🧠 CNN Animal Classifier

This project is part of the IronHack Machine Learning Specialization and focuses on building Convolutional Neural Network (CNN) models to classify animal images from the Animals10 Kaggle dataset. The final solution incorporates both custom CNN architectures and transfer learning, with model evaluation based on accuracy, F1-score, and confusion matrices.

---

Links

1. Report: Google Docs
2. Presentation: Google Slides
3. GitHub: GitHub Repo
4. Dataset: Animals10 - Kaggle

---

Bootstrap

1. Download the dataset by running the 1st cell.
2. Install python-dotenv module.
3. Create .env file.
4. Link your newly downloaded folder path to a variable called : DATASET_URI

---

📁 Repository Structure

This repository contains multiple notebooks created during the exploration and development process. The final report and conclusions are based primarily on the following:

• pre_processing.ipynb – Dataset loading, image resizing, splitting, and augmentation setup.
• M1_Joao.ipynb – Initial custom CNN architecture experimentation.
• final_notebook_JJ.ipynb – Final custom CNN model, training process, evaluation.
• transfer_model_MO.ipynb – Transfer learning implementation and final model selection.

Other notebooks in the repo may include additional experiments, visualizations, or helper scripts.

---

🧪 Dataset

• Source: Animals10 - Kaggle
• Classes: Dog, Cat, Horse, Spider, Butterfly, Chicken, Sheep, Cow, Squirrel, Elephant
• Image Count: ~28,000 images (medium quality)

---

📊 Final Model

After testing multiple CNN configurations, the best results were achieved using a transfer learning approach, which significantly improved validation performance and training efficiency.

• Framework: TensorFlow / Keras
• Preprocessing: Resizing, normalization, augmentation (flip, rotation, zoom, contrast)
• Final Accuracy Custom Model: ~73%
• Final Accuracy Transfer Learning Model: ~91%
• Evaluation Metrics: Precision, Recall, F1-score, Confusion Matrix

---

📝 Key Insights

• Transfer learning provided superior accuracy with less training time.
• Class weighting helped balance the model’s learning across underrepresented classes.
• Confusion matrices and classification reports revealed common misclassifications (e.g., cat ↔ cow, dog ↔ sheep).
• Time constraints and external commitments were limiting factors.
• Gladio deployment was optional and not completed due to time constraints.

---

🤝 Teamwork

"Working together as a group was pure bliss. We learned a lot from each other’s work and supported one another throughout the process."

---

🧠 Authors

• João [@JoaoPeseiro]
• JJ [@Jyok1m]
• MO [@mobenet]

---

📅 Project Date

April 2025 – IronHack Final CNN Project

---

📌 License

This project is for educational purposes only.