def forward(self, g, h, h_en):
"""Forward computation
"""
with g.local_scope():
h_src, h_dst = expand_as_pair(h)
h_src_en, h_dst_en = expand_as_pair(h_en)
g.srcdata['x'] = h_src
g.dstdata['x'] = h_dst
g.srcdata['en'] = h_src_en
g.dstdata['en'] = h_dst_en
if not self.batch_norm:
#g.update_all(self.message, fn.mean('e', 'x'))
g.apply_edges(self.message)
g.update_all(fn.copy_e('e', 'e'), fn.max('e', 'x'))
g.update_all(fn.copy_e('e_en', 'e_en'), fn.mean('e_en', 'en'))
else:
g.apply_edges(self.message)
g.edata['e'] = self.bn(g.edata['e'])
g.update_all(fn.copy_e('e', 'e'), fn.max('e', 'x'))
g.update_all(fn.copy_e('e_en', 'e_en'), fn.mean('e_en', 'en'))
return g.dstdata['x'], g.dstdata['en'] #+ h_en
def forward(self, g, feat):
"""
:param g: DGLGraph 二分图（只包含一种关系）
:param feat: tensor(N_src, d_in) or (tensor(N_src, d_in), tensor(N_dst, d_in)) 输入特征
:return: tensor(N_dst, K*d_out) 该关系关于目标顶点的表示
"""
with g.local_scope():
feat_src, feat_dst = expand_as_pair(feat, g)
feat_src = self.fc_src(self.feat_drop(feat_src)).view(-1, self.num_heads, self.out_dim)
feat_dst = self.fc_dst(self.feat_drop(feat_dst)).view(-1, self.num_heads, self.out_dim)
# a^T (z_u || z_v) = (a_l^T || a_r^T) (z_u || z_v) = a_l^T z_u + a_r^T z_v = el + er
el = (feat_src * self.attn_src[:, :self.out_dim]).sum(dim=-1, keepdim=True) # (N_src, K, 1)
er = (feat_dst * self.attn_src[:, self.out_dim:]).sum(dim=-1, keepdim=True) # (N_dst, K, 1)
g.srcdata.update({'ft': feat_src, 'el': el})
g.dstdata['er'] = er
g.apply_edges(fn.u_add_v('el', 'er', 'e'))
e = self.leaky_relu(g.edata.pop('e'))
g.edata['a'] = edge_softmax(g, e) # (E, K, 1)
# 消息传递
g.update_all(fn.u_mul_e('ft', 'a', 'm'), fn.sum('m', 'ft'))
ret = g.dstdata['ft'].view(-1, self.num_heads * self.out_dim)
if self.activation:
ret = self.activation(ret)
return ret
def __init__(self,
in_feats,
out_feats,
aggregator_type,
feat_drop=0.,
bias=True,
norm=None,
activation=None):
super(SAGEConvMLP, self).__init__()
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._aggre_type = aggregator_type
self.norm = norm
self.feat_drop = nn.Dropout(feat_drop)
self.activation = activation
# aggregator type: mean/pool/lstm/gcn
if aggregator_type == 'pool':
self.fc_pool = nn.Linear(self._in_src_feats, self._in_src_feats)
if aggregator_type == 'lstm':
self.lstm = nn.LSTM(self._in_src_feats,
self._in_src_feats,
batch_first=True)
if aggregator_type != 'gcn':
self.fc_self = nn.Linear(self._in_dst_feats, out_feats, bias=bias)
self.fc_neigh = nn.Linear(self._in_src_feats, out_feats, bias=bias)
self.reset_parameters()
def __init__(self,
in_feats: int,
attn_feats: int,
out_feats: int,
num_heads: int,
feat_drop: float = 0.,
attn_drop: float = 0.,
negative_slope: float = 0.2,
allow_zero_in_degree: bool = False,
is_increasing: bool = True,
reparam_method: str = 'ReLU'):
super(GATConv, self).__init__()
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self.fc = ReparameterizedLinear(in_features=self._in_src_feats,
out_features=out_feats * num_heads,
bias=False,
is_increasing=is_increasing,
reparam_method=reparam_method)
self.attn_l = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
self.fc_attn = nn.Linear(in_features=attn_feats,
out_features=out_feats * num_heads,
bias=False)
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
self.reset_parameters()
def forward(self, graph, feat):
"""Compute Graph Isomorphism Network layer.
