def test_row_norms():
X = np.random.RandomState(42).randn(100, 100)
for dtype in (np.float32, np.float64):
if dtype is np.float32:
precision = 4
else:
precision = 5
X = X.astype(dtype)
sq_norm = (X ** 2).sum(axis=1)
assert_array_almost_equal(sq_norm, row_norms(X, squared=True),
precision)
assert_array_almost_equal(np.sqrt(sq_norm), row_norms(X), precision)
for csr_index_dtype in [np.int32, np.int64]:
Xcsr = sparse.csr_matrix(X, dtype=dtype)
# csr_matrix will use int32 indices by default,
# up-casting those to int64 when necessary
if csr_index_dtype is np.int64:
Xcsr.indptr = Xcsr.indptr.astype(csr_index_dtype)
Xcsr.indices = Xcsr.indices.astype(csr_index_dtype)
assert Xcsr.indices.dtype == csr_index_dtype
assert Xcsr.indptr.dtype == csr_index_dtype
assert_array_almost_equal(sq_norm, row_norms(Xcsr, squared=True),
precision)
assert_array_almost_equal(np.sqrt(sq_norm), row_norms(Xcsr),
precision)
def test_row_norms():
X = np.random.RandomState(42).randn(100, 100)
sq_norm = (X ** 2).sum(axis=1)
assert_array_almost_equal(sq_norm, row_norms(X, squared=True), 5)
assert_array_almost_equal(np.sqrt(sq_norm), row_norms(X))
Xcsr = sparse.csr_matrix(X, dtype=np.float32)
assert_array_almost_equal(sq_norm, row_norms(Xcsr, squared=True), 5)
assert_array_almost_equal(np.sqrt(sq_norm), row_norms(Xcsr))
def euclidean_distances(X, Y=None):
YY = row_norms(Y, squared=True)[np.newaxis, :]
if X is Y: # shortcut in the common case euclidean_distances(X, X)
XX = YY.T
else:
XX = row_norms(X, squared=True)[:, np.newaxis]
distances = np.dot(X, Y.T)
distances *= -2
distances += XX
distances += YY
np.maximum(distances, 0, out=distances)
return distances
def test_get_auto_step_size():
X = np.array([[1, 2, 3], [2, 3, 4], [2, 3, 2]], dtype=np.float64)
alpha = 1.2
fit_intercept = False
# sum the squares of the second sample because that's the largest
max_squared_sum = 4 + 9 + 16
max_squared_sum_ = row_norms(X, squared=True).max()
assert_almost_equal(max_squared_sum, max_squared_sum_, decimal=4)
for fit_intercept in (True, False):
step_size_sqr = 1.0 / (max_squared_sum + alpha + int(fit_intercept))
step_size_log = 4.0 / (max_squared_sum + 4.0 * alpha +
int(fit_intercept))
step_size_sqr_ = get_auto_step_size(max_squared_sum_, alpha, "squared",
fit_intercept)
step_size_log_ = get_auto_step_size(max_squared_sum_, alpha, "log",
fit_intercept)
assert_almost_equal(step_size_sqr, step_size_sqr_, decimal=4)
assert_almost_equal(step_size_log, step_size_log_, decimal=4)
msg = 'Unknown loss function for SAG solver, got wrong instead of'
assert_raise_message(ValueError, msg, get_auto_step_size,
max_squared_sum_, alpha, "wrong", fit_intercept)
def fit(self, X, y):
"""Fit factorization machine to training data.
Parameters
----------
X : array-like or sparse, shape = [n_samples, n_features]
Training vectors, where n_samples is the number of samples
and n_features is the number of features.
y : array-like, shape = [n_samples]
Target values.
Returns
-------
self : Estimator
Returns self.
"""
if self.degree > 3:
raise ValueError("FMs with degree >3 not yet supported.")
X, y = self._check_X_y(X, y)
X = self._augment(X)
n_features = X.shape[1] # augmented
X_col_norms = row_norms(X.T, squared=True)
dataset = get_dataset(X, order="fortran")
rng = check_random_state(self.random_state)
loss_obj = self._get_loss(self.loss)
if not (self.warm_start and hasattr(self, 'w_')):
self.w_ = np.zeros(n_features, dtype=np.double)
if self.fit_lower == 'explicit':
n_orders = self.degree - 1
else:
n_orders = 1
if not (self.warm_start and hasattr(self, 'P_')):
self.P_ = 0.01 * rng.randn(n_orders, self.n_components, n_features)
if not (self.warm_start and hasattr(self, 'lams_')):
if self.init_lambdas == 'ones':
self.lams_ = np.ones(self.n_components)
elif self.init_lambdas == 'random_signs':
self.lams_ = np.sign(rng.randn(self.n_components))
else:
raise ValueError("Lambdas must be initialized as ones "
"(init_lambdas='ones') or as random "
"+/- 1 (init_lambdas='random_signs').")
y_pred = self._get_output(X)
converged = _cd_direct_ho(self.P_, self.w_, dataset, X_col_norms, y,
y_pred, self.lams_, self.degree, self.alpha,
self.beta, self.fit_linear,
self.fit_lower == 'explicit', loss_obj,
self.max_iter, self.tol, self.verbose)
if not converged:
warnings.warn("Objective did not converge. Increase max_iter.")
return self
def compute_distances(self, x1, x2=None):
"""
The method
- extracts normalized continuous attributes and then uses `row_norms`
and `safe_sparse_do`t to compute the distance as x^2 - 2xy - y^2
(the trick from sklearn);
- calls a function in Cython that adds the contributions of discrete
columns
"""
if self.normalize:
x1 = x1 - self.means
x1 /= np.sqrt(2 * self.vars)
# adapted from sklearn.metric.euclidean_distances
xx = row_norms(x1.T, squared=True)[:, np.newaxis]
distances = safe_sparse_dot(x1.T, x1, dense_output=True)
distances *= -2
distances += xx
distances += xx.T
with np.errstate(invalid="ignore"): # Nans are fixed below
np.maximum(distances, 0, out=distances)
distances.flat[::distances.shape[0] + 1] = 0.0
fixer = _distance.fix_euclidean_cols_normalized if self.normalize \
else _distance.fix_euclidean_cols
fixer(distances, x1, self.means, self.vars)
return np.sqrt(distances)
def test_labels_assignment_and_inertia():
# pure numpy implementation as easily auditable reference gold
# implementation
rng = np.random.RandomState(42)
noisy_centers = centers + rng.normal(size=centers.shape)
labels_gold = - np.ones(n_samples, dtype=np.int)
mindist = np.empty(n_samples)
mindist.fill(np.infty)
for center_id in range(n_clusters):
dist = np.sum((X - noisy_centers[center_id]) ** 2, axis=1)
labels_gold[dist < mindist] = center_id
mindist = np.minimum(dist, mindist)
inertia_gold = mindist.sum()
assert_true((mindist >= 0.0).all())
assert_true((labels_gold != -1).all())
# perform label assignment using the dense array input
x_squared_norms = (X ** 2).sum(axis=1)
labels_array, inertia_array = _labels_inertia(
X, x_squared_norms, noisy_centers)
assert_array_almost_equal(inertia_array, inertia_gold)
assert_array_equal(labels_array, labels_gold)
# perform label assignment using the sparse CSR input
x_squared_norms_from_csr = row_norms(X_csr, squared=True)
labels_csr, inertia_csr = _labels_inertia(
X_csr, x_squared_norms_from_csr, noisy_centers)
assert_array_almost_equal(inertia_csr, inertia_gold)
assert_array_equal(labels_csr, labels_gold)
def compute_distances(self, x1, x2=None):
"""
The method
- extracts normalized continuous attributes and then uses `row_norms`
and `safe_sparse_do`t to compute the distance as x^2 - 2xy - y^2
(the trick from sklearn);
- calls a function in Cython that recomputes the distances between pairs
of rows that yielded nan
- calls a function in Cython that adds the contributions of discrete
columns
"""
if self.continuous.any():
data1, data2 = self.continuous_columns(
x1, x2, self.means, np.sqrt(2 * self.vars))
# adapted from sklearn.metric.euclidean_distances
xx = row_norms(data1, squared=True)[:, np.newaxis]
if x2 is not None:
yy = row_norms(data2, squared=True)[np.newaxis, :]
else:
yy = xx.T
distances = safe_sparse_dot(data1, data2.T, dense_output=True)
distances *= -2
distances += xx
distances += yy
with np.errstate(invalid="ignore"): # Nans are fixed below
np.maximum(distances, 0, out=distances)
if x2 is None:
distances.flat[::distances.shape[0] + 1] = 0.0
fixer = _distance.fix_euclidean_rows_normalized if self.normalize \
else _distance.fix_euclidean_rows
fixer(distances, data1, data2,
self.means, self.vars, self.dist_missing2_cont,
x2 is not None)
else:
distances = np.zeros((x1.shape[0],
(x2 if x2 is not None else x1).shape[0]))
if np.any(self.discrete):
data1, data2 = self.discrete_columns(x1, x2)
_distance.euclidean_rows_discrete(
distances, data1, data2, self.dist_missing_disc,
self.dist_missing2_disc, x2 is not None)
if x2 is None:
_distance.lower_to_symmetric(distances)
return np.sqrt(distances)
def get_kpp_init(X,n_clusters,random_state=None):
random_state = None
random_state = check_random_state(random_state)
x_squared_norms = row_norms(X, squared=True)
