def fit(self, X, y=None):
n_samples = X.shape[0]
self.n_samples_ = n_samples
self.cum_dist_ = 0
if n_samples > self.min_split_samples:
if self.cluster_method == "graspy-gmm":
cluster = GaussianCluster(
min_components=1,
max_components=2,
n_init=self.n_init,
covariance_type="all",
)
elif self.cluster_method == "auto-gmm":
cluster = AutoGMMCluster(
min_components=1, max_components=2, max_agglom_size=None
)
elif self.cluster_method == "vmm":
# cluster = VonMisesFisherMixture(n)
pass
else:
raise ValueError(f"`cluster_method` must be one of {valid_methods}")
cluster.fit(X)
pred_labels = cluster.predict(X)
self.pred_labels_ = pred_labels
self.model_ = cluster
if hasattr(cluster, "bic_"):
bics = cluster.bic_
self.bics_ = bics
bic_ratio = bics.loc[2].min() / bics.loc[1].min()
self.bic_ratio_ = bic_ratio
if cluster.n_components_ != 1: # recurse
indicator = pred_labels == 0
self.X_children_ = (X[indicator, :], X[~indicator, :])
children = []
for i, X_child in enumerate(self.X_children_):
child = DivisiveCluster(
name=self.name + str(i),
parent=self,
min_split_samples=self.min_split_samples,
n_init=self.n_init,
cluster_method=self.cluster_method,
)
child = child.fit(X_child)
children.append(child)
self.children = children
return self
def one_iteration(start_labels, class_key="Merge Class"):
# generate walks
data, bins, classes = random_walk_classes(start_labels,
seed=None,
class_key=class_key)
log_data = np.log10(data + 1)
# plot the clustermap
path_clustermap(log_data, classes, bins)
# embed and plot by known class
embedding = PCA(n_components=8).fit_transform(log_data)
pairplot(embedding, labels=classes, palette=CLASS_COLOR_DICT)
# cluster
agm = AutoGMMCluster(min_components=2,
max_components=20,
n_jobs=-1,
verbose=10)
pred_labels = agm.fit_predict(embedding)
plt.figure()
sns.scatterplot(data=agm.results_, x="n_components", y="bic/aic")
# plot embedding by cluster
pairplot(embedding, labels=pred_labels, palette=cc.glasbey_light)
# plot predicted clusters by known class
stacked_barplot(pred_labels, classes, color_dict=CLASS_COLOR_DICT)
return pred_labels
print("Finding pairwise jaccard distances")
pdist_sparse = pairwise_sparse_jaccard_distance(path_mat)
print(pdist_sparse.shape)
print("Embedding with MDS")
mds = ClassicalMDS(dissimilarity="precomputed")
# mds = MDS(dissimilarity="precomputed", n_components=6, n_init=16, n_jobs=-2)
jaccard_embedding = mds.fit_transform(pdist_sparse)
# %% [markdown]
# #
print("Clustering embedding")
agmm = AutoGMMCluster(min_components=10,
max_components=40,
affinity="euclidean",
linkage="single")
labels = agmm.fit_predict(jaccard_embedding)
pairplot(jaccard_embedding,
title="AGMM o CMDS o Jaccard o Sensorimotor Paths",
labels=labels)
savefig("AGMM-CMDS-jaccard-sm-path")
print("Finding mean paths")
mean_paths = []
uni_labels = np.unique(labels)
for ul in uni_labels:
inds = np.where(labels == ul)[0]
paths = path_mat[inds, :]
mean_path = np.array(np.mean(paths, axis=0))
palette=CLASS_COLOR_DICT,
)
stashfig("raw-response-pairs" + basename)
# %% [markdown]
# # Cluster each thing separately
from graspy.cluster import AutoGMMCluster
pred = []
for i, name in enumerate(from_group_names):
print(name)
o = log_collapsed_hist[:, i * n_bins:(i + 1) * n_bins]
agmm = AutoGMMCluster(
min_components=2,
max_components=20,
n_jobs=-2,
verbose=10,
affinity=["euclidean", "manhattan", "none"],