Parameters
----------
graph : DGLGraph
The graph.
feat : torch.Tensor or pair of torch.Tensor
If a torch.Tensor is given, the input feature of shape :math:`(N, D_{in})` where
:math:`D_{in}` is size of input feature, :math:`N` is the number of nodes.
If a pair of torch.Tensor is given, the pair must contain two tensors of shape
:math:`(N_{in}, D_{in})` and :math:`(N_{out}, D_{in})`.
If ``apply_func`` is not None, :math:`D_{in}` should
fit the input dimensionality requirement of ``apply_func``.
Returns
-------
torch.Tensor
The output feature of shape :math:`(N, D_{out})` where
:math:`D_{out}` is the output dimensionality of ``apply_func``.
If ``apply_func`` is None, :math:`D_{out}` should be the same
as input dimensionality.
"""
graph = graph.local_var()
feat_src, feat_dst = expand_as_pair(feat)
graph.srcdata['h'] = feat_src
graph.send_and_recv(graph.edges(), fn.copy_u('h', 'm'),
self._reducer('m', 'neigh'))
rst = (1 + self.eps) * feat_dst + graph.dstdata['neigh']
if self.apply_func is not None:
rst = self.apply_func(rst)
return rst
def forward(self, g, feat):
"""
:param g: DGLGraph 二分图（只包含一种关系）
:param feat: tensor(N_src, d_in) or (tensor(N_src, d_in), tensor(N_dst, d_in)) 输入特征
:return: tensor(N_dst, d_out) 目标顶点该关于关系的表示
"""
with g.local_scope():
feat_src, feat_dst = expand_as_pair(feat, g)
# (N_src, d_in) -> (N_src, d_out) -> (N_src, K, d_out/K)
k = self.k_linear(feat_src).view(-1, self.num_heads, self.d_k)
v = self.v_linear(feat_src).view(-1, self.num_heads, self.d_k)
q = self.q_linear(feat_dst).view(-1, self.num_heads, self.d_k)
# k[:, h] @= w_att[h] => k[n, h, j] = ∑(i) k[n, h, i] * w_att[h, i, j]
k = torch.einsum('nhi,hij->nhj', k, self.w_att)
v = torch.einsum('nhi,hij->nhj', v, self.w_msg)
g.srcdata.update({'k': k, 'v': v})
g.dstdata['q'] = q
g.apply_edges(fn.v_dot_u('q', 'k', 't')) # g.edata['t']: (E, K, 1)
attn = g.edata.pop('t').squeeze(dim=-1) * self.mu / math.sqrt(
self.d_k)
attn = edge_softmax(g, attn) # (E, K)
self.attn = attn.detach()
g.edata['t'] = attn.unsqueeze(dim=-1) # (E, K, 1)
g.update_all(fn.u_mul_e('v', 't', 'm'), fn.sum('m', 'h'))
out = g.dstdata['h'].view(-1, self.out_dim) # (N_dst, d_out)
return out
def forward(self, graph, feat, edge_weight=None):
with graph.local_scope():
aggregate_fn = fn.copy_src('h', 'm')
if edge_weight is not None:
#assert edge_weight.shape[0] == graph.number_of_edges()
graph.edata['_edge_weight'] = edge_weight
aggregate_fn = fn.u_sub_e('h', '_edge_weight', 'm')
feat_src, feat_dst = expand_as_pair(feat, graph)
# aggregate first then mult W
graph.srcdata['h'] = feat_src
graph.apply_edges(
lambda edges: {'h': edges.data['_edge_weight'] * 2})
graph.update_all(aggregate_fn, fn.sum(msg='m', out='h'))
rst = graph.dstdata['h']
rst = th.matmul(feat_src, self.weight1) + th.matmul(
rst, self.weight2)
if self.activation is not None:
rst = self.activation(rst)
return rst
def forward(self, graph, feat, weight=None, alpha=None, gene_num=None):
self.alpha = alpha
self.gene_num = gene_num
with graph.local_scope():
if not self._allow_zero_in_degree:
if (graph.in_degrees() == 0).any():
raise DGLError('There are 0-in-degree nodes in the graph, '
'output for those nodes will be invalid. '
'This is harmful for some applications, '
'causing silent performance regression. '
'Adding self-loop on the input graph by '
'calling `g = dgl.add_self_loop(g)` will resolve '
'the issue. Setting ``allow_zero_in_degree`` '
'to be `True` when constructing this module will '
'suppress the check and let the code run.')
