def __init__(self,
in_feats: int,
hidden_size: int,
out_feats: int,
pooling_layer: int,
pooling_rates: List[float],
n_dropout: float = 0.5,
adj_dropout: float = 0.3,
activation: str = "elu",
improved: bool = False,
aug_adj: bool = False):
super(GraphUnet, self).__init__()
self.improved = improved
self.n_dropout = n_dropout
self.act = get_activation(activation)
assert pooling_layer <= len(pooling_rates)
pooling_rates = pooling_rates[:pooling_layer]
self.unet = GraphUnetLayer(hidden_size, pooling_layer, pooling_rates,
activation, n_dropout, aug_adj)
self.in_gcn = GraphConvolution(in_feats, hidden_size)
self.out_gcn = GraphConvolution(hidden_size, out_feats)
self.adj_drop = nn.Dropout(
adj_dropout) if adj_dropout > 0.001 else nn.Identity()
self.cache_edge_index = None
self.cache_edge_attr = None
def __init__(self,
in_feats,
hidden_size,
out_feats,
num_layers,
dropout,
activation="relu",
residual=True,
norm=None):
super(MoEGCN, self).__init__()
shapes = [in_feats] + [hidden_size] * num_layers
conv_func = GCNLayer
conv_params = {
"dropout": dropout,
"norm": norm,
"residual": residual,
"activation": activation,
}
self.layers = nn.ModuleList([
GraphConvBlock(
conv_func,
conv_params,
shapes[i],
shapes[i + 1],
dropout=dropout,
) for i in range(num_layers)
])
self.num_layers = num_layers
self.dropout = dropout
self.act = get_activation(activation)
self.final_cls = nn.Linear(hidden_size, out_feats)
def __init__(self, in_feats, hidden_size, out_feats, num_layers, dropout, activation="relu"):
super(PPRGoMLP, self).__init__()
self.dropout = dropout
self.nlayers = num_layers
shapes = [in_feats] + [hidden_size] * (num_layers - 1) + [out_feats]
self.layers = nn.ModuleList([nn.Linear(shapes[i], shapes[i + 1]) for i in range(num_layers)])
self.activation = get_activation(activation)
def __init__(
self,
in_feats: int,
hidden_size: int,
out_feats: int,
pooling_layer: int,
pooling_rates: List[float],
n_dropout: float = 0.5,
adj_dropout: float = 0.3,
activation: str = "elu",
improved: bool = False,
aug_adj: bool = False,
):
super(GraphUnet, self).__init__()
self.improved = improved
self.n_dropout = n_dropout
self.adj_dropout = adj_dropout
self.act = get_activation(activation)
assert pooling_layer <= len(pooling_rates)
pooling_rates = pooling_rates[:pooling_layer]
pooling_rates = [float(x) for x in pooling_rates]
self.unet = GraphUnetLayer(hidden_size, pooling_layer, pooling_rates,
activation, n_dropout, aug_adj)
self.in_gcn = GCNLayer(in_feats, hidden_size)
self.out_gcn = GCNLayer(hidden_size, out_feats)
def __init__(
self,
hidden_size: int,
pooling_layer: int,
pooling_rates: List[float],
activation: str = "elu",
dropout: float = 0.5,
aug_adj: bool = False,
):
super(GraphUnetLayer, self).__init__()
self.dropout = dropout
self.activation = activation
self.pooling_layer = pooling_layer
self.gcn = GCNLayer(hidden_size, hidden_size)
self.act = get_activation(activation)
self.down_gnns = nn.ModuleList(
[GCNLayer(hidden_size, hidden_size) for _ in range(pooling_layer)])
self.up_gnns = nn.ModuleList(
[GCNLayer(hidden_size, hidden_size) for _ in range(pooling_layer)])
self.poolings = nn.ModuleList([
Pool(hidden_size, pooling_rates[i], aug_adj, dropout)
for i in range(pooling_layer)
])
self.unpoolings = nn.ModuleList(
[UnPool() for _ in range(pooling_layer)])
def __init__(self,
in_feats,
out_feats,
hidden_size,
num_layers,
dropout=0.0,
activation="relu",
norm=None):
super(MLP, self).__init__()
self.norm = norm
self.activation = get_activation(activation)
self.dropout = dropout
shapes = [in_feats] + [hidden_size] * (num_layers - 1) + [out_feats]
self.mlp = nn.ModuleList([
nn.Linear(shapes[layer], shapes[layer + 1])
for layer in range(num_layers)
])
if norm is not None and num_layers > 1:
if norm == "layernorm":
self.norm_list = nn.ModuleList(
nn.LayerNorm(x) for x in shapes[1:-1])
elif norm == "batchnorm":
self.norm_list = nn.ModuleList(
nn.BatchNorm1d(x) for x in shapes[1:-1])
else:
raise NotImplementedError(
f"{norm} is not implemented in CogDL.")
