def __init__(self,
node_dim,
edge_dim,
hidden_dim,
residual: bool = True,
pairwise_distances: bool = False,
activation: Union[Callable, str] = "relu",
last_activation: Union[Callable, str] = "none",
mid_batch_norm: bool = False,
last_batch_norm: bool = False,
propagation_depth: int = 5,
dropout: float = 0.0,
posttrans_layers: int = 1,
pretrans_layers: int = 1,
**kwargs):
super(MPNNGNN, self).__init__()
self.node_input_net = MLP(
in_dim=node_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
layers=1,
mid_activation='relu',
dropout=dropout,
last_activation=last_activation,
)
if edge_dim > 0:
self.edge_input = MLP(
in_dim=edge_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
layers=1,
mid_activation='relu',
dropout=dropout,
last_activation=last_activation,
)
self.mp_layers = nn.ModuleList()
for _ in range(propagation_depth):
self.mp_layers.append(MPNNLayer(in_dim=hidden_dim,
out_dim=int(hidden_dim),
in_dim_edges=edge_dim,
pairwise_distances=pairwise_distances,
residual=residual,
dropout=dropout,
activation=activation,
last_activation=last_activation,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
posttrans_layers=posttrans_layers,
pretrans_layers=pretrans_layers,
),
)
def __init__(self, hidden_dim, target_dim, projection_dim = 3, distance_net=False, projection_layers=1, **kwargs):
super(DistancePredictor, self).__init__()
self.gnn = PNAGNN(hidden_dim=hidden_dim, **kwargs)
if projection_dim > 0:
self.node_projection_net = MLP(in_dim=hidden_dim, hidden_size=32,
mid_batch_norm=True, out_dim=projection_dim,
layers=projection_layers)
else:
self.node_projection_net = None
if distance_net:
self.distance_net = MLP(in_dim=hidden_dim * 2, hidden_size=projection_dim,
mid_batch_norm=True, out_dim=target_dim,
layers=projection_layers)
else:
self.distance_net = None
def __init__(self, node_dim, edge_dim, batch_norm_momentum, residual,
in_feat_dropout, dropout, layer_norm, batch_norm, gamma,
full_graph, GT_hidden_dim, GT_n_heads, GT_out_dim, GT_layers,
LPE_n_heads, LPE_layers, LPE_dim, **kwargs):
super().__init__()
self.residual = residual
self.layer_norm = layer_norm
self.batch_norm = batch_norm
self.in_feat_dropout = nn.Dropout(in_feat_dropout)
self.embedding_h = MLP(in_dim=node_dim,
hidden_size=node_dim,
layers=1,
out_dim=GT_hidden_dim - LPE_dim)
self.embedding_e_real = MLP(in_dim=edge_dim,
hidden_size=edge_dim,
layers=1,
out_dim=GT_hidden_dim)
self.embedding_e_fake = MLP(in_dim=edge_dim,
hidden_size=edge_dim,
layers=1,
out_dim=GT_hidden_dim)
self.linear_A = nn.Linear(2, LPE_dim)
encoder_layer = nn.TransformerEncoderLayer(d_model=LPE_dim,
nhead=LPE_n_heads)
self.PE_Transformer = nn.TransformerEncoder(encoder_layer,
num_layers=LPE_layers)
self.layers = nn.ModuleList([
GraphTransformerLayer(gamma, GT_hidden_dim, GT_hidden_dim,
GT_n_heads, full_graph, dropout,
self.layer_norm, self.batch_norm,
self.residual) for _ in range(GT_layers - 1)
])
self.layers.append(
GraphTransformerLayer(gamma, GT_hidden_dim, GT_out_dim, GT_n_heads,
full_graph, dropout, self.layer_norm,
self.batch_norm, self.residual))
def __init__(self,
in_dim: int,
out_dim: int,
in_dim_edges: int,
activation: Union[Callable, str] = "relu",
last_activation: Union[Callable, str] = "none",
dropout: float = 0.0,
residual: bool = True,
pairwise_distances: bool = False,
mid_batch_norm: bool = False,
last_batch_norm: bool = False,
posttrans_layers: int = 2,
pretrans_layers: int = 1, ):
super(MPNNLayer, self).__init__()
