Struct DecLayerJ 

pub struct DecLayerJ { /* private fields */ }

DecLayer applied one dimension deeper, over LigandMPNN’s ligand graph.

Port of DecLayerJ (model_utils.py). Structurally identical to DecLayer — same weight names, same W3(act(W2(act(W1)))) message and same two residual norms — and the reference comments the single difference as “the only difference”: the node expansion.

DecLayer carries h_V as (B, L, C) and broadcasts it over K neighbours. Here h_V is Y_nodes, already (B, L, M, C) — one node per (residue, ligand atom) pair — and it broadcasts over the ligand graph’s own M edges to (B, L, M, M, C).

Implementations

impl DecLayerJ

pub fn load( vb: VarBuilder<'_>, config: &ProteinMPNNConfig, layer: i32, ) -> Result<Self>

pub fn forward( &self, h_v: &Tensor, h_e: &Tensor, mask_v: Option<&Tensor>, mask_attend: Option<&Tensor>, ) -> Result<Tensor>

h_v is (B, L, M, C); h_e is (B, L, M, M, C).

Trait Implementations

impl Clone for DecLayerJ

fn clone(&self) -> DecLayerJ

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DecLayerJ

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

Formats the value using the given formatter.