def _preprocess_features(self, features):
""" The propagation step is made here"""
if self.normalize_features:
features = row_normalize(features)
for i in range(self.num_layers):
features = self.graph_adj @ features
return to_sparse_tensor(features)
def _preprocess_adj(self, graph_adj):
adj_with_self_loops = add_self_loops(graph_adj)
self.adj_dense_shape = adj_with_self_loops.shape
adj_with_self_loops_tensor = to_sparse_tensor(adj_with_self_loops)
adj_with_self_loops_coo = adj_with_self_loops.tocoo()
self.degrees = adj_with_self_loops.sum(axis=1).astype(np.float32)
if self.aggregator == 'maxpool':
# maxpool aggregator requires special preprocessing:
# Fill a matrix with up to max_degree neighbours. max_degree is the maximum degree appearing in the
# graph. If a node has less than max_degree neighbours, the remaining entries in the matrix are set
# to 0. This finally refers to node 0. However, for the maxpool aggregator, duplicates do not matter
# since the max-operation is invariant to duplicates in the input.
neighbours_matrix = np.zeros(
(self.num_nodes, int(np.max(self.degrees))), dtype=np.int32)
insert_index = 0
self_node_old = 0
for i, self_node in enumerate(adj_with_self_loops_coo.row):
if self_node != self_node_old:
insert_index = 0
neighbours_matrix[
self_node, insert_index] = adj_with_self_loops_coo.col[i]
insert_index += 1
self_node_old = self_node
self.adj_with_self_loops_indices = neighbours_matrix
else:
# extract the coordinates of all the edges
# since both row and column coordinates are ordered, row[0] corresponds to col[0] etc.
self.adj_with_self_loops_indices = np.mat(
[adj_with_self_loops_coo.row, adj_with_self_loops_coo.col])
return adj_with_self_loops_tensor
def _preprocess_adj(self, graph_adj):
adj_with_self_loops = add_self_loops(graph_adj)
self.adj_dense_shape = adj_with_self_loops.shape
adj_with_self_loops_tensor = to_sparse_tensor(adj_with_self_loops)
adj_with_self_loops_coo = adj_with_self_loops.tocoo()
# extract the coordinates of all the edges
# since both row and column coordinates are ordered, row[0] corresponds to col[0] etc.
self.adj_with_self_loops_indices = np.mat(
[adj_with_self_loops_coo.row, adj_with_self_loops_coo.col])
return adj_with_self_loops_tensor
def _preprocess_features(self, features):
""" The propagation step is made here"""
if self.normalize_features:
features = row_normalize(features)
initial_features = features.copy()
alpha = np.float32(self.teleport_prob)
for i in range(self.num_layers):
features = (
1 -
alpha) * self.graph_adj @ features + alpha * initial_features
return to_sparse_tensor(features)
def _preprocess_features(self, features):
if self.normalize_features:
features = row_normalize(features)
return to_sparse_tensor(features)
def _preprocess_adj(self, graph_adj):
return to_sparse_tensor(graph_adj)
def _preprocess_adj(self, graph_adj):
return to_sparse_tensor(renormalize_adj(graph_adj))
def _preprocess_adj(self, graph_adj):
self.degrees = graph_adj.sum(axis=1).astype(np.float32)
if self.prop_type == 'smoothed':
graph_adj = normalize_adj(graph_adj)
return to_sparse_tensor(graph_adj)
def _preprocess_features(self, features):
return to_sparse_tensor(features)