self.edge_features = in_dim_edges > 0
self.activation = activation
self.pairwise_distances = pairwise_distances
self.residual = residual
if in_dim != out_dim:
self.residual = False
self.pretrans = MLP(
in_dim=(2 * in_dim + in_dim_edges + 1) if self.pairwise_distances else (2 * in_dim + in_dim_edges),
hidden_size=in_dim,
out_dim=in_dim,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
layers=pretrans_layers,
mid_activation='relu',
dropout=dropout,
last_activation=last_activation,
)
self.posttrans = MLP(
in_dim=in_dim,
hidden_size=out_dim,
out_dim=out_dim,
layers=posttrans_layers,
mid_activation=self.activation,
last_activation=last_activation,
dropout=dropout,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
)
def __init__(self, GT_out_dim, readout_hidden_dim, readout_batchnorm,
readout_aggregators, target_dim, readout_layers,
batch_norm_momentum, **kwargs):
super().__init__()
self.readout_aggregators = readout_aggregators
self.gnn = SAN_NodeLPE(GT_out_dim=GT_out_dim,
batch_norm_momentum=batch_norm_momentum,
**kwargs)
self.output = MLP(in_dim=GT_out_dim * len(self.readout_aggregators),
hidden_size=readout_hidden_dim,
mid_batch_norm=readout_batchnorm,
out_dim=target_dim,
layers=readout_layers,
batch_norm_momentum=batch_norm_momentum)
def __init__(self,
hidden_dim,
target_dim,
batch_norm=False,
readout_batchnorm=True,
batch_norm_momentum=0.1,
dropout=0.0,
readout_layers: int = 2,
readout_hidden_dim=None,
fourier_encodings=0,
activation: str = 'SiLU',
**kwargs):
super(DistanceEncoder, self).__init__()
self.fourier_encodings = fourier_encodings
input_dim = 1 if fourier_encodings == 0 else 2 * fourier_encodings + 1
self.input_net = MLP(
in_dim=input_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=batch_norm,
last_batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
layers=1,
mid_activation=activation,
dropout=dropout,
last_activation=activation,
)
if readout_hidden_dim == None:
readout_hidden_dim = hidden_dim
self.output = MLP(in_dim=hidden_dim * 4,
hidden_size=readout_hidden_dim,
mid_batch_norm=readout_batchnorm,
batch_norm_momentum=batch_norm_momentum,
out_dim=target_dim,
layers=readout_layers)
def __init__(self,
metric_dim,
repeats,
dropout=0.8,
mid_batch_norm=True,
last_batch_norm=True,
**kwargs):
super(BasicCritic, self).__init__()
self.repeats = repeats
self.dropout = dropout
self.criticise = MLP(in_dim=metric_dim * repeats,
hidden_size=metric_dim * repeats,
mid_batch_norm=mid_batch_norm,
out_dim=metric_dim * repeats,
last_batch_norm=last_batch_norm,
dropout=0,
layers=2)
def __init__(self,
node_dim,
edge_dim,
hidden_dim,
aggregators: List[str],
scalers: List[str],
activation: Union[Callable, str] = "relu",
last_activation: Union[Callable, str] = "none",
mid_batch_norm: bool = False,
last_batch_norm: bool = False,
propagation_depth: int = 5,
dropout: float = 0.0,
posttrans_layers: int = 1,
pretrans_layers: int = 1,
**kwargs):
super(DGNGNN, self).__init__()
self.node_input_net = MLP(
in_dim=node_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
layers=1,
mid_activation='relu',
dropout=dropout,
last_activation=last_activation,
)
self.mp_layers = nn.ModuleList()
for _ in range(propagation_depth):
self.mp_layers.append(DGNMessagePassingLayer(in_dim=hidden_dim,
out_dim=hidden_dim,
in_dim_edges=edge_dim,
aggregators=aggregators,
scalers=scalers,
dropout=dropout,
activation=activation,
last_activation=last_activation,
avg_d={"log": 1.0},
