Struct Sam2FpnConfig 

pub struct Sam2FpnConfig {
    pub d_model: usize,
    pub backbone_channel_list: Vec<usize>,
    pub fpn_top_down_levels: Vec<usize>,
    pub interpolation_nearest: bool,
}

FPN neck configuration. Mirrors FpnNeck in the reference.

SAM 2’s neck takes the per-stage outputs from Hiera (finest → coarsest, i.e. stage 0..3) and runs a top-down pyramid:

1. Each level gets a 1×1 lateral conv to d_model=256.
2. Going coarse → fine, levels listed in fpn_top_down_levels also receive a nearest-neighbour ×2 upsample of the next- coarser level summed in.

The published _b+ / _l configs use fpn_top_down_levels=[2, 3], meaning only the two coarsest levels actually fuse with their neighbours; the two finest levels are emitted as plain laterals.

Note on indexing: backbone_channel_list and fpn_top_down_levels are stored coarse-to-fine (i.e. [stage3_dim, …, stage0_dim]) to match the reference YAML and let conv weight keys (image_encoder.neck.convs.{n-i}…) line up 1:1 with the checkpoint.

Fields

d_model: usize

backbone_channel_list: Vec<usize>

Per-stage Hiera output channels, coarse → fine order: [stage3_dim, stage2_dim, stage1_dim, stage0_dim].

fpn_top_down_levels: Vec<usize>

Backbone stage indices (in the same coarse-to-fine ordering as backbone_channel_list) that participate in the top-down sum. [2, 3] in every published config, i.e. only the two coarsest levels (note: indices into the reversed level enumeration the reference uses — kept as-is here for checkpoint compatibility).

interpolation_nearest: bool

Interpolation mode for the top-down upsample. Reference uses "nearest", which we lower as cheap host-side replicate.

Implementations

impl Sam2FpnConfig

pub fn for_hiera(cfg: &Sam2HieraConfig) -> Sam2FpnConfig

Trait Implementations

impl Clone for Sam2FpnConfig

fn clone(&self) -> Sam2FpnConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Sam2FpnConfig

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

Formats the value using the given formatter.