Animesh Karnewar and Oliver Wang
CVPR 2020
This folder provides a re-implementation of this paper in PyTorch, developed as part of the course METU CENG 796 - Deep Generative Models. The re-implementation is provided by:
-
Deniz A. Acar, denizalperacar@gmail.com
-
Yavuz Durmazkeser, yavuz.durmazkeser@metu.edu.tr
Please see the jupyter notebook file main.ipynb for a summary of paper, the implementation notes and our experimental results.
Solution to CelebA Download Error: In case the code fails to download CelebA Dataset, download this folder and place it under data/.
Call the train.py from terminal to train the model in the following way:
$ python3 train.py --save_dir test/ --continue_checkpoint 1 --num_epochs 1 --lr 0.0001 --batch_size 16 --latent_dim 128 --num_blocks 4 --use_gpu 1 --n_disc 1 --dataset CIFAR10 --lamda 10.0--save_dir <saveDir> directory that the model is saved to.
--continue_checkpoint loads the model in saveDir
--num_epochs 1 num_epochs < 0 for convergence termination else terminates at num_epochs iterations
--lr 0.0001 specify the learning rate for the optimizer
--batch_size 16 specify the batch size for the training
--latent_dim 128 specify the latent vector size of the model
--num_blocks 4 specify the number of blocks to be used in the generator and discriminator. https://arxiv.org/pdf/1903.06048.pdf Table 6-7
--use_gpu 1 true to use GPU
--n_disc 1 Number of discriminator optimization steps before one generator optimizer step
--dataset CIFAR10 CIFAR10 to load Cifar 10 dataset CelebA to load CelebA dataset
--lamda 10.0 WGAN-GP lambda value. \