def prop(self, params):
prop_types = [
key for key, value in params.items()
if value == 1 and key in self.prop_types
]
if not prop_types:
print(" -- dropped -- ")
return None, None
prop_result_list = []
for selected_prop in prop_types:
prop_result = propagate(self.adj, self.emb, selected_prop, params)
prop_result_list.append(prop_result)
if self.loss_type == "infomax":
neg_prop_result = []
pmt = self.permutation
for s_prop in prop_types:
neg_prop = propagate(self.adj, self.emb[pmt], s_prop, params)
neg_prop[pmt] = neg_prop
neg_prop_result.append(neg_prop)
return np.array(prop_result_list), np.array(neg_prop_result)
elif self.loss_type == "infonce":
return np.array(prop_result_list), None
elif self.loss_type == "sparse":
return np.array(prop_result_list), None
else:
raise ValueError(
"use 'infonce', 'infomax' or 'sparse' loss, currently using {}"
.format(self.loss_type))
def __call__(self, emb, adj):
if len(self.filter_types) == 1 and self.filter_types[0] == "identity":
return emb
dim = emb.shape[1]
prop_result = []
for tp in self.filter_types:
prop_result.append(propagate(adj, emb, tp))
prop_result_emb = np.concatenate(prop_result, axis=1)
if self.svd:
prop_result_emb = get_embedding_dense(prop_result_emb, dim)
return prop_result_emb
def test_prone_module():
from cogdl.layers.prone_module import propagate
import scipy.sparse as sp
data = np.ones(400)
edge_index = np.random.randint(0, 100, (2, 200))
row = np.concatenate((edge_index[0], edge_index[1]))
col = np.concatenate((edge_index[1], edge_index[0]))
print(row.shape, col.shape)
matrix = sp.csr_matrix((data, (row, col)), shape=(100, 100))
emb = np.random.randn(100, 20)
for module in ["heat", "ppr", "gaussian", "prone", "sc"]:
res = propagate(matrix, emb, module)
assert res.shape == emb.shape