Skip to content

Invert and perturb GAN images for test-time ensembling

License

Notifications You must be signed in to change notification settings

chail/gan-ensembling

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GAN Ensembling

Project Page | Paper | Bibtex

Ensembling with Deep Generative Views.
Lucy Chai, Jun-Yan Zhu, Eli Shechtman, Phillip Isola, Richard Zhang
CVPR 2021

Prerequisites

  • Linux
  • Python 3
  • NVIDIA GPU + CUDA CuDNN

Table of Contents:

  1. Colab - run a limited demo version without local installation
  2. Setup - download required resources
  3. Quickstart - short demonstration code snippet
  4. Notebooks - jupyter notebooks for visualization
  5. Pipeline - details on full pipeline

We project an input image into the latent space of a pre-trained GAN and perturb it slightly to obtain modifications of the input image. These alternative views from the GAN are ensembled at test-time, together with the original image, in a downstream classification task.

To synthesize deep generative views, we first align (Aligned Input) and reconstruct an image by finding the corresponding latent code in StyleGAN2 (GAN Reconstruction). We then investigate different approaches to produce image variations using the GAN, such as style-mixing on fine layers (Style-mix Fine), which predominantly changes color, or coarse layers (Style-mix Coarse), which changes pose.

Colab

This Colab Notebook demonstrates the basic latent code perturbation and classification procedure in a simplified setting on the aligned cat dataset.

  • Clone this repo:
git clone https://github.com/chail/gan-ensembling.git
cd gan-ensembling
  • Install dependencies:

    • we provide a Conda environment.yml file listing the dependencies. You can create the Conda environment using:
     conda env create -f environment.yml
  • Download resources:

    • We provide a script for downloading associated resources.
    • It will download precomputed latent codes (cat: 291M, car: 121M, celebahq: 1.8G, cifar10: 883M), a subset of trained models (592M), precomputed results (1.3G), and associated libraries.
    • Fetch the resources by running
     bash resources/download_resources.sh
    • Note, Optional: to run the StyleGAN ID-invert models, the models need to be downloaded separately. Follow the directions here to obtain styleganinv_ffhq256_encoder.pth and styleganinv_ffhq256_encoder.pth, and place them in models/pretrain
    • Note, Optional: the download script downloads a subset of the pretrained models for the demo notebook. For further experiments, the additional pretrained models (total 7.0G) can be downloaded here; it includes 40 binary face attribute classifiers, and classifiers trained on the different perturbation methods for the remaining datasets.
  • Download external datasets:

    • CelebA-HQ: Follow the instructions here to create the CelebA-HQ dataset and place CelebA-HQ images in directory dataset/celebahq/images/images.
    • Cars: This dataset is a subset of Cars196. Download the images from here and the devkit from here. (We are subsetting their training images into train/test/val partitions). Place the images in directory dataset/cars/images/images and the devkit in dataset/cars/devkit.
    • The processed and aligned cat images are downloaded with the above resources, and cifar10 dataset is downloaded via the PyTorch wrapper.

An example of the directory organization is below:

dataset/celebahq/
	images/images/
		000004.png
		000009.png
		000014.png
		...
	latents/
	latents_idinvert/
dataset/cars/
	devkit/
		cars_meta.mat
		cars_test_annos.mat
		cars_train_annos.mat
		...
	images/images/
		00001.jpg
		00002.jpg
		00003.jpg
		...
	latents/
dataset/catface/
	images/
	latents/
dataset/cifar10/
	cifar-10-batches-py/
	latents/

Once the datasets and precomputed resources are downloaded, the following code snippet demonstrates how to perturb GAN images. Additional examples are contained in notebooks/demo.ipynb.

import data
from networks import domain_generator

dataset_name = 'celebahq'
generator_name = 'stylegan2'
attribute_name = 'Smiling'
val_transform = data.get_transform(dataset_name, 'imval')
dset = data.get_dataset(dataset_name, 'val', attribute_name, load_w=True, transform=val_transform)
generator = domain_generator.define_generator(generator_name, dataset_name)

index = 100
original_image = dset[index][0][None].cuda()
latent = dset[index][1][None].cuda()
gan_reconstruction = generator.decode(latent)
mix_latent = generator.seed2w(n=4, seed=0)
perturbed_im = generator.perturb_stylemix(latent, 'fine', mix_latent, n=4)

Important: First, set up symlinks required for notebooks: bash notebooks/setup_notebooks.sh, and add the conda environment to jupyter kernels: python -m ipykernel install --user --name gan-ensembling.

The provided notebooks are:

  1. notebooks/demo.ipynb: basic usage example
  2. notebooks/evaluate_ensemble.ipynb: plot classification test accuracy as a function of ensemble weight
  3. notebooks/plot_precomputed_evaluations.ipynb: notebook to generate figures in paper

The full pipeline contains three main parts:

  1. optimize latent codes
  2. train classifiers
  3. evaluate the ensemble of GAN-generated images.

Examples for each step of the pipeline are contained in the following scripts:

bash scripts/optimize_latent/examples.sh
bash scripts/train_classifier/examples.sh
bash scripts/eval_ensemble/examples.sh

To add to the pipeline:

  • Data: in the data/ directory, add the dataset in data/__init__.py and create the dataset class and transformation functions. See data/data_*.py for examples.
  • Generator: modify networks/domain_generators.py to add the generator in domain_generators.define_generator. The perturbation ranges for each dataset and generator are specified in networks/perturb_settings.py.
  • Classifier: modify networks/domain_classifiers.py to add the classifier in domain_classifiers.define_classifier

Acknowledgements

We thank the authors of these repositories:

Citation

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

@inproceedings{chai2021ensembling,
  title={Ensembling with Deep Generative Views.},
  author={Chai, Lucy and Zhu, Jun-Yan and Shechtman, Eli and Isola, Phillip and Zhang, Richard},
  booktitle={CVPR},
  year={2021}
 }