Contrastive Learning of Generalized Game Representations

IEEE CoG 2021 Paper | Sports10 Dataset | Pre-Trained Models

This repository is the official code implementation of the paper "Contrastive Learning of Generalized Game Representations" by Chintan Trivedi, Antonios Liapis and Georgios Yannakakis.

Installation

git clone git@github.com:ChintanTrivedi/contrastive-game-representations.git
cd contrastive-game-representations
pip install -r requirements.txt

Note that this code has been developed/tested with a single NVIDIA GPU using CUDA 10.1 and Tensorflow 2.3 on Windows platform. It is expected to work as-is on Linux/Colab platforms (untested).

Download Sports10 Dataset

We present a new dataset containing 100,000 Gameplay Images of 175 Video Games across 10 Sports Genres. The games are also divided into three visual styling categories: RETRO (arcade-style, 1990s and earlier), MODERN (roughly 2000s) and PHOTOREAL (roughly late 2010s).

• Download the full dataset (~10GB) from google drive and extract the zip file's contents to the datasets directory in the project.
• Meta-data is also available with the full list of games.
• For more information on the dataset, please refer our paper.

Training models from scratch

After downloading and unzipping the dataset, learn game representations with a ResNet50 encoder by running the following commands on your terminal.

1. Plain Supervised Learning

python train_fulsup.py --dataset_directory="./datasets/Sports10" --epochs=10 --img_shape=224 --batch_size=64

2. Supervised Contrastive Learning

python train_supcon.py --dataset_directory="./datasets/Sports10" --epochs=10 --img_shape=224 --batch_size=64

For additional training arguments, check train_fulsup.py or train_supcon.py

Download Pre-Trained Model

Alternatively, you can download our pre-trained models from this google drive link. Models can be loaded using tensorflow.keras.models.load_model('$MODEL_FILENAME.h5') and fine-tuned for any downstream task involving RL or GANs.

Citation

@inproceedings{trivedi2021contrastive,
  title={Contrastive Learning of Generalized Game Representations},
  author={Trivedi, Chintan and Liapis, Antonios and Yannakakis, Georgios N},
  booktitle={2021 IEEE Conference on Games (CoG)},
  year={2021},
  organization={IEEE}
}

Acknowledgements

1. Improving Image Classifiers With Supervised Contrastive Learning by Sayak Paul
2. Contrastive Loss Functions by Zichen Wang