VideoGPT - Paper

This repo is to reproduce the results of the original paper. For a cleaner version that include pretrained models on BAIR, UCF-101, and Kinetics, please refer to this repo

Installation

Install the docker file in docker/Dockerfile. Commands below are run in a Docker container using Singularity. Once inside the docker environment, install the repo as an editable package using pip install -e VideoGPT-Paper.

Datasets

Note that you will need to edit the DATA_DIR variable in videogpt/dataset.py to correctly access you datasets.

We include functionality for two of the main datasets: BAIR Robot Pushing dataset and UCF-101. To install the BAIR dataset, run the script ./preprocess_bair/create_bair_dataset.sh. Then, create a bair_pushing directory in your datasets directory and copy the two .hdf5 files into the bair_pushing directory.

In order to install UCF-101, you can download the necessary files here. The code assumes a ucf101 directory with the following structure

ucf101/
    UCF-101/
        ApplyEyeMakeup/
            v1.avi
            ...
        ...
        YoYo/
            v1.avi
            ...
    ucfTrainTestlist

Training VQ-VQVAE

Execute python train_vqvae.py -h for all the tunable parameters. An example vq-vae training command is as follows

python train_vqvae.py -o <output_dir> --cfg vae_res64_ds422 --dataset bair_pushing

Training VideoGPT

Execute python train_videogpt.py -h for all the tunable parameters. An example VideoGPT training command is as follows

python train_videogpt.py -o <output_dir> --cfg gpt_small --vqvae_ckpt /path/to/vqvae/ckpt --dataset bair_pushing

You can add the --amp flag for automatic mixed precision training if you have the support hardware.

Misc

Other relevant scripts to compute FVD can be found in scripts/