# (BarclayII) For RGCN on heterogeneous graphs we need to support GCN on bipartite.
feat_src, feat_dst = expand_as_pair(feat, graph)
# print(f"feat_src : {feat_src.shape}, feat_dst {feat_dst.shape}")
if self._norm == 'both':
degs = graph.out_degrees().float().clamp(min=1)
norm = torch.pow(degs, -0.5)
shp = norm.shape + (1,) * (feat_src.dim() - 1)
norm = torch.reshape(norm, shp)
feat_src = feat_src * norm
if weight is not None:
if self.weight is not None:
raise DGLError('External weight is provided while at the same time the'
' module has defined its own weight parameter. Please'
' create the module with flag weight=False.')
else:
weight = self.weight
if weight is not None:
feat_src = torch.matmul(feat_src, weight)
graph.srcdata['h'] = feat_src
graph.update_all(self.edge_selection_simple,
fn.sum(msg='m', out='h'))
rst = graph.dstdata['h']
if self._norm != 'none':
degs = graph.in_degrees().float().clamp(min=1)
if self._norm == 'both':
norm = torch.pow(degs, -0.5)
else:
norm = 1.0 / degs
shp = norm.shape + (1,) * (feat_dst.dim() - 1)
norm = torch.reshape(norm, shp)
rst = rst * norm
if self.bias is not None:
rst = rst + self.bias
if self._activation is not None:
rst = self._activation(rst)
return rst
def __init__(
self,
edge_feats,
num_etypes,
in_feats,
out_feats,
num_heads,
feat_drop=0.0,
attn_drop=0.0,
negative_slope=0.2,
residual=False,
activation=None,
allow_zero_in_degree=False,
bias=False,
alpha=0.0,
):
super(myGATConv, self).__init__()
self._edge_feats = edge_feats
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self.edge_emb = nn.Embedding(num_etypes, edge_feats)
if isinstance(in_feats, tuple):
self.fc_src = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_dst = nn.Linear(self._in_dst_feats,
out_feats * num_heads,
bias=False)
else:
self.fc = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_e = nn.Linear(edge_feats, edge_feats * num_heads, bias=False)
self.attn_l = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_e = nn.Parameter(
th.FloatTensor(size=(1, num_heads, edge_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer("res_fc", None)
self.reset_parameters()
self.activation = activation
self.bias = bias
if bias:
self.bias_param = nn.Parameter(th.zeros((1, num_heads, out_feats)))
self.alpha = alpha
def __init__(self, in_feats, out_feats):
super(SAGEConv, self).__init__()
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self.fc_self = nn.Linear(self._in_dst_feats, out_feats, bias=False)
self.fc_neigh = nn.Linear(self._in_src_feats, out_feats)
self.reset_parameters()
def forward(self, graph, feat, weight=None, edge_weight=None):
with graph.local_scope():
if not self._allow_zero_in_degree:
if (graph.in_degrees() == 0).any():
raise DGLError(
'There are 0-in-degree nodes in the graph, '
'output for those nodes will be invalid. '
'This is harmful for some applications, '
'causing silent performance regression. '
'Adding self-loop on the input graph by '
'calling `g = dgl.add_self_loop(g)` will resolve '
'the issue. Setting ``allow_zero_in_degree`` '
'to be `True` when constructing this module will '
'suppress the check and let the code run.')
#aggregate_fn = fn.copy_src('h', 'm')
if edge_weight is not None:
assert edge_weight.shape[0] == graph.number_of_edges()
graph.edata['_edge_weight'] = edge_weight
aggregate_fn = fn.u_mul_e('h', '_edge_weight', 'm')