centers = sklearn.cluster.k_means_._k_init(X, n_clusters, random_state=random_state,x_squared_norms=x_squared_norms) # n_clusters x D
W = np.transpose( centers ) # D x D^(1)
W_tf = tf.constant(W)
return centers,W,W_tf
def _kmeans_spark(X, n_clusters, max_iter=300, worker_nums=10, init='k-means++', random_state=None, tol=1e-4):
from pyspark import SparkContext, SparkConf
conf = SparkConf().setAppName('K-Means_Spark').setMaster('local[%d]'%worker_nums)
sc = SparkContext(conf=conf)
data = sc.parallelize(X)
data.cache()
random_state = check_random_state(random_state)
best_labels, best_inertia, best_centers = None, None, None
x_squared_norms = row_norms(X, squared=True)
# x_squared_norms = data.map(lambda x: (x*x).sum(axis=0)).collect()
# x_squared_norms = np.array(x_squared_norms, dtype='float64')
centers = _init_centroids(X, n_clusters, init, random_state, x_squared_norms=x_squared_norms)
bs = X.shape[0]/worker_nums
data_temp = []
for i in range(worker_nums-1):
data_temp.append(X[i*bs:(i+1)*bs])
data_temp.append(X[(worker_nums-1)*bs:])
data_temp = np.array(data_temp, dtype='float64')
data_temp = sc.parallelize(data_temp)
data_temp.cache()
for i in range(max_iter):
centers_old = centers.copy()
all_distances = data_temp.map(lambda x: euclidean_distances(centers, x, squared=True)).collect()
temp_all_distances = all_distances[0]
for i in range(1, worker_nums):
temp_all_distances = np.hstack((temp_all_distances, all_distances[i]))
all_distances = temp_all_distances
# all_distances = data.map(lambda x: euclidean_distances(centers, x, squared=True)).collect()
# # reshape, from (1, n_samples, k) to (k, n_samples)
# all_distances = np.asarray(all_distances, dtype="float64").T[0]
# Assignment, also called E-step of EM
labels, inertia = _labels_inertia(X, x_squared_norms, centers, all_distances=all_distances)
# re-computation of the centroids, also called M-step of EM
centers = _centers(X, labels, n_clusters)
if best_inertia is None or inertia < best_inertia:
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
shift = squared_norm(centers_old - centers)
if shift <= tol:
break
return best_centers, best_labels, best_inertia
def test_row_norms():
X = np.random.RandomState(42).randn(100, 100)
for dtype in (np.float32, np.float64):
if dtype is np.float32:
precision = 4
else:
precision = 5
X = X.astype(dtype)
sq_norm = (X ** 2).sum(axis=1)
assert_array_almost_equal(sq_norm, row_norms(X, squared=True),
precision)
assert_array_almost_equal(np.sqrt(sq_norm), row_norms(X), precision)
Xcsr = sparse.csr_matrix(X, dtype=dtype)
assert_array_almost_equal(sq_norm, row_norms(Xcsr, squared=True),
precision)
assert_array_almost_equal(np.sqrt(sq_norm), row_norms(Xcsr), precision)
def get_auto_step_size(X, alpha, loss, gamma=None, sample_weight=None):
"""Compute automatic step size for SAG solver
Stepsize computed using the following objective:
minimize_w 1 / n_samples * \sum_i loss(w^T x_i, y_i)
+ alpha * 0.5 * ||w||^2_2
Parameters
----------
X : ndarray
Array of samples x_i.
alpha : float
Constant that multiplies the l2 penalty term.
loss : string, in {"log", "squared"}
The loss function used in SAG solver.
Returns
-------
step_size : float
Step size used in SAG/SAGA solver.
"""
if sample_weight is None:
weighted_norms = row_norms(X, squared=True)
else:
weighted_norms = sample_weight * row_norms(X, squared=True)
L = np.max(weighted_norms)
n_samples = X.shape[0]
if loss == 'log':
# inverse Lipschitz constant for log loss
lipschitz_constant = 0.25 * L + alpha
elif loss == 'squared':
lipschitz_constant = L + alpha
elif loss == 'modified_huber':
lipschitz_constant = 2 * L + alpha
elif loss == 'smooth_hinge':
lipschitz_constant = L + gamma + alpha
elif loss == 'squared_hinge':
lipschitz_constant = 2 * L + alpha
else:
raise ValueError("`auto` stepsize is only available for `squared` or "
"`log` losses (got `%s` loss). Please specify a "
"stepsize." % loss)
return 1.0 / lipschitz_constant
def fit(self, X):
x_squared_norms = row_norms(X, squared=True)
rng = np.random.RandomState(self.random_state)
if self.init == "kmeans++":
# Private function of sklearn.cluster.k_means_, to get the initial centers.
init_centers = _k_init(X, self.n_clusters, x_squared_norms, rng)
elif self.init == "random":
random_samples = rng.random_integers(0, X.shape[0], size=self.n_clusters)
init_centers = X[random_samples, :]
else:
raise ValueError("init should be either kmeans++ or random")
# Assign initial labels. skip norm of x**2
init_distances = np.sum(init_centers**2, axis=1) - 2 * np.dot(X, init_centers.T)
init_labels = np.argmin(init_distances, axis=1)
self.labels_ = init_labels
self.centers_ = init_centers
self.n_samples_ = np.zeros(self.n_clusters)
# Count the number of samples in each cluster.
for i in range(self.n_clusters):
self.n_samples_[i] = np.sum(self.labels_ == i)
for i, (sample, label) in enumerate(zip(X, self.labels_)):
curr_label = label
max_cost = np.inf
while max_cost > 0:
distances = x_squared_norms[i] - 2 * np.dot(sample, self.centers_.T) + np.sum(self.centers_**2, axis=1)
curr_distance = distances[curr_label]
other_distance = np.delete(distances, curr_label)
curr_n_samples = self.n_samples_[curr_label]
other_n_samples = np.delete(self.n_samples_, curr_label)
cost = (curr_n_samples / (curr_n_samples - 1) * curr_distance) - (other_n_samples / (other_n_samples + 1) * other_distance)
max_cost_ind = np.argmax(cost)
max_cost = cost[max_cost_ind]
if max_cost > 0:
# We deleted the label index from other_n_samples
if max_cost_ind > curr_label:
max_cost_ind += 1
# Reassign the clusters
self.labels_[i] = max_cost_ind
self.centers_[curr_label] = (curr_n_samples * self.centers_[curr_label] - sample) / (curr_n_samples - 1)
moved_n_samples = self.n_samples_[max_cost_ind]
self.centers_[max_cost_ind] = (moved_n_samples * self.centers_[max_cost_ind] + sample) / (moved_n_samples + 1)
self.n_samples_[curr_label] -= 1
self.n_samples_[max_cost_ind] += 1
curr_label = max_cost_ind
def prepare_data(x):
if self.discrete.any():
data = Cosine.discrete_to_indicators(x, self.discrete)
else:
data = x.copy()
for col, mean in enumerate(self.means):
column = data[:, col]
column[np.isnan(column)] = mean
if self.axis == 0:
data = data.T
data /= row_norms(data)[:, np.newaxis]
return data
def kmeans_subsample(X, n_clusters, random_state=None, n_local_trials=10):
random_state = check_random_state(random_state)
n_samples, n_features = X.shape
x_squared_norms = row_norms(X, squared=True)
centers = np.empty((n_clusters, n_features))
# Pick first center randomly
center_id = random_state.randint(n_samples)
centers[0] = X[center_id]
# Initialize list of closest distances and calculate current potential
closest_dist_sq = euclidean_distances(centers[0].reshape(1, -1), X, Y_norm_squared=x_squared_norms, squared=True)
current_pot = closest_dist_sq.sum()
# Pick the remaining n_clusters-1 points
for c in range(1, n_clusters):
# Choose center candidates by sampling with probability proportional
# to the squared distance to the closest existing center
rand_vals = random_state.random_sample(n_local_trials) * current_pot
candidate_ids = np.searchsorted(closest_dist_sq.cumsum(), rand_vals)
# Compute distances to center candidates
distance_to_candidates = euclidean_distances(X[candidate_ids], X, Y_norm_squared=x_squared_norms, squared=True)
# Decide which candidate is the best
best_candidate = None
best_pot = None
best_dist_sq = None
for trial in range(n_local_trials):
# Compute potential when including center candidate
new_dist_sq = np.minimum(closest_dist_sq, distance_to_candidates[trial])
new_pot = new_dist_sq.sum()
# Store result if it is the best local trial so far
if (best_candidate is None) or (new_pot < best_pot):
best_candidate = candidate_ids[trial]
best_pot = new_pot
best_dist_sq = new_dist_sq
# Permanently add best center candidate found in local tries
centers[c] = X[best_candidate]
current_pot = best_pot
closest_dist_sq = best_dist_sq
return centers
def kmeanspp(X, k, seed):
# That we need to do this is a bug in _init_centroids
x_squared_norms = row_norms(X, squared=True)
# Use k-means++ to initialise the centroids
centroids = _init_centroids(X, k, 'k-means++', random_state=seed, x_squared_norms=x_squared_norms)
