Struct DetrDecoderLayer 

pub struct DetrDecoderLayer {
    pub config: DetrConfig,
    pub weights: DetrDecoderLayerWeights,
}

A single DETR decoder layer.

Fields

config: DetrConfig

Decoder configuration (n_queries, embed_dim, n_heads, …).

weights: DetrDecoderLayerWeights

Learned weights for this layer.

Implementations

impl DetrDecoderLayer

pub fn new(cfg: DetrConfig, rng: &mut LcgRng) -> Self

Construct a new decoder layer with Xavier-initialised weights.

pub fn forward( &self, queries: &[f32], encoder_feats: &[f32], n_enc_tokens: usize, ) -> VisionResult<Vec<f32>>

Forward pass for one decoder layer.

Pre-norm residual scheme:

q1  = self_attn(LN1(queries)) + queries
q2  = cross_attn(LN2(q1), key=encoder, val=encoder) + q1
out = FFN(LN3(q2)) + q2

Parameters

• queries: flat [n_queries × embed_dim].
• encoder_feats: flat [n_enc_tokens × embed_dim].
• n_enc_tokens: number of encoder feature tokens.

Returns

Updated queries: flat [n_queries × embed_dim].

Errors

• DimensionMismatch if input tensor lengths are inconsistent.
• NonFinite if NaN/Inf appear in attention output.