def evaluate(self, metric):
if self.true_labels == 0:
print("Error: A true_labels.txt file is needed")
return
if metric == "nmi":
print("NMI: %f" % nmi(self.true_labels, self.predicted_labels))
elif metric == "purity":
print("Purity: %f" % purity(self.true_labels, self.predicted_labels))
elif metric == "both":
print("NMI: %f" % nmi(self.true_labels, self.predicted_labels))
print("Purity: %f" % purity(self.true_labels, self.predicted_labels))
else:
print("Error: This metric is not available. Choose among the following options: 'nmi', 'purity', 'both'")
def nmi(self):
"""
Return the Normalized Mutual Information (NMI) of the clustering
solution with respect to the ground-truth labels given.
"""
z = np.argmax(self.pi_nk, 1)
return nmi(z, self.label)
def cluster(o, n_clusters=3):
df = pd.read_pickle(o)
reprsn = df['embedding'].values
node_idx = df['node_id'].values
labels = ry1()
reprsn = [np.asarray(row, dtype='float32') for row in reprsn]
reprsn = np.array(reprsn, dtype='float32')
true_labels = [labels[int(node)] for node in node_idx]
data = reprsn
km = KMeans(init='k-means++', n_clusters=n_clusters, n_init=10)
km.fit(data)
km_means_labels = km.labels_
km_means_cluster_centers = km.cluster_centers_
km_means_labels_unique = np.unique(km_means_labels)
'''colors_ = cycle(colors.cnames.keys())
m,initial_dim = np.shape(data)
data_2 = tsne(data, 2, initial_dim, 30)
plt.figure(figsize=(12, 6))
plt.scatter(data_2[:, 0], data_2[:, 1], c=true_labels)
plt.title('True Labels')
plt.show()'''
nmiv = nmi(true_labels, km_means_labels)
print ' NMI value ', nmiv
return nmiv #cal the accuracy of valdata
def nmi_matrix(df):
mat = pd.DataFrame(index=df.columns, columns=df.columns, data=0)
for i in df.columns:
for j in df.columns:
mat.loc[i, j] = nmi(df[i], df[j])
return mat
def benchmark(self, name: str, features: np.ndarray,
labels: np.ndarray) -> Tuple[str, Dict]:
"""
Returns the clustering performance results in str and dict format.
The metrics used are as follows:
1. Duration
2. Adjusted RAND Score
3. Normalized Mutual Information
4. Davies-Bouldin Index
5. Silhouette Score
6. Calinski-Harabasz Score
7. Clustering Accuracy
Parameters
----------
name: str
The name of the benchmark.
features: np.ndarray
The test instances to cluster.
labels: np.ndarray
The test labels.
Returns
-------
str
The formatted string of the benchmark results.
results: Dict
The dictionary of benchmark results.
"""
start_time = time.time()
predictions = self.predict(features)
results = {}
results["name"] = name
results["duration"] = time.time() - start_time
results["ari"] = ari(labels_true=labels, labels_pred=predictions)
results["nmi"] = nmi(labels_true=labels, labels_pred=predictions)
results["dbi"] = davies_bouldin_score(features, predictions)
results["silhouette"] = silhouette_score(features,
predictions,
metric="euclidean")
results["ch_score"] = calinski_harabasz_score(features, predictions)
results["clustering_accuracy"] = clustering_accuracy(
target=labels, prediction=predictions)
return (
"%-9s\t%.2fs\t%.3f\t\t%.3f\t\t%.3f\t\t%.3f\t\t%.3f\t\t%.3f" % (
results.get("name"),
results.get("duration"),
results.get("dbi"),
results.get("silhouette"),
results.get("ch_score"),
results.get("nmi"),
results.get("ari"),
results.get("clustering_accuracy"),
),
results,
)
def get_accuracy(cluster_assignments, y_true, n_clusters):
'''
Computes the accuracy based on the provided kmeans cluster assignments
and true labels, using the Munkres algorithm
cluster_assignments: array of labels, outputted by kmeans
y_true: true labels
n_clusters: number of clusters in the dataset
returns: a tuple containing the accuracy and confusion matrix,
in that order
'''
y_pred, confusion_matrix = get_y_preds(cluster_assignments, y_true,
n_clusters)
from sklearn.metrics import normalized_mutual_info_score as nmi
nmi_score = nmi(y_true, y_pred)
print('NMI: ' + str(np.round(nmi_score, 4)))
from sklearn.metrics import adjusted_rand_score as ari
ari_score = ari(y_true, y_pred)
print('ARI: ' + str(np.round(ari_score, 4)))
# with open('C:/Users/mals6571/Desktop/SpectralNet-master/src/applications/Results.txt','a') as my_file:
ProjectDir = get_project_root()
with open(os.path.join(ProjectDir, 'Results.txt'), 'a') as my_file:
my_file.write("\n")
my_file.write('NMI: ' + str(np.round(nmi_score, 4)))
my_file.write("\n")
my_file.write('ARI: ' + str(np.round(ari_score, 4)))
my_file.write("\n")
# calculate the accuracy
return np.mean(y_pred == y_true), confusion_matrix
def run_net(data, params):
#
# UNPACK DATA
#
x_train_unlabeled, y_train_unlabeled, x_val, y_val, x_test, y_test = data[
'spectral']['train_and_test']
inputs_vae = Input(shape=(params['img_dim'], params['img_dim'], 1),
name='inputs_vae')
ConvAE = Conv.ConvAE(inputs_vae, params)
ConvAE.vae.load_weights('vae_mnist.h5')
lh = LearningHandler(lr=params['spec_lr'],
drop=params['spec_drop'],
lr_tensor=ConvAE.learning_rate,
patience=params['spec_patience'])
lh.on_train_begin()
losses_vae = np.empty((500, ))
for i in range(500):
# if i==0:
x_val_y = ConvAE.vae.predict(x_val)[2]
losses_vae[i] = ConvAE.train_vae(x_val, x_val_y, params['batch_size'])
x_val_y = ConvAE.vae.predict(x_val)[2]
y_sp = x_val_y.argmax(axis=1)
print_accuracy(y_sp, y_val, params['n_clusters'])
print("Epoch: {}, loss={:2f}".format(i, losses_vae[i]))
# else:
# losses_vae[i] = ConvAE.train_vae(x_val, x_val_y,params['batch_size'])
# x_val_y = ConvAE.vae.predict(x_val)[2]
# y_sp = x_val_y.argmax(axis=1)
# print_accuracy(y_sp, y_val, params['n_clusters'])
# print("Epoch: {}, loss={:2f}".format(i, losses_vae[i]))
if i > 1:
if np.abs(losses_vae[i] - losses_vae[i - 1]) < 0.0001:
print('STOPPING EARLY')
break