def __init__(
self, in_feats, out_feats, nhead=1, alpha=0.2, attn_drop=0.5, activation=None, residual=False, norm=None
):
super(GATLayer, self).__init__()
self.in_features = in_feats
self.out_features = out_feats
self.alpha = alpha
self.nhead = nhead
self.W = nn.Parameter(torch.FloatTensor(in_feats, out_feats * nhead))
self.a_l = nn.Parameter(torch.zeros(size=(1, nhead, out_feats)))
self.a_r = nn.Parameter(torch.zeros(size=(1, nhead, out_feats)))
self.edge_softmax = EdgeSoftmax()
self.mhspmm = MultiHeadSpMM()
self.dropout = nn.Dropout(attn_drop)
self.leakyrelu = nn.LeakyReLU(self.alpha)
self.act = None if activation is None else get_activation(activation)
self.norm = None if norm is None else get_norm_layer(norm, out_feats * nhead)
if residual:
self.residual = nn.Linear(in_feats, out_feats * nhead)
else:
self.register_buffer("residual", None)
self.reset_parameters()
def __init__(
self, in_feats: int, pooling_rate: float, aug_adj: bool = False, dropout: float = 0.5, activation: str = "tanh"
):
super(Pool, self).__init__()
self.aug_adj = aug_adj
self.pooling_rate = pooling_rate
self.act = get_activation(activation)
self.proj = nn.Linear(in_feats, 1)
self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity
def __init__(self, in_feats, hidden_size, out_feats, num_layers, dropout, activation="relu"):
super(PPRGoMLP, self).__init__()
self.dropout = dropout
self.nlayers = num_layers
shapes = [hidden_size] * (num_layers - 1) + [out_feats]
self.layers = nn.ModuleList()
self.layers.append(LinearLayer(in_feats, hidden_size, bias=False))
for i in range(num_layers - 1):
self.layers.append(nn.Linear(shapes[i], shapes[i + 1], bias=False))
self.activation = get_activation(activation)
def __init__(
self,
in_feats: int,
out_feats: int,
num_layers: int,
activation: str = "relu",
):
super(GraceEncoder, self).__init__()
shapes = [in_feats] + [2 * out_feats] * (num_layers - 1) + [out_feats]
self.layers = nn.ModuleList([GraphConvolution(shapes[i], shapes[i + 1]) for i in range(num_layers)])
self.activation = get_activation(activation)
def __init__(self, in_ft, out_ft, act, bias=True):
super(GCN, self).__init__()
self.fc = nn.Linear(in_ft, out_ft, bias=False)
self.act = nn.PReLU() if act == "prelu" else get_activation(act)
if bias:
self.bias = nn.Parameter(torch.FloatTensor(out_ft))
self.bias.data.fill_(0.0)
else:
self.register_parameter("bias", None)
for m in self.modules():
self.weights_init(m)
def __init__(
self,
in_feats,
hidden_size,
out_feats,
num_layers,
group=2,
alpha=0.2,
nhead=1,
dropout=0.5,
attn_drop=0.5,
activation="relu",
norm="batchnorm",
):
super(RevGAT, self).__init__()
self.dropout = dropout
self.num_layers = num_layers
self.layers = nn.ModuleList()
self.norm = get_norm_layer(norm, hidden_size * nhead)
self.act = get_activation(activation)
for i in range(num_layers):
if i == 0:
self.layers.append(
GATLayer(
in_feats,
hidden_size,
nhead,
alpha,
attn_drop,
residual=True,
)
)
elif i == num_layers - 1:
self.layers.append(GATLayer(hidden_size * nhead, out_feats, 1, alpha, attn_drop, residual=True))
else:
conv = GATLayer(
hidden_size * nhead // group,
hidden_size // group,
nhead=nhead,
alpha=alpha,
attn_drop=attn_drop,
)
res_conv = ResGNNLayer(
conv,
hidden_size * nhead // group,
activation=activation,
norm=norm,
out_norm=norm,
out_channels=hidden_size * nhead // group,
)
self.layers.append(RevGNNLayer(res_conv, group))
def __init__(
self,
in_feat,
hidden_size,
out_feat,
num_layers,
connection="res+",
activation="relu",
dropout=0.0,
aggr="max",
beta=1.0,
p=1.0,
learn_beta=False,
learn_p=False,
learn_msg_scale=True,
use_msg_norm=False,
):
super(DeeperGCN, self).__init__()
self.dropout = dropout
self.feat_encoder = nn.Linear(in_feat, hidden_size)
self.layers = nn.ModuleList()
self.layers.append(GENConv(hidden_size, hidden_size))
for i in range(num_layers - 1):
self.layers.append(
DeepGCNLayer(
in_feat=hidden_size,