# OK, we should just short-circuit and get these from k-means++...
# quick and dirty solution
nns = NearestNeighbors()
nns.fit(X)
centroid_candidatess = nns.radius_neighbors(X=centroids, radius=0, return_distance=False)
# Account for "degenerated" solutions: serveral voxels at distance 0, each becoming a centroid
centroids = set()
for centroid_candidates in centroid_candidatess:
centroid_candidates = set(centroid_candidates) - centroids
if len(set(centroid_candidates) - centroids) == 0:
raise Exception('Cannot get an unambiguous set of centers;'
'theoretically this cannot happen, so check for bugs')
centroids.add(centroid_candidates.pop())
return np.array(sorted(centroids))
def _init_centroids(X, k, init, random_state, x_squared_norms=None):
random_state = check_random_state(random_state)
n_samples = X.shape[0]
if x_squared_norms is None:
x_squared_norms = row_norms(X, squared=True)
if n_samples < k:
raise ValueError("n_samples=%d should be larger than k=%d"%(n_samples, k))
if init == 'k-means++':
centers = _k_init(X, k, random_state=random_state,
x_squared_norms=x_squared_norms)
elif init == 'random':
seeds = random_state.permutation(n_samples)[:k]
centers = X[seeds]
return centers
def test_get_auto_step_size():
X = np.array([[1, 2, 3], [2, 3, 4], [2, 3, 2]], dtype=np.float64)
alpha = 1.2
fit_intercept = False
# sum the squares of the second sample because that's the largest
max_squared_sum = 4 + 9 + 16
max_squared_sum_ = row_norms(X, squared=True).max()
n_samples = X.shape[0]
assert_almost_equal(max_squared_sum, max_squared_sum_, decimal=4)
for saga in [True, False]:
for fit_intercept in (True, False):
if saga:
L_sqr = (max_squared_sum + alpha + int(fit_intercept))
L_log = (max_squared_sum + 4.0 * alpha +
int(fit_intercept)) / 4.0
mun_sqr = min(2 * n_samples * alpha, L_sqr)
mun_log = min(2 * n_samples * alpha, L_log)
step_size_sqr = 1 / (2 * L_sqr + mun_sqr)
step_size_log = 1 / (2 * L_log + mun_log)
else:
step_size_sqr = 1.0 / (max_squared_sum +
alpha + int(fit_intercept))
step_size_log = 4.0 / (max_squared_sum + 4.0 * alpha +
int(fit_intercept))
step_size_sqr_ = get_auto_step_size(max_squared_sum_, alpha,
"squared",
fit_intercept,
n_samples=n_samples,
is_saga=saga)
step_size_log_ = get_auto_step_size(max_squared_sum_, alpha, "log",
fit_intercept,
n_samples=n_samples,
is_saga=saga)
assert_almost_equal(step_size_sqr, step_size_sqr_, decimal=4)
assert_almost_equal(step_size_log, step_size_log_, decimal=4)
msg = 'Unknown loss function for SAG solver, got wrong instead of'
assert_raise_message(ValueError, msg, get_auto_step_size,
max_squared_sum_, alpha, "wrong", fit_intercept)
def predict(self, X):
"""Predict the closest cluster each sample in X belongs to.
In the vector quantization literature, `cluster_centers_` is called
the code book and each value returned by `predict` is the index of
the closest code in the code book.
Parameters
----------
X : {array-like, sparse matrix}, shape = [n_samples, n_features]
New data to predict.
Returns
-------
labels : array, shape [n_samples,]
Index of the cluster each sample belongs to.
"""
#check_is_fitted(self, 'cluster_centers_')
X = self._check_test_data(X)
x_squared_norms = row_norms(X, squared=True)
return _labels_inertia(X, x_squared_norms, self.cluster_centers_)[0]
def run_step(self,run_number,step_size,howlong):
df_slot = self.get_input_slot('df')
df_slot.update(run_number, buffer_created=True, buffer_updated=True)
if df_slot.has_deleted():
self.reset()
df_slot.reset()
df_slot.update(run_number)
input_df = df_slot.data()
columns = self.get_columns(input_df)
if input_df is None or len(input_df)==0:
return self._return_run_step(self.state_blocked, steps_run=0)
indices = df_slot.next_created(step_size)
steps = indices_len(indices)
step_size -= steps
steps_run = steps
if steps != 0:
indices = fix_loc(indices)
self._buffer.append(input_df.loc[indices])
self._df = self._buffer.df()
self._df.loc[indices,self.UPDATE_COLUMN] = run_number
if step_size > 0 and df_slot.has_updated():
indices = df_slot.next_updated(step_size,as_slice=False)
steps = indices_len(indices)
if steps != 0:
steps_run += steps
indices = fix_loc(indices) # no need, but stick to the stereotype
updated = self.filter_columns(input_df, indices)
df = self.filter_columns(self._df, indices)
norms = row_norms(updated-df)
selected = (norms > (self._delta*self.get_scale()))
indices = indices[selected]
if selected.any():
logger.debug('updating at %d', run_number)
self._df.loc[indices, self._columns] = updated.loc[indices, self._columns]
self._df.loc[indices, self.UPDATE_COLUMN] = run_number
else:
logger.debug('Not updating at %d', run_number)
return self._return_run_step(df_slot.next_state(), steps_run=steps_run)
def _kmeans_single(X, n_clusters, max_iter=300, init='k-means++', random_state=None, tol=1e-4):
random_state = check_random_state(random_state)
best_labels, best_inertia, best_centers = None, None, None
# init
x_squared_norms = row_norms(X, squared=True)
centers = _init_centroids(X, n_clusters, init, random_state, x_squared_norms=x_squared_norms)
# distances = np.zeros(shape=(X.shape[0],), dtype=np.float64)
# iterations
for i in range(max_iter):
centers_old = centers.copy()
# Assignment, also called E-step of EM
labels, inertia = _labels_inertia(X, x_squared_norms, centers)
# re-computation of the centroids, also called M-step of EM
centers = _centers(X, labels, n_clusters)
if best_inertia is None or inertia < best_inertia:
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
shift = squared_norm(centers_old - centers)
if shift <= tol:
break
if shift > 0:
# rerun E-step in case of non-convergence so that predicted labels
# match cluster centers
best_labels, best_inertia = \
_labels_inertia(X, x_squared_norms, best_centers)
return best_centers, best_labels, best_inertia
from sklearn.utils import check_random_state
from sklearn.cluster import KMeans as skKMeans
# %%
# data = pd.read_csv('s1.csv', sep=',')
data = pd.read_csv('s1.csv', sep=',')
# %%
scaler = MinMaxScaler()
data_scaled = pd.DataFrame(scaler.fit_transform(data))
# %%
# K-means++ für erste Cluster
N = 15
random_state = 42
fit_data = np.asarray(data_scaled)
x_squared_norms = row_norms(fit_data, squared=True)
random = check_random_state(random_state)
initial_clusters = k_init(fit_data, N, x_squared_norms, random)
max_norm = KMeans(N, initial_clusters, order=np.inf)
max_norm.fit(fit_data)
manhattan = KMeans(N, initial_clusters, order=1)
manhattan.fit(fit_data)
euclid = KMeans(N, initial_clusters, order=2)
euclid.fit(fit_data)
# %%
fig, axs = plt.subplots(2, 3)
data_scaled['max'] = max_norm.labels
center = max_norm.centroids
axs[0, 0].scatter(data_scaled[0],
data_scaled[1],
data = data_original[idx]
labels = labels_original[idx]
x_train, x_test, y_train, y_test = train_test_split(data, labels,test_size=0.20)
#Scaling of features
scaler = StandardScaler()
scaled_x_train = scaler.fit_transform(x_train)
scaled_x_test = scaler.transform(x_test)
scaled_data_original = scaler.transform(data_original)
random_state = 0
K = 5896
gamma = 177.82
random_state = check_random_state(random_state)
x_squared_norms = row_norms(scaled_x_train, squared=True)
if not sp.issparse(scaled_x_train):
scaled_x_train_mean = scaled_x_train.mean(axis=0)
scaled_x_train -= scaled_x_train_mean
if not sp.issparse(scaled_x_test):
scaled_x_test_mean = scaled_x_test.mean(axis=0)
scaled_x_test -= scaled_x_test_mean
if not sp.issparse(scaled_data_original):
scaled_data_original_mean = scaled_data_original.mean(axis=0)
scaled_data_original -= scaled_data_original_mean
#Initializing the centers using k-means++ algorithm implementation of sklearn
centers = _k_init(scaled_x_train, K, random_state=random_state, x_squared_norms=x_squared_norms)
from sklearn import metrics
from sklearn.cluster import KMeans
import numpy as np
from time import time
from sklearn.utils.extmath import row_norms, squared_norm
from sklearn.metrics.pairwise import euclidean_distances
k = 10
csv = np.genfromtxt('census_50k.csv', delimiter=",")
sh = csv.shape
mu = np.ones((k, sh[1]))
x_squared_norms = row_norms(csv, squared=True)
t0 = time()
for i in range(100):
all_distances = euclidean_distances(mu, csv, x_squared_norms, squared=True)
mu = mu + 1
t = time() - t0
print t
def convert_sklearn_kmeans(scope, operator, container):
"""
Computation graph of distances to all centroids for a batch of examples.
Note that a centriod is just the center of a cluster. We use ``[]`` to
denote the dimension of a variable; for example, ``X[3, 2]`` means that
*X* is a *3-by-2* tensor. In addition, for a matrix *X*, $X'$ denotes its
transpose.
Symbols:
* *l*: # of examples.
* *n*: # of features per input example.
* *X*: input examples, l-by-n tensor.
* *C*: centroids, k-by-n tensor.
* $C^2$: 2-norm of all centriod vectors, its shape is ``[k]``.
* *Y*: 2-norm of difference between examples and centroids,
*l-by-k* tensor. The value at i-th row and k-th column row,
``Y[i,k]``,is the distance from example *i* to centroid *k*.