# if self.lh.on_epoch_end(i, val_losses[i]):
# print('STOPPING EARLY')
# break
# print training status
# ConvAE.vae.save_weights('IJCAI_mnist2.h5')
# spectral_net.net.save_weight('save.h5')
# spectral_net.save
print("finished training")
x_val_y = ConvAE.vae.predict(x_val)[2]
# x_val_y = ConvAE.classfier.predict(x_val_lp)
y_sp = x_val_y.argmax(axis=1)
print_accuracy(y_sp, y_val, params['n_clusters'])
from sklearn.metrics import normalized_mutual_info_score as nmi
nmi_score1 = nmi(y_sp, y_val)
print('NMI: ' + str(np.round(nmi_score1, 4)))
def evaluate_model(self):
"""calculates NMI
Arguments:
ground_truth,
predicted_values
"""
nmi_eval = nmi(self.ground_truth, self.predicted)
print(f"NMI Accuracy is: {nmi_eval}")
def metriques(model, y_true, y_pred):
pred1 = model.row_labels_
nmi_ = nmi(y_true, pred1)
ari_ = ari(y_true, pred1)
accuracy = ACCURACY(y_true, pred1)
print("NMI: {}\nARI: {} ".format(nmi_, ari_))
print("ACCURACY: %s" % accuracy)
return nmi_, ari_, accuracy
def run_experiment(ae_model_path):
logger.info(
f"++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++"
)
logger.info(
f"++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++"
)
logger.info(f"Working now on {ae_model_path.name}")
logger.info(
f"++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++"
)
logger.info(
f"++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++"
)
new_seed = random.randint(0, 1000)
logger.info(f"Seed value for this is: {new_seed}")
set_random_seed(new_seed)
ae_module = stacked_ae(pt_data.shape[1], [500, 500, 2000, 10],
weight_initalizer=torch.nn.init.xavier_normal_,
activation_fn=lambda x: F.relu(x),
loss_fn=None,
optimizer_fn=None)
model_data = torch.load(ae_model_path, map_location='cpu')
ae_module.load_state_dict(model_data)
ae_module = ae_module.cuda()
# Get embedded data
embedded_data = None
for batch_data in torch.utils.data.DataLoader(pt_data,
batch_size=256,
shuffle=False):
embedded_batch_np = ae_module.forward(
batch_data.cuda())[0].detach().cpu().numpy()
if embedded_data is None:
embedded_data = embedded_batch_np
else:
embedded_data = np.concatenate([embedded_data, embedded_batch_np],
0)
del ae_module
# Perform k-means
k_means_labels = k_means(embedded_data, n_clusters, n_init=20)[1]
k_means_nmi_value = nmi(gold_labels,
k_means_labels,
average_method='arithmetic')
k_means_acc_value = cluster_acc(gold_labels, k_means_labels)[0]
result_file = Path(f"{result_dir}/results_ae_kmeans_{dataset_name}.txt")
result_file_exists = result_file.exists()
f = open(result_file, "a+")
if not result_file_exists:
f.write("#\"ae_model_name\"\t\"NMI\"\t\"ACC\"\n")
f.write(
f"{ae_model_path.name}\t{k_means_nmi_value}\t{k_means_acc_value}\n")
f.close()
def tensfact_baseline():
n_clusters = 81
f = open('buzz_user_tensor_45.npy')
X_buzz = np.load(f)
print X_buzz.shape
X_buzz = X_buzz[buzz_ground.keys()]
buzz_ground1 = buzz_ground.values()
km = KMeans(n_clusters=81, init='k-means++', n_init=1, verbose=False)
sc = 0.0
sc1 = 0.0
sc2 = 0.0
for i in xrange(10):
km.fit(X_buzz)
sc += nmi(buzz_ground1, km.labels_)
sc1 += ari(buzz_ground1, km.labels_)
sc2 += ami(buzz_ground1, km.labels_)
print "BUZZ"
print "nmi score %f" % (sc / float(10))
print "ari score %f" % (sc1 / float(10))
print "ami score %f" % (sc2 / float(10))
f = open('poli_user_tensor_75.npy')
X_poli = np.load(f)
print X_poli.shape
X_poli = X_poli[poli_ground.keys()]
poli_ground1 = poli_ground.values()
sc = 0.0
sc1 = 0.0
km1 = KMeans(n_clusters=310, init='k-means++', n_init=1, verbose=False)
sc = 0.0
sc1 = 0.0
sc2 = 0.0
for i in xrange(10):
km1.fit(X_poli)
sc += nmi(poli_ground1, km1.labels_)
sc1 += ari(poli_ground1, km1.labels_)
sc2 += ami(poli_ground1, km1.labels_)
print "poli"
print "nmi score %f" % (sc / float(10))
print "ari score %f" % (sc1 / float(10))
print "ami score %f" % (sc2 / float(10))
def NMI(y_true, y_pred):
b, _, _, ch = y_true.shape
output = np.zeros((b, ch))
for b_idx in range(b):
for ch_idx in range(ch):
true_max = y_true[b_idx, ..., ch_idx].max()
pred_vmin = y_pred[b_idx, ..., ch_idx].max()
output[b_idx, ch_idx] = nmi((y_true[b_idx, ..., ch_idx]/true_max*255).astype('uint8').ravel(), (y_pred[b_idx, ..., ch_idx]/pred_vmin*255).astype('uint8').ravel())
return output
def execute_algo(model, model_name, X, y, verbose=True):
print("##############\n# {}\n##############".format(model_name))
model.fit(X)
res_nmi = nmi(model.row_labels_, y)
res_ari = ari(model.row_labels_, y)
res_acc = accuracy(model.row_labels_, y)
if verbose:
print("NMI =", res_nmi)
print("ARI =", res_ari)
print("ACC =", res_acc)
return res_nmi, res_ari, res_acc
def train(self):
x, y = np.load('images/64px_image_x.npy'), np.load(
'images/64px_image_y.npy')
x = np.reshape(x, (40000, 64, 64, 1))
kmeans = KMeans(n_clusters=2, n_init=20)
y_pred = kmeans.fit_predict(self.encoder.predict(x))
y_pred_last = np.copy(y_pred)
self.model.get_layer(name='clustering').set_weights(
[kmeans.cluster_centers_])
loss = 0
ae_loss = 0
index = 0
maxiter = 80000
update_interval = 100
index_array = np.arange(x.shape[0])
batch_size = 16
tol = 0.001
# model.load_weights('DEC_model_final.h5')
for ite in range(int(maxiter)):
if ite % update_interval == 0:
q = self.model.predict(x, verbose=0)
# update the auxiliary target distribution p
p = self.target_distribution(q)
# evaluate the clustering performance
y_pred = q.argmax(1)
if y is not None:
acc = np.round(metrics.acc(y, y_pred), 5)
nmi = np.round(metrics.nmi(y, y_pred), 5)
ari = np.round(metrics.ari(y, y_pred), 5)
loss = np.round(loss, 5)
print(
'Iter %d: acc = %.5f, nmi = %.5f, ari = %.5f, loss=%.5f'
% (ite, acc, nmi, ari, loss))
# check stop criterion - model convergence
delta_label = np.sum(y_pred != y_pred_last).astype(
np.float32) / y_pred.shape[0]
y_pred_last = np.copy(y_pred)
if ite > 0 and delta_label < tol:
print('delta_label ', delta_label, '< tol ', tol)
print('Reached tolerance threshold. Stopping training.')