out_feat=hidden_size,
conv=GENConv(
in_feat=hidden_size,
out_feat=hidden_size,
aggr=aggr,
beta=beta,
p=p,
learn_beta=learn_beta,
learn_p=learn_p,
use_msg_norm=use_msg_norm,
learn_msg_scale=learn_msg_scale,
),
connection=connection,
activation=activation,
dropout=dropout,
checkpoint_grad=(num_layers > 3)
and ((i + 1) == num_layers // 2),
))
self.norm = nn.BatchNorm1d(hidden_size, affine=True)
self.activation = get_activation(activation)
self.fc = nn.Linear(hidden_size, out_feat)
def __init__(
self,
in_feats,
hidden_size,
out_feats,
num_layers,
group=2,
activation="relu",
norm="batchnorm",
last_norm="batchnorm",
dropout=0.0,
aggr="softmax_sg",
beta=1.0,
p=1.0,
learn_beta=False,
learn_p=False,
learn_msg_scale=True,
use_msg_norm=False,
edge_attr_size: Optional[list] = None,
one_hot_emb: bool = False,
):
super(RevGEN, self).__init__()
self.input_fc = nn.Linear(in_feats, hidden_size)
self.output_fc = nn.Linear(hidden_size, out_feats)
self.layers = nn.ModuleList()
for _ in range(num_layers):
conv = GENConv(
hidden_size // group,
hidden_size // group,
aggr,
beta,
p,
learn_beta,
learn_p,
use_msg_norm,
learn_msg_scale,
residual=True,
edge_attr_size=edge_attr_size,
)
res_conv = ResGNNLayer(conv, hidden_size // group, norm=norm, activation=activation, residual=False)
self.layers.append(RevGNNLayer(res_conv, group))
self.activation = get_activation(activation)
self.norm = get_norm_layer(last_norm, hidden_size)
self.dropout = dropout
if one_hot_emb:
self.one_hot_encoder = nn.Linear(in_feats // 2, in_feats // 2)
self.use_one_hot_emb = one_hot_emb
def __init__(
self,
in_feat,
hidden_size,
out_feat,
num_layers,
activation="relu",
dropout=0.0,
aggr="max",
beta=1.0,
p=1.0,
learn_beta=False,
learn_p=False,
learn_msg_scale=True,
use_msg_norm=False,
edge_attr_size=None,
):
super(DeeperGCN, self).__init__()
self.dropout = dropout
self.feat_encoder = nn.Linear(in_feat, hidden_size)
self.layers = nn.ModuleList()
for i in range(num_layers - 1):
self.layers.append(
ResGNNLayer(
conv=GENConv(
in_feats=hidden_size,
out_feats=hidden_size,
aggr=aggr,
beta=beta,
p=p,
learn_beta=learn_beta,
learn_p=learn_p,
use_msg_norm=use_msg_norm,
learn_msg_scale=learn_msg_scale,
edge_attr_size=edge_attr_size,
),
in_channels=hidden_size,
activation=activation,
dropout=dropout,
checkpoint_grad=False,
)
)
self.norm = nn.BatchNorm1d(hidden_size, affine=True)
self.activation = get_activation(activation)
self.fc = nn.Linear(hidden_size, out_feat)
def __init__(
self,
in_feat,
out_feat,
conv,
connection="res",
activation="relu",
dropout=0.0,
checkpoint_grad=False,
):
super(DeepGCNLayer, self).__init__()
self.conv = conv
self.activation = get_activation(activation)
self.dropout = dropout
self.connection = connection
self.norm = nn.BatchNorm1d(out_feat, affine=True)
self.checkpoint_grad = checkpoint_grad
def __init__(
self,
edge_feats,
num_etypes,
in_features,
out_features,
nhead,
feat_drop=0.0,
attn_drop=0.5,
negative_slope=0.2,
residual=False,
activation=None,
alpha=0.0,
):
super(myGATConv, self).__init__()
self.edge_feats = edge_feats
self.in_features = in_features
self.out_features = out_features
self.nhead = nhead
self.edge_emb = nn.Parameter(torch.zeros(
size=(num_etypes,
edge_feats))) # nn.Embedding(num_etypes, edge_feats)
self.W = nn.Parameter(
torch.FloatTensor(in_features, out_features * nhead))
self.W_e = nn.Parameter(
torch.FloatTensor(edge_feats, edge_feats * nhead))
self.a_l = nn.Parameter(torch.zeros(size=(1, nhead, out_features)))
self.a_r = nn.Parameter(torch.zeros(size=(1, nhead, out_features)))
self.a_e = nn.Parameter(torch.zeros(size=(1, nhead, edge_feats)))
self.mhspmm = MultiHeadSpMM()
self.feat_drop = nn.Dropout(feat_drop)
self.dropout = nn.Dropout(attn_drop)
self.leakyrelu = nn.LeakyReLU(negative_slope)
self.act = None if activation is None else get_activation(activation)
if residual:
self.residual = nn.Linear(in_features, out_features * nhead)
