def gen_layer(gw, nfeat, efeat, hidden_size, name):
def _send_func(src_feat, dst_feat, edge_feat):
h = src_feat['h'] + edge_feat['h']
h = L.relu(h)
return h
def _recv_func(msg):
return L.sequence_pool(msg, "sum")
beta = L.create_parameter(
shape=[1],
dtype='float32',
default_initializer=F.initializer.ConstantInitializer(value=1.0),
name=name + '_beta')
# message passing
msg = gw.send(_send_func,
nfeat_list=[("h", nfeat)],
efeat_list=[("h", efeat)])
output = gw.recv(msg, message_passing.softmax_agg(beta))
output = message_passing.msg_norm(nfeat, output, name)
output = nfeat + output
output = mlp(output, [hidden_size * 2, hidden_size],
norm="layer_norm",
name=name)
return output
def gen_conv(gw,
feature,
name,
beta=None):
"""Implementation of GENeralized Graph Convolution (GENConv), see the paper
"DeeperGCN: All You Need to Train Deeper GCNs" in
https://arxiv.org/pdf/2006.07739.pdf
Args:
gw: Graph wrapper object (:code:`StaticGraphWrapper` or :code:`GraphWrapper`)
feature: A tensor with shape (num_nodes, feature_size).
beta: [0, +infinity] or "dynamic" or None
name: deeper gcn layer names.
Return:
A tensor with shape (num_nodes, feature_size)
"""
if beta == "dynamic":
beta = L.create_parameter(
shape=[1],
dtype='float32',
default_initializer=
fluid.initializer.ConstantInitializer(value=1.0),
name=name + '_beta')
# message passing
msg = gw.send(message_passing.copy_send, nfeat_list=[("h", feature)])
output = gw.recv(msg, message_passing.softmax_agg(beta))
# msg norm
output = message_passing.msg_norm(feature, output, name)
output = feature + output
output = L.fc(output,
feature.shape[-1],
bias_attr=False,
act="relu",
param_attr=fluid.ParamAttr(name=name + '_weight1'))
output = L.fc(output,
feature.shape[-1],
bias_attr=False,
param_attr=fluid.ParamAttr(name=name + '_weight2'))
return output