* *L*: the id of the nearest centroid for each input example,
its shape is ``[l]``.
::
.------------------------------------------------------.
| |
| v
X [l, n] --> ReduceSumSquare -> X^2 [l] Gemm (alpha=-2, transB=1) <- C [k, n]
| |
| v
`------> Add <-- -2XC' [l, k]
|
v
C^2 [k] --------> Add <----- Z [l, k]
|
v
L [l] <-- ArgMin <-- Y2 [l, k] --> Sqrt --> Y2 [l, k]
*scikit-learn* code:
::
X = data
Y = model.cluster_centers_
XX = row_norms(X, squared=True)
YY = row_norms(Y, squared=True)
distances = safe_sparse_dot(X, Y.T, dense_output=True)
distances *= -2
distances += XX[:, numpy.newaxis]
distances += YY[numpy.newaxis, :]
numpy.sqrt(distances, out=distances)
"""
op = operator.raw_operator
variable = operator.inputs[0]
N = variable.type.shape[0]
# centroids
shapeC = list(op.cluster_centers_.shape)
nameC = scope.get_unique_variable_name('centroid')
container.add_initializer(nameC, onnx_proto.TensorProto.FLOAT,
shapeC, op.cluster_centers_.flatten())
nameX2 = scope.get_unique_variable_name('X2')
nameX = operator.inputs[0].full_name
container.add_node('ReduceSumSquare', [nameX], [nameX2], axes=[1],
keepdims=1,
name=scope.get_unique_operator_name('ReduceSumSquare'))
# Compute -2XC'
zero_name = scope.get_unique_variable_name('zero')
zeros = np.zeros((N, ))
container.add_initializer(zero_name, onnx_proto.TensorProto.FLOAT,
list(zeros.shape), zeros)
nameXC2 = scope.get_unique_variable_name('XC2')
apply_gemm(scope, [nameX, nameC, zero_name], [nameXC2], container,
alpha=-2., transB=1)
# Compute Z = X^2 - 2XC'
nameZ = scope.get_unique_variable_name("Z")
apply_add(scope, [nameXC2, nameX2], [nameZ], container)
# centroids ^2
nameC2 = scope.get_unique_variable_name('C2')
c2 = row_norms(op.cluster_centers_, squared=True)
container.add_initializer(nameC2, onnx_proto.TensorProto.FLOAT,
[1, shapeC[0]], c2.flatten())
# Compute Y2 = Z + C^2
nameY2 = scope.get_unique_variable_name('Y2')
apply_add(scope, [nameZ, nameC2], [nameY2], container)
# Compute Y = sqrt(Y2)
nameY = operator.outputs[1].full_name
apply_sqrt(scope, [nameY2], [nameY], container)
# Compute the most-matched cluster index, L
nameL = operator.outputs[0].full_name
container.add_node('ArgMin', [nameY2], [nameL],
name=scope.get_unique_operator_name('ArgMin'),
axis=1, keepdims=0)
def row_norms(X, squared=False):
if isinstance(X, np.ndarray):
return skm.row_norms(X, squared=squared)
return X.map_blocks(
skm.row_norms, chunks=(X.chunks[0],), drop_axis=1, squared=squared
)
def k_means_gpu_sparsity(weight_vector,
n_clusters,
ratio=0.5,
verbosity=0,
seed=int(time.time()),
gpu_id=0):
if ratio == 0:
return k_means_gpu(weight_vector=weight_vector,
n_clusters=n_clusters,
verbosity=verbosity,
seed=seed,
gpu_id=gpu_id)
if ratio == 1:
if n_clusters == 1:
mean_sample = np.mean(weight_vector, axis=0)
weight_vector = np.tile(mean_sample, (weight_vector.shape[0], 1))
return weight_vector
elif weight_vector.shape[0] == n_clusters:
return weight_vector
else:
weight_vector_1_mean = np.mean(weight_vector, axis=0)
weight_vector_compress = np.zeros(
(weight_vector.shape[0], weight_vector.shape[1]),
dtype=np.float32)
for v in weight_vector.shape[0]:
weight_vector_compress[v, :] = weight_vector_1_mean
return weight_vector_compress
else:
if n_clusters == 1:
mean_sample = np.mean(weight_vector, axis=0)
weight_vector = np.tile(mean_sample, (weight_vector.shape[0], 1))
return weight_vector
elif weight_vector.shape[0] == n_clusters:
return weight_vector
elif weight_vector.shape[1] == 1:
return k_means_sparsity(weight_vector,
n_clusters,
ratio,
seed=seed)
else:
num_samples = weight_vector.shape[0]
mean_sample = np.mean(weight_vector, axis=0)
center_cluster_index = np.argsort(
np.linalg.norm(weight_vector - mean_sample,
axis=1))[:int(num_samples * ratio)]
weight_vector_1_mean = np.mean(
weight_vector[center_cluster_index, :], axis=0)
remaining_cluster_index = np.asarray([
i for i in np.arange(num_samples)
if i not in center_cluster_index
])
weight_vector_train = weight_vector[remaining_cluster_index, :]
init_centers = k_means_._k_init(X=weight_vector_train,
n_clusters=n_clusters - 1,
x_squared_norms=row_norms(
weight_vector_train,
squared=True),
random_state=RandomState(seed))
centers, labels = kmeans_cuda(samples=weight_vector_train,
clusters=n_clusters - 1,
init=init_centers,
yinyang_t=0,
seed=seed,
device=gpu_id,
verbosity=verbosity)
weight_vector_compress = np.zeros(
(weight_vector.shape[0], weight_vector.shape[1]),
dtype=np.float32)
for v in center_cluster_index:
weight_vector_compress[v, :] = weight_vector_1_mean
for i, v in enumerate(remaining_cluster_index):
weight_vector_compress[v, :] = centers[labels[i], :]
return weight_vector_compress
def _initialize_nrkmeans_parameters(X, n_clusters, V, m, P, centers, max_iter,
random_state):
"""
Initialize the input parameters form NrKmeans. This means that all input values which are None must be defined.
Also all input parameters which are not None must be checked, if a correct execution is possible.
:param X: input data
:param n_clusters: list containing number of clusters for each subspace
:param V: orthogonal rotation matrix
:param m: list containing number of dimensionalities for each subspace
:param P: list containing projections for each subspace
:param centers: list containing the cluster centers for each subspace
:param max_iter: maximum number of iterations for the algorithm
:param random_state: use a fixed random state to get a repeatable solution
:return: checked V, m, P, centers, random_state, number of subspaces, labels, scatter_matrices
"""
data_dimensionality = X.shape[1]
random_state = check_random_state(random_state)
# Check if n_clusters is a list
if not type(n_clusters) is list:
raise ValueError(
"Number of clusters must be specified for each subspace and therefore be a list.\nYour input:\n"
+ str(n_clusters))
# Check if n_clusters contains negative values
if len([x for x in n_clusters if x < 1]) > 0:
raise ValueError(
"Number of clusters must not contain negative values or 0.\nYour input:\n"
+ str(n_clusters))
# Check if n_clusters contains more than one noise space
nr_noise_spaces = len([x for x in n_clusters if x == 1])
if nr_noise_spaces > 1:
raise ValueError(
"Only one subspace can be the noise space (number of clusters = 1).\nYour input:\n"
+ str(n_clusters))
# Check if noise space is not the last member in n_clusters
if nr_noise_spaces != 0 and n_clusters[-1] != 1:
raise ValueError(
"Noise space (number of clusters = 1) must be the last entry in n_clusters.\nYour input:\n"
+ str(n_clusters))
# Get number of subspaces
subspaces = len(n_clusters)
# Check if V is orthogonal
if V is None:
V = ortho_group.rvs(dim=data_dimensionality, random_state=random_state)
if not _is_matrix_orthogonal(V):
raise Exception("Your input matrix V is not orthogonal.\nV:\n" +
str(V))
# Calculate dimensionalities m
if m is None and P is None:
m = [int(data_dimensionality / subspaces)] * subspaces
if data_dimensionality % subspaces != 0:
choices = random_state.choice(range(subspaces),
data_dimensionality - sum(m))
for choice in choices:
m[choice] += 1
# If m is None but P is defined use P's dimensionality
elif m is None:
m = [len(x) for x in P]
if not type(m) is list or not len(m) is subspaces:
raise ValueError(
"A dimensionality list m must be specified for each subspace.\nYour input:\n"
+ str(m))
# Calculate projections P
if P is None:
possible_projections = list(range(data_dimensionality))
P = []
for dimensionality in m:
choices = random_state.choice(possible_projections,
dimensionality,
replace=False)
P.append(choices)
possible_projections = list(
set(possible_projections) - set(choices))
if not type(P) is list or not len(P) is subspaces:
raise ValueError(
"Projection lists must be specified for each subspace.\nYour input:\n"
+ str(P))
else:
# Check if the length of entries in P matches values of m
used_dimensionalities = []
for i, dimensionality in enumerate(m):
used_dimensionalities.extend(P[i])
if not len(P[i]) == dimensionality:
raise ValueError(
"Values for dimensionality m and length of projection list P do not match.\nDimensionality m:\n"
+ str(dimensionality) + "\nDimensionality P:\n" +
str(P[i]))
# Check if every dimension in considered in P
if sorted(used_dimensionalities) != list(range(data_dimensionality)):
raise ValueError(
"Projections P must include all dimensionalities.\nYour used dimensionalities:\n"
+ str(used_dimensionalities))
# Define initial cluster centers with kmeans++ for each subspace
if centers is None:
centers = []
for i in range(subspaces):
k = n_clusters[i]
if k > 1:
P_subspace = P[i]
cropped_X = np.matmul(X, V[:, P_subspace])
centers_cropped = kpp(cropped_X, k,
row_norms(cropped_X, squared=True),
random_state)
labels, _ = pairwise_distances_argmin_min(
X=cropped_X,
Y=centers_cropped,
metric='euclidean',
metric_kwargs={'squared': True})
centers_sub = np.zeros((k, X.shape[1]))
# Update cluster parameters
for center_id, _ in enumerate(centers_sub):
# Get points in this cluster
points_in_cluster = np.where(labels == center_id)[0]
# Update center
centers_sub[center_id] = np.average(X[points_in_cluster],
axis=0)
centers.append(centers_sub)
else:
centers.append(np.expand_dims(np.average(X, axis=0), 0))
if not type(centers) is list or not len(centers) is subspaces:
raise ValueError(
"Cluster centers must be specified for each subspace.\nYour input:\n"
+ str(centers))
else:
# Check if number of centers for subspaces matches value in n_clusters
for i, subspace_centers in enumerate(centers):
if not n_clusters[i] == len(subspace_centers):
raise ValueError(
"Values for number of clusters n_clusters and number of centers do not match.\nNumber of clusters:\n"
+ str(n_clusters[i]) + "\nNumber of centers:\n" +
str(len(subspace_centers)))
# Check max iter
if max_iter is None or type(max_iter) is not int or max_iter <= 0:
raise ValueError(
"Max_iter must be an integer larger than 0. Your Max_iter:\n" +
str(max_iter))
# Initial labels and scatter matrices
labels = [None] * subspaces
scatter_matrices = [None] * subspaces
return V, m, P, centers, random_state, subspaces, labels, scatter_matrices
def _sparse_encode(X, dictionary, gram, cov=None, algorithm='lasso_lars',
regularization=None, copy_cov=True,
init=None, max_iter=1000):
"""Generic sparse coding
Each column of the result is the solution to a Lasso problem.