# (BarclayII) For RGCN on heterogeneous graphs we need to support GCN on bipartite.
feat_src, feat_dst = expand_as_pair(feat, graph)
if self._norm == 'both':
degs = graph.out_degrees().float().clamp(min=1)
norm = th.pow(degs, -0.5)
shp = norm.shape + (1, ) * (feat_src.dim() - 1)
norm = th.reshape(norm, shp)
feat_src = feat_src * norm
feat_sumsrc = th.matmul(feat_src, self.w1)
feat_prodsrc = (th.matmul(
th.cat(
(feat_src, th.ones([feat_src.shape[0], 1]).to('cuda:0')),
1), self.w2))
graph.srcdata['h_sum'] = feat_sumsrc
graph.srcdata['h_prod'] = feat_prodsrc
graph.update_all(fn.copy_src('h_sum', 'm_sum'),
self._elementwise_sum)
graph.update_all(fn.copy_src('h_prod', 'm_prod'),
self._elementwise_product)
rst = graph.dstdata['h_sum'] + th.matmul(graph.dstdata['h_prod'],
self.v)
if self._norm != 'none':
degs = graph.in_degrees().float().clamp(min=1)
if self._norm == 'both':
norm = th.pow(degs, -0.5)
else:
norm = 1.0 / degs
shp = norm.shape + (1, ) * (feat_dst.dim() - 1)
norm = th.reshape(norm, shp)
rst = rst * norm
if self._activation is not None:
rst = self._activation(rst)
return rst
def __init__(self,
in_feats,
out_feats,
num_heads,
num_nodes,
layerid,
feat_drop=0.,
attn_drop=0.,
negative_slope=0.2,
residual=False,
activation=None,
allow_zero_in_degree=False,
fix=False):
super(GATConvs, self).__init__()
self._num_heads = num_heads
self.num_nodes = num_nodes
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self.layerid = layerid
if isinstance(in_feats, tuple):
self.fc_src = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_dst = nn.Linear(self._in_dst_feats,
out_feats * num_heads,
bias=False)
else:
self.fc1 = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc2 = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc3 = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.attn_l = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer('res_fc', None)
self.reset_parameters()
self.activation = activation
self.fix = fix
def __init__(self,
in_feats,
out_feats,
K=3,
num_heads=1,
feat_drop=0.0,
edge_drop=0.0,
attn_drop=0.0,
negative_slope=0.2,
use_attn_dst=True,
residual=False,
activation=None,
allow_zero_in_degree=False,
norm='sym'):
super(GATHAConv, self).__init__()
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self._K = K
self._norm = norm
self.attn_l = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
if use_attn_dst:
self.attn_r = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
else:
self.register_buffer("attn_r", None)
self.fc = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.position_emb = nn.Parameter(
torch.FloatTensor(size=(K + 1, num_heads, out_feats)))
self.hop_attn_l = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
self.hop_attn_r = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
self.hop_attn_bias_l = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, 1)))
self.hop_attn_bias_r = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, 1)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.edge_drop = edge_drop
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.register_buffer("res_fc", None)
self.reset_parameters()
self._activation = activation
def __init__(
self,
in_feats,
# out_feats,
aggregator_type):
super(SAGEConvAggLoc, self).__init__()
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._aggre_type = aggregator_type
def __init__(self,
in_feats,
out_feats,
num_heads,
basis,
attn_basis,
basis_coef,
feat_drop=0.,
attn_drop=0.,
negative_slope=0.2,
residual=False,
activation=None,
allow_zero_in_degree=False,
bias=True,
use_checkpoint=False):
super(GATConv, self).__init__()
self._basis = basis
self._basis_coef = basis_coef
self._attn_basis = attn_basis
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
# if isinstance(in_feats, tuple):
# self.fc_src = nn.Linear(
# self._in_src_feats, out_feats * num_heads, bias=False)
# self.fc_dst = nn.Linear(
# self._in_dst_feats, out_feats * num_heads, bias=False)
# else:
# self.fc = nn.Linear(
# self._in_src_feats, out_feats * num_heads, bias=False)
# self.attn_l = nn.Parameter(th.FloatTensor(size=(1, num_heads, out_feats)))
# self.attn_r = nn.Parameter(th.FloatTensor(size=(1, num_heads, out_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if bias:
self.bias = nn.Parameter(th.FloatTensor(size=(num_heads * out_feats,)))
else:
self.register_buffer('bias', None)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(
self._in_dst_feats, num_heads * out_feats, bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer('res_fc', None)
self.reset_parameters()
self.activation = activation
self.dummy_tensor = th.ones(1, dtype=th.float32, requires_grad=True)
self.use_checkpoint = use_checkpoint
def __init__(
self,
in_feats,
out_feats,
num_heads=1,
feat_drop=0.0,
attn_drop=0.0,
edge_drop=0.0,
negative_slope=0.2,
use_attn_dst=True,
residual=False,
activation=None,
allow_zero_in_degree=False,
use_symmetric_norm=False,
):
super(GATConv, self).__init__()
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self._use_symmetric_norm = use_symmetric_norm
if isinstance(in_feats, tuple):
self.fc_src = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_dst = nn.Linear(self._in_dst_feats,
out_feats * num_heads,
bias=False)
else:
self.fc = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.attn_l = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
if use_attn_dst:
self.attn_r = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
else:
self.register_buffer("attn_r", None)
self.feat_drop = nn.Dropout(feat_drop)
assert feat_drop == 0.0 # not implemented
self.attn_drop = nn.Dropout(attn_drop)
assert attn_drop == 0.0 # not implemented
self.edge_drop = edge_drop
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.register_buffer("res_fc", None)
self.reset_parameters()
self._activation = activation
def forward(self, graph, feat, weight=None, edge_weight=None):
with graph.local_scope():
aggregate_fn = fn.copy_src('h', 'm')
if edge_weight is not None:
assert edge_weight.shape[0] == graph.number_of_edges()
graph.edata['_edge_weight'] = edge_weight
aggregate_fn = fn.u_mul_e('h', '_edge_weight', 'm')
