A complete deep learning pipeline — from raw images to fine-tuned predictions —
built with TensorFlow, MobileNetV2, and a GitHub-dark aesthetic throughout.

---

🗂️ Table of Contents

• Overview
• Demo
• Project Structure
• Pipeline
• Results
• Setup & Run
• Tech Stack
• How It Works
• License

---

🔭 Overview

This project implements a binary image classifier (Cat 🐱 vs Dog 🐶) using transfer learning on top of Google's MobileNetV2 pretrained on ImageNet.

The notebook walks through two training phases:

Phase	Strategy	Learning Rate
Feature Extraction	MobileNetV2 frozen, only custom head trained	1e-3
Fine-Tuning	Top 30 layers of MobileNetV2 unfrozen	1e-5

The result is a lightweight, highly accurate classifier that trains in minutes on a free GPU (Google Colab).

---

🎬 Demo

Sample Grid	Augmentation

Training Curves	Confusion Matrix

---

📁 Project Structure

📦 animal-classifier/
├── 📓 Animal_Classifier_MobileNetV2.ipynb   ← Main notebook (10 tasks)
├── 🖼️  sample_grid.png                       ← Dataset preview
├── 🔄  augmentation.png                      ← Augmentation examples
├── 📊  feature_extraction_model_curves.png   ← Phase 1 training curves
├── 📊  fine-tuned_model_curves.png           ← Phase 2 training curves
├── 🔲  confusion_matrix.png                  ← Evaluation heatmap
├── 🎯  predictions.png                       ← Test image predictions
└── 📄  README.md

---

🔄 Pipeline

Raw Images (Cats vs Dogs — 25,000 images)
        │
        ▼
┌──────────────────────────────────────┐
│   Preprocessing & Augmentation       │
│   • Resize → 224×224                 │
│   • Random Flip / Rotate / Zoom      │
│   • MobileNetV2 normalization        │
│   • tf.data  cache → shuffle → batch │
└──────────────────────────────────────┘
        │
        ▼
┌──────────────────────────────────────┐
│   MobileNetV2  (ImageNet weights)    │  ← Frozen
│   + GlobalAveragePooling2D           │
│   + Dense(256, relu)                 │
│   + Dropout(0.3)                     │
│   + Dense(1, sigmoid)                │  ← Trainable
└──────────────────────────────────────┘
        │
        ▼
  Phase 1: Feature Extraction  (lr = 1e-3)
        │
        ▼
  Phase 2: Fine-Tuning          (lr = 1e-5, top 30 layers unfrozen)
        │
        ▼
  Evaluation → Confusion Matrix + Classification Report
        │
        ▼
  Visualised Predictions on Test Set

---

📊 Results

Metric	Feature Extraction	Fine-Tuned
Train Accuracy	~87%	~93%
Val Accuracy	~85%	~90%
Test Accuracy	TBD (run notebook)	TBD (run notebook)

Results will vary slightly due to random seed and hardware differences.

---

🚀 Setup & Run

▶️ Run on Google Colab (Recommended — free GPU)

Replace YOUR_USERNAME with your GitHub username after uploading.

💻 Run Locally

# 1. Clone the repo
git clone https://github.com/YOUR_USERNAME/animal-classifier.git
cd animal-classifier

# 2. Create a virtual environment
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

# 3. Install dependencies
pip install tensorflow tensorflow-datasets seaborn scikit-learn matplotlib jupyter

# 4. Launch the notebook
jupyter notebook Animal_Classifier_MobileNetV2.ipynb

📦 Requirements

tensorflow>=2.10
tensorflow-datasets>=4.9
matplotlib>=3.6
seaborn>=0.12
scikit-learn>=1.2
numpy>=1.23

---

🛠️ Tech Stack

Tool	Purpose
TensorFlow 2.x	Deep learning framework
Keras	High-level model API
MobileNetV2	Pretrained backbone (ImageNet)
TensorFlow Datasets	cats_vs_dogs dataset (25k images)
Matplotlib	Training curves & image grids
Seaborn	Confusion matrix heatmap
scikit-learn	Classification report & metrics

---

🧠 How It Works

1. Transfer Learning

Instead of training from scratch, we load MobileNetV2 pretrained on 1.2M ImageNet images. Its convolutional layers already know how to detect edges, textures, and shapes — all transferable to our cat/dog task.

2. Feature Extraction (Phase 1)

The entire MobileNetV2 base is frozen (weights locked). Only the new classification head — Dense(256) → Dropout → Dense(1, sigmoid) — learns. This is fast and avoids destroying the pretrained features.

3. Fine-Tuning (Phase 2)

After Phase 1 converges, the top 30 layers of MobileNetV2 are unfrozen and retrained at a very small learning rate (1e-5). This lets the higher-level features adapt specifically to cats and dogs, squeezing out extra accuracy.

4. Data Augmentation

Random horizontal flips, rotations (±15°), and zooms (±10%) are applied on-the-fly during training, effectively multiplying the dataset size and improving generalisation.

---

📄 License

This project is licensed under the MIT License — feel free to use, modify, and distribute.

---

Made with ❤️ and TensorFlow

⭐ Star this repo if you found it useful!