posttrans_layers=posttrans_layers,
pretrans_layers=pretrans_layers,
),
)
def __init__(
self,
model_type,
model_parameters,
predictor_layers=1,
predictor_hidden_size=256,
predictor_batchnorm=False,
metric_dim=256,
ma_decay=0.99, #moving average decay
**kwargs):
super(BYOLwrapper, self).__init__()
self.student = globals()[model_type](**model_parameters, **kwargs)
self.teacher = copy.deepcopy(self.student)
self.predictor_layers = predictor_layers
if predictor_layers > 0:
self.predictor = MLP(in_dim=model_parameters['target_dim'],
hidden_size=predictor_hidden_size,
mid_batch_norm=predictor_batchnorm,
out_dim=metric_dim,
layers=predictor_layers)
self.ma_decay = ma_decay
for p in self.teacher.parameters():
p.requires_grad = False
def __init__(self,
node_dim,
edge_dim,
hidden_dim,
target_dim,
aggregators: List[str],
scalers: List[str],
readout_aggregators: List[str],
frozen_readout_aggregators: List[str],
latent3d_dim: int = 256,
readout_batchnorm: bool = True,
readout_hidden_dim=None,
readout_layers: int = 2,
residual: bool = True,
pairwise_distances: bool = False,
activation: Union[Callable, str] = "relu",
last_activation: Union[Callable, str] = "none",
mid_batch_norm: bool = False,
last_batch_norm: bool = False,
propagation_depth: int = 5,
dropout: float = 0.0,
posttrans_layers: int = 1,
pretrans_layers: int = 1,
**kwargs):
super(PNAFrozenCombined, self).__init__()
# the pretrained GNN
self.node_gnn = PNAGNN(node_dim=node_dim,
edge_dim=edge_dim,
hidden_dim=hidden_dim,
aggregators=aggregators,
scalers=scalers,
residual=residual,
pairwise_distances=pairwise_distances,
activation=activation,
last_activation=last_activation,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
propagation_depth=propagation_depth,
dropout=dropout,
posttrans_layers=posttrans_layers,
pretrans_layers=pretrans_layers)
self.node_gnn2D = PNAGNN(node_dim=node_dim,
edge_dim=edge_dim,
hidden_dim=hidden_dim,
aggregators=aggregators,
scalers=scalers,
residual=residual,
pairwise_distances=pairwise_distances,
activation=activation,
last_activation=last_activation,
mid_batch_norm=mid_batch_norm,
last_batch_norm=last_batch_norm,
propagation_depth=propagation_depth,
dropout=dropout,
posttrans_layers=posttrans_layers,
pretrans_layers=pretrans_layers)
if readout_hidden_dim == None:
readout_hidden_dim = hidden_dim
self.frozen_readout_aggregators = frozen_readout_aggregators
self.output = MLP(in_dim=hidden_dim *
len(self.frozen_readout_aggregators),
hidden_size=latent3d_dim,
mid_batch_norm=False,
out_dim=latent3d_dim,
layers=1)
self.readout_aggregators = readout_aggregators
self.output2D = MLP(in_dim=hidden_dim * len(self.readout_aggregators) +
latent3d_dim,
hidden_size=readout_hidden_dim,
mid_batch_norm=readout_batchnorm,
out_dim=target_dim,
layers=readout_layers)
def __init__(self, hidden_dim, target_dim, **kwargs):
super(GraphRepresentation, self).__init__()
self.gnn = PNAGNN(hidden_dim=hidden_dim, **kwargs)
self.distance_net = MLP(in_dim=hidden_dim * 2, hidden_size=32,
mid_batch_norm=True, out_dim=target_dim,
layers=2)
def __init__(self,
node_dim,
edge_dim,
hidden_dim,
target_dim,
readout_aggregators: List[str],
batch_norm=False,
node_wise_output_layers=2,
readout_batchnorm=True,
batch_norm_momentum=0.1,
reduce_func='sum',
dropout=0.0,
propagation_depth: int = 4,
readout_layers: int = 2,
readout_hidden_dim=None,
fourier_encodings=0,
activation: str = 'SiLU',
update_net_layers=2,
message_net_layers=2,