else:
self.register_buffer("residual", None)
self.reset_parameters()
self.alpha = alpha
def __init__(
self,
in_feats,
out_feats,
hidden_size,
num_layers,
dropout=0.5,
drop_edge_rate=0.1,
activation="relu",
norm="batchnorm",
group=2,
):
super(RevGCN, self).__init__()
self.dropout = dropout
self.drop_edge_rate = drop_edge_rate
self.num_layers = num_layers
self.layers = nn.ModuleList()
self.norm = get_norm_layer(norm, hidden_size)
self.act = get_activation(activation)
for i in range(num_layers):
if i == 0:
self.layers.append(
GCNLayer(
in_feats,
hidden_size,
residual=True,
)
)
elif i == num_layers - 1:
self.layers.append(GCNLayer(hidden_size, out_feats, residual=True))
else:
conv = GCNLayer(
hidden_size // group,
hidden_size // group,
)
res_conv = ResGNNLayer(
conv,
hidden_size // group,
activation=activation,
norm=norm,
out_norm=norm,
out_channels=hidden_size // group,
)
self.layers.append(RevGNNLayer(res_conv, group))
def __init__(self,
in_feats,
out_feats,
normalize=False,
aggr="mean",
dropout=0.0,
norm=None,
activation=None,
residual=False):
super(SAGELayer, self).__init__()
self.in_feats = in_feats
self.out_feats = out_feats
self.fc = nn.Linear(2 * in_feats, out_feats)
self.normalize = normalize
if dropout > 0:
self.dropout = nn.Dropout(dropout)
else:
self.dropout = None
if aggr == "mean":
self.aggr = MeanAggregator()
elif aggr == "sum":
self.aggr = SumAggregator()
elif aggr == "max":
self.aggr = MaxAggregator()
else:
raise NotImplementedError
if activation is not None:
self.act = get_activation(activation, inplace=True)
else:
self.act = None
if norm is not None:
self.norm = get_norm_layer(norm, out_feats)
else:
self.norm = None
if residual:
self.residual = nn.Linear(in_features=in_feats,
out_features=out_feats)
else:
self.residual = None
def __init__(self,
in_features,
out_features,
dropout=0.0,
activation=None,
residual=False,
norm=None,
bias=True):
super(GCNLayer, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.weight = Parameter(torch.FloatTensor(in_features, out_features))
if dropout > 0:
self.dropout = nn.Dropout(dropout)
else:
self.dropout = None
if residual:
self.residual = nn.Linear(in_features, out_features)
else:
self.residual = None
if activation is not None:
self.act = get_activation(activation)
else:
self.act = None
if norm is not None:
if norm == "batchnorm":
self.norm = nn.BatchNorm1d(out_features)
elif norm == "layernorm":
self.norm = nn.LayerNorm(out_features)
else:
raise NotImplementedError
else:
self.norm = None
if bias:
self.bias = Parameter(torch.FloatTensor(out_features))
else:
self.register_parameter("bias", None)
self.reset_parameters()
def __init__(self,
in_features,
out_features,
dropout=0.0,
activation=None,
residual=False,
norm=None,
bias=True,
**kwargs):
super(GCNLayer, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.linear = nn.Linear(in_features, out_features, bias=bias)
if dropout > 0:
self.dropout = nn.Dropout(dropout)
else:
self.dropout = None
if residual:
self.residual = nn.Linear(in_features, out_features)
else:
self.residual = None
if activation is not None:
self.act = get_activation(activation, inplace=True)
else:
self.act = None
if norm is not None:
if norm == "batchnorm":
self.norm = nn.BatchNorm1d(out_features)
elif norm == "layernorm":
self.norm = nn.LayerNorm(out_features)
else:
raise NotImplementedError
else:
self.norm = None
self.reset_parameters()
def __init__(
self,
conv,
in_channels,
activation="relu",
norm="batchnorm",
dropout=0.0,
out_norm=None,
out_channels=-1,
residual=True,
checkpoint_grad=False,
):
super(ResGNNLayer, self).__init__()
self.conv = conv
self.activation = get_activation(activation)
self.dropout = dropout
self.norm = get_norm_layer(norm, in_channels)
self.residual = residual
if out_norm:
self.out_norm = get_norm_layer(norm, out_channels)
else:
self.out_norm = None
self.checkpoint_grad = checkpoint_grad