def draw_mlp_clustering_report(weight_directory,
double_clustering_results,
n_cluster=4,
title=None,
figsize=(20, 30)):
weight_paths = get_weights_paths(weight_directory)
fig, axes = plt.subplots(2, 2, figsize=figsize)
if title is not None:
fig.suptitle(title)
axes[0][0].set_title('Unpruned')
draw_clustered_mlp(
weight_paths[True], # True represents **un**pruned
double_clustering_results[True],
n_clusters=n_cluster,
ax=axes[0][0])
draw_cluster_by_layer(weight_paths[True],
double_clustering_results[True],
n_clusters=n_cluster,
ax=axes[1][0])
axes[0][1].set_title('Pruned')
draw_clustered_mlp(weight_paths[False],
double_clustering_results[False],
n_clusters=n_cluster,
ax=axes[0][1])
draw_cluster_by_layer(weight_paths[False],
double_clustering_results[False],
n_clusters=n_cluster,
ax=axes[1][1])
def build_weighted_dist_mat(model_path,
clustering_result,
normalize_in_out=False):
weight_path = get_weights_paths(model_path)[False] # pruned
G = build_cluster_graph(
weight_path,
clustering_result[False], # pruned
normalize_in_out=normalize_in_out)
df = pd.DataFrame([{
'start': start,
'end': end,
'dist': _compute_weighted_dist(G, start, end)
} for start, end in it.combinations(G.nodes, 2)])
df = df[df != 0].dropna()
# The distance is normalized to [0, 1] inside the paths between two specific layers
# The max weighted sitance is one.
df['layers'] = df.apply(
lambda r: r['start'].split('-')[0] + '-' + r['end'].split('-')[0],
axis=1)
df['normalized_dist'] = df['dist'] / df.groupby(
'layers')['dist'].transform('max')
mat = df.pivot('start', 'end', 'normalized_dist')
return mat
def plot_eigenvalue_report(weight_directory,
unpruned_n_eigenvalues=None,
pruned_n_eigenvalues=None,
figsize=(10, 5)):
weight_paths = get_weights_paths(weight_directory)
is_slice = (unpruned_n_eigenvalues is not None
or pruned_n_eigenvalues is not None)
n_rows = 2 if is_slice else 1
_, axes = plt.subplots(n_rows, 2, squeeze=False, figsize=figsize)
axes[0][0].set_title('Unpruned')
plot_eigenvalues(weight_paths[True], ax=axes[0][0])
if is_slice:
plot_eigenvalues(weight_paths[True],
unpruned_n_eigenvalues,
ax=axes[1][0])
axes[0][1].set_title('Pruned')
plot_eigenvalues(weight_paths[False], ax=axes[0][1])
if is_slice:
plot_eigenvalues(weight_paths[False],
pruned_n_eigenvalues,
ax=axes[1][1])
def run_double_spectral_cluster(
weight_directory,
with_shuffle=True,
n_clusters=4,
shuffle_method='layer',
n_samples=None,
n_workers=None,
with_shuffled_ncuts=False,
random_state=RANDOM_STATE,
):
weight_paths = get_weights_paths(weight_directory)
return {
is_unpruned:
run_spectral_cluster(weight_path, with_shuffle, n_clusters,
shuffle_method, n_samples, n_workers,
with_shuffled_ncuts, random_state)
for is_unpruned, weight_path in weight_paths.items()
}