def test_diag_indices():
mnp_res = mnp.diag_indices(0)
onp_res = onp.diag_indices(0)
match_all_arrays(mnp_res, onp_res)
mnp_res = mnp.diag_indices(3, 0)
onp_res = onp.diag_indices(3, 0)
match_all_arrays(mnp_res, onp_res)
mnp_res = mnp.diag_indices(5, 7)
onp_res = onp.diag_indices(5, 7)
match_all_arrays(mnp_res, onp_res)
def _to_dag(self, train_data):
"""
threshold: from weaker to stronger in A
1- If some entries of A less than equal threshold, mask them to zero
(This can happen with stochastic proximal gradient descent)
2- Remove edges by threshold until a DAG is obtained.
Parameters
----------
train_data: NormalizationData
train samples
"""
if self.jac_thresh:
jacobian_adj = compute_jacobian_avg(
self.model, train_data, train_data.n_samples).transpose()
else:
jacobian_adj = self.model.get_w_adj()
dia_index = msnp.diag_indices(jacobian_adj.shape[0])
jacobian_adj[dia_index] = 0
# Find the smallest threshold that removes all cycle-inducing edges
thresholds = np.unique(jacobian_adj.asnumpy())
epsilon = 1e-8
for step, t in enumerate(thresholds):
to_keep = (jacobian_adj > t + epsilon).astype(dtype=mstype.float32)
new_adj = self.model.adjacency * to_keep
if is_acyclic(new_adj):
self.model.adjacency = new_adj
break
return self.model