Hugging Face's logo

Transformers documentation

V-JEPA 2

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

This model was published in HF papers on 2025-06-11 and contributed to Hugging Face Transformers on 2025-06-11.

SDPA FlashAttention

V-JEPA 2

V-JEPA 2 (blog post) is a self-supervised approach to training video encoders developed by FAIR, Meta. Using internet-scale video data, V-JEPA 2 attains state-of-the-art performance on motion understanding and human action anticipation tasks. V-JEPA 2-AC is a latent action-conditioned world model post-trained from V-JEPA 2 (using a small amount of robot trajectory interaction data) that solves robot manipulation tasks without environment-specific data collection or task-specific training or calibration.

drawing

You can find all original V-JEPA2 checkpoints under the V-JEPA 2 collection.

This model was contributed by koustuvs, yonigozlan and qubvel. The original code can be found here.

Usage example

The snippet below shows how to load the V-JEPA 2 model for feature extraction using the AutoModel class.

import numpy as np
from torchcodec.decoders import VideoDecoder


processor = AutoVideoProcessor.from_pretrained("facebook/vjepa2-vitl-fpc64-256")
model = AutoModel.from_pretrained(
    "facebook/vjepa2-vitl-fpc64-256",
    device_map="auto",
    attn_implementation="sdpa"
)

video_url = "https://huggingface.co/datasets/nateraw/kinetics-mini/resolve/main/val/archery/-Qz25rXdMjE_000014_000024.mp4"

vr = VideoDecoder(video_url)
frame_idx = np.arange(0, 64) # choosing some frames. here, you can define more complex sampling strategy
video = vr.get_frames_at(indices=frame_idx).data  # T x C x H x W
video = processor(video, return_tensors="pt").to(model.device)
outputs = model(**video)

# V-JEPA 2 encoder outputs, same as calling `model.get_vision_features()`
encoder_outputs = outputs.last_hidden_state

# V-JEPA 2 predictor outputs
predictor_outputs = outputs.predictor_output.last_hidden_state

V-JEPA 2 can also be finetuned for video classification. In the following snippet, we show how to use the finetuned on Something-Something-V2 video classification model.

import numpy as np
import torch
from torchcodec.decoders import VideoDecoder

from transformers import AutoModelForVideoClassification, AutoVideoProcessor


# Load model and video preprocessor
hf_repo = "facebook/vjepa2-vitl-fpc16-256-ssv2"

model = AutoModelForVideoClassification.from_pretrained(hf_repo, device_map="auto")
processor = AutoVideoProcessor.from_pretrained(hf_repo)

# To load a video, sample the number of frames according to the model.
video_url = "https://huggingface.co/datasets/nateraw/kinetics-mini/resolve/main/val/bowling/-WH-lxmGJVY_000005_000015.mp4"
vr = VideoDecoder(video_url)
frame_idx = np.arange(0, model.config.frames_per_clip, 8) # you can define more complex sampling strategy
video = vr.get_frames_at(indices=frame_idx).data  # frames x channels x height x width

# Preprocess and run inference
inputs = processor(video, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model(**inputs)
logits = outputs.logits

print("Top 5 predicted class names:")
top5_indices = logits.topk(5).indices[0]
top5_probs = torch.softmax(logits, dim=-1).topk(5).values[0]
for idx, prob in zip(top5_indices, top5_probs):
    text_label = model.config.id2label[idx.item()]
    print(f" - {text_label}: {prob:.2f}")

VJEPA2Config

class transformers.VJEPA2Config

< >

( transformers_version: str | None = Nonearchitectures: list[str] | None = Noneoutput_hidden_states: bool | None = Falsereturn_dict: bool | None = Truedtype: str | torch.dtype | None = Nonechunk_size_feed_forward: int = 0is_encoder_decoder: bool = Falseid2label: dict[int, str] | dict[str, str] | None = Nonelabel2id: dict[str, int] | dict[str, str] | None = Noneproblem_type: Literal['regression', 'single_label_classification', 'multi_label_classification'] | None = Nonepatch_size: int | list[int] | tuple[int, int] = 16crop_size: int = 256frames_per_clip: int = 64tubelet_size: int = 2hidden_size: int = 1024in_chans: int = 3num_attention_heads: int = 16num_hidden_layers: int = 24drop_path_rate: float | int = 0.0mlp_ratio: int | float = 4.0layer_norm_eps: float = 1e-06qkv_bias: bool = Trueattention_probs_dropout_prob: float | int = 0.0hidden_act: str = 'gelu'initializer_range: float = 0.02attention_dropout: float | int = 0.0num_pooler_layers: int = 3pred_hidden_size: int = 384pred_num_attention_heads: int = 12pred_num_hidden_layers: int = 12pred_num_mask_tokens: int = 10pred_zero_init_mask_tokens: bool = Truepred_mlp_ratio: int | float = 4.0 )

