MolecularDiffusion.modules.models.ligandiff.conv_layer¶

GVP message passing. Verbatim port of LigandDiff’s src/conv_layer.py.

Classes¶

`GVPConvLayer`	One GVP conv block: message passing + residual feed-forward.
`GVPMessagePassing`	Edge-GVP message passing with optional scalar attention gating.

Module Contents¶

class MolecularDiffusion.modules.models.ligandiff.conv_layer.GVPConvLayer(node_dims: Tuple[int, int], edge_dims: Tuple[int, int], drop_rate: float = 0.0, activations=(F.relu, torch.sigmoid), vector_gate: bool = False, residual: bool = True, attention: bool = True, aggr: str = 'add', normalization_factor: float = 1.0)¶

Bases: GVPMessagePassing, abc.ABC

One GVP conv block: message passing + residual feed-forward.

forward(x: MolecularDiffusion.modules.models.ligandiff.gvp.s_V | torch.Tensor, edge_index: torch.Tensor, edge_attr: torch.Tensor) → MolecularDiffusion.modules.models.ligandiff.gvp.s_V¶

drop_rate = 0.0¶

dropout¶

ff_func¶

norm¶

residual = True¶

class MolecularDiffusion.modules.models.ligandiff.conv_layer.GVPMessagePassing(in_dims: Tuple[int, int], out_dims: Tuple[int, int], edge_dims: Tuple[int, int], hidden_dims: Tuple[int, int] | None = None, activations=(F.relu, torch.sigmoid), vector_gate: bool = False, attention: bool = True, aggr: str = 'add', normalization_factor: float = 1.0)¶

Bases: torch_geometric.nn.MessagePassing, abc.ABC

Edge-GVP message passing with optional scalar attention gating.

forward(x: MolecularDiffusion.modules.models.ligandiff.gvp.s_V, edge_index: torch.Tensor, edge_attr: torch.Tensor) → MolecularDiffusion.modules.models.ligandiff.gvp.s_V¶

message(s_i, s_j, V_i, V_j, edge_attr, v_dim)¶

reset_parameters() → None¶

update(aggr_out: torch.Tensor) → MolecularDiffusion.modules.models.ligandiff.gvp.s_V¶

attention = True¶

edge_gvps¶

hidden_scalar¶

hidden_vector¶

in_vector¶

normalization_factor = 1.0¶