def main():
clust_num = 3
data_shape = 2
mu_list, sigma_list, w_list = get_sparse_gmm_model(clust_num, data_shape)
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x ** spsa_gamma)
spsa_beta = lambda x: 15. / (x ** (spsa_gamma / 4))
# spsa_alpha = lambda x: 0.001
# spsa_beta = lambda x: 0.001
clustering = spsa_clustering.ClusteringSPSA(n_clusters=clust_num, data_shape=data_shape, Gammas=None, alpha=spsa_alpha,
beta=spsa_beta, norm_init=False, verbose=False, sparse=False, eta=None)
N = 3000
data_set = []
true_labels = []
for _ in range(N):
mix_ind = np.random.choice(len(w_list), p=w_list)
data_point = np.random.multivariate_normal(mu_list[mix_ind], np.identity(data_shape) * sigma_list[mix_ind])
data_set.append(data_point)
true_labels.append(mix_ind)
clustering.fit(data_point)
data_set = np.array(data_set)
dataset_name = 'good'
dataset_dir = os.path.join('datasets', dataset_name)
if not os.path.isdir(dataset_dir):
os.mkdir(dataset_dir)
np.save(os.path.join(dataset_dir, 'data.npy'), data_set)
np.save(os.path.join(dataset_dir, 'true.npy'), np.array(true_labels))
param = {'mu': mu_list, 'sigma': sigma_list, 'w': w_list}
with open(os.path.join(dataset_dir, 'param.pickle'), 'wb') as f:
pickle.dump(param, f)
utils.order_clust_centers(np.array(mu_list), clustering)
clustering.clusters_fill(data_set)
ari_spsa = metrics.adjusted_rand_score(true_labels, clustering.labels_)
print('ARI: {}'.format(ari_spsa))
print('Mean centers dist: {}'.format(utils.mean_cent_dist(np.array(mu_list), clustering)))
utils.plot_centers(np.array(mu_list), clustering)
# utils.plot_centers_converg(np.array(mu_list), clustering)
utils.plot_clustering(data_set, clustering.labels_, 'SPSA clustering partition')
utils.plot_clustering(data_set, true_labels, 'True partition')
plt.show()
def load_experiment(name='bad'):
dataset_dir = os.path.join('datasets', name)
data_set = np.load(os.path.join(dataset_dir, 'data.npy'))
true_labels = np.load(os.path.join(dataset_dir, 'true.npy'))
with open(os.path.join(dataset_dir, 'param.pickle'), 'rb') as f:
param = pickle.load(f)
mu_list, sigma_list, w_list = param['mu'], param['sigma'], param['w']
clust_num = len(mu_list)
data_shape = data_set[0].shape[0]
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x ** spsa_gamma)
spsa_beta = lambda x: 15. / (x ** (spsa_gamma / 4))
clustering = spsa_clustering.ClusteringSPSA(n_clusters=clust_num, data_shape=data_shape, Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta, norm_init=False, verbose=False, sparse=False, eta=None,
spsa_sigma=False)
rand_ind = np.random.permutation(data_set.shape[0])
for i in rand_ind:
clustering.fit(data_set[i])
# utils.order_clust_centers(np.array(mu_list), clustering)
clustering.clusters_fill(data_set[rand_ind])
ari_spsa = metrics.adjusted_rand_score(true_labels[rand_ind], clustering.labels_)
print('ARI: {}'.format(ari_spsa))
print('Mean centers dist: {}'.format(utils.mean_cent_dist(np.array(mu_list), clustering)))
utils.plot_centers(np.array(mu_list), clustering)
# utils.plot_centers_converg(np.array(mu_list), clustering)
# utils.plot_clustering(data_set[rand_ind], clustering.labels_, 'SPSA clustering partition')
# utils.plot_clustering(data_set[rand_ind], true_labels[rand_ind], 'True partition')
# for Gamma in clustering.Gammas:
# print(Gamma)
# for center in clustering.cluster_centers_:
# print(center)
# utils.plot_clustering_cov(data_set, clustering.labels_, 'SPSA clustering partition', clustering.cluster_centers_,
# clustering.Gammas)
plt.show()
ari_kmeans = np.zeros(n_run)