break
idx = np.random.randint(low=0, high=x.shape[0], size=batch_size)
# ae_loss = ae.train_on_batch(x=x[idx], y=x[idx])
loss = self.model.train_on_batch(x=x[idx], y=p[idx])
index = index + 1 if (index + 1) * batch_size <= x.shape[0] else 0
self.model.save_weights('DEC_model_final_64px.h5')
self.test_model()
def on_epoch_end(self, epoch, logs=None):
if int(epochs/10) != 0 and epoch % int(epochs/10) != 0:
return
feature_model = Model(self.model.input,
self.model.get_layer(
'encoder_%d' % (int(len(self.model.layers) / 2) - 1)).output)
features = feature_model.predict(self.x)
km = KMeans(n_clusters=len(np.unique(self.y)), n_init=20, n_jobs=4)
y_pred = km.fit_predict(features)
# print()
print(' '*8 + '|==> acc: %.4f, nmi: %.4f <==|'
% (metrics.acc(self.y, y_pred), metrics.nmi(self.y, y_pred)))
def _are_labels_equal(labels_new, labels_old):
"""
Check if the old labels and new labels are equal. Therefore check the nmi for each subspace. If all are 1, labels
have not changed.
:param labels_new: new labels list
:param labels_old: old labels list
:return: True if labels for all subspaces are the same
"""
if labels_new is None or labels_old is None:
return False
return all([
nmi(labels_new[i], labels_old[i], average_method='arithmetic') == 1
for i in range(len(labels_new))
])
def evaluate(self, metric):
if self.true_labels == 0:
print("Error: A true_labels.csv file is needed")
return
true_labels = list(self.true_labels.values())
if metric == "nmi":
print("NMI: %f" % nmi(true_labels, self.predicted_labels))
elif metric == "purity":
print("Purity: %f" % purity(true_labels, self.predicted_labels))
elif metric == "confusion matrix":
print("Confusion matrix:")
print(confusion_matrix(true_labels, self.predicted_labels))
elif metric == "all":
print("NMI: %f" % nmi(true_labels, self.predicted_labels))
print("Purity: %f" % purity(true_labels, self.predicted_labels))
print("Confusion matrix:")
print(confusion_matrix(true_labels, self.predicted_labels))
else:
print(
"Error: This metric is not available. Choose among the following options: 'nmi', 'purity', 'both'"
)
def run_trial(X, labels, k):
errors = '"'
# Run our dbscan
start = time()
"""
if metric == 'seuclidean':
db = KMeans(eps,minPts,metric=metric,metric_params={'V':V})
else:
db = kmean(,minPts,metric=metric)
"""
db = KMeans(k, n_jobs=12)
pred_labels = db.fit_predict(X)
elapsed = time() - start
try:
ari_score = ari(pred_labels, labels)
except Exception as e:
errors += str(e) + '; '
ari_score = np.nan
try:
nmi_score = nmi(pred_labels, labels, average_method='arithmetic')
except Exception as e:
errors += str(e) + '; '
nmi_score = np.nan
try:
ss_score = ss(X, pred_labels)
except Exception as e:
errors += str(e) + '; '
ss_score = np.nan
try:
vrc_score = vrc(X, pred_labels)
except Exception as e:
errors += str(e) + '; '
vrc_score = np.nan
try:
dbs_score = dbs(X, pred_labels)
except Exception as e:
errors += str(e) + '; '
dbs_score = np.nan
errors += '"'
return [
k, elapsed, ari_score, nmi_score, ss_score, vrc_score, dbs_score,
errors
]
def run_net(data, params):
#
# UNPACK DATA
#
x_train_unlabeled, y_train_unlabeled, x_val, y_val, x_test, y_test = data[
'spectral']['train_and_test']
inputs_vae = Input(shape=(params['img_dim'], params['img_dim'], 1),
name='inputs_vae')
ConvAE = Conv.ConvAE(inputs_vae, params)
ConvAE.vae.load_weights(
'/home/stu2/Signal-1/Deep-Spectral-Clustering-using-Dual-Autoencoder-Network-master/src/applications/vae_mnist.h5'
)
lh = LearningHandler(lr=params['spec_lr'],
drop=params['spec_drop'],
lr_tensor=ConvAE.learning_rate,
patience=params['spec_patience'])
lh.on_train_begin()
losses_vae = np.empty((500, ))
for i in range(100):
# if i==0:
x_val_y = ConvAE.vae.predict(x_val)[2] #得到y
losses_vae[i] = ConvAE.train_vae(x_val, x_val_y, params['batch_size'])
x_val_y = ConvAE.vae.predict(x_val)[2]
y_sp = x_val_y.argmax(axis=1)
print_accuracy(y_sp, y_val, params['n_clusters'])
print("Epoch: {}, loss={:2f}".format(i, losses_vae[i]))
if i > 1:
if np.abs(losses_vae[i] - losses_vae[i - 1]) < 0.0001:
print('STOPPING EARLY')
break
print("finished training")
x_val_y = ConvAE.vae.predict(x_val)[2]
# x_val_y = ConvAE.classfier.predict(x_val_lp)
y_sp = x_val_y.argmax(axis=1)
print_accuracy(y_sp, y_val, params['n_clusters'])
from sklearn.metrics import normalized_mutual_info_score as nmi
nmi_score1 = nmi(y_sp, y_val)
print('NMI: ' + str(np.round(nmi_score1, 4)))
def cluster0(n_clusters=3):
labels = ry1()
true_labels = np.asarray(labels, dtype='int32')
data = rx1()
km = KMeans(init='k-means++', n_clusters=n_clusters, n_init=10)
km.fit(data)
km_means_labels = km.labels_
km_means_cluster_centers = km.cluster_centers_
km_means_labels_unique = np.unique(km_means_labels)
colors_ = cycle(colors.cnames.keys())
nmiv = nmi(true_labels, km_means_labels)
print 'nmi value ', nmiv
return nmiv #cal the accuracy of valdata
def evaluate(train_round_idx, ae_module, cluster_module):
test_loader = torch.utils.data.DataLoader(
torch.utils.data.TensorDataset(pt_data), batch_size=256)
pred_labels = np.zeros(pt_data.shape[0], dtype=np.int)
index = 0
n_batches = 0
for batch_data in test_loader:
batch_data = batch_data[0].cuda()
n_batches += 1
batch_size = batch_data.shape[0]
embedded_data, reconstructed_data = ae_module.forward(batch_data)
labels = cluster_module.prediction_hard_np(embedded_data)
pred_labels[index:index + batch_size] = labels
index = index + batch_size
nmi_value = nmi(gold_labels, pred_labels, average_method='arithmetic')
acc_value = cluster_acc(gold_labels, pred_labels)[0]
logger.info(
f"{train_round_idx} Evaluation: NMI: {nmi_value} ACC: {acc_value}")
return nmi_value, acc_value
def run(dataset_name='seeds', n_clusters=3, mode='cpu'):
X, Y = load_dataset(dataset_name)
X = normalize_dataset(X)
for i in range(1):
C, V, m = sub_kmeans(X, n_clusters, mode)