Parameters
----------
X: array of shape (n_samples, n_features)
Data matrix.
dictionary: array of shape (n_components, n_features)
The dictionary matrix against which to solve the sparse coding of
the data. Some of the algorithms assume normalized rows.
gram: None | array, shape=(n_components, n_components)
Precomputed Gram matrix, dictionary * dictionary'
gram can be None if method is 'threshold'.
cov: array, shape=(n_components, n_samples)
Precomputed covariance, dictionary * X'
algorithm: {'lasso_lars', 'lasso_cd', 'lars', 'omp', 'threshold'}
lars: uses the least angle regression method (linear_model.lars_path)
lasso_lars: uses Lars to compute the Lasso solution
lasso_cd: uses the coordinate descent method to compute the
Lasso solution (linear_model.Lasso). lasso_lars will be faster if
the estimated components are sparse.
omp: uses orthogonal matching pursuit to estimate the sparse solution
threshold: squashes to zero all coefficients less than regularization
from the projection dictionary * data'
regularization : int | float
The regularization parameter. It corresponds to alpha when
algorithm is 'lasso_lars', 'lasso_cd' or 'threshold'.
Otherwise it corresponds to n_nonzero_coefs.
init: array of shape (n_samples, n_components)
Initialization value of the sparse code. Only used if
`algorithm='lasso_cd'`.
max_iter: int, 1000 by default
Maximum number of iterations to perform if `algorithm='lasso_cd'`.
copy_cov: boolean, optional
Whether to copy the precomputed covariance matrix; if False, it may be
overwritten.
Returns
-------
code: array of shape (n_components, n_features)
The sparse codes
See also
--------
sklearn.linear_model.lars_path
sklearn.linear_model.orthogonal_mp
sklearn.linear_model.Lasso
SparseCoder
"""
if X.ndim == 1:
X = X[:, np.newaxis]
n_samples, n_features = X.shape
if cov is None and algorithm != 'lasso_cd':
# overwriting cov is safe
copy_cov = False
cov = np.dot(dictionary, X.T)
if algorithm == 'lasso_admm':
alpha = float(regularization) / n_features # account for scaling
try:
err_mgt = np.seterr(all='ignore')
code, dictionary = lasso_admm(X.T, dictionary.T,
gamma=alpha,
gram=gram, cov=cov,
max_iter=max_iter)
new_code = code.T
finally:
np.seterr(**err_mgt)
elif algorithm == 'lasso_lars':
alpha = float(regularization) / n_features # account for scaling
try:
err_mgt = np.seterr(all='ignore')
lasso_lars = LassoLars(alpha=alpha, fit_intercept=False,
verbose=False, normalize=False,
precompute=gram, fit_path=False)
lasso_lars.fit(dictionary.T, X.T, Xy=cov)
new_code = lasso_lars.coef_
finally:
np.seterr(**err_mgt)
elif algorithm == 'lasso_cd':
alpha = float(regularization) / n_features # account for scaling
clf = Lasso(alpha=alpha, fit_intercept=False, precompute=gram,
max_iter=max_iter, warm_start=True)
clf.coef_ = init
clf.fit(dictionary.T, X.T)
new_code = clf.coef_
elif algorithm == 'lars':
try:
err_mgt = np.seterr(all='ignore')
lars = Lars(fit_intercept=False, verbose=False, normalize=False,
precompute=gram, n_nonzero_coefs=int(regularization),
fit_path=False)
lars.fit(dictionary.T, X.T, Xy=cov)
new_code = lars.coef_
finally:
np.seterr(**err_mgt)
elif algorithm == 'threshold':
new_code = ((np.sign(cov) *
np.maximum(np.abs(cov) - regularization, 0)).T)
elif algorithm == 'omp':
new_code = orthogonal_mp_gram(gram, cov, regularization, None,
row_norms(X, squared=True),
copy_Xy=copy_cov).T
else:
raise ValueError('Sparse coding method must be "lasso_lars" '
'"lasso_cd", "lasso", "threshold" or "omp", got %s.'
% algorithm)
return new_code
def k_means_gpu_sparsity(weight_vector,
n_clusters,
ratio=0.5,
verbosity=0,
seed=int(time.time()),
gpu_id=0):
# print(n_clusters)
if ratio == 0:
return k_means_gpu(weight_vector=weight_vector,
n_clusters=n_clusters,
verbosity=verbosity,
seed=seed,
gpu_id=gpu_id)
if ratio == 1:
if n_clusters == 1:
mean_sample = np.mean(weight_vector, axis=0)
weight_vector = np.tile(mean_sample, (weight_vector.shape[0], 1))
return weight_vector
elif weight_vector.shape[0] == n_clusters:
return weight_vector
else:
# mean_sample = np.mean(weight_vector, axis=0)
weight_vector_1_mean = np.mean(weight_vector, axis=0)
weight_vector_compress = np.zeros(
(weight_vector.shape[0], weight_vector.shape[1]),
dtype=np.float32)
for v in weight_vector.shape[0]:
weight_vector_compress[v, :] = weight_vector_1_mean
return weight_vector_compress
else:
if n_clusters == 1:
mean_sample = np.mean(weight_vector, axis=0)
weight_vector = np.tile(mean_sample, (weight_vector.shape[0], 1))
return weight_vector
elif weight_vector.shape[0] == n_clusters:
return weight_vector
elif weight_vector.shape[1] == 1:
return k_means_sparsity(weight_vector,
n_clusters,
ratio,
seed=seed)
else:
# print('n_clusters', n_clusters)
# print('weight_vector.shape',weight_vector.shape)
# print('kmeans++ init start')
num_samples = weight_vector.shape[0]
mean_sample = np.mean(weight_vector, axis=0)
center_cluster_index = np.argsort(
np.linalg.norm(weight_vector - mean_sample,
axis=1))[:int(num_samples * ratio)]
# weight_vector_1 = weight_vector[min_index, :]
weight_vector_1_mean = np.mean(
weight_vector[center_cluster_index, :], axis=0)
remaining_cluster_index = np.asarray([
i for i in np.arange(num_samples)
if i not in center_cluster_index
])
weight_vector_train = weight_vector[remaining_cluster_index, :]
# weight_vector_train = [element for i, element in enumerate(weight_vector) if i not in min_index]
# weight_vector = np.tile(mean_sample, (weight_vector.shape[0], 1))
init_centers = sklearn.cluster.k_means_._k_init(
X=weight_vector_train,
n_clusters=n_clusters - 1,
x_squared_norms=row_norms(weight_vector_train, squared=True),
random_state=RandomState(seed))
# # print('kmeans++ init finished')
# # print('init_centers.shape',init_centers.shape)
centers, labels = kmeans_cuda(samples=weight_vector_train,
clusters=n_clusters - 1,
init=init_centers,
yinyang_t=0,
seed=seed,
device=gpu_id,
verbosity=verbosity)
# print(np.unique(labels, axis=0).shape[0]+1)
# centers, labels = kmeans_cuda(samples = weight_vector, clusters = n_clusters, init="k-means++", yinyang_t=0, seed=seed, device=gpu_id, verbosity=verbosity)
# centers, labels = kmeans_cuda(samples = weight_vector, clusters = n_clusters, init="random", yinyang_t=0, seed=seed, device=gpu_id, verbosity=verbosity)
# centers, labels = kmeans_cuda(samples = weight_vector, clusters = n_clusters, init="afk-mc2", yinyang_t=0, seed=seed, device=gpu_id, verbosity=verbosity)
weight_vector_compress = np.zeros(
(weight_vector.shape[0], weight_vector.shape[1]),
dtype=np.float32)
for v in center_cluster_index:
weight_vector_compress[v, :] = weight_vector_1_mean
for i, v in enumerate(remaining_cluster_index):
weight_vector_compress[v, :] = centers[labels[i], :]
# weight_compress = np.reshape(weight_vector_compress, (filters_num, filters_channel, filters_size, filters_size))
# print(np.unique(weight_vector_compress, axis=0).shape[0])
# print(n_clusters, '\n')
# assert np.unique(weight_vector_compress, axis=0).shape[0]==n_clusters, "cluster number mismatch"
return weight_vector_compress
def fit(self, X):
rng = np.random.RandomState(self.random_state)
new_cluster_centers = np.zeros((self.n_clusters, X.shape[1]))
n_samples_arrays = np.arange(X.shape[0])
if self.return_cost_per_iteration:
self.cost_array_ = np.zeros(self.max_iter)
if self.n_clusters > 20:
raise ValueError("Group clustering not supported yet")
if self.init == "random":
old_cluster_centers_ = X[rng.randint(0, X.shape[0], self.n_clusters), :]
else:
raise ValueError("wait till we support other initializations.")
