def compute_join_dist_mat(self, list_expert_id):
""" Reset global variables. """
joint_data = []
joint_structures = []
for i in list_expert_id:
joint_data += self.training_data[i]
# joint_structures += self.training_structures[expert_id]
_global_training_data[f"{self.name}_join"] = joint_data
_global_training_structures[f"{self.name}_join"] = joint_labels
kmat = kernel_distance_mat(
self.hyps,
self.name + "_join",
self.energy_kernel,
self.cutoffs,
self.hyps_mask,
self.n_cpus,
self.n_sample,
)
for i in range(kmat.shape[0]):
norm = np.sqrt(kmat[i, i])
kmat[i, :] /= norm
kmat[:, i] /= norm
print(np.min(kmat), np.max(kmat))
kmat = np.max(kmat) - kmat
del _global_training_data[f"{self.name}_join"]
del _global_training_structures[f"{self.name}_join"]
del joint_data
del joint_structures
return kmat
def redistribute_training_data(self, figure_name="default"):
""" redistribute data """
joint_data = []
joint_structures = []
joint_labels = []
for i in range(self.n_experts):
joint_data += self.training_data[i]
joint_labels += self.training_labels[i]
self.unsync_experts_data(i)
self.n_experts = 1
self.reset_container()
_global_training_data[f"{self.name}_join"] = joint_data
_global_training_structures[f"{self.name}_join"] = joint_structures
kmat = kernel_distance_mat(
self.hyps,
self.name + "_join",
self.energy_kernel,
self.cutoffs,
self.hyps_mask,
self.n_cpus,
self.n_sample,
)
for i in range(kmat.shape[0]):
norm = np.sqrt(kmat[i, i])
kmat[i, :] /= norm
kmat[:, i] /= norm
kmat = np.max(kmat) - kmat
print(np.min(kmat), np.max(kmat))
plot_mat(kmat, f"{figure_name}_unsorted_dmat.png")
# np.save(f"{figure_name}_unsorted_dmat.npy", kmat)
iu1 = np.triu_indices(kmat.shape[0], 1)
upper_triang = kmat[iu1]
Z = linkage(upper_triang, method="average")
K = fcluster(Z, t=0.5, criterion="distance")
ncluster = int(np.max(K))
new_indices = []
sizes = []
for i in range(ncluster):
index = np.where(K == (i + 1))[0]
new_indices += [index]
sizes += [len(index)]
new_indices = np.hstack(new_indices)
plot_mat(sort_matrix(kmat, new_indices),
f"{figure_name}_sorted_dmat.png")
# np.save(f"{figure_name}_sorted_dmat.npy", kmat)
for i in new_indices:
self.add_one_env(joint_data[i],
joint_labels[i],
train=False,
expert_id=None)
del _global_training_data[f"{self.name}_join"]
del _global_training_structures[f"{self.name}_join"]
del joint_data
del joint_structures
del joint_labels
self.set_L_alpha()
return kmat