BigGAN implemented by tensorflow 2.0 RC version.
!!!!!!!! NOT CONFIRMED TO WORK !!!!!!!!
- numpy==1.17.0
- opencv-python==4.1.0.25
- scipy==1.3.1
- tensorboard==1.14.0
- tensorflow==2.0.0rc0
- tqdm==4.35.0
# Some related libraries will also be installed.
pip install -r requirements.txt
The officially reported results are obtained by quite huge dataset, here I trained on a relatively small cat dataset with single GPU.
# Download cat dataset, this script finally creates "cats_bigger_than_128x128" directory.
sh setting_up_script.sh
CUDA_VISIBLE_DEVICES=0 python train.py -d Cat -dd /path/to/cats_bigger_than_128x128 -i 100000 -b 32 --resize_shape 128 128 --flip_lr
-d and -dd indicate the target dataset name and the corresponding directory.
Other configs are stated in utils.py. Hyperparameters are almost the same as the paper.
Training log can be watched in TensorBoard.
tensorboard --logdir=logs
You can try to train on the other dataset by inherit the Base dataset class in datasets.py.
- Official TensorFlow documents: https://www.tensorflow.org/beta
- The author's PyTorch implementation of BigGAN: https://github.com/ajbrock/BigGAN-PyTorch
- Using Cat dataset for GAN: https://github.com/AlexiaJM/Deep-learning-with-cats
- Sophisticated solution in Kaggle competition: https://github.com/bestfitting/kaggle/tree/master/gandogs
- Easy-to-read tensorflow implementation of BigGAN: https://github.com/taki0112/BigGAN-Tensorflow