Struct ClipTextConfig 

pub struct ClipTextConfig {
    pub vocab_size: usize,
    pub n_ctx: usize,
    pub width: usize,
    pub depth: usize,
    pub n_heads: usize,
    pub mlp_ratio: usize,
    pub embed_dim: usize,
    pub eot_token: usize,
}

Hyper-parameters for the CLIP text Transformer.

Fields

vocab_size: usize

Vocabulary size (number of distinct token ids).

n_ctx: usize

Maximum context length (number of positional embeddings).

width: usize

Transformer width (token embedding / residual-stream dimension).

depth: usize

Number of transformer blocks.

n_heads: usize

Number of attention heads. Must divide width.

mlp_ratio: usize

MLP hidden-dim multiplier (mlp_dim = mlp_ratio * width).

embed_dim: usize

Output joint-embedding dimension (after the text projection).

eot_token: usize

Id used as the end-of-text marker. When pooling, the position of the last occurrence of this id (or, if absent, the highest id present) is used. CLIP itself uses argmax over ids because <|endoftext|> is the largest BPE id; we follow that convention via Self::eot_token.

Implementations

impl ClipTextConfig

pub fn new( vocab_size: usize, n_ctx: usize, width: usize, depth: usize, n_heads: usize, mlp_ratio: usize, embed_dim: usize, eot_token: usize, ) -> VisionResult<Self>

Validate and construct a config.

Errors

• VisionError::InvalidEmbedDim if width == 0 or embed_dim == 0.
• VisionError::InvalidNumHeads if n_heads == 0.
• VisionError::HeadDimMismatch if n_heads does not divide width.
• VisionError::Internal if vocab_size, n_ctx, or depth is 0.

pub fn tiny() -> Self

A tiny config suitable for unit tests.

vocab_size = 64, n_ctx = 16, width = 32, depth = 2, n_heads = 4, mlp_ratio = 4, embed_dim = 24, eot_token = 63.

pub fn head_dim(&self) -> usize

Per-head dimension.

pub fn mlp_dim(&self) -> usize

MLP hidden dimension.

Trait Implementations

impl Clone for ClipTextConfig

fn clone(&self) -> ClipTextConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ClipTextConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl PartialEq for ClipTextConfig

fn eq(&self, other: &ClipTextConfig) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for ClipTextConfig