**kwargs):
super(Net3D, self).__init__()
self.fourier_encodings = fourier_encodings
self.input = MLP(
in_dim=node_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=batch_norm,
last_batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
layers=1,
mid_activation=activation,
dropout=dropout,
last_activation=activation,
)
edge_in_dim = 1 if fourier_encodings == 0 else 2 * fourier_encodings + 1
self.edge_input = MLP(
in_dim=edge_in_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=batch_norm,
last_batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
layers=1,
mid_activation=activation,
dropout=dropout,
last_activation=activation,
)
self.mp_layers = nn.ModuleList()
for _ in range(propagation_depth):
self.mp_layers.append(
Net3DLayer(node_dim,
edge_dim=hidden_dim,
hidden_dim=hidden_dim,
batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
dropout=dropout,
mid_activation=activation,
reduce_func=reduce_func,
message_net_layers=message_net_layers,
update_net_layers=update_net_layers))
self.node_wise_output_layers = node_wise_output_layers
if self.node_wise_output_layers > 0:
self.node_wise_output_network = MLP(
in_dim=hidden_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=batch_norm,
last_batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
layers=node_wise_output_layers,
mid_activation=activation,
dropout=dropout,
last_activation='None',
)
self.node_embedding = nn.Parameter(torch.empty((hidden_dim, )))
nn.init.normal_(self.node_embedding)
if readout_hidden_dim == None:
readout_hidden_dim = hidden_dim
self.readout_aggregators = readout_aggregators
self.output = MLP(in_dim=hidden_dim * len(self.readout_aggregators),
hidden_size=readout_hidden_dim,
mid_batch_norm=readout_batchnorm,
batch_norm_momentum=batch_norm_momentum,
out_dim=target_dim,
layers=readout_layers)
def __init__(self,
node_dim,
edge_dim,
hidden_dim,
target_dim,
readout_aggregators: List[str],
batch_norm=False,
node_wise_output_layers=2,
readout_batchnorm=True,
batch_norm_momentum=0.1,
reduce_func='sum',
dropout=0.0,
readout_layers: int = 2,
readout_hidden_dim=None,
fourier_encodings=0,
activation: str = 'SiLU',
**kwargs):
super(DistanceAggregator, self).__init__()
self.fourier_encodings = fourier_encodings
if reduce_func == 'sum':
self.reduce_func = fn.sum
elif reduce_func == 'mean':
self.reduce_func = fn.mean
else:
raise ValueError('reduce function not supportet: ', reduce_func)
edge_in_dim = 1 if fourier_encodings == 0 else 2 * fourier_encodings + 1
self.edge_input = MLP(
in_dim=edge_in_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=batch_norm,
last_batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
layers=1,
mid_activation=activation,
dropout=dropout,
last_activation=activation,
)
self.node_wise_output_layers = node_wise_output_layers
if self.node_wise_output_layers > 0:
self.node_wise_output_network = MLP(
in_dim=hidden_dim,
hidden_size=hidden_dim,
out_dim=hidden_dim,
mid_batch_norm=batch_norm,
last_batch_norm=batch_norm,
batch_norm_momentum=batch_norm_momentum,
layers=node_wise_output_layers,
mid_activation=activation,
dropout=dropout,
last_activation='None',
)
if readout_hidden_dim == None:
readout_hidden_dim = hidden_dim
self.readout_aggregators = readout_aggregators
self.output = MLP(in_dim=hidden_dim * len(self.readout_aggregators),
hidden_size=readout_hidden_dim,
mid_batch_norm=readout_batchnorm,
batch_norm_momentum=batch_norm_momentum,
out_dim=target_dim,
layers=readout_layers)
