MolecularDiffusion.modules.models.gfmdiff.equignn

GFMDiff Dual-Track Transformer network body.

Near-verbatim port of GFMDiff/models/diff/dtn.py::EquiGNN. Only the import paths changed (blocks now live under MolecularDiffusion.modules.layers.gfmdiff.blocks); the math/forward logic is unchanged from the source repo.

Per the approved integration plan, the diffused “degree” (valence) channel and the Geometric-Facilitated Loss are out of scope for this integration: include_de is hardcoded to False below (the constructor no longer reads it from model_config), while include_an (atomic-number channel) is kept, matching every other backbone already integrated in this repo (in_node_nf = len(atom_vocab) + n_dim_extra + 1).

Classes

Module Contents

class MolecularDiffusion.modules.models.gfmdiff.equignn.EquiGNN(model_config, data_config)

Bases: torch.nn.Module

abstractmethod forward(z_t, t, node_mask, pair_mask, context=None)
unwrap_forward()
wrap_forward(node_mask, edge_mask, context)
act
activate
add_time
angle_emb
block_calc
context_dim
convs
dropout
edge_emb
emb_dim
hidden_dim
include_an
include_de = False
node_nf
node_pair_fuse
num_heads
num_layers
pair_scale
xh_embedding
xh_embedding_out