Cat vs Dog Image Classification – CNN Project Overview

---

📑 Table of Contents

• 👤 Author’s Note
• 📘 Project Summary
• 📂 Dataset
• 🧠 Model Variants and Numbering
• 🏆 Best Results
• 🖼️ Visual Results
• ▶️ How to Run
• 📄 License

---

👤 Author’s Note

Hi, I’m Bo Fu – an experienced software engineer transitioning into the field of AI. I’ve been accepted into the MSc Artificial Intelligence program at City, University of London.

To build a strong foundation before the program starts, I’m actively developing deep learning projects from scratch — combining self-study with hands-on experimentation. This project reflects my personal exploration of Convolutional Neural Networks, from data preparation to model architecture and final evaluation. It is structured for clarity and reproducibility — both for myself and others who might learn from or improve upon it.

🔗 LinkedIn
📅 Completed: June 2025

This notebook is a work in progress — feedback, suggestions, or pull requests are welcome.

---

📘 Project Summary

This notebook builds and evaluates several Convolutional Neural Network (CNN) models to classify cat and dog images. The training pipeline includes:

• Multiple input image resolutions (50×50, 96×96, 160×160, 224×224, 256×256)
• Different data augmentation strategies
• Training with early stopping and history tracking
• Final evaluation on an unseen test dataset
• Also available as a Kaggle Notebook

---

📂 Dataset

• Source: The dataset is obtained from the Kaggle Dogs vs. Cats competition.
• Original Format: All images are stored in a single folder and named as cat.#.jpg or dog.#.jpg.

Custom Split:

• Training + Validation:

  • 10,000 cat images: cat.0.jpg to cat.9999.jpg
  • 10,000 dog images: dog.0.jpg to dog.9999.jpg
  • These were placed in train_val_ds/cats/ and train_val_ds/dogs/

• Test Set:

  • 2,500 cat images: cat.10000.jpg to cat.12499.jpg
  • 2,500 dog images: dog.10000.jpg to dog.12499.jpg
  • These were placed in test_ds/cats/ and test_ds/dogs/

🧾 Note: Due to file size limitations and licensing restrictions, the image dataset is not included in this repository. Please download it from Kaggle and organize the images as described above. The folders train_val_ds/ and test_ds/ are included as empty placeholders to preserve the project’s directory structure—please populate them with the appropriate images before running the code.

---

🧠 Model Variants and Numbering

Model No.	Model Name	Architecture	Image Size	Augmentation	Epochs	Callback
No. 1	model_v1_img50	create_model_v1	50×50	None	10	None
No. 2	model_v1_img96	create_model_v1	96×96	None	10	None
No. 3	model_v1_img50_augv1	create_model_v1	50×50	data_augmentation_v1	10	None
No. 4	model_v1_img96_augv1	create_model_v1	96×96	data_augmentation_v1	10	None
No. 5	model_v1_img96_augv1_e20_es3	create_model_v1	96×96	data_augmentation_v1	20	early_stop_es3
No. 6	model_v1_img96_augv2_e20_es3	create_model_v1	96×96	data_augmentation_v2	20	early_stop_es3
No. 7	model_v2_img96_augv1_e20_es3	create_model_v2	96×96	data_augmentation_v1	20	early_stop_es3
No. 8	model_v2_img96_augv2_e20_es3	create_model_v2	96×96	data_augmentation_v2	20	early_stop_es3
No. 9	model_v2_img96_augv1_e25_es8	create_model_v2	96×96	data_augmentation_v1	25	early_stop_es8
No. 10	model_v2_img96_augv2_e25_es8	create_model_v2	96×96	data_augmentation_v2	25	early_stop_es8
No. 11	model_v3_img160_augv1_e30_es8	create_model_v3	160×160	data_augmentation_v1	30	early_stop_es8
No. 12	model_v3_img160_augv2_e30_es8	create_model_v3	160×160	data_augmentation_v2	30	early_stop_es8
No. 13	model_v3_img224_augv2_e40_es10	create_model_v3	224×224	data_augmentation_v2	40	early_stop_es10
No. 14	model_v3_img256_augv2_e45_es12	create_model_v3	256×256	data_augmentation_v2	45	early_stop_es12

⚠️ All models are saved using ModelCheckpoint to retain the best validation accuracy.

---

🏆 Best Results

Model No.	Model Name	Train Accuracy	Val Accuracy	Precision	F1-Score
No. 13	model_v3_img224_augv2_e40_es10	95.1%	92.0%	95.7%	94.9%

---

🖼️ Visual Results

📈 Training History
Model 13 – Accuracy & Loss over 40 epochs

---

📊 Confusion Matrix
Model 13 – Evaluated on 5,000 unseen test images

---

🐾 Visual Predictions
Model 13 – Predictions on test images (labels shown above each image)

---

▶️ How to Run

1. Prepare the Datasets:

   • Place the training set (with cats/, dogs/) inside the train_val_ds/ folder
   • Place the test set (with cats/, dogs/) inside the test_ds/ folder

2. Set Parameters:

   • Configure the environment at the top of the notebook (LOCAL or KAGGLE)
   • All paths and hyperparameters (e.g., SEED, BATCH_SIZE) are defined centrally

3. Modularized Codebase:

   • data_utils.py: data loading + preprocessing
   • model_creation.py: model definitions
   • train.py: training pipeline
   • visualization.py: plots
   • evaluation.py: metrics + visual predictions

4. Review the Results:

   • Training curves (accuracy/loss)
   • Confusion matrix with precision and F1-score
   • Visual sample predictions

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

🔼 Back to Top