Parameters

  • patch_size (Union[int, list[int], tuple[int, int]], optional, defaults to 16) — The size (resolution) of each patch.
  • crop_size (int, optional, defaults to 256) — Input resolution of the model
  • frames_per_clip (int, optional, defaults to 64) — The number of frames the model has been pretrained with. Does not impact inference.
  • tubelet_size (int, optional, defaults to 2) — The number of temporal frames used for a single rastor, check paper for more information.
  • hidden_size (int, optional, defaults to 1024) — Dimension of the hidden representations.
  • in_chans (int, optional, defaults to 3) — The number of input channels.
  • num_attention_heads (int, optional, defaults to 16) — Number of attention heads for each attention layer in the Transformer decoder.
  • num_hidden_layers (int, optional, defaults to 24) — Number of hidden layers in the Transformer decoder.
  • drop_path_rate (Union[float, int], optional, defaults to 0.0) — Drop path rate for the patch fusion.
  • mlp_ratio (Union[int, float], optional, defaults to 4.0) — Ratio of the MLP hidden dim to the embedding dim.
  • layer_norm_eps (float, optional, defaults to 1e-06) — The epsilon used by the layer normalization layers.
  • qkv_bias (bool, optional, defaults to True) — Whether to add a bias to the queries, keys and values.
  • attention_probs_dropout_prob (Union[float, int], optional, defaults to 0.0) — The dropout ratio for the attention probabilities.
  • hidden_act (str, optional, defaults to gelu) — The non-linear activation function (function or string) in the decoder. For example, "gelu", "relu", "silu", etc.
  • initializer_range (float, optional, defaults to 0.02) — The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
  • attention_dropout (Union[float, int], optional, defaults to 0.0) — The dropout ratio for the attention probabilities.
  • num_pooler_layers (int, optional, defaults to 3) — The number of self-attention layers in the pooler.
  • pred_hidden_size (int, optional, defaults to 384) — Dimensionality of the predictor layers
  • pred_num_attention_heads (int, optional, defaults to 12) — Number of attention heads for each attention layer in the Predictor
  • pred_num_hidden_layers (int, optional, defaults to 12) — Number of hidden layers in the Predictor
  • pred_num_mask_tokens (int, optional, defaults to 10) — Define the number of mask tokens to use in the Predictor
  • pred_zero_init_mask_tokens (bool, optional, defaults to True) — Initialize the mask tokens in the predictor with 0.
  • pred_mlp_ratio (float, optional, defaults to 4.0) — Ratio of the hidden size of the MLPs used in Predictor relative to the pred_hidden_size.

This is the configuration class to store the configuration of a VJEPA2Model. It is used to instantiate a Vjepa2 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the facebook/vjepa2-vitl-fpc64-256

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

Example:

>>> from transformers import VJEPA2Config, VJEPA2Model

>>> # Initializing a VJEPA2 vjepa2-vitl-fpc64-256 style configuration
>>> configuration = VJEPA2Config()

>>> # Initializing a model (with random weights) from the vjepa2-vitl-fpc64-256  style configuration
>>> model = VJEPA2Model(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

VJEPA2Model

class transformers.VJEPA2Model

< >

( config: VJEPA2Config )

Parameters

  • config (VJEPA2Config) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The bare Vjepa2 Model outputting raw hidden-states without any specific head on top.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( pixel_values_videos: Tensorcontext_mask: list[torch.Tensor] | None = Nonetarget_mask: list[torch.Tensor] | None = Noneskip_predictor: bool = False**kwargs: *<class 'transformers.utils.generic.TransformersKwargs'> ) → VJEPA2WithMaskedInputModelOutput or tuple(torch.FloatTensor)