# (BarclayII) For RGCN on heterogeneous graphs we need to support GCN on bipartite.
feat_src, feat_dst = expand_as_pair(feat, graph)
if self._norm == 'both':
degs = graph.out_degrees().float().clamp(min=1)
norm = torch.pow(degs, -0.5)
shp = norm.shape + (1,) * (feat_src.dim() - 1)
norm = torch.reshape(norm, shp)
feat_src = feat_src * norm
weight = self.weight
if self._in_feats > self._out_feats:
# mult W first to reduce the feature size for aggregation.
if weight is not None:
feat_src = torch.tanh(torch.matmul(feat_src, weight))
graph.srcdata['h'] = feat_src
graph.update_all(aggregate_fn, self._elementwise_product)
rst = graph.dstdata['h']
else:
# aggregate first then mult W
graph.srcdata['h'] = feat_src
graph.update_all(aggregate_fn, self._elementwise_product)
rst = graph.dstdata['h']
if weight is not None:
rst = torch.matmul(rst, weight)
if self._norm != 'none':
degs = graph.in_degrees().float().clamp(min=1)
if self._norm == 'both':
norm = torch.pow(degs, -0.5)
else:
norm = 1.0 / degs
shp = norm.shape + (1,) * (feat_dst.dim() - 1)
norm = torch.reshape(norm, shp)
rst = rst * norm
if not self._out:
rst = torch.matmul(rst, self.weight2)
if self._activation is not None:
rst = self._activation(rst)
return rst
def __init__(
self,
node_feats,
edge_feats,
out_feats,
n_heads=1,
attn_drop=0.0,
edge_drop=0.0,
negative_slope=0.2,
residual=True,
activation=None,
use_attn_dst=True,
allow_zero_in_degree=True,
use_symmetric_norm=False,
):
super(GATConv, self).__init__()
self._n_heads = n_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(node_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self._use_symmetric_norm = use_symmetric_norm
# feat fc
self.src_fc = nn.Linear(self._in_src_feats,
out_feats * n_heads,
bias=False)
if residual:
self.dst_fc = nn.Linear(self._in_src_feats, out_feats * n_heads)
self.bias = None
else:
self.dst_fc = None
self.bias = nn.Parameter(out_feats * n_heads)
# attn fc
self.attn_src_fc = nn.Linear(self._in_src_feats, n_heads, bias=False)
if use_attn_dst:
self.attn_dst_fc = nn.Linear(self._in_src_feats,
n_heads,
bias=False)
else:
self.attn_dst_fc = None
if edge_feats > 0:
self.attn_edge_fc = nn.Linear(edge_feats, n_heads, bias=False)
else:
self.attn_edge_fc = None
self.attn_drop = nn.Dropout(attn_drop)
self.edge_drop = edge_drop
self.leaky_relu = nn.LeakyReLU(negative_slope, inplace=True)
self.activation = activation
self.reset_parameters()
def __init__(self,
in_feats,
out_feats,
num_heads,
n_nodes,
local_n_nodes,
apply_gather=False,
no_remote=True,
feat_drop=0.,
attn_drop=0.,
negative_slope=0.2,
residual=False,
activation=None):
super(GATConv, self).__init__()
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._n_nodes = n_nodes
self._local_n_nodes = local_n_nodes
self._no_remote = no_remote
self._apply_gather = apply_gather
if isinstance(in_feats, tuple):
assert False # FIXME
self.fc_src = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_dst = nn.Linear(self._in_dst_feats,
out_feats * num_heads,
bias=False)
else:
self.fc = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.attn_l = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer('res_fc', None)
self.reset_parameters()
self.activation = activation
def __init__(
self,
in_feats,
out_feats,
num_heads=1,
feat_drop=0.0,
attn_drop=0.0,
negative_slope=0.2,
residual=False,
activation=None,
allow_zero_in_degree=False,
norm="none",
):
super(GATConv, self).__init__()
if norm not in ("none", "both"):
raise DGLError('Invalid norm value. Must be either "none", "both".'
