def __init__(self, in_feats, hid_feats, out_feats, rel_names):
super().__init__()
self.conv1 = HeteroGraphConv({
rel: GraphConv(in_feats, hid_feats) for rel in rel_names
}, aggregate='sum')
self.conv2 = HeteroGraphConv({
rel: GraphConv(hid_feats, out_feats) for rel in rel_names
}, aggregate='sum')
def __init__(self,
in_dim,
out_dim,
ntypes,
etypes,
activation=None,
dropout=0.0):
"""R-GCN层（用于异构图）
:param in_dim: 输入特征维数
:param out_dim: 输出特征维数
:param ntypes: List[str] 顶点类型列表
:param etypes: List[str] 边类型列表
:param activation: callable, optional 激活函数，默认为None
:param dropout: float, optional Dropout概率，默认为0
"""
super().__init__()
self.activation = activation
self.dropout = nn.Dropout(dropout)
self.conv = HeteroGraphConv(
{
etype: GraphConv(in_dim, out_dim, norm='right', bias=False)
for etype in etypes
}, 'sum')
self.loop_weight = nn.ModuleDict({
ntype: nn.Linear(in_dim, out_dim, bias=False)
for ntype in ntypes
})
def __init__(self,
in_dim,
out_dim,
num_heads,
ntypes,
etypes,
dropout=0.2,
use_norm=True):
"""HGT层
: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 ntypes: List[str] 顶点类型列表
:param etypes: List[(str, str, str)] 规范边类型列表
:param dropout: dropout: float, optional Dropout概率，默认为0.2
:param use_norm: bool, optional 是否使用层归一化，默认为True
"""
super().__init__()
d_k = out_dim // num_heads
k_linear = {ntype: nn.Linear(in_dim, out_dim) for ntype in ntypes}
q_linear = {ntype: nn.Linear(in_dim, out_dim) for ntype in ntypes}
v_linear = {ntype: nn.Linear(in_dim, out_dim) for ntype in ntypes}
w_att = {
r[1]: nn.Parameter(torch.Tensor(num_heads, d_k, d_k))
for r in etypes
}
w_msg = {
r[1]: nn.Parameter(torch.Tensor(num_heads, d_k, d_k))
for r in etypes
}
mu = {r[1]: nn.Parameter(torch.ones(num_heads)) for r in etypes}
self.reset_parameters(w_att, w_msg)
self.conv = HeteroGraphConv(
{
etype: HGTAttention(out_dim, num_heads, k_linear[stype],
q_linear[dtype], v_linear[stype],
w_att[etype], w_msg[etype], mu[etype])
for stype, etype, dtype in etypes
}, 'mean')
self.a_linear = nn.ModuleDict(
{ntype: nn.Linear(out_dim, out_dim)
for ntype in ntypes})
self.skip = nn.ParameterDict(
{ntype: nn.Parameter(torch.ones(1))
for ntype in ntypes})
self.drop = nn.Dropout(dropout)
self.use_norm = use_norm
if use_norm:
self.norms = nn.ModuleDict(
{ntype: nn.LayerNorm(out_dim)
for ntype in ntypes})
def __init__(self,
in_dim,
out_dim,
rel_names,
num_bases=None,
weight=True,
self_loop=True,
activation=None,
dropout=0.0):
"""R-GCN层（用于异构图）
:param in_dim: 输入特征维数
:param out_dim: 输出特征维数
:param rel_names: List[str] 关系名称
:param num_bases: int, optional 基的个数，默认使用关系个数
:param weight: bool, optional 是否进行线性变换，默认为True
:param self_loop: 是否包括自环消息，默认为True
:param activation: callable, optional 激活函数，默认为None
:param dropout: float, optional Dropout概率，默认为0
"""
super().__init__()
self.rel_names = rel_names
self.self_loop = self_loop
self.activation = activation
self.dropout = nn.Dropout(dropout)
self.conv = HeteroGraphConv({
rel: GraphConv(in_dim,
out_dim,
norm='right',
weight=False,
bias=False)
for rel in rel_names
})
self.use_weight = weight
if not num_bases:
num_bases = len(rel_names)
self.use_basis = weight and 0 < num_bases < len(rel_names)
if self.use_weight:
if self.use_basis:
self.basis = WeightBasis((in_dim, out_dim), num_bases,
len(rel_names))
else:
self.weight = nn.Parameter(
torch.Tensor(len(rel_names), in_dim, out_dim))
nn.init.xavier_uniform_(self.weight,
nn.init.calculate_gain('relu'))
if self.self_loop:
self.loop_weight = nn.Parameter(torch.Tensor(in_dim, out_dim))
nn.init.xavier_uniform_(self.loop_weight,
nn.init.calculate_gain('relu'))