Parameters

  • pixel_values_videos (torch.Tensor of shape (batch_size, num_frames, num_channels, frame_size, frame_size)) — The tensors corresponding to the input video. Pixel values for videos can be obtained using VJEPA2VideoProcessor. See VJEPA2VideoProcessor.call() for details (processor_class uses VJEPA2VideoProcessor for processing videos).
  • context_mask (torch.Tensor with shape [batch_size, patch_size, 1], optional) — The mask position ids indicating which encoder output patches are going to be exposed to the predictor. By default, this mask is created as torch.arange(N).unsqueeze(0).repeat(B,1), indicating full context available to the predictor.
  • target_mask (torch.Tensor with shape [batch_size, patch_size, 1], optional) — The mask position ids indicating which encoder output patches are going to be used as a prediction target for the predictor. By default, this mask is created as torch.arange(N).unsqueeze(0).repeat(B,1), indicating that the predictor should predict all encoder patches.
  • skip_predictor (bool, optional, defaults to False) — flag to skip the predictor forward, useful if you just need the encoder outputs

Returns

VJEPA2WithMaskedInputModelOutput or tuple(torch.FloatTensor)

A VJEPA2WithMaskedInputModelOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (VJEPA2Config) and inputs.

The VJEPA2Model forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.

  • masked_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional, returned when context_mask is provided which is applied on VJEPA2Encoder outputs) — The masked hidden state of the model.

  • hidden_states (tuple[torch.FloatTensor, ...], optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

    Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.

  • attentions (tuple[torch.FloatTensor, ...], optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

    Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

  • predictor_output (VJEPA2WithMaskedInputPredictorOutput, optional) — The output from the Predictor module.

VJEPA2ForVideoClassification

class transformers.VJEPA2ForVideoClassification

< >

( config: VJEPA2Config )

Parameters

  • config (VJEPA2Config) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

V-JEPA 2 Model transformer with a video classification head on top (a linear layer on top of the attentive pooler).

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( pixel_values_videos: Tensorlabels: typing.Optional[torch.Tensor] = None**kwargs: *<class 'transformers.utils.generic.TransformersKwargs'> ) → ImageClassifierOutput or tuple(torch.FloatTensor)

Parameters

  • pixel_values_videos (torch.Tensor of shape (batch_size, num_frames, num_channels, frame_size, frame_size)) — The tensors corresponding to the input video. Pixel values for videos can be obtained using VJEPA2VideoProcessor. See VJEPA2VideoProcessor.call() for details (processor_class uses VJEPA2VideoProcessor for processing videos).
  • labels (torch.LongTensor of shape (batch_size,), optional) — Labels for computing the image classification/regression loss. Indices should be in [0, ..., config.num_labels - 1]. If config.num_labels == 1 a regression loss is computed (Mean-Square loss), If config.num_labels > 1 a classification loss is computed (Cross-Entropy).

Returns

ImageClassifierOutput or tuple(torch.FloatTensor)

A ImageClassifierOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (VJEPA2Config) and inputs.

The VJEPA2ForVideoClassification forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification (or regression if config.num_labels==1) loss.

  • logits (torch.FloatTensor of shape (batch_size, config.num_labels)) — Classification (or regression if config.num_labels==1) scores (before SoftMax).

  • hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each stage) of shape (batch_size, sequence_length, hidden_size). Hidden-states (also called feature maps) of the model at the output of each stage.

  • attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, patch_size, sequence_length).

    Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Examples:

>>> import torch
>>> import numpy as np
>>> from transformers import AutoVideoProcessor, VJEPA2ForVideoClassification

>>> device = "cuda"

>>> video_processor = AutoVideoProcessor.from_pretrained("facebook/vjepa2-vitl-fpc16-256-ssv2")
>>> model = VJEPA2ForVideoClassification.from_pretrained("facebook/vjepa2-vitl-fpc16-256-ssv2").to(device)

>>> video = np.ones((64, 256, 256, 3))  # 64 frames, 256x256 RGB
>>> inputs = video_processor(video, return_tensors="pt").to(device)

>>> # For inference
>>> with torch.no_grad():
...     outputs = model(**inputs)
>>> logits = outputs.logits

>>> predicted_label = logits.argmax(-1).item()
>>> print(model.config.id2label[predicted_label])

