Pokemon Type Classifier

Authors: Saigaurav Bettela, Jeron Perey, Nick Perlich, Sean McCormick

Motivation

We aim to classify a Pokémon’s primary type (18 classes) from an image using transfer learning with a pretrained CNN.

Dataset access link

https://www.kaggle.com/datasets/vishalsubbiah/pokemon-images-and-types?select=pokemon.csv

Problem

Given a Pokémon image, predict its primary type (e.g., Fire, Water, Electric) using learned visual features.

Approach

• Supervised learning using a Kaggle dataset (images + type labels)
• Transfer learning (e.g., ResNet / EfficientNet)
• Train head first (freeze backbone), then fine-tune upper layers

Evaluation

• Macro F1 (primary metric)
• Top-k accuracy (top-1 and top-3/top-5)
• Train/val/test split: 70/10/20

Repo Structure

• data/ dataset (not tracked)
• classifiers/ model implementations
  • pytorch/ PyTorch-based classifiers
  • tensorflow/ TensorFlow-based classifiers
  • baseline/ baseline models + EDA
• evaluations/ evaluation code + results
• utils/ shared utilities (dataset, preprocessing)
• scripts/ helper scripts
• requirements/ dependency files (base, cpu, cu126)

Setup

1. Create a virtual environment and install dependencies:

   python -m venv venv
   source venv/bin/activate
   # For CPU-only:
   pip install -r requirements/requirements.cpu.txt
   # For CUDA 12.6:
   pip install -r requirements/requirements.cu126.txt

2. Download the dataset from the Kaggle link above and place the files so the structure looks like:

   data/
   ├── pokemon.csv
   └── images/
       ├── bulbasaur.png
       ├── charmander.png
       └── ...
   

3. Train
4. Evaluate