Awesome
Temporal Shift GAN for Large Scale Video Generation
This repository contains the official Pytorch implementation of Temporal Shift GAN for Large Scale Video Generation. Further details can be found here: https://arxiv.org/abs/2004.01823
To reproduce results download the pre-trained weights, then start training.
Pre-trained ImageNet weights: https://drive.google.com/open?id=1QvbsuhVB036wbtc4pKJRHhli0q_eOUGn
Usage
To use any dataset you should extract the frames of all videos and create a .txt file with three columns in this exact order: Location of the frames of the video, Number of frames in the folder and Class it belongs to.
Train
Run the script run_video_gan.sh.
Test
The script run create_dataset.sh will automatically create N videos of each class on the dataset.
Architecture
<img src="https://github.com/amunozgarza/mvt-gan/blob/master/demos/arch.png" width="625">Samples
MaisToy Real Samples (Click for samples)
Download the dataset right here: https://drive.google.com/file/d/15KqP9uxsJ9cMUY5WZyVvl2NlIQe-HQXz/view?usp=sharing
UCF-101 (Click for video)
20BN-Jester
Weizmann
