def add_distance_matrix(self, name: str, distance_matrix: sp.csr_matrix,
labels: Sequence):
"""Add a distance matrix.
Parameters
----------
name
Unique identifier of the distance matrix
distance_matrix
sparse distance matrix `D` in CSR format
labels
array with row/column names of the distance matrix.
`len(array) == D.shape[0] == D.shape[1]`
"""
if not (len(labels) == distance_matrix.shape[0] ==
distance_matrix.shape[1]):
raise ValueError("Dimension mismatch!")
if not isinstance(distance_matrix, csr_matrix):
raise TypeError(
"Distance matrix must be sparse and in CSR format. ")
# The class relies on zeros not being explicitly stored during reverse lookup.
distance_matrix.eliminate_zeros()
self.distance_matrices[name] = distance_matrix
self.distance_matrix_labels[name] = {
k: i
for i, k in enumerate(labels)
}
# The label "nan" does not have an index in the matrix
self.distance_matrix_labels[name]["nan"] = np.nan
def add_distance_matrix(
self,
name: str,
distance_matrix: sp.csr_matrix,
labels: Sequence,
labels2: Sequence = None,
):
"""Add a distance matrix.
Parameters
----------
name
Unique identifier of the distance matrix
distance_matrix
sparse distance matrix `D` in CSR format
labels
array with row names of the distance matrix.
`len(array) == D.shape[0]`
labels2
array with column names of the distance matrix.
Can be omitted if the distance matrix is symmetric.
`len(array) == D.shape[1]`.
"""
labels2 = labels if labels2 is None else labels2
if not len(labels) == distance_matrix.shape[0]:
raise ValueError("Dimensions mismatch alon axis 0")
if not len(labels2) == distance_matrix.shape[1]:
raise ValueError("Dimensions mismatch alon axis 1")
if not isinstance(distance_matrix, csr_matrix):
raise TypeError(
"Distance matrix must be sparse and in CSR format. ")
# The class relies on zeros not being explicitly stored during reverse lookup.
distance_matrix.eliminate_zeros()
self.distance_matrices[name] = distance_matrix
self.distance_matrix_labels[name] = {
k: i
for i, k in enumerate(labels)
}
self.distance_matrix_labels2[name] = {
k: i
for i, k in enumerate(labels2)
}
# The label "nan" does not have an index in the matrix
self.distance_matrix_labels[name]["nan"] = np.nan
self.distance_matrix_labels2[name]["nan"] = np.nan
def _setup_linear_propagator(self, struct):
"""Sets up the noise-independent, linear propagator
Returns:
The propagator as CSR Matrix
"""
i, j, a = hcons.setup_linear_propagator(struct.vecind,
struct.indab,
struct.indbl,
self._h_sys,
self._g,
self._gamma + 1.j*self._omega,
self._l_map,
self._with_terminator)
csr = CSRMatrix((a, j, i))
csr.sum_duplicates()
csr.eliminate_zeros()
return csr
def bias_knn(
self,
conn: csr_matrix,
pseudotime: np.ndarray,
n_jobs: Optional[int] = None,
backend: str = "loky",
show_progress_bar: bool = True,
**kwargs: Any,
) -> csr_matrix:
"""
Bias cell-cell connectivities of a KNN graph.
Parameters
----------
conn
Sparse matrix of shape ``(n_cells, n_cells)`` containing the nearest neighbor connectivities.
pseudotime
Pseudotemporal ordering of cells.
%(parallel)s
Returns
-------
The biased connectivities.
