def __init__(
self, in_dim, hidden_dim, out_dim, num_heads, dropout,
residual=False, activation=None):
"""GAT模型
:param in_dim: int 输入特征维数
:param hidden_dim: int 隐含特征维数
:param out_dim: int 输出特征维数
:param num_heads: List[int] 每一层的注意力头数，长度等于层数
:param dropout: float Dropout概率
:param residual: bool, optional 是否使用残差连接，默认为False
:param activation: callable, optional 输出层激活函数
:raise ValueError: 如果层数（即num_heads的长度）小于2
"""
super().__init__()
num_layers = len(num_heads)
if num_layers < 2:
raise ValueError('层数至少为2，实际为{}'.format(num_layers))
self.layers = nn.ModuleList()
self.layers.append(GATConv(
in_dim, hidden_dim, num_heads[0], dropout, dropout, residual=residual, activation=F.elu
))
for i in range(1, num_layers - 1):
self.layers.append(GATConv(
num_heads[i - 1] * hidden_dim, hidden_dim, num_heads[i], dropout, dropout,
residual=residual, activation=F.elu
))
self.layers.append(GATConv(
num_heads[-2] * hidden_dim, out_dim, num_heads[-1], dropout, dropout,
residual=residual, activation=activation
))
def __init__(self,
g, #DGL的图对象
n_layers, #层数
in_feats, #输入特征维度
n_hidden, #隐层特征维度
n_classes, #类别数
heads, #多头注意力的数量
activation, #激活函数
in_drop, #输入特征的Dropout比例
at_drop, #注意力特征的Dropout比例
negative_slope, #注意力计算中Leaky ReLU的a值
):
super( GAT, self ).__init__( )
self.g = g
self.num_layers = n_layers
self.activation = activation
self.gat_layers = nn.ModuleList()
self.gat_layers.append( GATConv(
in_feats, n_hidden, heads[0],
in_drop, at_drop, negative_slope, activation=self.activation ) )
for l in range(1, n_layers):
self.gat_layers.append( GATConv(
n_hidden * heads[l-1], n_hidden, heads[l],
in_drop, at_drop, negative_slope, activation=self.activation))
self.gat_layers.append( GATConv(
n_hidden * heads[-2], n_classes, heads[-1],
in_drop, at_drop, negative_slope, activation=None) )
def __init__(self, input_dim, hidden_dim, output_dim, num_layers, gnn_type):
super(GNN, self).__init__()
self.hidden_dim = hidden_dim
self.num_layers = num_layers
self.gnn_type = gnn_type
self.convs = nn.ModuleList()
if gnn_type == "gcn":
self.convs.append(GraphConv(input_dim, hidden_dim))
elif gnn_type == "sage":
self.convs.append(SAGEConv(input_dim, hidden_dim, "gcn"))
elif gnn_type == "gat":
self.convs.append(GATConv(input_dim, hidden_dim, num_heads=3))
else:
raise ValueError("Invalid gnn_type")
for i in range(num_layers - 2):
if gnn_type == "gcn":
self.convs.append(GraphConv(hidden_dim, hidden_dim))
elif gnn_type == "sage":
self.convs.append(SAGEConv(hidden_dim, hidden_dim, "gcn"))
elif gnn_type == "gat":
self.convs.append(GATConv(hidden_dim, hidden_dim, num_heads=3))
if gnn_type == "gcn":
self.convs.append(GraphConv(hidden_dim, output_dim))
elif gnn_type == "sage":
self.convs.append(SAGEConv(hidden_dim, output_dim, "gcn"))
elif gnn_type == "gat":
self.convs.append(GATConv(hidden_dim, output_dim, num_heads=3))
def __init__(self,
num_layers,
in_dim,
num_hidden,
num_classes,
heads,
activation,
feat_drop=0,
attn_drop=0,
negative_slope=0.2,
residual=False):
super(GAT, self).__init__()
self.num_layers = num_layers
self.gat_layers = nn.ModuleList()
self.activation = activation
# input projection (no residual)
self.gat_layers.append(
GATConv(in_dim, num_hidden, heads[0], feat_drop, attn_drop,
negative_slope, False, self.activation))
# hidden layers
for l in range(1, num_layers):
# due to multi-head, the in_dim = num_hidden * num_heads
self.gat_layers.append(
GATConv(num_hidden * heads[l - 1], num_hidden, heads[l],
feat_drop, attn_drop, negative_slope, residual,
self.activation))
# output projection
self.gat_layers.append(
GATConv(num_hidden * heads[-2], num_classes, heads[-1], feat_drop,
attn_drop, negative_slope, residual, None))
def __init__(self, model_config):
super(GAT, self).__init__()
# adj = np.load("/root/workdir/input/fnc_adj.npy")
self.G = dgl.DGLGraph(np.where(np.ones([53, 53]) == 1))
self.linear1 = nn.Linear(400, 256)
