Get cache directory for LPIPS weights.

Get cleaning function by type.

Get input dtype based on precision string.

Get features from intermediate layers. Args: image: Input images, shape (B, C, H, W). n: Number of last layers to ret

Get features from the last layer of vision encoder. This returns the original high-dimensional features without bottleneck, which is

Get SSIM metric if torchmetrics is available.

Get tokenizer instance. Args: model_name: Model name (for compatibility, not used in simple implementation) context_length:

Prune layers not required for specified intermediates.

Scale learning rate based on batch size.

Setup linear classifiers with different configurations.

Setup logger that outputs to both console and file (rank 0 only).

Run linear probing evaluation. Args: model_path: Path to VTP HuggingFace model directory imagenet_root: Path to ImageNet dataset

Test ImageNet reconstruction. Args: model_path: Path to VTP HuggingFace model directory data_path: Path to ImageNet validation da

Train for one epoch.

Calculate LPIPS between two images. Args: img1, img2: Images in range [-1, 1] with shape (B, C, H, W)

Returns the tokenized representation of given input string(s). Parameters ---------- texts : Union[str, List[str]]

Get scheduled value for given iteration. Args: it (int): Current iteration Returns: float: Schedule

Initialize VTP Tokenizer. Args: hf_model_path: Path to HuggingFace VTPModel directory img_size: Input image size

Args: vtp_config: Pre-constructed configuration (takes priority). config_path: Path to load YAML from when vtp_config

Initialize cosine scheduler. Args: base_value (float): Initial value after warmup final_value (float): Final

Basic, remove whitespace, lower case.

Basic, remove whitespace.

Initialize weights following VTP conventions.

Convert 3D patch tokens (B, N, C) to 4D format (B, C, H, W). Args: patch_tokens: Patch tokens with shape (B, N, C) patch_siz

Base DINOv3-based pixel decoder

Large DINOv3-based pixel decoder

Small DINOv3-based pixel decoder

Encode images to latent features.

Fix random seeds.

Extract intermediate layer features. Returns: List of (patch_tokens, cls_token) tuples for each requested layer