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

Bridging Composite and Real: Towards End-to-end Deep Image Matting

《Towards High Fidelity Face Relighting with Realistic Shadows》(CVPR 2021)

Trading Gym is an open source project for the development of reinforcement learning algorithms in the context of trading.

PyVideoAI: Action Recognition Framework

Speech recognition tool to convert audio to text transcripts, for Linux and Raspberry Pi.

Hardware-accelerated DNN model inference ROS2 packages using NVIDIA Triton/TensorRT for both Jetson and x86_64 with CUDA-capable GPU

An official implementation of the paper Exploring Sequence Feature Alignment for Domain Adaptive Detection Transformers

[CVPR-2021] UnrealPerson: An adaptive pipeline for costless person re-identification

Conformer: Local Features Coupling Global Representations for Visual Recognition

A PyTorch Implementation of ViT (Vision Transformer)

An end-to-end project on customer segmentation

Joint detection and tracking model named DEFT, or ``Detection Embeddings for Tracking.

A self-supervised learning framework for audio-visual speech

Segmentation-Aware Convolutional Networks Using Local Attention Masks

A curated list and survey of awesome Vision Transformers.

DA2Lite is an automated model compression toolkit for PyTorch.

Synthetic Scene Text from 3D Engines

StocksMA is a package to facilitate access to financial and economic data of Moroccan stocks.

PyTorch implementation of the WarpedGANSpace: Finding non-linear RBF paths in GAN latent space (ICCV 2021)

Affine / perspective transformation in Pose Estimation with Tensorflow 2