data

VJEPA2Config

extendsModelConfig

VJEPA2Config(image_size: int = 256, patch_size: int = 16, tubelet_size: int = 2, num_frames: int = 64, in_channels: int = 3, num_classes: int = 400, dim: int = 1024, depth: int = 24, num_heads: int = 16, mlp_ratio: float = 4.0, layer_norm_eps: float = 1e-06, qkv_bias: bool = True, init_std: float = 0.02, use_silu: bool = False, wide_silu: bool = True, use_rope: bool = True, rope_base: float = 10000.0, uniform_power: bool = False, drop_rate: float = 0.0, attn_drop_rate: float = 0.0, drop_path_rate: float = 0.0, predictor_dim: int = 384, predictor_out_dim: int | None = None, predictor_depth: int = 12, predictor_heads: int = 12, predictor_mlp_ratio: float = 4.0, predictor_num_mask_tokens: int = 10, zero_init_mask_tokens: bool = True, predictor_return_all_tokens: bool = False, num_pooler_layers: int = 3, normalize_targets: bool = True, objective: str = 'l1', ema: tuple[float, float] = (0.99925, 0.99925), ema_schedule_scale: float = 1.25)

source edit

Frozen architecture configuration for V-JEPA 2.

The defaults are the released ViT-L/16 checkpoint geometry. The public factories select the other paper variants by overriding dim, depth and num_heads; callers may still use a small explicit override for tests or experiments.

Parameters

image_sizeint= 256

Height and width of each video frame.

patch_sizeint= 16

Spatial side of one patch.

tubelet_sizeint= 2

Number of adjacent frames in one token.

num_framesint= 64

Clip length used by the released pretraining configuration.

dimint= 1024, 24, 16

Width, block count and head count of the video encoder.

depthint= 1024, 24, 16

Width, block count and head count of the video encoder.

num_headsint= 1024, 24, 16

Width, block count and head count of the video encoder.

use_ropebool= True

Use the released three-axis rotary attention geometry.

predictor_dimint= 384

Width of the latent predictor.

predictor_depthint= 12

Number of predictor blocks.

predictor_headsint= 12

Predictor attention heads.

predictor_mlp_ratiofloat= 4.0

Predictor feed-forward expansion. The ViT-g encoder ratio is wider than its predictor ratio in the released checkpoint.

predictor_num_mask_tokensint= 10

Learned mask-token slots used by the official mask generator.

num_pooler_layersint= 3

Self-attention blocks the attentive probe runs before its query attends. Every released classifier checkpoint carries three.

ematuple of float= (0.99925, 0.99925)

Target-encoder momentum at the start and end of the schedule. The released pretraining configuration holds it constant, which is where V-JEPA 2 departs from V-JEPA 1's ramp.

ema_schedule_scalefloat= 1.25

Stretches the momentum horizon past the run's own length, as the released ipe_scale does.

Used by 3

Constructors

dunder

init

→None

__init__(image_size: int = 256, patch_size: int = 16, tubelet_size: int = 2, num_frames: int = 64, in_channels: int = 3, num_classes: int = 400, dim: int = 1024, depth: int = 24, num_heads: int = 16, mlp_ratio: float = 4.0, layer_norm_eps: float = 1e-06, qkv_bias: bool = True, init_std: float = 0.02, use_silu: bool = False, wide_silu: bool = True, use_rope: bool = True, rope_base: float = 10000.0, uniform_power: bool = False, drop_rate: float = 0.0, attn_drop_rate: float = 0.0, drop_path_rate: float = 0.0, predictor_dim: int = 384, predictor_out_dim: int | None = None, predictor_depth: int = 12, predictor_heads: int = 12, predictor_mlp_ratio: float = 4.0, predictor_num_mask_tokens: int = 10, zero_init_mask_tokens: bool = True, predictor_return_all_tokens: bool = False, num_pooler_layers: int = 3, normalize_targets: bool = True, objective: str = 'l1', ema: tuple[float, float] = (0.99925, 0.99925), ema_schedule_scale: float = 1.25)

source edit

Properties

prop

num_tokens

→int

num_tokens: int

source edit

Number of tokens in a full configured clip.

