Landmark-Classification

This project demonstrates how to train and evaluate Convolutional Neural Networks (CNNs) for image classification using PyTorch.
It includes three main workflows: training a CNN from scratch, applying transfer learning, and deploying a simple prediction app.

---

📂 Project Structure

.
├── cnn_from_scratch.ipynb   # Training a custom CNN from scratch
├── transfer_learning.ipynb  # Transfer learning with pretrained ResNet
├── app.ipynb                # Simple inference and demo app
├── src/
│   ├── helpers.py           # Project utilities
│   ├── model.py             # Custom CNN architecture
│   ├── transfer.py          # Transfer learning model builder
│   ├── train.py             # Training, validation, testing loops
│   ├── optimization.py      # Loss functions and optimizers
│   ├── predictor.py         # Inference helper
│   └── data.py              # Data loading and preprocessing

---

🚀 Workflows

1. CNN From Scratch

• Implements a custom CNN defined in model.py.
• Architecture: 5 convolutional blocks (Conv2D → BatchNorm → ReLU → MaxPool), followed by fully connected layers with dropout.
• Trained with CrossEntropyLoss and optimizers (SGD/Adam).
• Pros: full control of the architecture.
• Cons: requires more training data and computation.

2. Transfer Learning

• Uses a ResNet18 pretrained on ImageNet (torchvision.models).
• Freezes all backbone layers (requires_grad=False).
• Replaces the final fully connected layer with a custom linear classifier for our dataset (50 classes by default).
• Much faster training and better performance on small datasets.

3. Inference App

• Loads a trained model checkpoint.
• Uses predictor.py to preprocess input images and predict class labels.
• Demonstrates how to move from training to deployment.

---

⚙️ Installation

1. Clone the repository:

   git clone https://github.com/CarlosYazid/Landmark-Classification.git
   cd cnn-image-classification

2. Install dependencies:

   pip install -r requirements.txt

   Main dependencies:

   • PyTorch
   • Torchvision
   • Matplotlib
   • tqdm
   • livelossplot

---

🛠 Usage

Training from scratch

jupyter notebook cnn_from_scratch.ipynb

Transfer learning

jupyter notebook transfer_learning.ipynb

Run inference app

jupyter notebook app.ipynb

---

📊 Results

• From Scratch: flexible architecture, but requires more epochs to converge.
• Transfer Learning: achieves higher accuracy with fewer epochs by leveraging pretrained features.
• App: demonstrates how to perform predictions on unseen images.

---

🔬 Key Files

• src/model.py → Custom CNN (MyModel)
• src/transfer.py → Transfer learning setup with pretrained ResNet
• src/train.py → Training/validation loops, early stopping, learning rate scheduling
• src/optimization.py → Loss and optimizer selection
• src/predictor.py → Preprocessing and prediction logic

---

📌 Notes

• Images are resized to 224×224 pixels for consistency.
• Data augmentation (flips, crops, rotations) improves generalization when training from scratch.
• Validation loss is tracked, and the model is checkpointed when improvement is detected.
• Learning rate scheduling (ReduceLROnPlateau) is used to stabilize training.

---

📖 License

This project is for educational purposes. You are free to use and adapt it under the MIT license.