ari_spsa = np.zeros(n_run)
ari_spsa_cov = np.zeros(n_run)
ari_gmm = np.zeros(n_run)
ari_pam = np.zeros(n_run)
centers_dist_kmeans = np.zeros(n_run)
centers_dist_spsa = np.zeros(n_run)
centers_dist_spsa_cov = np.zeros(n_run)
centers_dist_gmm = np.zeros(n_run)
for i in range(n_run):
print('Run {0}'.format(i))
clustering = spsa_clustering.ClusteringSPSA(n_clusters=clust_means.shape[0], data_shape=2, Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta, norm_init=False, verbose=False, noise=noises[j])
clustering_cov = spsa_clustering.ClusteringSPSA(n_clusters=clust_means.shape[0], data_shape=2, Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta, norm_init=False, noise=noises[j],
eta=3000, verbose=False)
data_set = []
true_labels = []
for _ in range(N):
mix_ind = np.random.choice(len(mix_prob), p=mix_prob)
data_point = np.random.multivariate_normal(clust_means[mix_ind], clust_gammas[mix_ind])
data_set.append(data_point)
true_labels.append(mix_ind)
# clustering.fit(data_point)
# data_generator.save_example()
train_generator = data_generator.generate('train')
centers_fname = '/home/a.boiarov/Projects/spsa_clustering_gmm_log/centers.npy'
if not args.only_clf:
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x**spsa_gamma)
spsa_beta = lambda x: 15. / (x**(spsa_gamma / 4))
n_filters = 500
clustering = spsa_clustering.ClusteringSPSA(n_clusters=n_filters,
data_shape=patch_size *
patch_size,
Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta,
norm_init=False,
eta=900)
# spsa_train_num = data_generator.train_number
spsa_train_num = 1500
num = 0
for _ in range(spsa_train_num):
print(num)
num += 1
train_data = next(train_generator)
for patch in train_data[0]:
patch = patch.flatten()
clustering.fit(patch)
noise_5 = spsa_clustering.Noise(func=lambda x: 10 * (np.random.rand() * 4 - 2),
name='random')
noise_6 = spsa_clustering.Noise(func=lambda x: 0.1 * np.sin(x) + 19 * np.sign(50 - x % 100),
name='irregular')
noise_7 = spsa_clustering.Noise(func=lambda x: 20, name='constant')
experiment_noise = noise_0
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x ** spsa_gamma)
spsa_beta = lambda x: 15. / (x ** (spsa_gamma / 4))
# spsa_alpha = lambda x: 0.0001
# spsa_beta = lambda x: 0.0001
clustering = spsa_clustering.ClusteringSPSA(n_clusters=10, data_shape=784, Gammas=None, alpha=spsa_alpha,
beta=spsa_beta, norm_init=False, noise=experiment_noise)
data_set = []
true_labels = []
# init_ind = []
# for label in range(10):
# ind = np.random.choice(df.index[df['label'] == label].tolist(), 1)
# row = df.loc[ind[0], :]
# true_labels.append(row[0])
# data_point = np.array(row[1:].tolist(), dtype=float)
# data_set.append(data_point)
# clustering.fit(data_point)
# init_ind.append(ind)
index = list(range(df.shape[0]))
clust_means = np.array([[0, 0], [2, 2], [-3, 6]])
clust_gammas = np.array([[[1, -0.7], [-0.7, 1]],
np.eye(2), [[1, 0.8], [0.8, 1]]])
data_set = []
true_labels = []
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x**spsa_gamma)
spsa_beta = lambda x: 15. / (x**(spsa_gamma / 4))
# spsa_alpha = lambda x: 0.001
# spsa_beta = lambda x: 0.001
clustering = spsa_clustering.ClusteringSPSA(n_clusters=clust_means.shape[0],
data_shape=2,
Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta,
norm_init=False)
for _ in range(N):
mix_ind = np.random.choice(len(mix_prob), p=mix_prob)
data_point = np.random.multivariate_normal(clust_means[mix_ind],
clust_gammas[mix_ind])
data_set.append(data_point)
true_labels.append(mix_ind)
clustering.fit(data_point)
