def test_to_disjoint():
# TODO test e_list
x, i = to_disjoint(x_list, None)
a, i = to_disjoint(None, a_list)
x, a, i = to_disjoint(x_list, a_list)
assert x.ndim == 2
assert a.ndim == 2
assert x.shape[0] == a.shape[0] == a.shape[1]
def collate(self, batch):
packed = self.pack(batch, return_dict=True)
y = None
if "y" in self.dataset.signature:
y = packed.pop("y_list")
if self.node_level:
if len(np.shape(y[0])) == 1:
y = [y_[:, None] for y_ in y]
y = np.vstack(y)
else:
if len(np.shape(y[0])) == 0:
y = [np.array([y_]) for y_ in y]
y = np.array(y)
output = to_disjoint(**packed)
# Sparse matrices to SparseTensors
output = list(output)
for i in range(len(output)):
if sp.issparse(output[i]):
output[i] = sp_matrix_to_sp_tensor(output[i])
output = tuple(output)
if y is None:
return output
else:
return output, y
def test_modes_ops():
ns = [10, 5, 8]
x_list = [np.random.randn(n, 3) for n in ns]
a_list = [sp.csr_matrix(np.ones((n, n))) for n in ns]
x, a, i = to_disjoint(x_list, a_list)
a = sp_matrix_to_sp_tensor(a)
expected_x = np.zeros((len(ns), max(ns), 3))
for i_, n in enumerate(ns):
expected_x[i_, :n] = x_list[i_]
expected_a = np.zeros((len(ns), max(ns), max(ns)))
for i_, n in enumerate(ns):
expected_a[i_, :n, :n] = a_list[i_].A
# Disjoint signal to batch
result = ops.disjoint_signal_to_batch(x, i).numpy()
assert expected_x.shape == result.shape
assert np.allclose(expected_x, result, atol=tol)
# Disjoint adjacency to batch
result = ops.disjoint_adjacency_to_batch(a, i).numpy()
assert expected_a.shape == result.shape
assert np.allclose(expected_a, result, atol=tol)
def collate(self, batch):
packed = self.pack(batch)
y = packed.pop("y_list", None)
if y is not None:
y = collate_labels_disjoint(y, node_level=self.node_level)
output = to_disjoint(**packed)
output = sp_matrices_to_sp_tensors(output)
if len(output) == 1:
output = output[0]
if y is None:
return output
else:
return output, y
def collate(self, batch):
packed = self.pack(batch, return_dict=True)
y = None
if "y" in self.dataset.signature:
y = packed.pop("y_list")
y = np.vstack(y) if self.node_level else np.array(y)
output = to_disjoint(**packed)
output = list(output)
for i in range(len(output)):
if sp.issparse(output[i]):
output[i] = sp_matrix_to_sp_tensor(output[i])
output = tuple(output)
if y is None:
return output
else:
return output, y
def collate(self, batch):
packed = self.pack(batch)
y = packed.pop("y_list", None)
if y is not None:
y = collate_labels_disjoint(y, node_level=True)
output = to_disjoint(**packed)
output = output[:-1] # Discard batch index
output = sp_matrices_to_sp_tensors(output)
if len(output) == 1:
output = output[0]
output = (output, )
if y is not None:
output += (y, )
if self.sample_weights is not None:
output += (self.sample_weights, )
if len(output) == 1:
output = output[0] # Again, in case there are no targets and no SW
return output