Pokémon Type Classifier: Multi-Stage Deep Learning Challenge

Universidade Nova de Lisboa - Deep Learning Course (2024-2025)

This repository contains the solution for a Pokémon Image Classification Challenge, exploring various deep learning approaches from a Multilayer Perceptron (MLP) to Convolutional Neural Networks (CNN) and Transfer Learning/Fine-Tuning techniques. It was also a kaggle competition, we get the Best Private Leaderboard Score Achieved: 0.97 (Task 2).

1. Project Overview

The core objective was to classify Pokémon images into 9 primary types. The project was structured into three incremental tasks:

• Task 1: MLP (Multilayer Perceptron (4 layers: 12288 → 256 → 128 → 9 neurons)). Achieved with 0.93 validation accuracy due to highly effective preprocessing (background removal and image centering).
• Task 2: CNN from Scratch: Custom CNN (3 Conv layers, 2 FC layers). Achieved with 0.97 private score and outperformed the MLP, demonstrating superior feature extraction from image data.
• Task 3: Transfer Learning: Pre-trained EfficientNet with Transfer Learning and Fine-Tuning.Tested against a complex dataset without the custom preprocessing. Fine-Tuning achieved the best performance among the transfer learning models (best F1: 0.2596).

1.1. Critical Preprocessing Step

A major factor in the high performance for the first two tasks was the custom preprocessing pipeline: undersampling to balance the dataset, followed by background removal (using boundary fixing, thresholding, and morphological operations) and centering the Pokémon within the $64\times64$ image. This drastically improved the MLP's validation accuracy from ~0.30 to ~0.93.

1.2. Project Structure

The repository organization reflects the progression of the tasks:

• data/: Contains raw and processed image data for all tasks.
• data/Test_processed, data/Train_processed: Processed (centered & background-removed) images used in Tasks 1 & 2.
• data/Train_task3, data/Test_task3: Image data for the more complex Transfer Learning challenge (Task 3).
• notebooks/: Main workspace for model training and analysis. (MLP implementation and initial data preprocessing, Custom CNN implementation and optimization, Transfer Learning and Fine-Tuning pipeline comparison (EfficientNet, ResNet, VGG).
• train_labels.csv, train_labels_task3.csv: Labels for the respective tasks.
• slides_presentation/: Source presentation files.

2. Installation and Setup

This project requires Python and standard deep learning libraries, primarily PyTorch.

1. Clone the repository:

   git clone [URL_OF_YOUR_REPOSITORY]
   cd pokemon_classifier

2. Create and activate a virtual environment (Recommended):

   python -m venv venv
   source venv/bin/activate  # On Linux/macOS
   # .\venv\Scripts\activate  # On Windows

3. Install dependencies:

   # Assuming you are using PyTorch and other common packages
   pip install torch torchvision numpy matplotlib scikit-learn pandas jupyter

---

3. How to Run the Code

All training, evaluation, and analysis scripts are contained within the Jupyter Notebooks located in the notebooks/ directory.

1. Start Jupyter Notebook Server:

   jupyter notebook

2. Navigate to the notebooks/ directory:
   • To explore the MLP model, open and run Task1.ipynb.
   • To explore the Custom CNN model, open and run Task2.ipynb.
   • To explore the Transfer Learning and Fine-Tuning experiments, open and run Task3.ipynb.

GPU Usage

The notebooks are configured to automatically check for and use an available GPU (CUDA) to accelerate training. This significantly reduces training time.

4. Authors

This project was a collaborative effort by Team LLMΗ for the Deep Learning Challenge.

• GALLO Lorenzo
• LEKBOURI Lina
• LICHTNER Marc
• WERCK Hugo