hw2

Conditional GAN Practice

Report

by Zhi-Yi Chin

In this homework, we apply two conditional GANs: cGAN on the MNIST dataset and pix2pix on the Facades dataset. The goal of cGAN is to perform an image generation task, which is given a condition (give a number range from 0 to 9) and a random latent vector; it can generate a digit image specified by the condition. The goal of pix2pix is to perform an image-to-image translation, which is given an input image (building segmentation mask) and the target image (real building photo); based on the input image, it can generate an output image similar to the target image.

Getting the code

You can download a copy of all the files in this repository by cloning the repository:

svn export https://github.com/joycenerd/Image_Manipulation_Techniques_Visual_Effects/trunk/hw2

Dependencies

You need to have Anaconda or Miniconda installed in your environment. You can create a virtual environment using conda and install all reequired dependencies in it by:

conda env create -f environment.yml
conda activate gan

Dataset

MNIST for cGAN

You don't need to download this explicitly. In cgan.py, torchvision.dataset will download the data before training and will be saved in the path specified by --dataroot argument

Facades for pix2pix

You can download the facades dataset via this Google drive link: facades.zip

Cityscapes for pix2pix

You can download the cityscapes dataset by following the website link: pix2pix/datasets

You will need to pre-process the cityscapes dataset because this dataset doesn't have its own validation set. We split the whole training set into 80% for training and 20% for validation.

1. rename the train/ folder in cityscapes to all_train/
2. split all_train to train and val

   python train_valid_split.py --data-root <data_dir>/cityscapes/all_train --ratio 0.2
   

   • input: all_train/ directory
   • output: train/ and val/ directories in cityscapes

Reproducing the results

cGAN on MNIST

Training

python cgan.py --dataroot <mnist_save_path> --logdir <output_path> --device cuda:0 --epochs 100 --mode train

• output:
  • MNIST data saves in --dataroot
  • checkpoints/, logs/ and results/ in --logdir
  • cGSN_loss.png

Testing

python cgan.py --logdir <log_root_path> --device cuda:0 --mode test

• input: logging root directory same as training --logdir
• output:
  • result.jpg: 10x10 grid of generated digits save in results/ of the logging root directory
  • result_condition.jpg: digit 0-9 save in results/ of the logging root directory

pix2pix on facades

Training

python pix2pix.py --dataroot <data_dir>/facades --logdir <output_path> --device cuda:0 --epochs 100 --bs 1 --mode train

• input: facades dataset
• output:
  • checkpoints/, logs/ and results/ in --logdir
  • pix2pix_loss.png

Testing

python pix2pix.py --dataroot <data_dir>/facades --logdir <output_path> --device cuda:0 --mode test

• input:
  • facades dataset
  • ogging root directory same as training --logdir
• output: test result images in results/ in --logdir

pix2pix on cityscapes

You only need to change the facades dataset to cityscapes dataset in --dataroot.

SAGAN on MNIST

python sagan.py \
    --batch_size 32 \
    --dataroot <mnist_save_path> \
    --epochs 200 \
    --logdir <output_pzth>

• output:
  • MNIST data saves in --dataroot
  • checkpoints/ and logs/ in --logdir
  • sagan_loss.png

GitHub Acknowledgement

We thank the authors of these repositories:

• christiancosgrove/pytorch-sagan