' But got "{}".'.format(norm))
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
self._norm = norm
if isinstance(in_feats, tuple):
self.fc_src = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_dst = nn.Linear(self._in_dst_feats,
out_feats * num_heads,
bias=False)
else:
self.fc = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.attn_l = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(
torch.FloatTensor(size=(1, num_heads, out_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer("res_fc", None)
self.reset_parameters()
self._activation = activation
def forward(self, graph, feat):
r"""Compute Graph Isomorphism Network layer.
Parameters
----------
graph : DGLGraph
The graph.
feat : torch.Tensor or pair of torch.Tensor
If a torch.Tensor is given, the input feature of shape :math:`(N, D_{in})` where
:math:`D_{in}` is size of input feature, :math:`N` is the number of nodes.
If a pair of torch.Tensor is given, the pair must contain two tensors of shape
:math:`(N_{in}, D_{in})` and :math:`(N_{out}, D_{in})`.
If ``apply_func`` is not None, :math:`D_{in}` should
fit the input dimensionality requirement of ``apply_func``.
Returns
-------
torch.Tensor
The output feature of shape :math:`(N, D_{out})` where
:math:`D_{out}` is the output dimensionality of ``apply_func``.
If ``apply_func`` is None, :math:`D_{out}` should be the same
as input dimensionality.
"""
graph = graph.local_var()
dgl_context = dgl.utils.to_dgl_context(feat.device)
g = graph._graph.get_immutable_gidx(dgl_context)
feat_src, feat_dst = expand_as_pair(feat)
with self.cm.zoomIn(namespace=[self, torch],
graph=g,
node_feats={
'fsrc': feat_src,
'fdst': feat_dst
}) as v:
if self.aggregator_type == 'sum':
rst = sum([nb.fsrc for nb in v.innbs])
elif self.aggregator_type == 'mean':
rst = self.cm.mean([nb.fsrc for nb in v.innbs])
elif self.aggregator_type == 'max':
rst = self.cm.max([nb.fsrc for nb in v.innbs])
else:
raise NotImplementedError('Cannot find aggregator typoe',
self.aggregator_type)
# Temp workaround for rst = (1 + self.eps) * v.fdst + rst
rst = v.fdst + self.eps * v.fdst + rst
self.cm.collect_output(rst)
rst = self.cm.zoomOut()
if self.apply_func is not None:
rst = self.apply_func(rst)
return rst
def __init__(self,
in_feats,
out_feats,
num_heads,
feat_drop=0.,
attn_drop=0.,
negative_slope=0.2,
residual=True,
allow_zero_in_degree=False):
super(GATConvFF, self).__init__()
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._allow_zero_in_degree = allow_zero_in_degree
if isinstance(in_feats, tuple):
self.fc_src = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.fc_dst = nn.Linear(self._in_dst_feats,
out_feats * num_heads,
bias=False)
else:
self.fc = nn.Linear(self._in_src_feats,
out_feats * num_heads,
bias=False)
self.attn_l = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(
th.FloatTensor(size=(1, num_heads, out_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(self._in_dst_feats,
num_heads * out_feats,
bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer('res_fc', None)
self.reset_parameters()
self.layer_norm = th.nn.LayerNorm(out_feats * num_heads)
self.ff_layer_norm = th.nn.LayerNorm(out_feats * num_heads)
self.activation = PositionwiseFeedForward(
model_dim=out_feats * num_heads,
d_hidden=4 * out_feats * num_heads)
def __init__(self,
in_feats,
out_feats,
aggr,
feat_drop=0.,
activation=None):
super(SAGEConv, self).__init__()
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._aggr = aggr
self.feat_drop = nn.Dropout(feat_drop)
self.activation = activation
self.fc_self = nn.Linear(self._in_dst_feats, out_feats, bias=False)
self.fc_neigh = nn.Linear(self._in_src_feats, out_feats)
self.reset_parameters()
def __init__(self,
in_feats,
out_feats,
edata_channels):
super(EGATLayer, self).__init__()