"""
res = parallelize(
self._bias_knn_helper,
np.arange(conn.shape[0]),
as_array=False,
unit="cell",
n_jobs=n_jobs,
backend=backend,
show_progress_bar=show_progress_bar,
)(conn, pseudotime, **kwargs)
data, indices, indptr = zip(*res)
conn = csr_matrix((np.concatenate(data), np.concatenate(indices),
np.concatenate(indptr)))
conn.eliminate_zeros()
return conn
def _eliminate(matrix: sp.csr_matrix, user_indices, item_indices):
matrix = matrix.copy()
# `lil_matrix` is too slow
matrix[list(user_indices), list(item_indices)] = 0
matrix.eliminate_zeros()
return matrix
def _eliminate_subzeroes(self, m: sparse.csr_matrix, epsilon):
m.data = np.where(abs(m.data) < epsilon, 0, m.data)
m.eliminate_zeros()
def _eliminate(matrix: sp.csr_matrix, user_indices, item_indices):
matrix = matrix.copy()
matrix[user_indices, item_indices] = 0
matrix.eliminate_zeros()
return matrix
def mask_test_edges(adj: sp.csr_matrix):
# Function to build test set with 2% positive links
# Remove diagonal elements
adj = adj - sp.dia_matrix((adj.diagonal()[np.newaxis, :], [0]), shape=adj.shape)
adj.eliminate_zeros()
adj_triu = sp.triu(adj)
adj_tuple = sparse_to_tuple(adj_triu)
edges = adj_tuple[0]
edges_all = sparse_to_tuple(adj)[0]
num_test = int(np.floor(edges.shape[0] / 10.))
num_val = int(np.floor(edges.shape[0] / 10.))
# split to get the training, valid and test set.
all_edge_idx = range(edges.shape[0])
all_edge_idx = list(all_edge_idx)
np.random.shuffle(all_edge_idx)
val_edge_idx = all_edge_idx[:num_val]
test_edge_idx = all_edge_idx[num_val:(num_val + num_test)]
test_edges = edges[test_edge_idx]
val_edges = edges[val_edge_idx]
train_edges = np.delete(edges, np.hstack([test_edge_idx, val_edge_idx]), axis=0)
def ismember(a, b):
rows_close = np.all((a - b[:, None]) == 0, axis=-1)
return np.any(rows_close)
test_edges_false = []
while len(test_edges_false) < len(test_edges):
n_rnd = len(test_edges) - len(test_edges_false)
# num_nodes = adj.shape[0]
rnd = np.random.randint(0, adj.shape[0], size=2 * n_rnd)
idxs_i = rnd[:n_rnd]
idxs_j = rnd[n_rnd:]
for i in range(n_rnd):
idx_i = idxs_i[i]
idx_j = idxs_j[i]
if idx_i == idx_j:
continue
if ismember([idx_i, idx_j], edges_all):
continue
if test_edges_false:
if ismember([idx_j, idx_i], np.array(test_edges_false)):
continue
if ismember([idx_i, idx_j], np.array(test_edges_false)):
continue
test_edges_false.append([idx_i, idx_j])
val_edges_false = []
while len(val_edges_false) < len(val_edges):
n_rnd = len(val_edges) - len(val_edges_false)
rnd = np.random.randint(0, adj.shape[0], size=2 * n_rnd)
idxs_i = rnd[:n_rnd]
idxs_j = rnd[n_rnd:]
for i in range(n_rnd):
idx_i = idxs_i[i]
idx_j = idxs_j[i]
if idx_i == idx_j:
continue
if ismember([idx_i, idx_j], train_edges):
continue
if ismember([idx_j, idx_i], train_edges):
continue
if ismember([idx_i, idx_j], val_edges):
continue
if ismember([idx_j, idx_i], val_edges):
continue
if val_edges_false:
if ismember([idx_j, idx_i], np.array(val_edges_false)):
continue
if ismember([idx_i, idx_j], np.array(val_edges_false)):
continue
val_edges_false.append([idx_i, idx_j])
# Re-build adj matrix
data = np.ones(train_edges.shape[0])
adj_train = sp.csr_matrix((data, (train_edges[:, 0], train_edges[:, 1])), shape=adj.shape)
adj_train = adj_train + adj_train.T
return adj_train, train_edges, val_edges, val_edges_false, test_edges, test_edges_false
