eigs.py

import scipy.sparse as sps
from scipy.spatial.distance import pdist, squareform

def get_normalized_laplacian(sq_distances, sigma_features):
  squareform_sq_distances = np.exp(-squareform(sq_distances) / (2 * sigma_features**2))
  lap = sps.csgraph.laplacian(squareform_sq_distances, normed=True, return_diag=False)
  return lap

def get_square_distances(output):
  # output: [512, 128, 10, 18]
  output = output.reshape(512, -1).transpose()
  sq_distances = pdist(output, 'sqeuclidean')
  return sq_distances

num_clusters = 50
sigma_features = 3.0
sq_distances = get_square_distances(output)
norm_laplace = get_normalized_laplacian(sqdistances, sigma_features=sigma_features)
eigenvalues, eigenvectors = sps.linalg.eigsh(norm_laplace, k=num_clusters, which='SA')