# Run K-Means for the first time.
# Don't do cluster.KMeans().fit(X) because of input_validation etc.
dot_product = 2 * np.dot(X, old_cluster_centers_.T)
cluster_norms = row_norms(old_cluster_centers_, squared=True).reshape(1, -1)
self.distances_ = row_norms(X, squared=True).reshape(-1, 1) - dot_product + cluster_norms
# Remove the closest and the second closest cluster.
upper_and_lower_bounds = np.argpartition(self.distances_, 1, axis=1)
self.labels_ = upper_and_lower_bounds[:, 0]
self.almost_labels_ = upper_and_lower_bounds[:, 1]
self.upper_and_lower_bounds_ = self.distances_[n_samples_arrays.reshape(-1, 1), upper_and_lower_bounds]
# Update cluster centers
for i in range(self.n_clusters):
new_cluster_centers[i] = np.mean(X[self.labels_ == i], axis=0)
self.cluster_centers_ = new_cluster_centers
for n_iter in range(self.max_iter):
if self.return_cost_per_iteration:
self.cost_array_[n_iter] = _calculate_cost(X, self.labels_, self.cluster_centers_)
# Calculate how much each center has drifted.
drift = ((old_cluster_centers_ - self.cluster_centers_)**2).sum(axis=1)
if np.sum(drift) < self.tol:
break
old_cluster_centers_ = np.copy(self.cluster_centers_)
# Add the drift to the upper bounds and subtract the drift from the lower bounds.
for i in range(self.n_clusters):
mask = self.labels_ == i
self.upper_and_lower_bounds_[:, 0][mask] += drift[i]
self.upper_and_lower_bounds_[:, 1][mask] -= drift[i]
# If the previously second_largest_bound is now lesser than the largest bound
# set the upper bound to the distance between the largest_bound
# This is based on d(old_center, new_center) + d(old_center, X) > d(X, new_center)
mask_changed_bounds = self.upper_and_lower_bounds_[:, 1] < self.upper_and_lower_bounds_[:, 0]
#XXX: Vectorize?
for i in range(self.n_clusters):
cluster = self.cluster_centers_[i]
new_mask = np.logical_and(mask_changed_bounds, self.labels_ == i)
distances = np.sum((X[new_mask] - cluster)**2, axis=1)
self.upper_and_lower_bounds_[:, 0][new_mask] = distances
# Now we can be sure that the second closest center is actually the closest.
# Reassign the labels.
mask_changed_bounds = self.upper_and_lower_bounds_[:, 1] < self.upper_and_lower_bounds_[:, 0]
tmp = self.labels_[mask_changed_bounds]
self.labels_[mask_changed_bounds] = self.almost_labels_[mask_changed_bounds]
self.almost_labels_[mask_changed_bounds] = tmp
self.upper_and_lower_bounds_[:, 1][mask_changed_bounds] = self.upper_and_lower_bounds_[:, 0][mask_changed_bounds]
#XXX: Vectorize?
for i in range(self.n_clusters):
cluster = self.cluster_centers_[i]
new_mask = np.logical_and(mask_changed_bounds, self.labels_ == i)
distances = np.sum((X[new_mask] - cluster)**2, axis=1)
self.upper_and_lower_bounds_[:, 0][new_mask] = distances
# TODO: Optimize this step.
for i in range(self.n_clusters):
mask = self.labels_ == i
self.cluster_centers_[i] = np.mean(X[mask], axis=0)
self.n_iter_ = n_iter
def spherical_k_means(X,
n_clusters,
init='k-means++',
n_init=10,
max_iter=300,
verbose=False,
tol=1e-4,
random_state=None,
copy_x=True,
n_jobs=1,
algorithm="auto",
return_n_iter=False):
"""Modified from sklearn.cluster.k_means_.k_means.
"""
if n_init <= 0:
raise ValueError("Invalid number of initializations."
" n_init=%d must be bigger than zero." % n_init)
random_state = check_random_state(random_state)
if max_iter <= 0:
raise ValueError('Number of iterations should be a positive number,'
' got %d instead' % max_iter)
best_inertia = np.infty
X = as_float_array(X, copy=copy_x)
tol = _tolerance(X, tol)
if hasattr(init, '__array__'):
init = check_array(init, dtype=X.dtype.type, copy=True)
_validate_center_shape(X, n_clusters, init)
if n_init != 1:
warnings.warn(
'Explicit initial center position passed: '
'performing only one init in k-means instead of n_init=%d' %
n_init,
RuntimeWarning,
stacklevel=2)
n_init = 1
# precompute squared norms of data points
x_squared_norms = row_norms(X, squared=True)
if n_jobs == 1:
# For a single thread, less memory is needed if we just store one set
# of the best results (as opposed to one set per run per thread).
for it in range(n_init):
# run a k-means once
labels, inertia, centers, n_iter_ = _spherical_kmeans_single_lloyd(
X,
n_clusters,
max_iter=max_iter,
init=init,
verbose=verbose,
tol=tol,
x_squared_norms=x_squared_norms,
random_state=random_state)
# determine if these results are the best so far
if best_inertia is None or inertia < best_inertia:
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
best_n_iter = n_iter_
else:
# parallelisation of k-means runs
seeds = random_state.randint(np.iinfo(np.int32).max, size=n_init)
results = Parallel(n_jobs=n_jobs, verbose=0)(
delayed(_spherical_kmeans_single_lloyd)(
X,
n_clusters,
max_iter=max_iter,
init=init,
verbose=verbose,
tol=tol,
x_squared_norms=x_squared_norms,
# Change seed to ensure variety
random_state=seed) for seed in seeds)
# Get results with the lowest inertia
labels, inertia, centers, n_iters = zip(*results)
best = np.argmin(inertia)
best_labels = labels[best]
best_inertia = inertia[best]
best_centers = centers[best]
best_n_iter = n_iters[best]
if return_n_iter:
return best_centers, best_labels, best_inertia, best_n_iter
else:
return best_centers, best_labels, best_inertia
def mbkmean(self, options, n_clusters, n_init, batch_size, n_iter,
n_samples, labels_true, k_means, X):
#to do with online MBK_mean
#Compute clustering with MiniBatchKMeans
mbk = cluster.MiniBatchKMeans(init=self.init,
n_clusters=n_clusters,
batch_size=batch_size,
n_init=10,
max_no_improvement=n_iter,
verbose=0)
#INIT THREADs
try:
if options[2] == '-pp' or options[3] == '-pp':
thread_1 = afficheur('starting threads', labels_true, mbk,
k_means, X, n_clusters)
thread_1.start()
except IndexError:
pass
try:
if options[2] == '-s':
#init state
n_batches = int(np.ceil(float(n_samples) / batch_size))
max_iter = 100
tol = 0
_, n_features = X.shape
old_center_buffer = np.zeros(n_features, dtype=X.dtype)
random_state = check_random_state(None)
init_size = 3 * batch_size
if init_size > n_samples:
init_size = n_samples
validation_indices = random_state.randint(
0, n_samples, init_size)
X_valid = X[validation_indices]
x_squared_norms = row_norms(X, squared=True)
x_squared_norms_valid = x_squared_norms[validation_indices]
counts = np.zeros(n_clusters, dtype=np.int32)
best_inertia = None
cluster_centers = None
for init_idx in range(n_init):
cluster_centers = cluster._init_centroids(
X,
n_clusters,
self.init,
random_state=random_state,
x_squared_norms=x_squared_norms,
init_size=init_size)
batch_inertia, centers_squared_diff = cluster._mini_batch_step(
X_valid,
x_squared_norms[validation_indices],
cluster_centers,
counts,
old_center_buffer,
False,
distances=None,
verbose=False)
_, inertia = cluster._labels_inertia(
X_valid, x_squared_norms_valid, cluster_centers)
if best_inertia is None or inertia < best_inertia:
mbk.cluster_centers_ = cluster_centers
mbk.counts_ = counts
best_inertia = inertia
print('best inertia %d' % best_inertia)
while (True):
thread_1 = afficheur('starting threads', labels_true, mbk,
k_means, X, n_clusters)
thread_1.start()
t0 = time.time()
for iteration_idx in range(n_iter):
minibatch_indices = random_state.randint(
0, n_samples, batch_size)
mbk = mbk.partial_fit(X[minibatch_indices])
thread_1.update(mbk)
t_mini_batch = time.time() - t0
thread_1.stop()
thread_1.join()
n_iter = self.input_num("Iterations suivante : ")
if n_iter == "stop":
return mbk, t_mini_batch
break
if isinstance(n_iter, int) == False:
print('error integer is required !!! type %s' %
type(n_iter))
break
except IndexError:
pass
try:
if options[2] == '-pp':
