The official PyTorch implementation for NCSNv2 (NeurIPS 2020)

Overview

Improved Techniques for Training Score-Based Generative Models

This repo contains the official implementation for the paper Improved Techniques for Training Score-Based Generative Models.

by Yang Song and Stefano Ermon, Stanford AI Lab.

Note: The method has been extended by the subsequent work Score-Based Generative Modeling through Stochastic Differential Equations (code) that allows better sample quality and exact log-likelihood computation.


We significantly improve the method proposed in Generative Modeling by Estimating Gradients of the Data Distribution. Score-based generative models are flexible neural networks trained to capture the score function of an underlying data distribution—a vector field pointing to directions where the data density increases most rapidly. We present new techniques to improve the performance of score-based generative models, scaling them to high resolution images that are previously impossible. Without requiring adversarial training, they can produce sharp and diverse image samples that rival GANs.

samples

(From left to right: Our samples on FFHQ 256px, LSUN bedroom 128px, LSUN tower 128px, LSUN church_outdoor 96px, and CelebA 64px.)

Running Experiments

Dependencies

Run the following to install all necessary python packages for our code.

pip install -r requirements.txt

Project structure

main.py is the file that you should run for both training and sampling. Execute python main.py --help to get its usage description:

usage: main.py [-h] --config CONFIG [--seed SEED] [--exp EXP] --doc DOC
               [--comment COMMENT] [--verbose VERBOSE] [--test] [--sample]
               [--fast_fid] [--resume_training] [-i IMAGE_FOLDER] [--ni]

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       Path to the config file
  --seed SEED           Random seed
  --exp EXP             Path for saving running related data.
  --doc DOC             A string for documentation purpose. Will be the name
                        of the log folder.
  --comment COMMENT     A string for experiment comment
  --verbose VERBOSE     Verbose level: info | debug | warning | critical
  --test                Whether to test the model
  --sample              Whether to produce samples from the model
  --fast_fid            Whether to do fast fid test
  --resume_training     Whether to resume training
  -i IMAGE_FOLDER, --image_folder IMAGE_FOLDER
                        The folder name of samples
  --ni                  No interaction. Suitable for Slurm Job launcher

Configuration files are in config/. You don't need to include the prefix config/ when specifying --config . All files generated when running the code is under the directory specified by --exp. They are structured as:

<exp> # a folder named by the argument `--exp` given to main.py
├── datasets # all dataset files
├── logs # contains checkpoints and samples produced during training
│   └── <doc> # a folder named by the argument `--doc` specified to main.py
│      ├── checkpoint_x.pth # the checkpoint file saved at the x-th training iteration
│      ├── config.yml # the configuration file for training this model
│      ├── stdout.txt # all outputs to the console during training
│      └── samples # all samples produced during training
├── fid_samples # contains all samples generated for fast fid computation
│   └── <i> # a folder named by the argument `-i` specified to main.py
│      └── ckpt_x # a folder of image samples generated from checkpoint_x.pth
├── image_samples # contains generated samples
│   └── <i>
│       └── image_grid_x.png # samples generated from checkpoint_x.pth       
└── tensorboard # tensorboard files for monitoring training
    └── <doc> # this is the log_dir of tensorboard

Training

For example, we can train an NCSNv2 on LSUN bedroom by running the following

python main.py --config bedroom.yml --doc bedroom

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

Sampling

If we want to sample from NCSNv2 on LSUN bedroom, we can edit bedroom.yml to specify the ckpt_id under the group sampling, and then run the following

python main.py --sample --config bedroom.yml -i bedroom

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

We can interpolate between different samples (see more details in the paper). Just set interpolation to true and an appropriate n_interpolations under the group of sampling in bedroom.yml. We can also perform other tasks such as inpainting. Usages should be quite obvious if you read the code and configuration files carefully.

Computing FID values quickly for a range of checkpoints

We can specify begin_ckpt and end_ckpt under the fast_fid group in the configuration file. For example, by running the following command, we can generate a small number of samples per checkpoint within the range begin_ckpt-end_ckpt for a quick (and rough) FID evaluation.

python main.py --fast_fid --config bedroom.yml -i bedroom

You can find samples in <exp>/fid_samples/bedroom.

Pretrained Checkpoints

Link: https://drive.google.com/drive/folders/1217uhIvLg9ZrYNKOR3XTRFSurt4miQrd?usp=sharing

You can produce samples using it on all datasets we tested in the paper. It assumes the --exp argument is set to exp.

