Struct VitPatchEmbed 

pub struct VitPatchEmbed { /* private fields */ }

ViT patch-embedding layer with [CLS] token and learnable positions.

Implementations

impl VitPatchEmbed

pub fn new(config: VitPatchConfig, rng: &mut VisionRng) -> VisionResult<Self>

Construct a new patch embedder with random parameters.

The projection weight is initialised N(0, 1/√patch_dim) (the fan-in scaling used by the reference implementations); biases are near-zero, the class token is N(0, 0.02) (ViT’s truncated-normal scale), and the positional embedding is N(0, 0.02).

Errors

Propagates VitPatchConfig::validate failures.

pub fn config(&self) -> &VitPatchConfig

Read-only access to the configuration.

pub fn n_patches(&self) -> usize

Number of image patches: (image_size / patch_size)².

pub fn forward(&self, image: &[f32]) -> VisionResult<Vec<f32>>

Embed an image into a token sequence.

image must be [n_channels × image_size × image_size] row-major (CHW). The returned sequence is [(n_patches + 1) × d_model]: a prepended [CLS] token followed by the n_patches projected patch tokens, with the learnable positional embedding added element-wise.

Errors

• VisionError::DimensionMismatch if image.len() does not equal n_channels · image_size².
• VisionError::NonFinite if a non-finite value is produced.