random_state = check_random_state(None)
t0 = time.time()
# Sample a minibatch from the full dataset
for iteration_idx in range(n_iter - 1):
minibatch_indices = random_state.randint(
0, n_samples, batch_size)
mbk = mbk.partial_fit(X[minibatch_indices])
thread_1.update(mbk)
t_mini_batch = time.time() - t0
thread_1.stop()
thread_1.join()
return mbk, t_mini_batch
except IndexError:
pass
try:
if options[2] == '-p':
random_state = check_random_state(None)
t0 = time.time()
for iteration_idx in range(n_iter):
minibatch_indices = random_state.randint(
0, n_samples, batch_size)
mbk = mbk.partial_fit(X[minibatch_indices])
t_mini_batch = time.time() - t0
return mbk, t_mini_batch
except IndexError:
pass
try:
if options[2] == '-n':
t0 = time.time()
mbk = mbk.fit(X)
t_mini_batch = time.time() - t0
return mbk, t_mini_batch
except IndexError:
pass
try:
if options[2] == None:
random_state = check_random_state(None)
# Sample a minibatch from the full dataset
t0 = time.time()
for iteration_idx in range(n_iter - 1):
minibatch_indices = random_state.randint(
0, n_samples, self.batch_size)
mbk = mbk.partial_fit(X,
minibatch_indices=minibatch_indices)
t_mini_batch = time.time() - t0
return mbk, t_mini_batch
except IndexError:
pass
try:
if options[2] == '-o':
n_batches = int(np.ceil(float(n_samples) / batch_size))
max_iter = 100
n_iter = int(max_iter * n_batches)
tol = 0
_, n_features = X.shape
old_center_buffer = np.zeros(n_features, dtype=X.dtype)
try:
# print('self.max_iter %d , n_batches %d '%(n_iter,n_batches))
if options[3] == '-pp':
#init state
random_state = check_random_state(None)
init_size = 3 * batch_size
if init_size > n_samples:
init_size = n_samples
validation_indices = random_state.randint(
0, n_samples, init_size)
X_valid = X[validation_indices]
x_squared_norms = row_norms(X, squared=True)
x_squared_norms_valid = x_squared_norms[
validation_indices]
counts = np.zeros(n_clusters, dtype=np.int32)
best_inertia = None
cluster_centers = None
#Random init with minimum inertia
for init_idx in range(n_init):
cluster_centers = cluster._init_centroids(
X,
n_clusters,
self.init,
random_state=random_state,
x_squared_norms=x_squared_norms,
init_size=init_size)
batch_inertia, centers_squared_diff = cluster._mini_batch_step(
X_valid,
x_squared_norms[validation_indices],
cluster_centers,
counts,
old_center_buffer,
False,
distances=None,
verbose=False)
_, inertia = cluster._labels_inertia(
X_valid, x_squared_norms_valid,
cluster_centers)
if best_inertia is None or inertia < best_inertia:
mbk.cluster_centers_ = cluster_centers
mbk.counts_ = counts
best_inertia = inertia
print('best inertia %d' % best_inertia)
convergence_context = {}
mbk.batch_inertia = batch_inertia
mbk.centers_squared_diff = centers_squared_diff
t0 = time.time()
for iteration_idx in range(n_iter):
minibatch_indices = random_state.randint(
0, n_samples, batch_size)
mbk = mbk.partial_fit(X[minibatch_indices])
tol = self._tolerance(X, tol)
thread_1.update(mbk)
# Monitor convergence and do early stopping if necessary
if cluster._mini_batch_convergence(
mbk,
iteration_idx,
n_iter,
tol,
n_samples,
mbk.centers_squared_diff,
mbk.batch_inertia,
convergence_context,
verbose=mbk.verbose):
t_mini_batch = time.time() - t0
thread_1.stop()
thread_1.join()
return mbk, t_mini_batch
break
elif options[3] == '-p':
random_state = check_random_state(None)
convergence_context = {}
t0 = time.time()
for iteration_idx in range(n_iter):
minibatch_indices = random_state.randint(
0, n_samples, batch_size)
mbk = mbk.partial_fit(X[minibatch_indices])
tol = self._tolerance(X, tol)
# Monitor convergence and do early stopping if necessary
if cluster._mini_batch_convergence(
mbk,
iteration_idx,
n_iter,
tol,
n_samples,
mbk.centers_squared_diff,
mbk.batch_inertia,
convergence_context,
verbose=False):
t_mini_batch = time.time() - t0
return mbk, t_mini_batch
break
except IndexError:
pass
except IndexError:
pass
def __init__(self,K,X,weights):
self.K = K
self.x_squared_norms = row_norms(X, squared=True)
self.X = X
self.weights = weights
def k_means_constrained(X,
n_clusters,
size_min=None,
size_max=None,
init='k-means++',
n_init=10,
max_iter=300,
verbose=False,
tol=1e-4,
random_state=None,
copy_x=True,
n_jobs=1,
return_n_iter=False):
"""K-Means clustering with minimum and maximum cluster size constraints.
Read more in the :ref:`User Guide <k_means>`.
Parameters
----------
X : array-like or sparse matrix, shape (n_samples, n_features)
The observations to cluster.
size_min : int, optional, default: None
Constrain the label assignment so that each cluster has a minimum
size of size_min. If None, no constrains will be applied
size_max : int, optional, default: None
Constrain the label assignment so that each cluster has a maximum
size of size_max. If None, no constrains will be applied
n_clusters : int
The number of clusters to form as well as the number of
centroids to generate.
init : {'k-means++', 'random', or ndarray, or a callable}, optional
Method for initialization, default to 'k-means++':
'k-means++' : selects initial cluster centers for k-mean
clustering in a smart way to speed up convergence. See section
Notes in k_init for more details.
'random': generate k centroids from a Gaussian with mean and
variance estimated from the data.
If an ndarray is passed, it should be of shape (n_clusters, n_features)
and gives the initial centers.
If a callable is passed, it should take arguments X, k and
and a random state and return an initialization.
n_init : int, optional, default: 10
Number of time the k-means algorithm will be run with different
centroid seeds. The final results will be the best output of
n_init consecutive runs in terms of inertia.
max_iter : int, optional, default 300
Maximum number of iterations of the k-means algorithm to run.
verbose : boolean, optional
Verbosity mode.
tol : float, optional
The relative increment in the results before declaring convergence.
random_state : int, RandomState instance or None, optional, default: None
If int, random_state is the seed used by the random number generator;
If RandomState instance, random_state is the random number generator;
If None, the random number generator is the RandomState instance used
by `np.random`.
copy_x : boolean, optional
When pre-computing distances it is more numerically accurate to center
the data first. If copy_x is True, then the original data is not
modified. If False, the original data is modified, and put back before
the function returns, but small numerical differences may be introduced
by subtracting and then adding the data mean.
n_jobs : int
The number of jobs to use for the computation. This works by computing
each of the n_init runs in parallel.
If -1 all CPUs are used. If 1 is given, no parallel computing code is
used at all, which is useful for debugging. For n_jobs below -1,
(n_cpus + 1 + n_jobs) are used. Thus for n_jobs = -2, all CPUs but one
are used.
return_n_iter : bool, optional
Whether or not to return the number of iterations.
Returns
-------
centroid : float ndarray with shape (k, n_features)
Centroids found at the last iteration of k-means.
label : integer ndarray with shape (n_samples,)
label[i] is the code or index of the centroid the
i'th observation is closest to.
inertia : float
The final value of the inertia criterion (sum of squared distances to
the closest centroid for all observations in the training set).
best_n_iter : int
Number of iterations corresponding to the best results.
Returned only if `return_n_iter` is set to True.
"""
if n_init <= 0:
raise ValueError("Invalid number of initializations."