prop

predictor_output_dim

→int

predictor_output_dim: int

source edit

Target width emitted by the predictor.

prop

token_grid

→tuple[int, int, int]

token_grid: tuple[int, int, int]

source edit

Number of tubelet positions along time, height and width.

data

VJEPA2Config

extendsModelConfig

VJEPA2Config(image_size: int = 256, patch_size: int = 16, tubelet_size: int = 2, num_frames: int = 64, in_channels: int = 3, num_classes: int = 400, dim: int = 1024, depth: int = 24, num_heads: int = 16, mlp_ratio: float = 4.0, layer_norm_eps: float = 1e-06, qkv_bias: bool = True, init_std: float = 0.02, use_silu: bool = False, wide_silu: bool = True, use_rope: bool = True, rope_base: float = 10000.0, uniform_power: bool = False, drop_rate: float = 0.0, attn_drop_rate: float = 0.0, drop_path_rate: float = 0.0, predictor_dim: int = 384, predictor_out_dim: int | None = None, predictor_depth: int = 12, predictor_heads: int = 12, predictor_mlp_ratio: float = 4.0, predictor_num_mask_tokens: int = 10, zero_init_mask_tokens: bool = True, predictor_return_all_tokens: bool = False, num_pooler_layers: int = 3, normalize_targets: bool = True, objective: str = 'l1', ema: tuple[float, float] = (0.99925, 0.99925), ema_schedule_scale: float = 1.25)

source edit

Frozen architecture configuration for V-JEPA 2.

Parameters

image_sizeint= 256

Height and width of each video frame.

patch_sizeint= 16

Spatial side of one patch.

tubelet_sizeint= 2

Number of adjacent frames in one token.

num_framesint= 64

Clip length used by the released pretraining configuration.

dimint= 1024, 24, 16

Width, block count and head count of the video encoder.

depthint= 1024, 24, 16

Width, block count and head count of the video encoder.

num_headsint= 1024, 24, 16

Width, block count and head count of the video encoder.

use_ropebool= True

Use the released three-axis rotary attention geometry.

predictor_dimint= 384

Width of the latent predictor.

predictor_depthint= 12

Number of predictor blocks.

predictor_headsint= 12

Predictor attention heads.

predictor_mlp_ratiofloat= 4.0

Predictor feed-forward expansion. The ViT-g encoder ratio is wider than its predictor ratio in the released checkpoint.

predictor_num_mask_tokensint= 10

Learned mask-token slots used by the official mask generator.

num_pooler_layersint= 3

Self-attention blocks the attentive probe runs before its query attends. Every released classifier checkpoint carries three.

ematuple of float= (0.99925, 0.99925)

Target-encoder momentum at the start and end of the schedule. The released pretraining configuration holds it constant, which is where V-JEPA 2 departs from V-JEPA 1's ramp.

ema_schedule_scalefloat= 1.25

Stretches the momentum horizon past the run's own length, as the released ipe_scale does.

Used by 3

Constructors

dunder

init

→None

__init__(image_size: int = 256, patch_size: int = 16, tubelet_size: int = 2, num_frames: int = 64, in_channels: int = 3, num_classes: int = 400, dim: int = 1024, depth: int = 24, num_heads: int = 16, mlp_ratio: float = 4.0, layer_norm_eps: float = 1e-06, qkv_bias: bool = True, init_std: float = 0.02, use_silu: bool = False, wide_silu: bool = True, use_rope: bool = True, rope_base: float = 10000.0, uniform_power: bool = False, drop_rate: float = 0.0, attn_drop_rate: float = 0.0, drop_path_rate: float = 0.0, predictor_dim: int = 384, predictor_out_dim: int | None = None, predictor_depth: int = 12, predictor_heads: int = 12, predictor_mlp_ratio: float = 4.0, predictor_num_mask_tokens: int = 10, zero_init_mask_tokens: bool = True, predictor_return_all_tokens: bool = False, num_pooler_layers: int = 3, normalize_targets: bool = True, objective: str = 'l1', ema: tuple[float, float] = (0.99925, 0.99925), ema_schedule_scale: float = 1.25)