)
pred_labels = agmm.fit_predict(o)
pairplot(o[:, :5],
labels=meta["Merge Class"].values,
palette=CLASS_COLOR_DICT)
stashfig(f"from-g{name}-known" + basename)
pairplot(o[:, :5], labels=pred_labels, palette=cc.glasbey_light)
stashfig(f"from-g{name}-predicted" + basename)
print(len(np.unique(pred_labels)))
pred.append(pred_labels)
print()
# generate_cascade_paths(start_ind, probs, 1, stop_inds=out_inds, max_depth=10)
for i, (fg, fg_name) in enumerate(zip(from_groups, from_group_names)):
print(f"Clustering for {fg_name}")
# run the clustering on histogram
hop_hist = fg_hop_hists[i]
X = hop_hist.T
if normalize:
sums = X.sum(axis=1)
sums[sums == 0] = 1
X = X / sums[:, None]
if log_cluster:
X = np.log10(X + 1)
agmm = AutoGMMCluster(**cluster_kws)
pred_labels = agmm.fit_predict(X)
results = agmm.results_
fg_col_meta[i]["pred_labels"] = pred_labels
fg_autoclusters.append(agmm)
ggmm = GaussianCluster(min_components=10,
max_components=40,
n_init=20,
covariance_type="diag")
ggmm.fit(X)
fg_graspyclusters.append(ggmm)
gbics = ggmm.bic_
fig, ax = plt.subplots(1, 1, figsize=(10, 5))
pc.set_zorder(10)
ax.plot(rng, cmds.singular_values_, "o-")
ax.legend()
stashfig("cmds-screeplot" + basename)
# %% [markdown]
# ##
pairplot(path_embed, alpha=0.02)
stashfig("cmds-pairs-all" + basename)
# %% [markdown]
# ##
print("Running AGMM on CMDS embedding")
n_components = 4
agmm = AutoGMMCluster(max_components=40, n_jobs=-2)
pred = agmm.fit_predict(path_embed[:, :n_components])
print(f"Number of clusters: {agmm.n_components_}")
# %% [markdown]
# ##
pairplot(
path_embed[:, :n_components],
alpha=0.02,
labels=pred,
palette=cc.glasbey_light,
legend_name="Cluster",
)
stashfig("pairplot-agmm-cmds" + basename)
screeplot(all_hop_hist, show_first=40)
stashfig("scree-first-40")
screeplot(all_hop_hist, show_first=None)
stashfig("scree-all")
screeplot(np.log10(all_hop_hist + 1), show_first=100)
screeplot(np.log10(all_hop_hist + 1), show_first=100, cumulative=True)
# %% [markdown]
# ##
from graspy.cluster import AutoGMMCluster
agmm = AutoGMMCluster(
min_components=2,
max_components=50,
affinity=["euclidean", "manhattan"],
max_agglom_size=3000,
n_jobs=-2,
verbose=10,
)
agmm.fit(all_hop_hist.T)
# %% [markdown]
# ##
from graspy.embed import select_dimension
select_dimension(all_hop_hist.T, n_elbows=5)
#%%
from graspy.embed import selectSVD
from graspy.plot import pairplot
side_mb_mg = side_mgs[side]
labels = side_mb_mg.meta["class1"].values
labels = np.vectorize(label_map.get)(labels)
plot_labels = side_mb_mg.meta["merge_class"].values
# embed
ase = AdjacencySpectralEmbed(n_components=None, algorithm="randomized")
embed = ase.fit_transform(pass_to_ranks(side_mb_mg.adj))
embed = np.concatenate(embed, axis=1)
# cluster using AutoGMM
method = "AutoGMM"
agmm = AutoGMMCluster(
min_components=2,
max_components=10,
affinity=["euclidean", "manhattan", "cosine"],
covariance_type=["full"],
n_jobs=-1,
)
agmm.fit(embed, labels)
agmm_results = agmm.results_.copy()
agmm_results.sort_values("bic/aic", inplace=True)
agmm_model = agmm.model_
agmm_pred_labels = agmm_model.predict(embed)
ari = adjusted_rand_score(labels, agmm_pred_labels)
ari_no_kc = adjusted_rand_score(
labels[labels != "KC"], agmm_pred_labels[labels != "KC"]
)
row = dict(
ari=ari,
ari_no_kc=ari_no_kc,