self.gat_conv1 = GATConv(256,
32,
num_heads=8,
feat_drop=model_config.dropout_rate,
residual=True)
self.bn1 = nn.BatchNorm1d(64)
self.gat_conv2 = GATConv(256,
32,
num_heads=8,
feat_drop=model_config.dropout_rate,
residual=True)
self.bn2 = nn.BatchNorm1d(64)
self.gat_conv3 = GATConv(256,
32,
num_heads=8,
feat_drop=model_config.dropout_rate,
residual=True)
self.bn3 = nn.BatchNorm1d(64)
self.loading_mlp = nn.Sequential(
nn.Linear(26, 128),
nn.LeakyReLU(),
nn.Dropout(model_config.dropout_rate),
)
self.last_linear = nn.Linear(256 * 3 + 128, model_config.num_classes)
def __init__(self,
num_layers,
in_dim,
num_hidden,
heads,
feat_drop=0,
attn_drop=0,
negative_slope=0.2,
activation=None,
residual=False):
super(GAT, self).__init__()
self.num_layers = num_layers
self.gat_layers = nn.ModuleList()
self.activation = activation
# input projection (no residual)
self.gat_layers.append(
GATConv(in_dim,
num_hidden,
heads[0],
feat_drop,
attn_drop,
negative_slope,
False,
None,
allow_zero_in_degree=True))
def __init__(self, in_dim, hid_dim, out_dim, num_heads=1, residual=False):
super(GeniePathConv, self).__init__()
self.breadth_func = GATConv(in_dim,
hid_dim,
num_heads=num_heads,
residual=residual)
self.depth_func = LSTM(hid_dim, out_dim)
def __init__(self,
data_info: dict,
embed_size: int = -1,
num_layers: int = 2,
hidden_size: int = 8,
heads: List[int] = [8, 8],
activation: str = "elu",
feat_drop: float = 0.6,
attn_drop: float = 0.6,
negative_slope: float = 0.2,
residual: bool = False):
"""Graph Attention Networks
Parameters
----------
data_info : dict
The information about the input dataset.
embed_size : int
The dimension of created embedding table. -1 means using original node embedding
hidden_size : int
Hidden size.
num_layers : int
Number of layers.
norm : str
GCN normalization type. Can be 'both', 'right', 'left', 'none'.
activation : str
Activation function.
feat_drop : float
Dropout rate for features.
attn_drop : float
Dropout rate for attentions.
negative_slope: float
Negative slope for leaky relu in GATConv
residual : bool
If true, the GATConv will use residule connection
"""
super(GAT, self).__init__()
self.data_info = data_info
self.embed_size = embed_size
self.num_layers = num_layers
self.gat_layers = nn.ModuleList()
self.activation = getattr(torch.nn.functional, activation)
if embed_size > 0:
self.embed = nn.Embedding(data_info["num_nodes"], embed_size)
in_size = embed_size
else:
in_size = data_info["in_size"]
for i in range(num_layers):
in_hidden = hidden_size * heads[i - 1] if i > 0 else in_size
out_hidden = hidden_size if i < num_layers - \
1 else data_info["out_size"]
use_residual = i == num_layers
activation = None if i == num_layers else self.activation
self.gat_layers.append(
GATConv(in_hidden, out_hidden, heads[i], feat_drop, attn_drop,
negative_slope, use_residual, activation))
def __init__(self,
in_dim,
hidden_dim,
layer_num,
sample_size,
device,
GNN_name="GIN"):
super(GNNStructEncoder, self).__init__()
self.n_distribution = 7 # How many gaussian distribution should exist
self.out_dim = hidden_dim
if GNN_name == "GIN":
self.linear1 = MLP(layer_num, hidden_dim, hidden_dim, hidden_dim)
self.graphconv1 = GINConv(apply_func=self.linear1,
aggregator_type='sum')
self.linear2 = MLP(layer_num, hidden_dim, hidden_dim, hidden_dim)
self.graphconv2 = GINConv(apply_func=self.linear2,
aggregator_type='sum')
elif GNN_name == "GCN":
self.graphconv1 = GraphConv(hidden_dim, hidden_dim)
self.graphconv2 = GraphConv(hidden_dim, hidden_dim)
else:
self.graphconv = GATConv(hidden_dim, hidden_dim, num_heads=10)
# self.neighbor_num_list = neighbor_num_list
self.linear_classifier = MLP(1, hidden_dim, hidden_dim,
self.n_distribution)
self.neighbor_generator = MLP_generator(hidden_dim, hidden_dim,
sample_size).to(device)
# Gaussian Means, and std
self.gaussian_mean = nn.Parameter(