trans = V.T.real
X_rotated = np.matmul(trans[None, :, :],
np.transpose(X[:, None, :], [0, 2, 1]))
X_rotated = X_rotated.squeeze(-1).T
acc = nmi(Y, C)
M = assign_markers(C)
K = assign_colors(Y)
print('')
print('[i] Results')
print('[*] m: %d' % m)
print('[*] NMI: %.5f' % acc)
data_points = zip(X_rotated[0], X_rotated[1], K, M)
for x_, y_, c_, m_ in data_points:
plt.scatter(x_, y_, c=c_, marker=m_, s=3)
plt.title('Seeds, m={:d}, NMI={:.3f}'.format(m, acc))
plt.savefig('{}.png'.format(dataset_name), dpi=300)
def get_scores(x, y, n, k, dtr, dev):
tx = tens0((n, k), dt=dtr, dev=dev)
ty = tens0((n, k), dt=dtr, dev=dev)
tx = tens_sel_set(tx, x, 1)
ty = tens_sel_set(ty, y, 1)
t = tx.t().matmul(ty)
del tx, ty
tt = t.max() - t
tt = tt.cpu().numpy()
row, col = ass(tt)
del tt
t = t.cpu().numpy()
t = t[row, col].sum()
t = t.tolist() / n
x = x.cpu().numpy()
y = y.cpu().numpy()
s = {
'nmi': nmi(x, y, average_method='geometric'),
'ari': ari(x, y),
'acc': t,
}
return s
def run_net(data, params):
#
# UNPACK DATA
#
x_train, y_train, x_val, y_val, x_test, y_test = data['spectral'][
'train_and_test']
x_train_unlabeled, y_train_unlabeled, x_train_labeled, y_train_labeled = data[
'spectral']['train_unlabeled_and_labeled']
x_val_unlabeled, y_val_unlabeled, x_val_labeled, y_val_labeled = data[
'spectral']['val_unlabeled_and_labeled']
if 'siamese' in params['affinity']:
pairs_train, dist_train, pairs_val, dist_val = data['siamese'][
'train_and_test']
x = np.concatenate((x_train, x_val, x_test), axis=0)
y = np.concatenate((y_train, y_val, y_test), axis=0)
if len(x_train_labeled):
y_train_labeled_onehot = OneHotEncoder().fit_transform(
y_train_labeled.reshape(-1, 1)).toarray()
else:
y_train_labeled_onehot = np.empty((0, len(np.unique(y))))
#
# SET UP INPUTS
#
# create true y placeholder (not used in unsupervised training)
y_true = tf.placeholder(tf.float32,
shape=(None, params['n_clusters']),
name='y_true')
batch_sizes = {
'Unlabeled': params['batch_size'],
'Labeled': params['batch_size'],
'Orthonorm': params.get('batch_size_orthonorm', params['batch_size']),
}
input_shape = x.shape[1:]
# spectralnet has three inputs -- they are defined here
inputs = {
'Unlabeled': Input(shape=input_shape, name='UnlabeledInput'),
'Labeled': Input(shape=input_shape, name='LabeledInput'),
'Orthonorm': Input(shape=input_shape, name='OrthonormInput'),
}
#
# DEFINE AND TRAIN SIAMESE NET
#
# run only if we are using a siamese network
if params['affinity'] == 'siamese':
siamese_net = networks.SiameseNet(inputs, params['arch'],
params.get('siam_reg'), y_true)
history = siamese_net.train(pairs_train, dist_train, pairs_val,
dist_val, params['siam_lr'],
params['siam_drop'],
params['siam_patience'], params['siam_ne'],
params['siam_batch_size'])
else:
siamese_net = None
#
# DEFINE AND TRAIN SPECTRALNET
#
spectral_net = networks.SpectralNet(inputs, params['arch'],
params.get('spec_reg'), y_true,
y_train_labeled_onehot,
params['n_clusters'],
params['affinity'],
params['scale_nbr'], params['n_nbrs'],
batch_sizes, siamese_net, x_train,
len(x_train_labeled))
spectral_net.train(x_train_unlabeled, x_train_labeled, x_val_unlabeled,
params['spec_lr'], params['spec_drop'],
params['spec_patience'], params['spec_ne'])
print("finished training")
#
# EVALUATE
#
# get final embeddings
x_spectralnet = spectral_net.predict(x)
# get accuracy and nmi
kmeans_assignments, km = get_cluster_sols(x_spectralnet,
ClusterClass=KMeans,
n_clusters=params['n_clusters'],
init_args={'n_init': 10})
y_spectralnet, _ = get_y_preds(kmeans_assignments, y, params['n_clusters'])
print_accuracy(kmeans_assignments, y, params['n_clusters'])
from sklearn.metrics import normalized_mutual_info_score as nmi
nmi_score = nmi(kmeans_assignments, y)
print('NMI: ' + str(np.round(nmi_score, 3)))
if params['generalization_metrics']:
x_spectralnet_train = spectral_net.predict(x_train_unlabeled)
x_spectralnet_test = spectral_net.predict(x_test)
km_train = KMeans(
n_clusters=params['n_clusters']).fit(x_spectralnet_train)
from scipy.spatial.distance import cdist
dist_mat = cdist(x_spectralnet_test, km_train.cluster_centers_)
closest_cluster = np.argmin(dist_mat, axis=1)
print_accuracy(closest_cluster, y_test, params['n_clusters'],
' generalization')
nmi_score = nmi(closest_cluster, y_test)
print('generalization NMI: ' + str(np.round(nmi_score, 3)))
return x_spectralnet, y_spectralnet
y[i] = 2
sparsity = 1 - (np.sum(T>0) / np.product(T.shape))
f, dt = CreateOutputFile("yelp", date = True)
output_path = f"./output/_yelp/" + dt[:10] + "_" + dt[11:13] + "." + dt[14:16] + "." + dt[17:19] + "/"
directory = os.path.dirname(output_path)
if not os.path.exists(directory):
os.makedirs(directory)
model = CoClust(np.sum(T.shape) * 10, optimization_strategy = alg, path = output_path)
model.fit(T)
tau = model.final_tau_
nmi_x = nmi(y, model.x_, average_method='arithmetic')
ari_x = ari(y, model.x_)
f.write(f"{T.shape[0]},{T.shape[1]},{T.shape[2]},{len(set(y))},,,,{tau[0]},{tau[1]},{tau[2]},{nmi_x},,,{ari_x},,,{model._n_clusters[0]},{model._n_clusters[1]},{model._n_clusters[2]},{model.execution_time_},{sparsity},{alg}\n")
f.close()
gx = open(output_path + alg + "_assignments_"+ tensor + "_x.txt", 'w')
for i in range(T.shape[0]):
gx.write(f"{i}\t{model._assignment[0][i]}\n")
gx.close()
gy = open(output_path + alg + "_assignments_"+ tensor + "_y.txt", 'w')
for i in range(T.shape[1]):
gy.write(f"{i}\t{model._assignment[1][i]}\n")
gy.close()
#feat_lbp.append(mahotas.features.lbp(img[i],1,8))
feat_lbp = scale(np.array(feat_lbp))
# PCA on LBP features
#pca = PCA(n_components=20)
#feat_lbp = pca.fit_transform(feat_lbp)
#print "Variance Ratio: ",sum(pca.explained_variance_ratio_)
# Normalization of features
#feat_lbp = scale(feat_lbp)
# Save LBP features
file_pkl = open("face_lbp.pkl","wb")
