website.

2. Preparing the Data

1. Create an animal_data/ folder in the root directory.
2. Inside animal_data/, create subfolders for each animal you want to classify (e.g., animal_data/Bear/, animal_data/Zebra/).
3. Place your raw .jpg or .png images into their respective folders.

3. Execution

Launch Jupyter Notebook and open Animal-Classifier.ipynb. Run the cells sequentially to build the dataset splits, instantiate the dataloaders, execute the training epochs, and run single-image inferences.

---

📈 Monitoring Performance & Results

During execution, keep an eye on the following outputs generated by the notebook:

• Loss Curves (Train vs Test): Watch for convergence and ensure the model isn't overfitting.
• Accuracy Thresholds: Expect an upward trajectory culminating in robust classification metrics.
• Single-Image Robustness: Take images completely outside of your dataset (from the internet) and pass them to the inference function to qualitatively test model generalization.

---

💡 Future Enhancements

• Data Augmentation Strategies: Implement dynamic rotation, flipping, and color jitter to further generalize the model.
• Experiment Tracking: Integrate tools like Weights & Biases (WandB) or Tensorboard.
• Deployment Preparation: Export the final PyTorch weights to ONNX format to serve the model via an API (FastAPI) or run it natively on Edge devices.