def idrnnclus(
ref,
modelin: IO,
assign_out: TextIO,
thresh: float,
):
"""
Identifies clusters by comparing against a reference and forcing a match
"""
estimator = pickle.load(modelin)
knn_index = estimator.named_steps["pynndescenttransformer"].index_
indices, distances = knn_index.query(ref.ref_embeddings, k=32)
rnndbscan = estimator.named_steps["rnndbscan"]
labels = rnndbscan.labels_
unique_labels = np.unique(labels)
num_clusts = len(unique_labels) - (1 if unique_labels[0] == -1 else 0)
dists = build_dists(indices, distances, labels, num_clusts, thresh)
dists = min_pool_sparse(dists, ref.ref_group_sizes)
ref_ids, clus_ids = min_weight_full_bipartite_matching(dists)
assign_out.write("label,clus\n")
ref_labels = list(ref.labels())
for ref_idx, clus_idx in zip(ref_ids, clus_ids):
assign_out.write("{},c{}\n".format(ref_labels[ref_idx], clus_idx))
def test_two_methods_give_same_result_on_many_sparse_inputs():
# As opposed to the test above, here we do not spell out the expected
# output; only assert that the two methods give the same result.
# Concretely, the below tests 100 cases of size 100x100, out of which
# 36 are infeasible.
np.random.seed(1234)
for _ in range(100):
lsa_raises = False
mwfbm_raises = False
sparse = random(100,
100,
density=0.06,
data_rvs=lambda size: np.random.randint(1, 100, size))
# In csgraph, zeros correspond to missing edges, so we explicitly
# replace those with infinities
dense = np.full(sparse.shape, np.inf)
dense[sparse.row, sparse.col] = sparse.data
sparse = sparse.tocsr()
try:
row_ind, col_ind = linear_sum_assignment(dense)
lsa_cost = dense[row_ind, col_ind].sum()
except ValueError:
lsa_raises = True
try:
row_ind, col_ind = min_weight_full_bipartite_matching(sparse)
mwfbm_cost = sparse[row_ind, col_ind].sum()
except ValueError:
mwfbm_raises = True
# Ensure that if one method raises, so does the other one.
assert lsa_raises == mwfbm_raises
if not lsa_raises:
assert lsa_cost == mwfbm_cost
def test_explicit_zero_causes_warning():
with pytest.warns(UserWarning):
biadjacency_matrix = csr_matrix(((2, 0, 3), (0, 1, 1), (0, 2, 3)))
min_weight_full_bipartite_matching(biadjacency_matrix)
def test_min_weight_full_matching_infeasible_problems(biadjacency_matrix):
with pytest.raises(ValueError):
min_weight_full_bipartite_matching(csr_matrix(biadjacency_matrix))
def test_min_weight_full_matching_trivial_graph(num_rows, num_cols):
biadjacency_matrix = csr_matrix((num_cols, num_rows))
row_ind, col_ind = min_weight_full_bipartite_matching(biadjacency_matrix)
assert len(row_ind) == 0
assert len(col_ind) == 0
def time_evaluation(self, *args):
min_weight_full_bipartite_matching(self.biadjacency_matrix)