Adversarial score matching and improved sampling for image generation

This repo contains the official implementation for the ICLR 2021 paper Adversarial score matching and improved sampling for image generation. It it a highly extended version of the original repo on score matching.

Discussion and more samples at https://ajolicoeur.wordpress.com/adversarial-score-matching-and-consistent-sampling.

---

Denoising score matching with Annealed Langevin Sampling (DSM-ALS) is a recent approach to generative modeling. Despite the convincing visual quality of samples, this method appears to perform worse than Generative Adversarial Networks (GANs) under the Frechet Inception Distance, a popular metric forgenerative models. We show that this apparent gap vanishes when denoising thefinal Langevin samples using the score network. In addition, we propose two improvements to DSM-ALS: 1) Consistent Annealed Sampling as a more stable alternative to Annealed Langevin Sampling, and 2) a hybrid training formulation, composed of both denoising score matching and adversarial objectives. By combining both of these techniques and exploring different network architectures, we elevate score matching methods and obtain results competitive with state-of-the-art image generation on CIFAR-10

---

Citation

If you find this code useful please cite us in your work:

@inproceedings{
  jolicoeur-martineau2021adversarial,
  title={Adversarial score matching and improved sampling for image generation},
  author={Alexia Jolicoeur-Martineau and R{\'e}mi Pich{\'e}-Taillefer and Ioannis Mitliagkas and Remi Tachet des Combes},
  booktitle={International Conference on Learning Representations},
  year={2021},
  url={https://openreview.net/forum?id=eLfqMl3z3lq}
}

Setup

Needed

• see requirements.txt (or do 'pip install numpy==1.16.0 lmdb torch torchvision jupyter matplotlib scipy tensorflow_gpu==2.1.0 tqdm PyYAML tensorboardX seaborn pillow setuptools==41.6.0 opencv-python')
• set your default directories in main.py and tests/training_sampling_fid.sh properly
• To use our pre-trained score betworks, download them (https://ln5.sync.com/dl/df5e312c0/krdqveya-prd8ur88-avv3sezg-x62ugj8f) and unzip to /exp
• To calculate the FID, download the FID statistics (https://ln5.sync.com/dl/ea2f51340/mqjixbhz-dkirwhic-cagmbzsw-5z2dnvd3) and unzip to /exp/datasets

Hyperparameter choice

• Use "main.py --compute_approximate_sigma_max" to choose model.sigma_begin based on the current dataset (based on Technique 1 from https://arxiv.org/abs/2006.09011)
• Use calculate_number_of_steps.R to choose model.num_classes (based on Technique 2 from https://arxiv.org/abs/2006.09011)
• tune sampling.step_lr manually for consistent or non-consistent with n_sigma=1 (see Appendix B for how to extrapolate to n_sigma > 1 from the step_lr at n_sigma = 1)
• Everything else can be left to default

To Replicate paper

• For images: run tests/training_sampling_fid.sh (Important: You should run each step separately, so save results and load them back as needed)
• For synthetic experiments: run the google colab

To train a non-adversarial score network

python main.py --config cifar10_9999ema.yml --doc cifar10_bs128_L2_9999ema --ni

Log files will be saved in <exp>/logs/cifar10_bs128_L2_9999ema.

To train an adversarial score network

python main.py --config cifar10_9999ema.yml --doc cifar10_bs128_L2_9999ema_adam0_9_adamD-5_9_LSGAN_ --ni  --adam --adam_beta 0 .9 --D_adam --D_adam_beta -.5 .9 --adversarial

Log files will be saved in <exp>/logs/cifar10_bs128_L2_9999ema_adam0_9_adamD-5_9_LSGAN_.

To sample from a pre-trained score network (ex: cifar10_bs128_L2_9999ema, consistent, nsigma=1)

python main.py --sample --config cifar10_9999ema.yml -i cifar10_bs128_L2_9999ema --ni --consistent --nsigma 1 --step_lr 5.6e-6 --batch_size 100 --begin_ckpt 250000

Samples will be saved in <exp>/image_samples/cifar10_bs128_L2_9999ema.

To compute the FID for a range of checkpoints from a pre-trained score network (ex: cifar10_bs128_L2_9999ema, at 100k to 300k iterations)

python main.py --fast_fid --config cifar10_9999ema.yml -i cifar10_bs128_L2_9999ema --ni --consistent --nsigma 1 --step_lr 5.6e-6 --batch_size 4000 --fid_num_samples 10000 --begin_ckpt 100000 --end_ckpt 300000

FIDs will be saved in {args.fid_folder}/log_FID.txt.

Pretrained Score Network Checkpoints

Link: https://ln5.sync.com/dl/df5e312c0/krdqveya-prd8ur88-avv3sezg-x62ugj8f

Download and unzip it to the exp folder.

FID statistics (for FID evaluation)

Link: https://ln5.sync.com/dl/ea2f51340/mqjixbhz-dkirwhic-cagmbzsw-5z2dnvd3

Download and unzip it to the exp/datasets folder.