self._edata_channels = edata_channels
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self.fc = nn.Linear(self._in_src_feats, out_feats * edata_channels, bias=False)
self.edge_fc = nn.Linear(self._edata_channels, self._edata_channels, bias=False)
self.nfeat_with_e_fc = nn.Linear(out_feats+1, out_feats , bias=False)
self.attn_l = nn.Parameter(th.FloatTensor(size=(1, edata_channels, out_feats)))
self.attn_r = nn.Parameter(th.FloatTensor(size=(1, edata_channels, out_feats)))
self.leaky_relu = nn.LeakyReLU(0.2)
self.reset_parameters()
def forward(self, graph, feat, edge_weight):
with graph.local_scope():
feat_src, feat_dst = expand_as_pair(feat, graph)
graph.srcdata['h'] = feat_src
graph.srcdata['norm_h'] = F.normalize(feat_src, p=2, dim=-1)
if isinstance(feat, tuple) or graph.is_block:
graph.dstdata['norm_h'] = F.normalize(feat_dst, p=2, dim=-1)
e = self.beta * edge_weight
graph.edata['p'] = edge_softmax(graph, e, norm_by='src')
graph.update_all(fn.u_mul_e('norm_h', 'p', 'm'), fn.sum('m', 'h'))
rst = graph.dstdata.pop('h')
rst = (1 + self.eps) * feat + rst
return rst
def forward(self, G: DGLBlock, feat):
feat_src, feat_dst = expand_as_pair(input_=feat, g=G)
# print(G)
with G.local_scope():
funcs = {}
for srctype, etype, dsttype in G.canonical_etypes:
k_linear = self.k_linears[self.node_dict[srctype]]
v_linear = self.v_linears[self.node_dict[srctype]]
q_linear = self.q_linears[self.node_dict[dsttype]]
G.srcnodes[srctype].data['k'] = k_linear(feat_src[srctype]).view(-1, self.n_heads, self.d_k)
G.srcnodes[srctype].data['v'] = v_linear(feat_src[srctype]).view(-1, self.n_heads, self.d_k)
G.dstnodes[dsttype].data['q'] = q_linear(feat_dst[dsttype]).view(-1, self.n_heads, self.d_k)
G.apply_edges(func=self.edge_attention, etype=etype)
if G.batch_num_edges(etype=etype).item() > 0:
funcs[etype] = (self.message_func, self.reduce_func)
# print("funcs", funcs.keys())
G.multi_update_all(funcs, cross_reducer='mean')
new_h = {}
for ntype in G.ntypes:
'''
Step 3: Target-specific Aggregation
x = norm( W[node_type] * gelu( Agg(x) ) + x )
'''
nty_id = self.node_dict[ntype]
alpha = torch.sigmoid(self.skip[nty_id])
# print(ntype, G.srcnodes[ntype].data.keys(), G.dstnodes[ntype].data.keys())
if "t" in G.dstnodes[ntype].data:
trans_out = self.dropout(self.a_linears[nty_id].forward(G.dstnodes[ntype].data['t']))
else:
trans_out = self.dropout(feat_dst[ntype])
trans_out = trans_out * alpha + feat_dst[ntype] * (1 - alpha)
if self.use_norm:
new_h[ntype] = self.norms[nty_id](trans_out)
else:
new_h[ntype] = trans_out
return new_h
def __init__(self, in_feats, out_feats, gnn_mlps, bias=True):
super(SAGEConv, self).__init__()
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
# din attention
self.atten_src = nn.utils.weight_norm(nn.Linear(self._in_src_feats, out_feats), name='weight')
self.atten_dst = nn.utils.weight_norm(nn.Linear(self._in_src_feats, out_feats), name='weight')
self.atten_sub = nn.utils.weight_norm(nn.Linear(self._in_src_feats, out_feats), name='weight')
self.atten_mul = nn.utils.weight_norm(nn.Linear(self._in_src_feats, out_feats), name='weight')
self.atten_out = nn.utils.weight_norm(nn.Linear(out_feats, 1), name='weight')
self.leaky_relu = nn.LeakyReLU(0.2)
# other
self.fc_pool = nn.utils.weight_norm(nn.Linear(self._in_src_feats, self._in_src_feats), name='weight')
self.fc_pool2 = nn.utils.weight_norm(nn.Linear(self._in_src_feats, self._in_src_feats), name='weight')
self.fc_self = nn.utils.weight_norm(nn.Linear(self._in_dst_feats, out_feats, bias=bias), name='weight')
self.fc_neigh = nn.utils.weight_norm(nn.Linear(self._in_src_feats, out_feats, bias=bias), name='weight')
self.fc_neigh2 = nn.utils.weight_norm(nn.Linear(self._in_src_feats, out_feats, bias=bias), name='weight')
