[Paper][Website][Colab] Integrated to Huggingface Spaces with Gradio. See demo:
We present VideoGPT: a conceptually simple architecture for scaling likelihood based generative modeling to natural videos. VideoGPT uses VQ-VAE that learns downsampled discrete latent representations of a raw video by employing 3D convolutions and axial self-attention. A simple GPT-like architecture is then used to autoregressively model the discrete latents using spatio-temporal position encodings. Despite the simplicity in formulation and ease of training, our architecture is able to generate samples competitive with state-of-the-art GAN models for video generation on the BAIR Robot dataset, and generate high fidelity natural images from UCF-101 and Tumbler GIF Dataset (TGIF). We hope our proposed architecture serves as a reproducible reference for a minimalistic implementation of transformer based video generation models.
Change the cudatoolkit
version compatible to your machine.
conda install --yes -c conda-forge cudatoolkit=11.7 cudnn
pip install -r requirements.txt
export PYTHONPATH=$PYTHONPATH:.
For limited compute scenarios, it may be beneficial to use sparse attention.
sudo apt-get install llvm-9-dev
DS_BUILD_SPARSE_ATTN=1 pip install deepspeed
After installng deepspeed
, you can train a sparse transformer by setting the flag --attn_type sparse
in scripts/train_videogpt.py
. The default supported sparsity configuration is an N-d strided sparsity layout, however, you can write your own arbitrary layouts to use.
The default code accepts data as an HDF5 file with the specified format in videogpt/data.py
. An example of such a dataset can be constructed from the BAIR Robot data by running the script:
sh scripts/preprocess/bair/create_bair_dataset.sh datasets/bair
Alternatively, the code supports a dataset with the following directory structure:
video_dataset/
train/
class_0/
video1.mp4
video2.mp4
...
class_1/
video1.mp4
...
...
class_n/
...
test/
class_0/
video1.mp4
video2.mp4
...
class_1/
video1.mp4
...
...
class_n/
...
An example of such a dataset can be constructed from UCF-101 data by running the script
sh scripts/preprocess/ucf101/create_ucf_dataset.sh datasets/ucf101
You may need to install unrar
and unzip
for the code to work correctly.
If you do not care about classes, the class folders are not necessary and the dataset file structure can be collapsed into train
and test
directories of just videos.
There are four available pre-trained VQ-VAE models. All strides listed with each model are downsampling amounts across THW for the encoders.
bair_stride4x2x2
: trained on 16 frame 64 x 64 videos from the BAIR Robot Pushing datasetucf101_stride4x4x4
: trained on 16 frame 128 x 128 videos from UCF-101kinetics_stride4x4x4
: trained on 16 frame 128 x 128 videos from Kinetics-600kinetics_stride2x4x4
: trained on 16 frame 128 x 128 videos from Kinetics-600, with 2x larger temporal latent codes (achieves slightly better reconstruction)
from torchvision.io import read_video
from videogpt import load_vqvae
from videogpt.data import preprocess
video_filename = 'path/to/video_file.mp4'
sequence_length = 16
resolution = 128
device = torch.device('cuda')
vqvae = load_vqvae('kinetics_stride2x4x4')
video = read_video(video_filename, pts_unit='sec')[0]
video = preprocess(video, resolution, sequence_length).unsqueeze(0).to(device)
encodings = vqvae.encode(video)
video_recon = vqvae.decode(encodings)
Use the scripts/train_vqvae.py
script to train a VQ-VAE. Execute python scripts/train_vqvae.py -h
for information on all available training settings. A subset of more relevant settings are listed below, along with default values.
--embedding_dim
: number of dimensions for codebooks embeddings--n_codes 2048
: number of codes in the codebook--n_hiddens 240
: number of hidden features in the residual blocks--n_res_layers 4
: number of residual blocks--downsample 4 4 4
: T H W downsampling stride of the encoder
--gpus 2
: number of gpus for distributed training--sync_batchnorm
: usesSyncBatchNorm
instead ofBatchNorm3d
when using > 1 gpu--gradient_clip_val 1
: gradient clipping threshold for training--batch_size 16
: batch size per gpu--num_workers 8
: number of workers for each DataLoader
--data_path <path>
: path to anhdf5
file or a folder containingtrain
andtest
folders with subdirectories of videos--resolution 128
: spatial resolution to train on--sequence_length 16
: temporal resolution, or video clip length
There are two available pre-trained VideoGPT models
bair_gpt
: single frame-conditional BAIR model using discrete encodings frombair_stride4x2x2
VQ-VAEucf101_uncond_gpt
: unconditional UCF101 model using discrete encodings fromucf101_stride4x4x4
VQ-VAE Note that both pre-trained models use sparse attention. For purposes of fine-tuning, you will need to install sparse attention, however, sampling does not required sparse attention to be installed.
You can download a pretrained VQ-VAE, or train your own. Afterwards, use the scripts/train_videogpt.py
script to train an VideoGPT model for sampling. Execute python scripts/train_videogpt.py -h
for information on all available training settings. A subset of more relevant settings are listed below, along with default values.
--vqvae kinetics_stride4x4x4
: path to a vqvae checkpoint file, OR a pretrained model name to download. Available pretrained models are:bair_stride4x2x2
,ucf101_stride4x4x4
,kinetics_stride4x4x4
,kinetics_stride2x4x4
. BAIR was trained on 64 x 64 videos, and the rest on 128 x 128 videos--n_cond_frames 0
: number of frames to condition on.0
represents a non-frame conditioned model--class_cond
: trains a class conditional model if activated--hidden_dim 576
: number of transformer hidden features--heads 4
: number of heads for multihead attention--layers 8
: number of transformer layers--dropout 0.2'
: dropout probability applied to features after attention and positionwise feedforward layers--attn_type full
:full
orsparse
attention. Refer to the Installation section for install sparse attention--attn_dropout 0.3
: dropout probability applied to the attention weight matrix
--gpus 4
: number of gpus for distributed training--gradient_clip_val 1
: gradient clipping threshold for training--batch_size 8
: batch size per gpu--num_workers 2
: number of workers for each DataLoader--amp_level O1
: for mixed precision training--precision 16
: for mixed precision training
--data_path <path>
: path to anhdf5
file or a folder containingtrain
andtest
folders with subdirectories of videos--resolution 128
: spatial resolution to train on--sequence_length 16
: temporal resolution, or video clip length
VideoGPT models can be sampled using the scripts/sample_videogpt.py
. You can specify a path to a checkpoint during training, or the name of a pretrained model. You may need to install ffmpeg
: sudo apt-get install ffmpeg
Evaluation is done primarily using Frechet Video Distance (FVD) for BAIR and Kinetics, and Inception Score for UCF-101. Inception Score can be computed by generating samples and using the code from the TGANv2 repo. FVD can be computed through python scripts/compute_fvd.py
, which runs a PyTorch-ported version of the original codebase
Note that this repo is primarily designed for simplicity and extending off of our method. Reproducing the full paper results can be done using code found at a separate repo. However, be aware that the code is not as clean.
Please consider using the follow citation when using our code:
@misc{yan2021videogpt,
title={VideoGPT: Video Generation using VQ-VAE and Transformers},
author={Wilson Yan and Yunzhi Zhang and Pieter Abbeel and Aravind Srinivas},
year={2021},
eprint={2104.10157},
archivePrefix={arXiv},
primaryClass={cs.CV}
}