Determine the absolute feature indices to 'take' from. Note: This function can be called in forward() so must be torchscript compatible, whi

Global sampling allocation aligned with dinov2: - Under DDP, compute global keep samples and distribute evenly across ranks - Return loca

Calculate FID between two image folders.

Get CLIP-projected image features for zero-shot tasks. Args: image: Input images, shape (B, C, H, W). normalize: Whet

Get CLIP-projected text features for zero-shot tasks. Args: text: Input token IDs, shape (B, L). normalize: Whether t

Decode latents to reconstructed images. Args: latents: Latent features, shape (B, C, H', W'). Returns: Recon

Get bottleneck latents for reconstruction. Args: image: Input images, shape (B, C, H, W). Returns: Latents i

Basic text cleaning.

Returns list of utf-8 byte and a corresponding list of unicode strings. The reversible bpe codes work on unicode strings. This means you

Center cropping implementation from ADM. https://github.com/openai/guided-diffusion/blob/8fb3ad9197f16bbc40620447b2742e13458d2831/guided_diff

Count the number of image files in a directory.

Create input for linear classifier from intermediate features.

Decode token IDs to text.

Generate 1D sincos position embedding from grid. Args: embed_dim: output dimension for each position pos: a list of position

Get autocast context manager based on precision.

Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length strings).

Convert input to 2-tuple.

Clean whitespace.

Build MLP for DINO head. Args: nlayers: Number of layers in_dim: Input dimension bottleneck_dim: Output dimension

Manual FID calculation using InceptionV3.

Convert 3D patch tokens (B, N, C) to 4D format (B, C, H, W) for pixel decoder. Args: patch_tokens: Patch tokens from DIN

Find BPE file in common locations.

Forward for CLIP mode.

Forward for feature extraction mode.

Forward for reconstruction mode.

Initialize text encoder and related components.

Initialize vision encoder and related components.

Convert patch tokens (B, N, C) to 4D format (B, C, H, W).

Setup normalization and inverse normalization transforms.

Calculate top-k accuracy.

Batch data into lists of length n. The last batch may be shorter.

Apply BPE encoding to a token.

Build zero-shot classifier weights. Args: model: VTPModel instance with get_clip_text_feature method. tokenizer: Text tokenizer.

Calculate PSNR metric. Args: original: Original image tensor (values in 0-255 range) processed: Reconstructed image tensor (value

Create ImageNet validation dataloader.

Get default BPE file path.

Encode text to token IDs.

Run zero-shot evaluation. Args: model: VTPModel instance. classifier: Zero-shot classifier weights, shape (embed_dim, num_classes

Evaluate all classifiers and return accuracies.