Data Science project: Convolutional Neural Networks (CNN)

Course: Data Science
University: University of Tehran
Instructors: Dr. Bahrak, Dr. Yaghoobzadeh

---

Overview

This project implements Convolutional Neural Networks (CNNs) for classifying images in the Flowers Multiclass Dataset. The assignment explores two main models:

• A VGG-style CNN model built from scratch.
• A fine-tuned pretrained ResNet50 model.

The implementation uses PyTorch and includes data loading, augmentation, training, evaluation, and visualization.

---

Project Goals

1. Load and preprocess the flower dataset.
2. Apply data augmentation to improve model generalization.
3. Train a custom CNN and evaluate its performance.
4. Fine-tune a pretrained ResNet50 model on the same dataset.
5. Compare both models using:
   • Accuracy
   • Precision
   • Recall
   • F1-score
   • AUC

---

Dataset

• Dataset: Flowers Multiclass Dataset (Kaggle)
• Directory structure (pre-divided):

  /train
  /validation
  /test
  

---

Key Features

• PyTorch-based training and evaluation
• Google Colab compatible
• Supports GPU acceleration
• Model checkpoints and training plots
• Well-structured pipeline for reproducibility

---

Running the Project

1. Clone the repository
2. Upload the notebook to Google Colab
3. Ensure you have access to Google Drive and the dataset is stored appropriately
4. Run all cells in the notebook

---

Output

• Trained VGG-style CNN model
• Trained ResNet50 model (fine-tuned)
• Performance metrics for both models
• Plots comparing accuracy and loss