embedding = PCA(n_components=8).fit_transform(raw_hist_data) pairplot(embedding, labels=dfs[0]["Merge Class"].values, palette=CLASS_COLOR_DICT) # %% [markdown] # # from sklearn.cluster import AgglomerativeClustering agg = AgglomerativeClustering(n_clusters=10, affinity="euclidean", linkage="average") labels = agg.fit_predict(raw_hist_data) pairplot(embedding, labels=labels, palette=cc.glasbey_light) # %% [markdown] # # from graspy.cluster import AutoGMMCluster agm = AutoGMMCluster(min_components=2, max_components=20, n_jobs=-1 agm.fit(embedding) # %% [markdown] # # # agm.results_.groupby(["affinity", "covariance_type", "linkage"]) sns.scatterplot(data=agm.results_, x='n_components', y='bic/aic') # %% [markdown] # # new_groups = agm.predict(embedding) stacked_barplot(new_groups, meta["Merge Class"].values, color_dict=CLASS_COLOR_DICT) # %% [markdown] # #
else:
fig, ax = plt.subplots(1, 1, figsize=(10, 10))
sns.scatterplot(path_embed[:, 0], path_embed[:, 1], s=30, alpha=0.2)
ax.axis("off")
stashfig(f"pairs-all" + basename)
# %% [markdown]
# ## Cluster and plot on the embedding
print("Running AGMM on path embedding")
n_components = elbows[0]
# n_components = 2d
print(f"Using {n_components} dimensions")
agmm = AutoGMMCluster(max_components=30, n_jobs=-2)
pred = agmm.fit_predict(path_embed[:, :n_components]
) # + np.random.normal(0, 0.01, size=path_embed.shape)
print(f"Number of clusters: {agmm.n_components_}")
pg = pairplot(
path_embed[:, :n_components],
alpha=0.1,
labels=pred,
palette=cc.glasbey_light,
legend_name="Cluster",
)
leg = pg._legend
for lh in leg.legendHandles:
lh.set_alpha(1)
heatmap(
mg.adj,
transform="simple-all",
title=f"MB, threshold={threshold}",
inner_hier_labels=true_labels,
hier_label_fontsize=10,
sort_nodes=True,
)
latent = ase(mg.adj, n_components, ptr=ptr)
# cluster = GaussianCluster(
# min_components=2, max_components=10, covariance_type="all", n_init=100
# )
cluster = AutoGMMCluster(min_components=2, max_components=10)
pred_labels = cluster.fit_predict(latent)
ari = adjusted_rand_score(true_labels, pred_labels)
row = {"ARI": ari, "Threshold": threshold, "Method": "GMMoASE"}
rows.append(row)
# do the MCMC
block_series = run_minimize_blockmodel(mg, weight_model="discrete-poisson")
ari = adjusted_rand_score(true_labels, block_series.values)
row = {"ARI": ari, "Threshold": threshold, "Method": "GT-dp"}
rows.append(row)
# do the MCMC
block_series = run_minimize_blockmodel(mg, weight_model=None)
ari = adjusted_rand_score(true_labels, block_series.values)
row = {"ARI": ari, "Threshold": threshold, "Method": "GT-None"}
cutoff = 8
base = f"-c{cutoff}-t{threshold}-{graph_type}"
base_path = Path(f"./maggot_models/notebooks/outs/{run_name}/csvs")
meta = pd.read_csv(base_path / str("meta" + base + ".csv"), index_col=0)
path_mat = pd.read_csv(base_path / str("prob-path-mat" + base + ".csv"),
index_col=0).values
base_path = Path(f"./maggot_models/notebooks/outs/{embed_name}/csvs")
embed_mat = pd.read_csv(base_path / str("euclid-mds-embed.csv"), index_col=0)
gmm = AutoGMMCluster(
min_components=10,
max_components=50,
affinity="all",
linkage="all",
covariance_type="all",
n_jobs=-2,
verbose=30,
)
labels = gmm.fit_predict(embed_mat.values)
label_df = pd.DataFrame(data=labels)
stashcsv(label_df, "labels")
print("Finding mean paths")
mean_paths = []
uni_labels = np.unique(labels)
for ul in uni_labels:
inds = np.where(labels == ul)[0]
paths = path_mat[inds, :]