pickle.dump(feat_lbp,file_pkl)
file_pkl.close()
# Compute affinity matrix
flag_sigma = 'global'
sigma_lbp, aff_lbp = compute_affinity(feat_lbp,flag_sigma=flag_sigma,\
sigma=100.,nn=8)
print "kernel computation finished"
label_pred_identity = spectral_clustering(aff_lbp,n_clusters=20)
nmi_identity = nmi(label_pred_identity,img_identity)
print "NMI with identity: ",nmi_identity
label_pred_pose = spectral_clustering(aff_lbp,n_clusters=4)
nmi_pose = nmi(label_pred_pose,img_pose)
print "NMI with pose: ",nmi_pose
def main():
path = './data/aucs_edgelist.txt'
# Declare each layer's graph
lunch = init_graph()
facebook = init_graph()
leisure = init_graph()
work = init_graph()
coauthor = init_graph()
table = {
'lunch': lunch,
'facebook': facebook,
'leisure': leisure,
'work': work,
'coauthor': coauthor,
}
truth, na = get_truth()
# Load data into graph
print(
"--------------------------------------------------Load multilayers graph--------------------------------------------------"
)
with open(path) as f:
for line in f:
line = line.strip().split(',')
name = line[2]
if line[0] in na or line[1] in na:
continue
else:
table[name].add_edge(line[0], line[1])
for name, graph in table.items():
print("\nGraph: {}".format(name))
print("\tNumber of nodes: {}".format(nx.number_of_nodes(graph)))
print("\tNumber of edges: {}".format(nx.number_of_edges(graph)))
graph_list = [lunch, work, coauthor, leisure]
node_list = list(lunch.nodes)
# # Tunning k
# print("--------------------------------------------------Perform k clusters selection--------------------------------------------------")
# sse_list = []
# range_k = np.arange(2, 15)
# for k in range_k:
# labels, sse = SCML(graph_list, k, 0.5)
# score = silhouette_score(matrix, labels, random_state=42)
# print("Number of clusters k = {}".format(k),
# ",Silhouette Score = {}".format(round(score, 5)))
# sse_list.append(sse)
# # Plot elbow method for k
# plot_elbow(range_k, sse_list, "Selection of k")
# Tunning alpha
print(
"--------------------------------------------------Perform alpha selection--------------------------------------------------"
)
range_a = np.arange(0.2, 1.1, 0.1)
den = []
nmi_list = []
for alpha in range_a:
labels = SCML(graph_list, 8, alpha)
partitions = get_partition(labels, node_list)
density = get_score(graph_list, partitions)
den.append(density)
print("\nAlpha = {}".format(round(alpha, 1)))
print("\tDensity = {}".format(density))
nmi_value = nmi(truth, labels)
print("\tNMI = {}".format(nmi_value))
nmi_list.append(nmi_value)
# Plot elbow method for alpha
plot_elbow(range_a, den, "Selection of alpha (Density)")
plot_elbow(range_a, nmi_list, "Selection of alpha (NMI)")
# Select the best model
print(
"--------------------------------------------------Multilayer Result---------------------------------------------------"
)
labels = SCML(graph_list, 8, 0.2)
partitions = get_partition(labels, node_list)
print("NMI: {}".format(nmi(truth, labels)))
purity = purity_score(truth, labels)
print("Purity: {}".format(purity))
print(
"--------------------------------------------------Single layer Result--------------------------------------------------"
)
for name, g in table.items():
print("\nLayer: {}".format(name))
labels = onelayer(g, 8)
# print(labels)
print("\tNMI: {}".format(nmi(truth, labels)))
purity = purity_score(truth, labels)
print("\tPurity: {}".format(purity))
x_init = np.zeros(nVars)
options = {'verbose':3}
data_mat = {'nInst':nInst, 'nVars':nVars, 'A':A, 'x':x, 'b':b, \
'x_init':x_init}
savemat("./Mark_Schmidt/minConf/minConf_SPG_input.mat",data_mat)
(x, f, funEvals, projects) = minConf_SPG(funObj, x_init, funProj, options)
elif flag_test == 8:
options_default = {'verbose':2, 'numDiff':0, 'optTol':1e-5, 'progTol':1e-9,\
'maxIter':500, 'suffDec':1e-4, 'interp':2, 'memory':10,\
'useSpectral':1,'curvilinear':0,'feasibleInit':0,'testOpt':1,\
'bbType':1}
options = {'verbose':100, 'interp':10}
options = setDefaultOptions(options, options_default)
elif flag_test == 9:
label = np.random.randint(0,10,100)
alpha = np.array([.95, .85])
beta = np.array([.7, .55])
[num_ML, num_CL] = [100, 100]
S = genConstraints(label, alpha, beta, num_ML, num_CL)
elif flag_test == 10:
tp = load_iris()
[X, Y] = [scale(tp['data']), tp['target']]
sim_mat = rbf_kernel(X)
Y_pred = my_spectral_clustering(sim_mat, n_clusters=3)
print nmi(Y_pred, Y)
else:
pass
file_csv = open("data_pheno.csv","wb")
csvwriter = csv.writer(file_csv)
for i in range(data.shape[0]):
csvwriter.writerow(list(data[i,:]))
file_csv.close()
label_pred_4 = np.loadtxt('label_pred_4.csv')
label_nbs_lf4 = np.loadtxt('label_nbs_lf4.csv')
nmi_1 = []
nmi_2 = []
nmi_3 = []
nmi_4 = []
for j in range(data.shape[1]):
nmi_1.append(nmi(data[:,j],label_pred_4))
nmi_2.append(nmi(data[:,j],label_nbs_lf4))
nmi_3.append(nmi(data[:,j],patient_label))
nmi_4.append(nmi(data[:,j],gold_stage))
index = np.arange(1,13)
bar_width = 0.2
labels = ['BD.FEV1','oxygen','ExacTrunc','BD.FEV.FVC','FracVol.950U',\
'Lowest.15.','Emphysema','Neutrophils','Lymphocytes',\
'Monocytes','Eosinophils','Basophils']
bar_1 = plt.bar(index,nmi_1,bar_width,color='b',label='Normalization+NMF')
bar_2 = plt.bar(index+bar_width,nmi_2,bar_width,color='r',label='NMF')
bar_3 = plt.bar(index+bar_width*2,nmi_3,bar_width,color='g',label='Case/Control')
bar_4 = plt.bar(index+bar_width*3,nmi_4,bar_width,color='y',label='Gold Stage')
plt.xlabel('Phenotype Features')
# Find U by eigendecomposition
aff = beta[0]*aff_pca+beta[1]*aff_gabor
eig_val,eig_vec = la.eig(aff)
# Sort eigenvalues and eigenvectors
idx = eig_val.argsort()[::-1]
eig_val = eig_val[idx]
eig_vec = eig_vec[:,idx]
U = eig_vec[:,0:40]
# Optimize beta
gamma = matrix([np.trace(aff_pca.dot(U).dot(U.T)),\
np.trace(aff_gabor.dot(U).dot(U.T))])
G = matrix([[-1.0,0.],[0.,-1.0]])
h = matrix([0.,0.])
