Forward pass with flexible layer control. Args: x: Input tensor no_last_layer: If True, skip the final projection lay

Unified forward with multiple modes. Args: image: Input images, shape (B, C, H, W). text: Input token IDs, shape (B,

Converts all `BatchNorm2d` and `SyncBatchNorm` layers of provided module into `FrozenBatchNorm2d`. If `module` is itself an instance of eithe

Create VTPConfig from legacy VTP YAML configuration file. Args: yaml_path: Path to the YAML configuration file. Returns:

Generate 2D sincos position embedding. Args: embed_dim: embedding dimension grid_size: int of the grid height and width

Compute CLIP similarity logits. Args: image: Input images, shape (B_img, C, H, W). text: Input token IDs, shape (B_tx

Retrieves intermediate layer features from the model, bypassing the feature bottleneck. This method directly calls the parent class's

Reinitialize weights.

Interpolate position embeddings for high-resolution.

Prune layers not required for specified intermediates.

Prune layers not required for specified intermediates.

Unwraps a model from its DataParallel wrapper if it is wrapped.

Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64