5th place solution for Kaggle's Generative Dog Images competition.
This solution is based on PyTorch BigGAN implementation by Andy Brock. Some of the original modules were replaced or removed. In case you want to use this model to solve other problems, please, start with the original repo.
Kaggle's kernel environment with GPU support was used to develop and run this solution. Approximate specs are listed bellow:
- Intel(R) Xeon(R) CPU @ 2.20GHz, 2 cores
- 13 Gb RAM
- 1 x NVIDIA Tesla P100
- Debian GNU/Linux 9 (stretch)
- Python 3.6.6
- CUDA 10.0.130
- cuddn 7.5.0
- NVIDIA drivers 418.67
The description of available settings can be found in the beginning of utils.py. Some of them (like parallel execution or half/mixed precision) won't work, because the required functionality was partially removed.
To change path to model inputs --data_root and --label_root parameters should be set to appropriate locations.
To start training a new model execute run_train.sh. You may want to edit it first to change some of the settings.
Execute run_sample.sh. To set the number of images to be generated change --sample_num parameter. Generated images will be stored in the --samples_root directory (./samples by default). Zip archive with generated images will be saved in the script root directory.
You can download it from this Kaggle's dataset (Kaggle's account required). Typically the script expects weights to be in ./weights/generative_dog_images directory, but you can change this behavior by setting --weights_root and --experiment_name parameters.
@inproceedings{
brock2018large,
title={Large Scale {GAN} Training for High Fidelity Natural Image Synthesis},
author={Andrew Brock and Jeff Donahue and Karen Simonyan},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=B1xsqj09Fm},
}