A = matrix([1.,1.],(1,2))
b = matrix(1.0)
res = qp(2*Q,-2*gamma,G,h,A,b)
beta = res['x'].T
residue_old = la.norm(aff-U_old.dot(U_old.T))
residue = la.norm(aff-U.dot(U.T))
n_iter = n_iter+1
print n_iter
print "beta: ",beta
print "-2*gamma: ",-2*gamma
print "Residue: ",residue
clf = KMeans(n_clusters=K,init='random')
label_u = clf.fit_predict(U)
nmi_u = nmi(label_u,label_true)
print nmi_u
from keras.datasets import mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x = np.concatenate((x_train, x_test))
y = np.concatenate((y_train, y_test))
x = x.reshape((x.shape[0], -1))
x = np.divide(x, 50.) # normalize as it does in DEC paper
print('MNIST samples', x.shape)
return x, y
db = 'mnist'
n_clusters = 10
x, y = load_mnist()
# define and train SAE model
sae = SAE(dims=[x.shape[-1], 500, 500, 2000, 10])
sae.fit(x=x, epochs=400)
sae.autoencoders.save_weights('weights_%s.h5' % db)
# extract features
print('Finished training, extracting features using the trained SAE model')
features = sae.extract_feature(x)
print('performing k-means clustering on the extracted features')
from sklearn.cluster import KMeans
km = KMeans(n_clusters, n_init=20)
y_pred = km.fit_predict(features)
from sklearn.metrics import normalized_mutual_info_score as nmi
print('K-means clustering result on extracted features: NMI =',
nmi(y, y_pred))
n_eye = 2
# Repeat KMeans 50 times to reduce randomness
label_tmp = []
inertia_tmp = []
for i in range(50):
clf = KMeans(n_clusters=n_pose,init='random')
clf.fit(U)
label_tmp.append(list(clf.labels_))
inertia_tmp.append(clf.inertia_)
idx_tmp = inertia_tmp.index(min(inertia_tmp))
label_u = label_tmp[idx_tmp]
inertia_vec.append(min(inertia_tmp))
score_vec.append(silhouette_score(U,np.array(label_u)))
nmi_identity.append(nmi(label_u,img_identity))
nmi_pose.append(nmi(label_u,img_pose))
n_show = n_pose
# Show mean image
img_avg = np.zeros((n_show,img.shape[1],img.shape[2]))
cnt_avg = np.zeros((n_show,1))
for i in range(len(label_u)):
img_avg[label_u[i]] += img[i]
cnt_avg[label_u[i]] += 1.
for i in range(n_show):
img_avg[i] = img_avg[i]/cnt_avg[i]
#plt.imshow(img_avg[i],cmap=cm.Greys_r)
#plt.show()
identity,pose,expression,eye],file_pkl)
file_pkl.close()
# Normalization of each image
for i in range(img.shape[0]):
img[i] = (img[i]-img[i].min())*1./(img[i].max()-img[i].min())
img = img.reshape(img.shape[0],img.shape[1]*img.shape[2])
img = scale(img)
# 'global','local','manual'
flag_sigma = 'global'
# Compute similarity matrix
sigma,aff_img = compute_affinity(img,flag_sigma=flag_sigma,sigma=100.,nn=7)
if flag_sigma == 'local':
sigma_init = sum(sigma**2)/len(sigma)
print "Average Sigma(local): ",sigma_init
K = 20
# Construct existing solution Y
Y = np.zeros((img.shape[0],20))
for i in range(img.shape[0]):
Y[i,img_identity[i]] = 1
val_lambda = 1.2
arr_tmp = val_lambda*Y.dot(Y.T)
label_pred_identity = spectral_clustering(aff_img,n_clusters=K)
nmi_identity = nmi(label_pred_identity,img_identity)
print nmi_identity
feat_lbp = pickle.load(file_lbp)
file_lbp.close()
# Compute similarity matrix for FFT and Gabor
flag_sigma = 'global'
sigma_fft, aff_fft = compute_affinity(feat_fft,flag_sigma=flag_sigma)
sigma_gabor, aff_gabor = compute_affinity(feat_gabor,flag_sigma=flag_sigma)
sigma_lbp, aff_lbp = compute_affinity(feat_lbp,flag_sigma=flag_sigma)
print "kernel computation finished"
# Spectral Clustering using FFT
K = 4
label_pred_fft = spectral_clustering(aff_fft,n_clusters=K)
label_pred_gabor = spectral_clustering(aff_gabor,n_clusters=K)
nmi_fft_identity = nmi(label_pred_fft,img_identity)
nmi_gabor_identity = nmi(label_pred_gabor,img_identity)
print "nmi_fft_identity: ", nmi_fft_identity
print "nmi_gabor_identity: ",nmi_gabor_identity
for alpha in np.arange(0.1,1.0,0.1):
aff_add = alpha*aff_fft+(1-alpha)*aff_gabor
label_pred_add = spectral_clustering(aff_add,n_clusters=K)
nmi_add_identity = nmi(label_pred_add,img_identity)
print (alpha,nmi_add_identity)
# Weighted summation
M = 2
Q = matrix([[np.trace(aff_fft.dot(aff_fft)),\
np.trace(aff_fft.dot(aff_gabor))],\
[np.trace(aff_gabor.dot(aff_fft)),\
def fPredict(test_ref, test_art, dParam, dHyper):
weights_file = dParam['sOutPath'] + os.sep + '{}.h5'.format(
dHyper['bestModel'])
patchSize = dParam['patchSize']
vae = createModel(patchSize, dHyper)
vae.compile(optimizer='adam', loss=None)
vae.load_weights(weights_file)
test_ref = np.expand_dims(test_ref, axis=1)
test_art = np.expand_dims(test_art, axis=1)
predict_ref, predict_art = vae.predict([test_ref, test_art],
dParam['batchSize'][0],
verbose=1)
test_ref = np.squeeze(test_ref, axis=1)
test_art = np.squeeze(test_art, axis=1)
predict_art = np.squeeze(predict_art, axis=1)
if dHyper['unpatch']:
test_ref = fRigidUnpatchingCorrection2D(dHyper['actualSize'], test_ref,
dParam['patchOverlap'])
test_art = fRigidUnpatchingCorrection2D(dHyper['actualSize'], test_art,
dParam['patchOverlap'])
predict_art = fRigidUnpatchingCorrection2D(dHyper['actualSize'],
predict_art,
dParam['patchOverlap'],
'average')
# pre TV processing
test_art_tv_1 = denoise_tv_chambolle(test_art, weight=1)
test_art_tv_3 = denoise_tv_chambolle(test_art, weight=3)
test_art_tv_5 = denoise_tv_chambolle(test_art, weight=5)
if dHyper['evaluate']:
if dParam['lSaveIndividual']:
fig = plt.figure()
plt.gray()
label = 'NRMSE: {:.2f}, SSIM: {:.3f}, NMI: {:.3f}'
for i in range(len(test_ref)):
ax = imshow(test_ref[i])
