Digan - Official PyTorch implementation of Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks

Last update: Dec 31, 2022

Overview

DIGAN (ICLR 2022)

Official PyTorch implementation of "Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks" by Sihyun Yu*, Jihoon Tack*, Sangwoo Mo*, Hyunsu Kim, Junho Kim, Jung-Woo Ha, Jinwoo Shin.

TL;DR: We make video generation scalable leveraging implicit neural representations.

Illustration of the (a) generator and (b) discriminator of DIGAN. The generator creates a video INR weight from random content and motion vectors, which produces an image that corresponds to the input 2D grids {(x, y)} and time t. Two discriminators determine the reality of each image and motion (from a pair of images and their time difference), respectively.

1. Environment setup

conda create -n digan python=3.8
conda activate digan

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html

pip install hydra-core==1.0.6
pip install tqdm scipy scikit-learn av ninja
pip install click gitpython requests psutil einops tensorboardX

2. Dataset

One should organize the video dataset as follows:

UCF-101

UCF-101
|-- train
    |-- class1
        |-- video1.avi
        |-- video2.avi
        |-- ...
    |-- class2
        |-- video1.avi
        |-- video2.avi
        |-- ...
    |-- ...

Dataset download

Link: UCF-101, Sky Time lapse, TaiChi-HD
For Kinetics-food dataset, read prepare_data/README.md

3. Training

To train the model, navigate to the project directory and run:

python src/infra/launch.py hydra.run.dir=. +experiment_name=<EXP_NAME> +dataset.name=<DATASET>

You may change training options via modifying configs/main.yml and configs/digan.yml.
Also the dataset list is as follows, <DATASET>: {UCF-101,sky,taichi,kinetics}

4. Evaluation (FVD and KVD)

python src/scripts/compute_fvd_kvd.py --network_pkl <MODEL_PATH> --data_path <DATA_PATH>

5. Video generation

Genrate and visualize videos (as gif and mp4):

python src/scripts/generate_videos.py --network_pkl <MODEL_PATH> --outdir <OUTPUT_PATH>

6. Results

Generated video results of DIGAN on TaiChi (top) and Sky (bottom) datasets.
More generated video results are available at the following site.

Citation

@inproceedings{
    yu2022generating,
    title={Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks},
    author={Yu, Sihyun and Tack, Jihoon and Mo, Sangwoo and Kim, Hyunsu and Kim, Junho and Ha, Jung-Woo and Shin, Jinwoo},
    booktitle={International Conference on Learning Representations},
    year={2022},
}

Reference

This code is mainly built upon StyleGAN2-ada and INR-GAN repositories.
We also used the code from following repositories: DiffAug, VideoGPT, MDGAN

Lisence

Copyright 2022-present NAVER Corp.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:

* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.

* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

Digan - Official PyTorch implementation of Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks

Related tags

Overview

DIGAN (ICLR 2022)

1. Environment setup

2. Dataset

UCF-101

Other video datasets (Sky Time lapse, TaiChi-HD, Kinetics-food)

Dataset download

3. Training

4. Evaluation (FVD and KVD)

5. Video generation

6. Results

Citation

Reference

Lisence

Owner

Sihyun Yu

BOVText: A Large-Scale, Multidimensional Multilingual Dataset for Video Text Spotting

AIR^2 for Interaction Prediction

We will release the code of "ConTNet: Why not use convolution and transformer at the same time?" in this repo

NAS Benchmark in "Prioritized Architecture Sampling with Monto-Carlo Tree Search", CVPR2021

Tutel MoE: An Optimized Mixture-of-Experts Implementation

TensorFlow implementation of the algorithm in the paper "Decoupled Low-light Image Enhancement"

Code for our ACL 2021 paper - ConSERT: A Contrastive Framework for Self-Supervised Sentence Representation Transfer

Attention Probe: Vision Transformer Distillation in the Wild

Predict bus arrival time using VertexAI and Nvidia's Jetson Nano

Reinforcement Learning for Portfolio Management

FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection

Train an imgs.ai model on your own dataset

IDRLnet, a Python toolbox for modeling and solving problems through Physics-Informed Neural Network (PINN) systematically.

Implementation of "Scaled-YOLOv4: Scaling Cross Stage Partial Network" using PyTorch framwork.

FastyAPI is a Stack boilerplate optimised for heavy loads.

This is the official pytorch implementation for the paper: Instance Similarity Learning for Unsupervised Feature Representation.

tmm_fast is a lightweight package to speed up optical planar multilayer thin-film device computation.

Official PyTorch implementation of N-ImageNet: Towards Robust, Fine-Grained Object Recognition with Event Cameras (ICCV 2021)

TrackTech: Real-time tracking of subjects and objects on multiple cameras

Repository for self-supervised landmark discovery