>>> # For training
>>> labels = torch.ones(1, dtype=torch.long, device=device)
>>> loss = model(**inputs, labels=labels).loss

VJEPA2VideoProcessor

class transformers.VJEPA2VideoProcessor

< >

( **kwargs: *<class 'transformers.processing_utils.VideosKwargs'> )

Parameters

  • do_convert_rgb (bool, kwargs, optional) — Whether to convert the image to RGB.
  • do_resize (bool, kwargs, optional, defaults to True) — Whether to resize the image.
  • size (Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, None], kwargs, defaults to {'shortest_edge' -- 292}): Describes the maximum input dimensions to the model.
  • default_to_square (bool, kwargs, optional, defaults to True) — Whether to default to a square image when resizing, if size is an int.
  • resample (Annotated[Union[int, PILImageResampling, NoneType], None], kwargs, defaults to 2) — Resampling filter to use if resizing the image. This can be one of the enum PILImageResampling. Only has an effect if do_resize is set to True.
  • do_rescale (bool, kwargs, optional, defaults to True) — Whether to rescale the image.
  • rescale_factor (float, kwargs, optional, defaults to 0.00392156862745098) — Rescale factor to rescale the image by if do_rescale is set to True.
  • do_normalize (bool, kwargs, optional, defaults to True) — Whether to normalize the image.
  • image_mean (Union[float, list[float], tuple[float, ...]], kwargs, optional, defaults to [0.485, 0.456, 0.406]) — Image mean to use for normalization. Only has an effect if do_normalize is set to True.
  • image_std (Union[float, list[float], tuple[float, ...]], kwargs, optional, defaults to [0.229, 0.224, 0.225]) — Image standard deviation to use for normalization. Only has an effect if do_normalize is set to True.
  • do_center_crop (bool, kwargs, optional, defaults to True) — Whether to center crop the image.
  • do_pad (bool, kwargs, optional) — Whether to pad the image. Padding is done either to the largest size in the batch or to a fixed square size per image. The exact padding strategy depends on the model.
  • crop_size (Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, None], kwargs, defaults to 256) — Size of the output image after applying center_crop.
  • data_format (Union[str, ~image_utils.ChannelDimension], kwargs, optional) — Only ChannelDimension.FIRST is supported. Added for compatibility with slow processors.
  • input_data_format (Union[str, ~image_utils.ChannelDimension], kwargs, optional) — The channel dimension format for the input image. If unset, the channel dimension format is inferred from the input image. Can be one of:

    • "channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format.
    • "channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format.
    • "none" or ChannelDimension.NONE: image in (height, width) format.
  • device (Annotated[Union[str, torch.device, NoneType], None], kwargs) — The device to process the videos on. If unset, the device is inferred from the input videos.
  • do_sample_frames (bool, kwargs, optional) — Whether to sample frames from the video before processing or to process the whole video.
  • video_metadata (Annotated[~video_utils.VideoMetadata | dict | list[dict | ~video_utils.VideoMetadata] | list[list[dict | ~video_utils.VideoMetadata]] | None, None], kwargs) — Metadata of the video containing information about total duration, fps and total number of frames. It will be used to sample frames from video or compute timestamps. Don’t pass any metadata unless you are trying to decode the video manually before processing
  • fps (Annotated[int | float | None, None], kwargs) — Target frames to sample per second when do_sample_frames=True.
  • num_frames (Annotated[int | None, None], kwargs) — Maximum number of frames to sample when do_sample_frames=True.
  • return_metadata (bool, kwargs, optional, defaults to False) — Whether to return video metadata or not. Video metadats is an object containing info about video duration, fps, decoding backend, etc.
  • return_tensors (Annotated[str | ~utils.generic.TensorType | None, None], kwargs) — If set, will return tensors of a particular framework. Acceptable values are:

    • 'pt': Return PyTorch torch.Tensor objects.
    • 'np': Return NumPy np.ndarray objects.
  • **kwargs (VideosKwargs, optional) — Additional processing options for each modality (text, images, videos, audio). Model-specific parameters are listed above; see the TypedDict class for the complete list of supported arguments.

Constructs a VJEPA2VideoProcessor video processor.

Update on GitHub

←TimeSformer VideoMAE→