plt.xticks([])
plt.yticks([])
ax.set_xlabel(
label.format(
nrmse(test_ref[i], test_ref[i]),
ssim(test_ref[i],
test_ref[i],
data_range=(test_ref[i].max() -
test_ref[i].min())),
nmi(test_ref[i].flatten(), test_ref[i].flatten())))
ax.set_title('reference image')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + 'reference_' + str(i) + '.png')
else:
plt.show()
ax = imshow(test_art[i])
plt.xticks([])
plt.yticks([])
ax.set_xlabel(
label.format(
nrmse(test_ref[i], test_art[i]),
ssim(test_ref[i],
test_art[i],
data_range=(test_art[i].max() -
test_art[i].min())),
nmi(test_ref[i].flatten(), test_art[i].flatten())))
ax.set_title('motion-affected image')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + 'art_' + str(i) + '.png')
else:
plt.show()
ax = imshow(predict_art[i])
plt.xticks([])
plt.yticks([])
ax.set_xlabel(
label.format(
nrmse(test_ref[i], predict_art[i]),
ssim(test_ref[i],
predict_art[i],
data_range=(predict_art[i].max() -
predict_art[i].min())),
nmi(test_ref[1].flatten(),
predict_art[i].flatten())))
ax.set_title('reconstructed image')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + 'recon_' + str(i) + '.png')
else:
plt.show()
ax = imshow(test_art_tv_1[i])
plt.xticks([])
plt.yticks([])
ax.set_xlabel(
label.format(
nrmse(test_ref[i], test_art_tv_1[i]),
ssim(test_ref[i],
test_art_tv_1[i],
data_range=(test_art_tv_1[i].max() -
test_art_tv_1[i].min())),
nmi(test_ref[i].flatten(),
test_art_tv_1[i].flatten())))
ax.set_title('TV weight 1')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + 'tv1_' + str(i) + '.png')
else:
plt.show()
ax = imshow(test_art_tv_3[i])
plt.xticks([])
plt.yticks([])
ax.set_xlabel(
label.format(
nrmse(test_ref[i], test_art_tv_3[i]),
ssim(test_ref[i],
test_art_tv_3[i],
data_range=(test_art_tv_3[i].max() -
test_art_tv_3[i].min())),
nmi(test_ref[i].flatten(),
test_art_tv_3[i].flatten())))
ax.set_title('TV weight 3')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + 'tv3_' + str(i) + '.png')
else:
plt.show()
ax = imshow(test_art_tv_5[i])
plt.xticks([])
plt.yticks([])
ax.set_xlabel(
label.format(
nrmse(test_ref[i], test_art_tv_5[i]),
ssim(test_ref[i],
test_art_tv_5[i],
data_range=(test_art_tv_5[i].max() -
test_art_tv_5[i].min())),
nmi(test_ref[i].flatten(),
test_art_tv_5[i].flatten())))
ax.set_title('TV weight 5')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + 'tv5_' + str(i) + '.png')
else:
plt.show()
else:
fig, axes = plt.subplots(nrows=2,
ncols=3,
figsize=(15, 10),
sharex=True,
sharey=True)
# fig, axes = plt.subplots(nrows=2, ncols=3, figsize=(15, 15), sharex=True, sharey=True)
ax = axes.ravel()
plt.gray()
label = 'NRMSE: {:.2f}, SSIM: {:.3f}, NMI: {:.3f}'
for i in range(len(test_ref)):
# orignal reconstructed images
ax[0].imshow(test_ref[i])
ax[0].set_xlabel(
label.format(
nrmse(test_ref[i], test_ref[i]),
ssim(test_ref[i],
test_ref[i],
data_range=(test_ref[i].max() -
test_ref[i].min())),
nmi(test_ref[i].flatten(), test_ref[i].flatten())))
ax[0].set_title('reference image')
ax[1].imshow(test_art[i])
ax[1].set_xlabel(
label.format(
nrmse(test_ref[i], test_art[i]),
ssim(test_ref[i],
test_art[i],
data_range=(test_art[i].max() -
test_art[i].min())),
nmi(test_ref[i].flatten(), test_art[i].flatten())))
ax[1].set_title('motion-affected image')
ax[2].imshow(predict_art[i])
ax[2].set_xlabel(
label.format(
nrmse(test_ref[i], predict_art[i]),
ssim(test_ref[i],
predict_art[i],
data_range=(predict_art[i].max() -
predict_art[i].min())),
nmi(test_ref[1].flatten(),
predict_art[i].flatten())))
ax[2].set_title('reconstructed image')
# TV denoiser
ax[3].imshow(test_art_tv_1[i])
ax[3].set_xlabel(
label.format(
nrmse(test_ref[i], test_art_tv_1[i]),
ssim(test_ref[i],
test_art_tv_1[i],
data_range=(test_art_tv_1[i].max() -
test_art_tv_1[i].min())),
nmi(test_ref[i].flatten(),
test_art_tv_1[i].flatten())))
ax[3].set_title('TV weight 1')
ax[4].imshow(test_art_tv_3[i])
ax[4].set_xlabel(
label.format(
nrmse(test_ref[i], test_art_tv_3[i]),
ssim(test_ref[i],
test_art_tv_3[i],
data_range=(test_art_tv_3[i].max() -
test_art_tv_3[i].min())),
nmi(test_ref[i].flatten(),
test_art_tv_3[i].flatten())))
ax[4].set_title('TV weight 3')
ax[5].imshow(test_art_tv_5[i])
ax[5].set_xlabel(
label.format(
nrmse(test_ref[i], test_art_tv_5[i]),
ssim(test_ref[i],
test_art_tv_5[i],
data_range=(test_art_tv_5[i].max() -
test_art_tv_5[i].min())),
nmi(test_ref[i].flatten(),
test_art_tv_5[i].flatten())))
ax[5].set_title('TV weight 5')
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + str(i) + '.png')
else:
plt.show()
else:
plt.figure()
plt.gray()
for i in range(predict_art.shape[0]):
plt.imshow(predict_art[i])
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' +
os.sep + str(i) + '.png',
dpi=300)
else:
plt.show()
else:
nPatch = predict_art.shape[0]
for i in range(nPatch // 4):
fig, axes = plt.subplots(nrows=4, ncols=2)
plt.gray()
cols_title = ['original_art', 'predicted_art']
for ax, col in zip(axes[0], cols_title):
ax.set_title(col)
for j in range(4):
axes[j, 0].imshow(test_art[4 * i + j])
axes[j, 1].imshow(predict_art[4 * i + j])
if dParam['lSave']:
plt.savefig(dParam['sOutPath'] + os.sep + 'result' + os.sep +
str(i) + '.png')
else:
plt.show()
return feature_model.predict(x, batch_size=self.batch_size)
if __name__ == "__main__":
"""
An example for how to use SAE model on MNIST dataset. In terminal run
python3 SAE.py
to see the result.
