mg = MetaGraph(thresh_g, weight="max_norm_weight")
meta = mg.meta
adj = mg.adj.copy()
# colsums = np.sum(adj, axis=0)
# colsums[colsums == 0] = 1
# adj = adj / colsums[np.newaxis, :]
adj = pass_to_ranks(adj)
if use_spl:
adj = graph_shortest_path(adj)
if plus_c:
adj += np.min(adj)
if embed == "lse":
latent = lse(adj, None, ptr=False)
elif embed == "ase":
latent = ase(adj, None, ptr=False)
rot_latent, diff = procrustes_match(latent, meta)
rot_latent = latent
n_components = latent.shape[1]
plot_df = pd.DataFrame(data=rot_latent)
plot_df["Class"] = mg["Class 1"]
fig, ax = plt.subplots(1, 1, figsize=(10, 10))
sns.scatterplot(x=0, y=1, data=plot_df, hue="Class", legend=False, ax=ax)
ax.set_title(f"Residual F. norm = {diff}, threshold = {threshold}")
left_paired_inds, right_paired_inds = get_paired_inds(meta)
temp_neigh_probs = compute_neighbors_at_k(
mg.make_lcc()
print(f"Removed {n_verts - mg.n_verts} when finding the LCC")
# old_n_verts = sym_adj.shape[0]
# sym_adj, class_labels, side_labels = preprocess_graph(
# sym_adj, class_labels, side_labels
# )
# n_verts = sym_adj.shape[0]
# print(f"Removed {old_n_verts - n_verts} nodes")
# %% [markdown]
# # Embedding
n_verts = mg.n_verts
sym_adj = mg.adj
side_labels = mg["Hemisphere"]
class_labels = mg["Merge Class"]
latent, laplacian = lse(sym_adj, N_COMPONENTS, regularizer=None, ptr=PTR)
latent_dim = latent.shape[1] // 2
screeplot(
laplacian,
title=f"Laplacian scree plot, R-DAD (ZG2 = {latent_dim} + {latent_dim})")
print(f"ZG chose dimension {latent_dim} + {latent_dim}")
plot_latent = np.concatenate(
(latent[:, :3], latent[:, latent_dim:latent_dim + 3]), axis=-1)
pairplot(plot_latent, labels=side_labels)
# take the mean for the paired cells, making sure to add back in the unpaired cells
sym_latent = (latent[:n_pairs] + latent[n_pairs:2 * n_pairs]) / 2
sym_latent = np.concatenate((sym_latent, latent[2 * n_pairs:]))
latent = sym_latent
# select the right hemisphere
if ONLY_RIGHT:
right_inds = np.where(side_labels == "R")[0]
adj = adj[np.ix_(right_inds, right_inds)]
class_labels = class_labels[right_inds]
skeleton_labels = skeleton_labels[right_inds]
adj, class_labels, skeleton_labels = preprocess_graph(adj, class_labels,
skeleton_labels)
known_inds = np.where(class_labels != "Unk")[0]
# %% [markdown]
# # Embedding
n_verts = adj.shape[0]
latent = lse(adj, N_COMPONENTS, regularizer=None, ptr=PTR)
# pairplot(latent, labels=class_labels, title=embed)
latent_dim = latent.shape[1] // 2
print(f"ZG chose dimension {latent_dim} + {latent_dim}")
# %% [markdown]
# # Fitting divisive cluster model
start = timer()
dc = DivisiveCluster(n_init=N_INIT, cluster_method=CLUSTER_METHOD)
dc.fit(latent)
end = end = timer()
print()
print(f"DivisiveCluster took {(end - start)/60.0} minutes to fit")
print()
dc.print_tree(print_val="bic_ratio")
pred_labels = dc.predict(latent)
sym_mg.adj,
inner_hier_labels=sym_mg["Class 1"],
outer_hier_labels=sym_mg["Hemisphere"],
figsize=(30, 30),
hier_label_fontsize=5,
transform="binarize",
cbar=False,
)
stashfig("heatmap-after-mods")
# %% [markdown]
# #
# ad_norm_mg, n_pairs = pair_augment(ad_norm_mg)
# ad_norm_mg = max_symmetrize(ad_norm_mg, n_pairs)
# ad_norm_mg.make_lcc()
ad_norm_lse_latent = lse(sym_mg.adj, n_components=None)
# plot_latent_sweep(ad_norm_lse_latent, n_pairs)
# remove_inds = [2, 7, 10, 15]
# ad_norm_lse_latent = remove_cols(ad_norm_lse_latent, remove_inds)
# %% [markdown]
# #
from scipy.linalg import orthogonal_procrustes
left_latent = latent[:n_pairs, :]
right_latent = latent[n_pairs:2 * n_pairs, :]
R, scalar = orthogonal_procrustes(left_latent, right_latent)
# %% [markdown]
# #
ad_raw_mg = load_metagraph("Gad")
degrees = degrees[d_sort]
plt.figure(figsize=(10, 5))
sns.scatterplot(x=range(len(degrees)), y=degrees, s=30, linewidth=0)
known_inds = np.where(class_labels != "Unk")[0]
# %% [markdown]
# #
from graspy.cluster import PartitionalGaussianCluster
# %% [markdown]
# # Run clustering using LSE on the sum graph
n_verts = adj.shape[0]
latent = lse(adj, n_components, regularizer=None, ptr=PTR)
pairplot(latent, labels=class_labels, title=embed)
# %% [markdown]
# #
class PartitionCluster:
def __init__(self):
self.min_split_samples = 5
def fit(self, X, y=None):
n_samples = X.shape[0]
if n_samples > self.min_split_samples:
cluster = GaussianCluster(min_components=1,
print((left_pair_ids == right_pair_ids).all())
sym_mg.make_lcc()
n_pairs = sym_mg.meta["Pair ID"].nunique() - 1
left_pair_ids = sym_mg["Pair ID"][:n_pairs]
right_pair_ids = sym_mg["Pair ID"][n_pairs:2 * n_pairs]
print((left_pair_ids == right_pair_ids).all())
uni_pair, counts = np.unique(sym_mg["Pair ID"], return_counts=True)
print(np.min(counts))
# %% [markdown]
# #
left_pair_ids = sym_mg["Pair ID"][:n_pairs]
right_pair_ids = sym_mg["Pair ID"][n_pairs:2 * n_pairs]
latent = lse(sym_mg.adj, n_components=None)
left_latent = latent[:n_pairs, :]
right_latent = latent[n_pairs:2 * n_pairs, :]
R, scalar = orthogonal_procrustes(left_latent, right_latent)
n_components = latent.shape[1]
class_labels = sym_mg["lineage"]
n_unique = len(np.unique(class_labels))
sym_mg.meta["Original index"] = range(len(sym_mg.meta))
left_df = sym_mg.meta[sym_mg.meta["Hemisphere"] == "L"]
left_inds = left_df["Original index"].values
left_latent = latent[left_inds, :]
left_latent = left_latent @ R
latent[left_inds, :] = left_latent
latent_cols = [f"dim {i}" for i in range(latent.shape[1])]
