Struct ViTModel 

pub struct ViTModel {
    pub config: ViTConfig,
    pub patch_embed: PatchEmbed,
    pub pos_embed: LearnablePosEmbed,
    pub encoder: ViTEncoder,
    pub weights: ViTModelWeights,
}

Full Vision Transformer model.

Fields

config: ViTConfig

Top-level model configuration.

patch_embed: PatchEmbed

Strided conv2d patch embedder.

pos_embed: LearnablePosEmbed

Learnable positional embeddings for seq_len positions (CLS + patches).

encoder: ViTEncoder

Transformer encoder stack.

weights: ViTModelWeights

Classification head weights.

Implementations

impl ViTModel

pub fn new(cfg: ViTConfig, rng: &mut LcgRng) -> VisionResult<Self>

Build and initialise a full ViT model.

All sub-modules share the same rng stream for reproducibility.

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

Forward pass.

image is flat CHW: [in_chans, img_size, img_size]. Returns logits: [n_classes].

Errors

Returns DimensionMismatch if image.len() does not match in_chans * img_size * img_size.