"""
import numpy as np
from load_mnist import load_mnist
x,y=load_mnist(sample_size=10000,seed=0)
db = 'mnist'
n_clusters = 10
# define and train SAE model
sae = SAE(dims=[x.shape[-1], 64,32])
sae.fit(x=x, epochs=400)
sae.autoencoders.save_weights('weights_%s.h5' % db)
# extract features
print ('Finished training, extracting features using the trained SAE model')
features = sae.extract_feature(x)
print ('performing k-means clustering on the extracted features')
from sklearn.cluster import KMeans
km = KMeans(n_clusters, n_init=20)
y_pred = km.fit_predict(features)
from sklearn.metrics import normalized_mutual_info_score as nmi
print ('K-means clustering result on extracted features: NMI =', nmi(y, y_pred))
print "Optimization fails",v_lambda_idx
# Repeat KMeans 50 times to reduce randomness
label_tmp = []
inertia_tmp = []
for i in range(50):
clf = KMeans(n_clusters=dim_q,init='random')
clf.fit(U)
label_tmp.append(list(clf.labels_))
inertia_tmp.append(clf.inertia_)
idx_tmp = inertia_tmp.index(min(inertia_tmp))
label_u = label_tmp[idx_tmp]
inertia_vec.append(min(inertia_tmp))
score_vec.append(silhouette_score(U,np.array(label_u)))
nmi_e.append(nmi(label_e,label_u))
beta_vec = np.array(beta_vec)
# Plot the result
plt.figure(0)
plt.plot(v_lambda_range,nmi_e,'r',label='nmi_e')
plt.xlabel("lambda(tradeoff between clustering quality and novelty)")
plt.ylabel("NMI value")
plt.legend(loc='upper left')
plt.figure(1)
plt.plot(v_lambda_range,inertia_vec)
plt.xlabel("lambda")
plt.ylabel("Inertia Value")
plt.figure(2)
def run_net(data, params):
"""run the network with the parameters."""
#
# UNPACK DATA
#
x_train, y_train, x_val, y_val, x_test, y_test = data['cnc']['train_and_test']
x_train_unlabeled, _, x_train_labeled, y_train_labeled = data['cnc'][
'train_unlabeled_and_labeled']
x_val_unlabeled, _, _, _ = data['cnc']['val_unlabeled_and_labeled']
if 'siamese' in params['affinity']:
pairs_train, dist_train, pairs_val, dist_val = data['siamese'][
'train_and_test']
x = np.concatenate((x_train, x_val, x_test), axis=0)
y = np.concatenate((y_train, y_val, y_test), axis=0)
if x_train_labeled:
y_train_labeled_onehot = OneHotEncoder().fit_transform(
y_train_labeled.reshape(-1, 1)).toarray()
else:
y_train_labeled_onehot = np.empty((0, len(np.unique(y))))
#
# SET UP INPUTS
#
# create true y placeholder (not used in unsupervised training)
y_true = tf.placeholder(
tf.float32, shape=(None, params['n_clusters']), name='y_true')
batch_sizes = {
'Unlabeled': params['batch_size'],
'Labeled': params['batch_size']
}
input_shape = x.shape[1:]
# inputs to CNC
inputs = {
'Unlabeled': Input(shape=input_shape, name='UnlabeledInput'),
'Labeled': Input(shape=input_shape, name='LabeledInput'),
}
#
# DEFINE AND TRAIN SIAMESE NET
# http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf.
# DEFINE AND TRAIN Siamese NET
if params['affinity'] == 'siamese':
siamese_net = networks.SiameseNet(inputs, params['siam_arch'],
params.get('siam_reg'),
params['main_path'], y_true)
siamese_net.train(pairs_train, dist_train, pairs_val, dist_val,
params['siam_lr'], params['siam_drop'],
params['siam_patience'], params['siam_ne'],
params['siam_batch_size'], params['dset'])
else:
siamese_net = None
#
# DEFINE AND TRAIN CNC NET
#
cnc_net = networks.CncNet(inputs, params['cnc_arch'], params.get('cnc_reg'),
y_true, y_train_labeled_onehot,
params['n_clusters'], params['affinity'],
params['scale_nbr'], params['n_nbrs'], batch_sizes,
params['result_path'], params['dset'], siamese_net,
x_train, params['cnc_lr'], params['cnc_tau'],
params['bal_reg'])
cnc_net.train(x_train_unlabeled, x_train_labeled, x_val_unlabeled,
params['cnc_drop'], params['cnc_patience'], params['min_tem'],
params['cnc_epochs'])
#
# EVALUATE
#
x_cncnet = cnc_net.predict(x)
prediction = np.argmax(x_cncnet, 1)
accuray_all = print_accuracy(prediction, y, params['n_clusters'])
nmi_score_all = nmi(prediction, y)
print('NMI: {0}'.format(np.round(nmi_score_all, 3)))
if params['generalization_metrics']:
x_cncnet_train = cnc_net.predict(x_train_unlabeled)
x_cncnet_test = cnc_net.predict(x_test)
prediction_train = np.argmax(x_cncnet_train, 1)
accuray_train = print_accuracy(prediction_train, y_train,
params['n_clusters'])
nmi_score_train = nmi(prediction_train, y_train)
print('TRAIN NMI: {0}'.format(np.round(nmi_score_train, 3)))
prediction_test = np.argmax(x_cncnet_test, 1)
accuray_test = print_accuracy(prediction_test, y_test, params['n_clusters'])
nmi_score_test = nmi(prediction_test, y_test)
print('TEST NMI: {0}'.format(np.round(nmi_score_test, 3)))
with gfile.Open(params['result_path'] + 'results', 'w') as f:
f.write(accuray_all + ' ' + accuray_train + ' ' + accuray_test + '\n')
f.write(
str(np.round(nmi_score_all, 3)) + ' ' +
str(np.round(nmi_score_train, 3)) + ' ' +
str(np.round(nmi_score_test, 3)) + '\n')
else:
with gfile.Open(params['result_path'] + 'results', 'w') as f:
f.write(accuray_all + ' ' + str(np.round(nmi_score_all, 3)) + '\n')
sigma_fac, aff_fac = compute_affinity(data_fac,flag_sigma=flag_sigma,\
sigma=422.6228,nn=8)
print "kernel computing finished"
if flag_sigma == 'local':
sigma_fou_init = sum(sigma_fou**2)/len(sigma_fou)
sigma_fac_init = sum(sigma_fac**2)/len(sigma_fac)
K = 10
label_true = []
for i in range(K):
for j in range(200):
label_true.append(i)
# Spectral Clustering: Fourier coefficient
label_fou = spectral_clustering(aff_fou,n_clusters=K)
nmi_fou = nmi(label_fou,label_true)
print "NMI(Source 1)",nmi_fou
# SC: Autocorrelation Profile
label_fac = spectral_clustering(aff_fac,n_clusters=K)
nmi_fac = nmi(label_fac,label_true)
print "NMI(Source 2)",nmi_fac
# kernel addition
for alpha in np.arange(0.1,1.0,0.1):
aff_add = alpha*aff_fou+(1-alpha)*aff_fac
label_add = spectral_clustering(aff_add,n_clusters=K)
nmi_add = nmi(label_add,label_true)
print "NMI(a*source_1+(1-a)*source_2)",(alpha,nmi_add)
# Parameter settings