GAN steerability

Project Page | Paper

On the "steerability" of generative adversarial networks.
Ali Jahanian*, Lucy Chai*, Phillip Isola

Prerequisites

• Linux
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Table of Contents:

1. Setup
   
2. Visualizations - plotting image panels, videos, and distributions
   
3. Training - pipeline for training your own walks
   
4. Notebooks - some jupyter notebooks, good place to start for trying your own transformations
   
5. PyTorch/Colab Demo - pytorch implementation in a colab notebook
   

Setup

• Clone this repo:

git clone https://github.com/ali-design/gan_steerability.git

• Install dependencies:
  • we provide a Conda environment.yml file listing the dependencies. You can create a Conda environment with the dependencies using:

conda env create -f environment.yml

• Download resources:
  • we provide a script for downloading associated resources (e.g. stylegan). Fetch these by running:

bash resources/download_resources.sh

Visualizations

Plotting image panels:

• Run python vis_image.py -h to list available visualization options. The key things to provide are a model checkpoint and a config yml file. We added some pretrained weights in the ./models_pretrained, but you can also use the models you train yourself.

python vis_image.py \
	models_pretrained/biggan_zoom_linear_lr0.0001_l2/model_20000_final.ckpt \
	models_pretrained/biggan_zoom_linear_lr0.0001_l2/opt.yml \
	--gpu 0 --num_samples 50 --noise_seed 20 --truncation 0.5 --category 207

python vis_image.py \
        models_pretrained/stylegan_color_linear_lr0.0001_l2_cats_w/model_2000_final.ckpt \
        models_pretrained/stylegan_color_linear_lr0.0001_l2_cats_w/opt.yml \
        --gpu 1 --num_samples 10 --noise_seed 20 

• By default this will save generated images to <output_dir>/images specified in the config yml, unless overwritten with the --output_dir option

To make a videos:

python vis_video.py [CHECKPOINT] [CONFIG] --gpu [GPU] --noise_seed [SEED] --sample [SAMPLE]

python vis_video.py models_pretrained/biggan_color_linear_lr0.001_l2/model_20000_final.ckpt \
	models_pretrained/biggan_color_linear_lr0.001_l2/opt.yml  --gpu 0 --sample 10 \
	--noise_seed 20 --truncation 0.5 --category 538 --min_alpha -1 --max_alpha 0

• By default this will save output to <output_dir>/video specified in the config yml, unless overwritten with the --output_dir and --filename options

To draw distributions:

To draw distributions, you will need to have downloaded the object detector through resources/download_resources.sh (for objects) or installed dlib through environment.yml (for faces).

python vis_distribution.py [CHECKPOINT] [CONFIG] --gpu [GPU]

python vis_distribution.py models_pretrained/biggan_shiftx_linear_lr0.001_l2/model_20000_final.ckpt \
	models_pretrained/biggan_shiftx_linear_lr0.001_l2/opt.yml  --gpu 0

• By default this will save output to <output_dir>/distribution specified in the config yml, unless overwritten with the --output_dir option

Training walks

• The current implementation covers these variants:
  • models: biggan, stylegan, pgan
  • transforms: color, colorlab, shiftx, shifty, zoom, rotate2d, rotate3d
  • walk_type: linear, NNz
  • losses: l2, lpips
• Some examples of commands for training walks:

# train a biggan NN walk for shiftx with lpips loss
python train.py --model biggan --transform shiftx --num_samples 20000 --learning_rate 0.0001 \
	--walk_type NNz --loss lpips --gpu 0 --eps 25 --num_steps 5

# train a stylegan linear walk with l2 loss using the w latent space
python train.py --model stylegan --transform color --num_samples 2000 --learning_rate 0.0001 \
	--walk_type linear --loss l2 --gpu 0 --latent w --model_save_freq 100

# train a pgan linear walk with l2 loss
python train.py --model pgan --transform color --num_samples 60000 --learning_rate 0.0001 \
	--walk_type linear --loss l2 --gpu 0 --dset celebahq --model_save_freq 1000

• Currently for pgan we only support the celebahq dataset, but you can easily train walks using other pretrained models by downloading the weights to resources/pgan_pretrained and adding the model information to graphs/pgan/constants.py

• Alternatively you can train using a config yml file, for example:

python train.py --config_file config/biggan_color_linear.yml

• Each training run will save a config file called opt.yml in the same directory as the weights, which can be used for rerunning experiments with the same settings. You may need to use the flag --overwrite_config for overwriting existing weights and config files.

• Run python train.py -h to list available training options

• You can use these walks that you train yourself for the above visualizations, by specifying the path to the corresponding model checkpoint and opt.yml config file.

Notebooks

• We provide some examples of jupyter notebooks illustrating the full training pipeline. See notebooks.
• It might be easiest to start here if you want to try your own transformations! The key things to modify are get_target_np and to scale alpha appropriately when feeding to graph.
• If using the provided conda environment, you'll need to add it to the jupyter kernel:

source activate gan_steerability
python -m ipykernel install --user --name gan_steerability

PyTorch

• You can find a google colab notebook implementation here. This uses a pytorch port of the tensorflow BigGAN-deep weights.
• An earlier version is on the pytorch BigGAN model is here.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{gansteerability,
  title={On the "steerability" of generative adversarial networks},
  author={Jahanian, Ali and Chai, Lucy and Isola, Phillip},
  booktitle={International Conference on Learning Representations},
  year={2020}
 }