# mlps
self.out_mlp = nn.ModuleList()
for i in range(gnn_mlps):
self.out_mlp.append(nn.utils.weight_norm(nn.Linear(out_feats, out_feats), name='weight'))
self.reset_parameters()
def __init__(self,
in_feats,
out_feats,
aggregator_type,
feat_drop=0.,
bias=True,
norm=None,
activation=None):
super(SAGEConv, self).__init__()
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self._aggre_type = aggregator_type
self.norm = norm
self.feat_drop = nn.Dropout(feat_drop)
self.activation = activation
# aggregator type: mean/pool/lstm/gcn
if self._aggre_type == 'cheb': # default using activation TODO
self._cheb_k = 2 # it should be consistent with the sampling since every block is a bipartitte graph
# in spectral method there's no fc_neigh or fc_self
self._cheb_linear = nn.Linear(self._cheb_k * self._in_src_feats,
out_feats)
if aggregator_type == 'ginmean':
self._gin_reducer = fn.mean
self.fc_gin = nn.Linear(
self._in_src_feats,
out_feats) # default apply_func is nn.linear
if aggregator_type == 'pool':
self.fc_pool = nn.Linear(self._in_src_feats, self._in_src_feats)
if aggregator_type == 'lstm':
self.lstm = nn.LSTM(self._in_src_feats,
self._in_src_feats,
batch_first=True)
if aggregator_type != 'gcn' and aggregator_type != 'ginmean' and aggregator_type != 'cheb':
self.fc_self = nn.Linear(self._in_dst_feats, out_feats, bias=bias)
if aggregator_type != 'ginmean' and aggregator_type != 'cheb':
self.fc_neigh = nn.Linear(self._in_src_feats, out_feats, bias=bias)
self.reset_parameters()
def __init__(self,
in_feats,
out_feats,
num_heads,
feat_drop=0.,
attn_drop=0.,
negative_slope=0.2,
residual=False,
activation=None):
super(Conv, self).__init__()
self._num_heads = num_heads
self._in_src_feats, self._in_dst_feats = expand_as_pair(in_feats)
self._out_feats = out_feats
self.cache_atte = None
if isinstance(in_feats, tuple):
self.fc_src = nn.Linear(
self._in_src_feats, out_feats * num_heads, bias=False)
self.fc_dst = nn.Linear(
self._in_dst_feats, out_feats * num_heads, bias=False)
else:
self.fc = nn.Linear(
self._in_src_feats, out_feats * num_heads, bias=False)
self.attn_l = nn.Parameter(th.FloatTensor(size=(1, num_heads, out_feats)))
self.attn_r = nn.Parameter(th.FloatTensor(size=(1, num_heads, out_feats)))
self.feat_drop = nn.Dropout(feat_drop)
self.attn_drop = nn.Dropout(attn_drop)
self.leaky_relu = nn.LeakyReLU(negative_slope)
if residual:
if self._in_dst_feats != out_feats:
self.res_fc = nn.Linear(
self._in_dst_feats, num_heads * out_feats, bias=False)
else:
self.res_fc = Identity()
else:
self.register_buffer('res_fc', None)
self.reset_parameters()
self.activation = activation
def forward(self, graph, feat, weight):
with graph.local_scope():
feat_src, feat_dst = expand_as_pair(feat, graph)
graph.srcdata['h'] = feat_src
graph.dstdata['h'] = feat_dst
if self._norm == 'both':
degs = graph.out_degrees().float().clamp(min=1)
norm = torch.pow(degs, -0.5)
shp = norm.shape + (1,) * (feat_src.dim() - 1)
norm = torch.reshape(norm, shp)
feat_src = feat_src * norm
rel_weight_feature = weight[graph.edata['etype']] # get rel weight
graph.edata['r'] = self.rel_dropout(rel_weight_feature)
# gcn
graph.srcdata['h'] = feat_src
graph.apply_edges(fn.u_mul_e('h', 'r', 'm'))
graph.update_all(fn.copy_e('m', 'm'),
fn.mean(msg='m', out='mix'))
rst = graph.dstdata['mix']
# self_loop
self_loop_h = torch.matmul(graph.dstdata['h'], self.loop_weight)
rst = self_loop_h + rst
# need to check if norm is needed
if self._norm != 'none':
degs = graph.in_degrees().float().clamp(min=1)
if self._norm == 'both':
norm = torch.pow(degs, -0.5)
else:
norm = 1.0 / degs
shp = norm.shape + (1,) * (feat_dst.dim() - 1)
norm = torch.reshape(norm, shp)
rst = rst * norm
self.ent_dropout(rst)
if self.activation is not None:
rst = self.activation(rst)
return rst