torch.FloatTensor(sample_size, self.n_distribution,
hidden_dim).uniform_(-0.5 / hidden_dim, 0.5 /
hidden_dim)).to(device)
self.gaussian_log_sigma = nn.Parameter(
torch.FloatTensor(sample_size, self.n_distribution,
hidden_dim).uniform_(-0.5 / hidden_dim, 0.5 /
hidden_dim)).to(device)
self.m = torch.distributions.Normal(
torch.zeros(sample_size, self.n_distribution, hidden_dim),
torch.ones(sample_size, self.n_distribution, hidden_dim))
# Before MLP Gaussian Means, and std
self.mlp_gaussian_mean = nn.Parameter(
torch.FloatTensor(hidden_dim).uniform_(-0.5 / hidden_dim, 0.5 /
hidden_dim)).to(device)
self.mlp_gaussian_log_sigma = nn.Parameter(
torch.FloatTensor(hidden_dim).uniform_(-0.5 / hidden_dim, 0.5 /
hidden_dim)).to(device)
self.mlp_m = torch.distributions.Normal(torch.zeros(hidden_dim),
torch.ones(hidden_dim))
# Decoders
self.degree_decoder = FNN(hidden_dim, hidden_dim, 1, 4)
# self.degree_loss_func = FocalLoss(int(max_degree_num) + 1)
self.degree_loss_func = nn.MSELoss()
self.pool = mp.Pool(1)
self.in_dim = in_dim
self.sample_size = sample_size
def __init__(self, d_model, d_inner, dropout=0.1):
super(GATEncoder, self).__init__()
# self.proj1 = nn.Linear(d_model, d_model)
# self.proj2 = nn.Linear(d_inner, d_model)
# self.layer_norm = LayerNormalization(d_model)
self.proj1 = GATConv(d_model, d_model, num_heads=3)
init.xavier_normal_(self.proj1.weight)
def __init__(self, G, hid_dims, num_layers, num_heads=3, multihot=True):
super().__init__(G, hid_dims, num_layers, multihot)
self.num_heads = num_heads
self.input_layer = GATConv(self.in_dims,
hid_dims,
num_heads=num_heads,
residual=True)
self.hidden_layers = [
GATConv(hid_dims * num_heads,
hid_dims,
num_heads=num_heads,
residual=True) for _ in range(num_layers)
]
self.output_layer = GATConv(hid_dims * num_heads,
self.out_dims,
num_heads=1,
residual=True)
def __init__(self, in_dim, out_dim, hid_dim=16, num_layers=2, num_heads=1, residual=False):
super(GeniePathLazy, self).__init__()
self.hid_dim = hid_dim
self.linear1 = nn.Linear(in_dim, hid_dim)
self.linear2 = th.nn.Linear(hid_dim, out_dim)
self.breaths = nn.ModuleList()
self.depths = nn.ModuleList()
for i in range(num_layers):
self.breaths.append(GATConv(hid_dim, hid_dim, num_heads=num_heads, residual=residual))
self.depths.append(LSTM(hid_dim*2, hid_dim))
def __init__(self, num_metapaths, in_dim, out_dim, num_heads, dropout):
"""HAN层
:param num_metapaths: int 元路径个数
:param in_dim: int 输入特征维数
:param out_dim: int 输出特征维数
:param num_heads: int 注意力头数K
:param dropout: float Dropout概率
"""
super().__init__()
# 顶点层次的注意力，每个GAT层对应一个元路径
self.gats = nn.ModuleList([
GATConv(in_dim, out_dim, num_heads, dropout, dropout, activation=F.elu)
for _ in range(num_metapaths)
])
# 语义层次的注意力
self.semantic_attention = SemanticAttention(in_dim=num_heads * out_dim)
def __init__(self, in_dim, hidden_dim, device):
super(GNNAnomalyDetctor, self).__init__()
self.in_dim = in_dim
self.gatconv = GATConv(in_dim, hidden_dim, num_heads=3)
self.nwr_gae = GNNStructEncoder(in_dim, hidden_dim, layer_num=2, sample_size=5, device=device)
self.linear = torch.nn.Linear(10, 10)
def __init__(self):
super(Net, self).__init__()
self.layer1 = NNConv(300, 150,edge_f, 'mean')
self.layer2 = GATConv()
def __init__(self, nfeat, nhid, nclass=2, dropout=False):
super(GAT, self).__init__()
self.conv1 = GATConv(nfeat, nhid, num_heads=4)
self.conv2 = GATConv(4 * nhid, nclass, num_heads=1)
def __init__(self, in_dim, out_dim, num_heads=4, is_final_layer=False):
super(GAT_Layer, self).__init__()
self.gat = GATConv(in_dim, out_dim, num_heads=num_heads)
self.batchnorm = nn.BatchNorm1d(out_dim * num_heads)
self.activation = nn.LeakyReLU()
self.is_final_layer = is_final_layer
def __init__(self, in_feats, h_feats):
super(GAT, self).__init__()
self.conv1 = GATConv(in_feats, h_feats // 8, 8)
self.conv2 = GATConv(h_feats, h_feats// 8, 8)
