Struct Sam2HieraConfig 

pub struct Sam2HieraConfig {
    pub embed_dim: usize,
    pub num_heads: usize,
    pub stages: Vec<usize>,
    pub global_att_blocks: Vec<usize>,
    pub window_pos_embed_bkg_spatial_size: [usize; 2],
    pub window_spec: [usize; 4],
    pub layer_norm_eps: f64,
    pub mlp_ratio: f64,
    pub qkv_bias: bool,
    pub fpn_out_chans: usize,
}

Hiera image-encoder configuration — Tiny, Small, Base+ or Large.

Field names mirror Hiera’s Python kwargs so the values map 1:1 to the published checkpoints.

Fields

embed_dim: usize

Stage-0 embedding dimension. Doubles after each Q-pool.

num_heads: usize

Stage-0 head count. Doubles after each Q-pool.

stages: Vec<usize>

Number of blocks per stage. stages.len() == 4 for every published Hiera variant (Tiny/Small/Base+/Large).

global_att_blocks: Vec<usize>

Indices (in the flattened block enumeration across stages) of blocks that use global attention rather than mask-unit (windowed) attention. Always exactly 3 entries in published configs.

window_pos_embed_bkg_spatial_size: [usize; 2]

Background-pos-embed spatial size: [Ph, Pw] for the learned-pos table that gets bilinear-interpolated to the current grid each forward.

window_spec: [usize; 4]

Mask-unit window size per stage (in units of post-Q-pool tokens at that stage). Always 4 entries, one per stage.

layer_norm_eps: f64

LayerNorm eps used throughout the encoder.

mlp_ratio: f64

MLP expansion ratio (FFN hidden = mlp_ratio · embed_dim).

qkv_bias: bool

QKV linear bias toggle (always true in published configs).

fpn_out_chans: usize

Output channels per FPN level (256 for every published config).

Implementations

impl Sam2HieraConfig

pub fn tiny() -> Sam2HieraConfig

sam2_hiera_tiny — ~30 M params.

pub fn small() -> Sam2HieraConfig

sam2_hiera_small — ~46 M params.

pub fn base_plus() -> Sam2HieraConfig

sam2_hiera_base_plus — ~80 M params, the recommended default.

pub fn large() -> Sam2HieraConfig

sam2_hiera_large — ~224 M params. The YAML overrides window_pos_embed_bkg_spatial_size to [7, 7] (vs [14, 14] for base+).

pub fn total_blocks(&self) -> usize

Total number of transformer blocks across all stages.

pub fn q_pool_block_indices(&self) -> Vec<usize>

Indices (in the flattened block enumeration) where a Q-pool happens — at the first block of every stage after stage 0.

Reference: in Hiera.__init__ the Q-pool boundaries are cumulative_sum(stages)[:-1], i.e. [s0, s0+s1, s0+s1+s2].

pub fn stage_of_block(&self, block_idx: usize) -> usize

Stage index for the i-th flattened block.

pub fn embed_dim_at_stage(&self, s: usize) -> usize

Embedding dimension at stage s (doubles per Q-pool).

pub fn num_heads_at_stage(&self, s: usize) -> usize

Number of heads at stage s (doubles per Q-pool).

pub fn window_size_at_stage(&self, s: usize) -> usize

Mask-unit window size at stage s.

pub fn grid_size_at_stage(&self, s: usize) -> usize

Per-axis spatial size of the token grid at stage s (before any Q-pool inside the stage — i.e. the size at the stage’s first post-Q-pool block, for s>0, or stage-0 patch grid for s=0).

Trait Implementations

impl Clone for Sam2HieraConfig

fn clone(&self) -> Sam2HieraConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Sam2HieraConfig

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Sam2HieraConfig

fn deserialize<__D>( __deserializer: __D, ) -> Result<Sam2HieraConfig, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.