PokéGAN: Gotta Generate 'em All!

I trained a Wasserstein GAN with gradient penalty (WGAN-GP) to generate images of Pokémon. The Jupyter notebook in this repository goes through the entire process of from obtaining the data using a webscraper, to training the WGAN-GP, to deploying the generator with a Flask web app. The web app itself is also checked into the repository.

Training the GAN required about 24 hours on a laptop with an Intel Core i7-6820HQ CPU @ 2.70GHz and an NVIDIA Quadro M2002M GPU (CUDA Compute Capability 5.0).

I wrote the code in Python 3.8 and used Fast.ai v2 to install most of the dependencies needed to run the notebook. A requirements.txt file is included for convenience.

The files of interest are:

• PokeGAN.ipynb The code to develop the webscraper, image pre-processing, and WGAN-GP architecture and training loop.
• app.py Runs the Flask web server app that generates random Pokémon images.

For results and discussion, refer to PokeGAN.ipynb.

The PokéGAN Webapp

You can try out the generator yourself! Install Flask, navigate to the "webapp" directory, open a console and type flask run. Navigate to 127.0.0.1:5000 in your browser and generate some Pokémon!

Note: I could not include the trained GAN models in this repo as they exceed GitHub's 100 MB file size limit. Run the Jupyter notebook to generate the models before running the web app.

Acknowledgments

Image Credits

Thanks to Font Meme for the PokéGAN image generated using their Pokémon font. Web app background image created by DesenhosdoReu. Training images scraped from the Pokémon Fandom wiki.

Conceptual Resources

Thanks to Christian Versloot for his explanations of transpose convolution and leaky ReLU. Thanks also to Mayank Agarwal for his informative illustration of normalization techniques. Finally, thanks to Francesco for his tutorial on how-and-when to use the various PyTorch containers.

Programming Resources

Thanks of course to Nathan Inkawhich, the author of the PyTorch DCGAN tutorial. While designing the WGAN-GP I referred to code written by Mathworks, Emilien Dupont, Erik Lindernoren, Martin Arjovsky, and Aladdin Persson.

References

1. I. J. Goodfellow et al., “Generative Adversarial Networks,” arXiv:1406.2661 [cs, stat], Jun. 2014, Accessed: Jun. 09, 2021. [Online]. Available: http://arxiv.org/abs/1406.2661
2. A. Radford, L. Metz, and S. Chintala, “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks,” arXiv:1511.06434 [cs], Jan. 2016, Accessed: Jun. 09, 2021. [Online]. Available: http://arxiv.org/abs/1511.06434
3. M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein GAN,” arXiv:1701.07875 [cs, stat], Dec. 2017, Accessed: Jun. 08, 2021. [Online]. Available: http://arxiv.org/abs/1701.07875
4. I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville, “Improved Training of Wasserstein GANs,” arXiv:1704.00028 [cs, stat], Dec. 2017, Accessed: Jun. 09, 2021. [Online]. Available: http://arxiv.org/abs/1704.00028