data_set = np.array(data_set)
utils.order_clust_centers(clust_means, clustering)
clustering.clusters_fill(data_set)
noise_7 = spsa_clustering.Noise(func=lambda x: [20] * x.shape[0],
name='constant')
experiment_noise = noise_3
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x**spsa_gamma)
spsa_beta = lambda x: 15. / (x**(spsa_gamma / 4))
# spsa_alpha = lambda x: 0.001
# spsa_beta = lambda x: 0.001
clustering = spsa_clustering.ClusteringSPSA(n_clusters=3,
data_shape=2,
Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta,
norm_init=False,
noise=experiment_noise,
eta=None)
clustering_cov = spsa_clustering.ClusteringSPSA(n_clusters=3,
data_shape=2,
Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta,
norm_init=False,
noise=experiment_noise,
eta=1000)
for _ in range(N):
mix_ind = np.random.choice(len(mix_prob), p=mix_prob)
def stat():
clust_num = 3
data_shape = 2
mu_list, sigma_list, w_list = get_sparse_gmm_model(clust_num, data_shape)
spsa_gamma = 1. / 6
spsa_alpha = lambda x: 0.25 / (x ** spsa_gamma)
spsa_beta = lambda x: 15. / (x ** (spsa_gamma / 4))
# spsa_alpha = lambda x: 0.001
# spsa_beta = lambda x: 0.001
n_run = 10
N = 3000
ari_spsa = np.zeros(n_run)
ari_kmeans = np.zeros(n_run)
ari_mb_kmeans = np.zeros(n_run)
ari_pam = np.zeros(n_run)
cent_dist = np.zeros(n_run)
cent_dist_kmeans = np.zeros(n_run)
cent_dist_mb_kmeans = np.zeros(n_run)
cent_dist_pam = np.zeros(n_run)
for i in tqdm(range(n_run)):
clustering = spsa_clustering.ClusteringSPSA(n_clusters=clust_num, data_shape=data_shape, Gammas=None,
alpha=spsa_alpha,
beta=spsa_beta, norm_init=False, verbose=False, sparse=True, eta=700,
spsa_sigma=False)
kmeans = cluster.KMeans(n_clusters=clust_num)
mb_kmeans = cluster.MiniBatchKMeans(n_clusters=clust_num, n_init=1, init='random', max_iter=1,
batch_size=1,
max_no_improvement=None)
data_set = []
true_labels = []
for _ in range(N):
mix_ind = np.random.choice(len(w_list), p=w_list)
data_point = np.random.multivariate_normal(mu_list[mix_ind], np.identity(data_shape) * sigma_list[mix_ind])
data_set.append(data_point)
true_labels.append(mix_ind)
# clustering.fit(data_point)
data_set = np.array(data_set)
# utils.order_clust_centers(np.array(mu_list), clustering)
# clustering.clusters_fill(data_set)
labels_pred_kmenas = kmeans.fit_predict(data_set)
labels_pred_mb_kmeans = mb_kmeans.fit_predict(data_set)
dist = pairwise_distances(data_set)
labels_pred_pam, pam_med = pam.cluster(dist, k=clust_num)
# ari_spsa[i] = metrics.adjusted_rand_score(true_labels, clustering.labels_)
# cent_dist[i] = utils.mean_cent_dist(np.array(mu_list), clustering)
ari_kmeans[i] = metrics.adjusted_rand_score(true_labels, labels_pred_kmenas)
ari_mb_kmeans[i] = metrics.adjusted_rand_score(true_labels, labels_pred_mb_kmeans)
ari_pam[i] = metrics.adjusted_rand_score(true_labels, labels_pred_pam)
cent_dist_kmeans[i] = utils.mean_cent_dist_(np.array(mu_list), kmeans.cluster_centers_)
cent_dist_mb_kmeans[i] = utils.mean_cent_dist_(np.array(mu_list), mb_kmeans.cluster_centers_)
cent_dist_pam[i] = utils.mean_cent_dist_(np.array(mu_list), data_set[pam_med])
print(ari_spsa.mean(), cent_dist.mean())
print('\nMean ARI k-means: {:f}, Mean L2: {:f}'.format(ari_kmeans.mean(), cent_dist_kmeans.mean()))
print('Mean ARI online k-means: {:f}, Mean L2: {:f}'.format(ari_mb_kmeans.mean(), cent_dist_mb_kmeans.mean()))
# print('Mean ARI SPSA clustering: {:f}, Mean L2: {:f}'.format(ari_spsa.mean(), cen))
print('\nMean ARI PAM: {:f}, Mean L2: {:f}'.format(ari_pam.mean(), cent_dist_pam.mean()))