References

If you find the code/idea useful for your research, please consider citing

@inproceedings{song2020improved,
  author    = {Yang Song and Stefano Ermon},
  editor    = {Hugo Larochelle and
               Marc'Aurelio Ranzato and
               Raia Hadsell and
               Maria{-}Florina Balcan and
               Hsuan{-}Tien Lin},
  title     = {Improved Techniques for Training Score-Based Generative Models},
  booktitle = {Advances in Neural Information Processing Systems 33: Annual Conference
               on Neural Information Processing Systems 2020, NeurIPS 2020, December
               6-12, 2020, virtual},
  year      = {2020}
}

and/or our previous work

@inproceedings{song2019generative,
  title={Generative Modeling by Estimating Gradients of the Data Distribution},
  author={Song, Yang and Ermon, Stefano},
  booktitle={Advances in Neural Information Processing Systems},
  pages={11895--11907},
  year={2019}
}
MatchGAN: A Self-supervised Semi-supervised Conditional Generative Adversarial Network

MatchGAN: A Self-supervised Semi-supervised Conditional Generative Adversarial Network This repository is the official implementation of MatchGAN: A S

Justin Sun 12 Dec 27, 2022
Third party Pytorch implement of Image Processing Transformer (Pre-Trained Image Processing Transformer arXiv:2012.00364v2)

ImageProcessingTransformer Third party Pytorch implement of Image Processing Transformer (Pre-Trained Image Processing Transformer arXiv:2012.00364v2)

61 Jan 01, 2023
[CVPR 2021] Few-shot 3D Point Cloud Semantic Segmentation

Few-shot 3D Point Cloud Semantic Segmentation Created by Na Zhao from National University of Singapore Introduction This repository contains the PyTor

117 Dec 27, 2022
Convert dog pictures into various painting styles. Try LimnPet

LimnPet Cartoon stylization service project Try our service » Home page · Team notion · Members 목차 프로젝트 소개 프로젝트 목표 사용한 기술스택과 수행도구 팀원 구현 기능 주요 기능 추가 기능

LiJell 7 Jul 14, 2022
GRaNDPapA: Generator of Rad Names from Decent Paper Acronyms

GRaNDPapA: Generator of Rad Names from Decent Paper Acronyms Trying to publish a new machine learning model and can't write a decent title for your pa

264 Nov 08, 2022
Key information extraction from invoice document with Graph Convolution Network

Key Information Extraction from Scanned Invoices Key information extraction from invoice document with Graph Convolution Network Related blog post fro

Phan Hoang 39 Dec 16, 2022
Official Code for VideoLT: Large-scale Long-tailed Video Recognition (ICCV 2021)

Pytorch Code for VideoLT [Website][Paper] Updates [10/29/2021] Features uploaded to Google Drive, for access please send us an e-mail: zhangxing18 at

Skye 26 Sep 18, 2022
Create images and texts with the First Order Generative Adversarial Networks

First Order Divergence for training GANs This repository contains code accompanying the paper First Order Generative Advesarial Netoworks The majority

Zalando Research 35 Dec 11, 2021
The official implementation of NeurIPS 2021 paper: Finding Optimal Tangent Points for Reducing Distortions of Hard-label Attacks

The official implementation of NeurIPS 2021 paper: Finding Optimal Tangent Points for Reducing Distortions of Hard-label Attacks

machen 11 Nov 27, 2022
The code for the NeurIPS 2021 paper "A Unified View of cGANs with and without Classifiers".

Energy-based Conditional Generative Adversarial Network (ECGAN) This is the code for the NeurIPS 2021 paper "A Unified View of cGANs with and without

sianchen 22 May 28, 2022
Pytorch Implementation of Various Point Transformers

Pytorch Implementation of Various Point Transformers Recently, various methods applied transformers to point clouds: PCT: Point Cloud Transformer (Men

Neil You 434 Dec 30, 2022
Pytorch implementation for M^3L

Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification (CVPR 2021) Introduction This is the Py

Yuyang Zhao 45 Dec 26, 2022
Code for DeepXML: A Deep Extreme Multi-Label Learning Framework Applied to Short Text Documents

DeepXML Code for DeepXML: A Deep Extreme Multi-Label Learning Framework Applied to Short Text Documents Architectures and algorithms DeepXML supports

Extreme Classification 49 Nov 06, 2022
Code of TVT: Transferable Vision Transformer for Unsupervised Domain Adaptation

TVT Code of TVT: Transferable Vision Transformer for Unsupervised Domain Adaptation Datasets: Digit: MNIST, SVHN, USPS Object: Office, Office-Home, Vi

37 Dec 15, 2022
vit for few-shot classification

Few-Shot ViT Requirements PyTorch (= 1.9) TorchVision timm (latest) einops tqdm numpy scikit-learn scipy argparse tensorboardx Pretrained Checkpoints

Martin Dong 26 Nov 30, 2022
Implicit Deep Adaptive Design (iDAD)

Implicit Deep Adaptive Design (iDAD) This code supports the NeurIPS paper 'Implicit Deep Adaptive Design: Policy-Based Experimental Design without Lik

Desi 12 Aug 14, 2022
Provide partial dates and retain the date precision through processing

Prefix date parser This is a helper class to parse dates with varied degrees of precision. For example, a data source might state a date as 2001, 2001

Friedrich Lindenberg 13 Dec 14, 2022
MultiMix: Sparingly Supervised, Extreme Multitask Learning From Medical Images (ISBI 2021, MELBA 2021)

MultiMix This repository contains the implementation of MultiMix. Our publications for this project are listed below: "MultiMix: Sparingly Supervised,

Ayaan Haque 27 Dec 22, 2022
3 Apr 20, 2022
DPT: Deformable Patch-based Transformer for Visual Recognition (ACM MM2021)

DPT This repo is the official implementation of DPT: Deformable Patch-based Transformer for Visual Recognition (ACM MM2021). We provide code and model

CASIA-IVA-Lab 111 Dec 21, 2022