V-JEPA: Video Joint Embedding Predictive Architecture

Official PyTorch codebase for the video joint-embedding predictive architecture, V-JEPA, a method for self-supervised learning of visual representations from video.

Meta AI Research, FAIR

Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, Nicolas Ballas

[Blog] [Paper] [Yannic Kilcher's Video]

V-JEPA models are trained by passively watching video pixels from the VideoMix2M dataset, and produce versatile visual representations that perform well on downstream video and image tasks, without adaption of the model’s parameters; e.g., using a frozen backbone and only a light-weight task-specific attentive probe.

Method

V-JEPA pretraining is based solely on an unsupervised feature prediction objective, and does not utilize pretrained image encoders, text, negative examples, human annotations, or pixel-level reconstruction.

     

Visualizations

As opposed to generative methods that have a pixel decoder, V-JEPA has a predictor that makes predictions in latent space. We train a conditional diffusion model to decode the V-JEPA feature-space predictions to interpretable pixels; the pretrained V-JEPA encoder and predictor networks are kept frozen in this process. The decoder is only fed the representations predicted for the missing regions of the video, and does not have access to the unmasked regions of the video.

The V-JEPA feature predictions are indeed grounded, and exhibit spatio-temporal consistency with the unmasked regions of the video.

MODEL ZOO

Pretrained models

model	patch size	resolution	iterations	batch size	data	download
ViT-L	2x16x16	224x224	90K	3072	VideoMix2M	checkpoint	configs
ViT-H	2x16x16	224x224	90K	3072	VideoMix2M	checkpoint	configs
ViT-H	2x16x16	384x384	90K	2400	VideoMix2M	checkpoint	configs

K400 Attentive probes

model	resolution	accuracy (16x8x3)	download
ViT-L/16	224x224	80.8	attentive probe checkpoint	configs
ViT-H/16	224x224	82.0	attentive probe checkpoint	configs
ViT-H/16	384x384	81.9	attentive probe checkpoint	configs

SSv2 Attentive probes

model	resolution	accuracy (16x2x3)	download
ViT-L/16	224x224	69.5	attentive probe checkpoint	configs
ViT-H/16	224x224	71.4	attentive probe checkpoint	configs
ViT-H/16	384x384	72.2	attentive probe checkpoint	configs

ImageNet1K Attentive probes

model	resolution	accuracy	download
ViT-L/16	224x224	74.8	attentive probe checkpoint	configs
ViT-H/16	224x224	75.9	attentive probe checkpoint	configs
ViT-H/16	384x384	77.4	attentive probe checkpoint	configs

Places205 Attentive probes

model	resolution	accuracy	download
ViT-L/16	224x224	60.3	attentive probe checkpoint	configs
ViT-H/16	224x224	61.7	attentive probe checkpoint	configs
ViT-H/16	384x384	62.8	attentive probe checkpoint	configs

iNat21 Attentive probes

model	resolution	accuracy	download
ViT-L/16	224x224	67.8	attentive probe checkpoint	configs
ViT-H/16	224x224	67.9	attentive probe checkpoint	configs
ViT-H/16	384x384	72.6	attentive probe checkpoint	configs

Code Structure

Config files: All experiment parameters are specified in config files (as opposed to command-line arguments). See the configs/ directory for example config files. Note, before launching an experiment, you must update the paths in the config file to point to your own directories, indicating where to save the logs and checkpoints and where to find the training data.

.
├── app                       # the only place where training loops are allowed
│   ├── vjepa                 #   Video JEPA pre-training
│   ├── main_distributed.py   #   entrypoint for launching app on slurm cluster
│   └── main.py               #   entrypoint for launching app locally on your machine for debugging
├── evals                     # the only place where evaluation of 'apps' are allowed
│   ├── image_classification  #   training an attentive probe for image classification with frozen backbone
│   ├── video_classification  #   training an attentive probe for video classification with frozen backbone
│   ├── main_distributed.py   #   entrypoint for launching distributed evaluations on slurm cluster
│   └── main.py               #   entrypoint for launching evaluations locally on your machine for debugging
├── src                       # the package
│   ├── datasets              #   datasets, data loaders, ...
│   ├── models                #   model definitions
│   ├── masks                 #   mask collators, masking utilities, ...
│   └── utils                 #   shared utilities
└── configs                   # the only place where config files are allowed (specify experiment params for app/eval runs)
    ├── evals                 #   configs for launching vjepa frozen evaluations
    └── pretrain              #   configs for launching vjepa pretraining

Data preparation

Video Datasets

V-JEPA pretraining and evaluations work with many standard video formats. To make a video dataset compatible with the V-JEPA codebase, you simply need to create a .csv file with the following format and then specify the path to this CSV file in your config.

/absolute_file_path.[mp4, webvid, etc.] $integer_class_label
/absolute_file_path.[mp4, webvid, etc.] $integer_class_label
/absolute_file_path.[mp4, webvid, etc.] $integer_class_label
...

Since V-JEPA is entirely unsupervised, the pretraining code will disregard the $integer_class_label in the CSV file. Thus, feel free to put a random value in this column. However, if you wish to run a supervised video classification evaluation on your video dataset, you must replace $integer_class_label with the ground truth label for each video.

Image Datasets

We use the standard PyTorch ImageFolder class in our image classification evals. Thus, to set up an image dataset for the image classification evaluation, first create a directory to store your image datasets $your_directory_containing_image_datasets. Next, download your image datasets into this directory in a format compatible with PyTorch ImageFolder.

For example, suppose we have a directory called my_image_datasets. We would then download our image datasets into this directory so that we end up with the following file tree ``` . └── /my_image_datasets/ # where we store image datasets ├── places205/121517/pytorch/ # Places205 │ └── [...] ├── iNaturalist-2021/110421/ # iNaturalist21 │ └── [...] ├── [...] # Other Image Datasets │ └── [...] └── imagenet_full_size/061417/ # ImageNet1k └── train │ ├── $class_1 │ │ ├── xxx.[png, jpeg, etc.] │ │ ├── [...] │ │ └── xxz.[png, jpeg, etc.] │ ├── [...] │ └── $class_n │ ├── abc.[png, jpeg, etc.] │ ├── [...] │ └── abz.[png, jpeg, etc.] └── val ├── $class_1 │ ├── xxx.[png, jpeg, etc.] │ ├── [...] │ └── xxz