AdversarialConsistentScoreMatching - GitLite

KO | EN

gitlite — search

#javascript #hacktoberfest #react #ai #typescript #llm #golang #android #linux #java #kubernetes #vue #pytorch #open-source

AdversarialConsistentScoreMatching

(★ 126)

Code for paper "Adversarial score matching and improved sampling for image generation"

파일 탐색기

최종 버전 다운로드 (.zip)

- __init__.py
- celeba.yml
- church.yml
- cifar10_9999ema.yml
- cifar10_unet_9999_1e-4.yml
- default.yml
- ffhq.yml
- stackedmnist.yml
- stackedmnist_ema9999.yml
- tower.yml
- __init__.py
- celeba.py
- ffhq.py
- stackedmnist.py
- utils.py
- vision.py
- fid_score.py
- inception-2015-12-05.tgz
- inception.py
- inception_score.py
- mnist_cnn.pt
- nearest_neighbor.py
- pt_inception-2015-12-05-6726825d.pth
- train_mnist_classifier.py
- - c_adv_n1_5k.npy
  - c_adv_n5_5k.npy
  - nc_adv_n1_5k.npy
  - nc_adv_n5_5k.npy
- - c_adv_n15_5k.npy
  - c_adv_n1_5k.npy
  - c_adv_n5_1k.npy
  - c_adv_n5_5k.npy
  - nc_adv_n1_5k.npy
  - nc_nadv_n1_5k.npy
- - c_adv_n1_2k.npy
  - c_adv_n5_2k.npy
  - nc_adv_n1_2k.npy
  - nc_adv_n5_2k.npy
- README.txt
- __init__.py
- adversarial.py
- dsm.py
- ssim.py
- __init__.py
- ema.py
- GAN_D.py
- inception.py
- layers.py
- layers_biggan.py
- models_biggan.py
- ncsnv2.py
- normalization.py
- UNet.py
- __init__.py
- abc_runner.py
- eval_runner.py
- fid_runner.py
- inception_runner.py
- sample_runner.py
- stackedmnist_runner.py
- train_runner.py
- training_sampling_fid.sh
.gitignore
calculate_number_of_steps.R
Clean_Basic_code_GAN_with_DSM.ipynb
LICENSE
main.py
README.md
requirements.txt

# 설치 가이드

1. 코드 내려받기

git clone https://github.com/AlexiaJM/AdversarialConsistentScoreMatching

깃허브에서 프로젝트 코드 전체를 내 컴퓨터로 내려받습니다.

cd AdversarialConsistentScoreMatching

방금 내려받은 프로젝트 폴더 안으로 이동합니다.

2. Python

쉬움 추천

사전 준비물

Git GitHub에서 프로젝트 코드를 내려받으려면 필요합니다.
Python 3 설치 시 'Add Python to PATH' 옵션을 꼭 체크하세요 (Windows).

pip install -r requirements.txt

requirements.txt 등에 명시된 파이썬 라이브러리를 설치합니다.

python main.py

파이썬 스크립트(또는 모듈)를 실행합니다.

✅ 에러 메시지 없이 실행되고 터미널에 안내 문구가 출력되면 정상입니다.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 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](https://arxiv.org/abs/2009.05475). It it a highly extended version of [the original repo on score matching](https://github.com/ermongroup/ncsnv2).

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

-----------------------------------------------------------------------------------------

[Denoising score matching with Annealed Langevin Sampling (DSM-ALS)](https://arxiv.org/abs/2006.09011) 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

![Adversarial LSUN-Churches](https://ajolicoeur.files.wordpress.com/2020/09/image.png?w=662)

-----------------------------------------------------------------------------------------

## 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](https://github.com/AlexiaJM/AdversarialConsistentScoreMatching/blob/master/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](https://github.com/AlexiaJM/AdversarialConsistentScoreMatching/blob/master/Clean_Basic_code_GAN_with_DSM.ipynb)

## To train a non-adversarial score network

```bash
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

```bash
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)

```bash
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)

```bash
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.