pokedec

This package can detect pokemon species based on images.

Project Description

Pokémon are fascinating creatures beloved by millions worldwide, and their diversity makes them an exciting challenge for image classification. Detecting and classifying Pokémon images is not just a fun exercise, it also provides a valuable opportunity to explore and demonstrate simple machine learning techniques.

In this project, we aim to fine-tune a pre-trained model of the ResNet structure, which we found on PyTorch Image Models (TIMM). This is called ResNet50-D, which was trained on the ImageNet-1k dataset and has 25.6 million parameters. The new task will be to classify a Pokémon species from an image. Additionally, an important part of the project focuses on ensuring reproducibility and robustness and setting up an ML operations framework focusing on continuous integration.

We utilize the “1000 Pokemon Dataset” from Kaggle, which consists of 26,539 images of size 128 by 128 pixels comprising 1000 different Pokémon classes. The dataset is further split into 33% training, 33% validation and 33% testing. To ensure reliable classifications the data is augmented in order to generate more diverse training data. This is done using the TIMM framework or alternatively the built-in PyTorch augmentation functionality.

Website

The deployed webpage can be found here. When classifying the first image it will take some time as the model is also loaded at that time. Afterwards, all following classifications are made immediately.

Project structure

The directory structure of the project looks like this:

├── .dvc/                     # Dvc config to remote
├── .github/                  # Github actions and dependabot
│   ├── dependabot.yaml
│   └── workflows/
│       └── *.yaml
├── configs/                  # Configuration files
│       └── *.yaml
├── data/                     # Data directory after pulling from dvc
│   ├── processed
│   └── raw
├── dockerfiles/              # Dockerfiles
│   └── *.dockerfile
├── docs/                     # Documentation
│   ├── mkdocs.yml
│   └── source/
│       └── index.md
├── models/                   # Trained models
├── reports/                  # Reports
│   ├── figures/
│   └── README.md             # Report and answers to questions
├── src/                      # Source code
│   ├── pokedec/
│   │   ├── __init__.py
│   │   ├── backend.py
│   │   ├── data_drift.py
│   │   ├── data_statistics.py
│   │   ├── data.py
│   │   ├── evaluate.py
│   │   ├── frontend.py
│   │   ├── image_analysis.py
│   │   ├── models.py
│   │   └── train.py
├── tests/                    # Tests
│   ├── __init__.py
│   ├── integration/
│   ├── performance/
│   ├── tests_data/
│   └── unit/
├── .dvcignore
├── .gcloudignore
├── .gitignore
├── .pre-commit-config.yaml
├── data.dvc
├── LICENSE
├── pokemon_labels.json
├── pyproject.toml            # Python project file
├── README.md                 # Project README
├── requirements_backend.txt  # Backend requirements
├── requirements_datadrift.txt# Datadrift requirements
├── requirements_dev.txt      # Development requirements
├── requirements_frontend.txt # Frontend requirements
├── requirements.txt          # Project requirements
├── tasks.py                  # Project tasks
└── uv.lock                   # uv dependency versions

Created using mlops_template, a cookiecutter template for getting started with Machine Learning Operations (MLOps).