" n_init=%d must be bigger than zero." % n_init)
random_state = check_random_state(random_state)
if max_iter <= 0:
raise ValueError('Number of iterations should be a positive number,'
' got %d instead' % max_iter)
X = as_float_array(X, copy=copy_x)
tol = _tolerance(X, tol)
# Validate init array
if hasattr(init, '__array__'):
init = check_array(init, dtype=X.dtype.type, copy=True)
_validate_center_shape(X, n_clusters, init)
if n_init != 1:
warnings.warn(
'Explicit initial center position passed: '
'performing only one init in k-means instead of n_init=%d' %
n_init,
RuntimeWarning,
stacklevel=2)
n_init = 1
# subtract of mean of x for more accurate distance computations
if not sp.issparse(X):
X_mean = X.mean(axis=0)
# The copy was already done above
X -= X_mean
if hasattr(init, '__array__'):
init -= X_mean
# precompute squared norms of data points
x_squared_norms = row_norms(X, squared=True)
best_labels, best_inertia, best_centers = None, None, None
if n_jobs == 1:
# For a single thread, less memory is needed if we just store one set
# of the best results (as opposed to one set per run per thread).
for it in range(n_init):
# run a k-means once
labels, inertia, centers, n_iter_ = kmeans_constrained_single(
X,
n_clusters,
size_min=size_min,
size_max=size_max,
max_iter=max_iter,
init=init,
verbose=verbose,
tol=tol,
x_squared_norms=x_squared_norms,
random_state=random_state)
# determine if these results are the best so far
if best_inertia is None or inertia < best_inertia:
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
best_n_iter = n_iter_
else:
# parallelisation of k-means runs
seeds = random_state.randint(np.iinfo(np.int32).max, size=n_init)
results = Parallel(n_jobs=n_jobs, verbose=0)(
delayed(kmeans_constrained_single)(
X,
n_clusters,
size_min=size_min,
size_max=size_max,
max_iter=max_iter,
init=init,
verbose=verbose,
tol=tol,
x_squared_norms=x_squared_norms,
# Change seed to ensure variety
random_state=seed) for seed in seeds)
# Get results with the lowest inertia
labels, inertia, centers, n_iters = zip(*results)
best = np.argmin(inertia)
best_labels = labels[best]
best_inertia = inertia[best]
best_centers = centers[best]
best_n_iter = n_iters[best]
if not sp.issparse(X):
if not copy_x:
X += X_mean
best_centers += X_mean
if return_n_iter:
return best_centers, best_labels, best_inertia, best_n_iter
else:
return best_centers, best_labels, best_inertia
def k_means(X,
n_clusters,
init='k-means++',
precompute_distances='auto',
n_init=10,
max_iter=300,
verbose=False,
tol=1e-4,
random_state=None,
copy_x=True,
n_jobs=1,
return_n_iter=False,
sample_weight=None):
"""K-means clustering algorithm.
Read more in the :ref:`User Guide <k_means>`.
Parameters
----------
X : array-like or sparse matrix, shape (n_samples, n_features)
The observations to cluster.
n_clusters : int
The number of clusters to form as well as the number of
centroids to generate.
max_iter : int, optional, default 300
Maximum number of iterations of the k-means algorithm to run.
n_init : int, optional, default: 10
Number of time the k-means algorithm will be run with different
centroid seeds. The final results will be the best output of
n_init consecutive runs in terms of inertia.
init : {'k-means++', 'random', or ndarray, or a callable}, optional
Method for initialization, default to 'k-means++':
'k-means++' : selects initial cluster centers for k-mean
clustering in a smart way to speed up convergence. See section
Notes in k_init for more details.
'random': generate k centroids from a Gaussian with mean and
variance estimated from the data.
If an ndarray is passed, it should be of shape (n_clusters, n_features)
and gives the initial centers.
If a callable is passed, it should take arguments X, k and
and a random state and return an initialization.
precompute_distances : {'auto', True, False}
Precompute distances (faster but takes more memory).
'auto' : do not precompute distances if n_samples * n_clusters > 12
million. This corresponds to about 100MB overhead per job using
double precision.
True : always precompute distances
False : never precompute distances
tol : float, optional
The relative increment in the results before declaring convergence.
verbose : boolean, optional
Verbosity mode.
random_state : integer or numpy.RandomState, optional
The generator used to initialize the centers. If an integer is
given, it fixes the seed. Defaults to the global numpy random
number generator.
copy_x : boolean, optional
When pre-computing distances it is more numerically accurate to center
the data first. If copy_x is True, then the original data is not
modified. If False, the original data is modified, and put back before
the function returns, but small numerical differences may be introduced
by subtracting and then adding the data mean.
n_jobs : int
The number of jobs to use for the computation. This works by computing
each of the n_init runs in parallel.
If -1 all CPUs are used. If 1 is given, no parallel computing code is
used at all, which is useful for debugging. For n_jobs below -1,
(n_cpus + 1 + n_jobs) are used. Thus for n_jobs = -2, all CPUs but one
are used.
return_n_iter : bool, optional
Whether or not to return the number of iterations.
Returns
-------
centroid : float ndarray with shape (k, n_features)
Centroids found at the last iteration of k-means.
label : integer ndarray with shape (n_samples,)
label[i] is the code or index of the centroid the
i'th observation is closest to.
inertia : float
The final value of the inertia criterion (sum of squared distances to
the closest centroid for all observations in the training set).
best_n_iter: int
Number of iterations corresponding to the best results.
Returned only if `return_n_iter` is set to True.
"""
if n_init <= 0:
raise ValueError("Invalid number of initializations."
" n_init=%d must be bigger than zero." % n_init)
random_state = check_random_state(random_state)
if max_iter <= 0:
raise ValueError('Number of iterations should be a positive number,'
' got %d instead' % max_iter)
best_inertia = np.infty
X = as_float_array(X, copy=copy_x)
tol = _tolerance(X, tol)
# If the distances are precomputed every job will create a matrix of shape
# (n_clusters, n_samples). To stop KMeans from eating up memory we only
# activate this if the created matrix is guaranteed to be under 100MB. 12
# million entries consume a little under 100MB if they are of type double.
if precompute_distances == 'auto':
n_samples = X.shape[0]
precompute_distances = (n_clusters * n_samples) < 12e6
elif isinstance(precompute_distances, bool):
pass
else:
raise ValueError("precompute_distances should be 'auto' or True/False"
", but a value of %r was passed" %
precompute_distances)
# subtract of mean of x for more accurate distance computations
if not sp.issparse(X) or hasattr(init, '__array__'):
X_mean = X.mean(axis=0)
if not sp.issparse(X):
# The copy was already done above
X -= X_mean
if hasattr(init, '__array__'):
init = check_array(init, dtype=np.float64, copy=True)
_validate_center_shape(X, n_clusters, init)
init -= X_mean
if n_init != 1:
warnings.warn(
'Explicit initial center position passed: '
'performing only one init in k-means instead of n_init=%d' %
n_init,
RuntimeWarning,
stacklevel=2)
n_init = 1
# precompute squared norms of data points
x_squared_norms = row_norms(X, squared=True)
best_labels, best_inertia, best_centers = None, None, None
if n_jobs == 1:
# For a single thread, less memory is needed if we just store one set
# of the best results (as opposed to one set per run per thread).
for it in range(n_init):
# run a k-means once
labels, inertia, centers, n_iter_ = _kmeans_single(
X,
n_clusters,
max_iter=max_iter,
init=init,
verbose=verbose,
precompute_distances=precompute_distances,
tol=tol,
x_squared_norms=x_squared_norms,
random_state=random_state,
sample_weight=sample_weight)
# determine if these results are the best so far
if best_inertia is None or inertia < best_inertia:
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
best_n_iter = n_iter_
else:
# parallelisation of k-means runs
seeds = random_state.randint(np.iinfo(np.int32).max, size=n_init)
results = Parallel(n_jobs=n_jobs, verbose=0)(
delayed(_kmeans_single)(
X,
n_clusters,
max_iter=max_iter,
init=init,
verbose=verbose,
tol=tol,
precompute_distances=precompute_distances,
x_squared_norms=x_squared_norms,
# Change seed to ensure variety
random_state=seed) for seed in seeds)
# Get results with the lowest inertia
labels, inertia, centers, n_iters = zip(*results)
best = np.argmin(inertia)
best_labels = labels[best]
best_inertia = inertia[best]
best_centers = centers[best]
best_n_iter = n_iters[best]
if not sp.issparse(X):
if not copy_x:
X += X_mean
best_centers += X_mean
if return_n_iter:
return best_centers, best_labels, best_inertia, best_n_iter
else:
return best_centers, best_labels, best_inertia
def setup(self):
self.X = _china_dataset()
self.n_clusters = 64
self.x_squared_norms = row_norms(self.X, squared=True)
def _estimate_log_gaussian_prob(X, means, precisions_chol, covariance_type):
"""Estimate the log Gaussian probability.
Parameters
----------
X : array-like, shape (n_samples, n_features)
means : array-like, shape (n_components, n_features)
precisions_chol : array-like
Cholesky decompositions of the precision matrices.
'full' : shape of (n_components, n_features, n_features)
'tied' : shape of (n_features, n_features)
'diag' : shape of (n_components, n_features)
'spherical' : shape of (n_components,)
covariance_type : {'full', 'tied', 'diag', 'spherical'}
Returns
-------
log_prob : array, shape (n_samples, n_components)
"""
n_samples, n_features = X.shape
n_components, _ = means.shape
# det(precision_chol) is half of det(precision)
log_det = _compute_log_det_cholesky(
precisions_chol, covariance_type, n_features)
# print(log_det)
if covariance_type == 'full':
log_prob = np.empty((n_samples, n_components))
for k, (mu, prec_chol) in enumerate(zip(means, precisions_chol)):
# import time
# t0 = time.time()
y = torch.mm(torch.Tensor(X).cuda(), torch.Tensor(prec_chol).cuda()) - \
torch.mv(torch.Tensor(prec_chol).t().cuda(), torch.Tensor(mu).cuda())
# t1 = time.time()
log_prob[:, k] = torch.sum(y**2, dim=1).cpu().numpy()
# t00 = time.time()
# y = np.dot(X, prec_chol) - np.dot(mu, prec_chol)
# t11 = time.time()
# log_prob[:, k] = np.sum(np.square(y), axis=1)
# print(time.time() - t11, t11-t00, 'cpus', t00 - t1, t1 - t0, 'gpus')
elif covariance_type == 'tied':
log_prob = np.empty((n_samples, n_components))
for k, mu in enumerate(means):
y = np.dot(X, precisions_chol) - np.dot(mu, precisions_chol)
log_prob[:, k] = np.sum(np.square(y), axis=1)
elif covariance_type == 'diag':
precisions = precisions_chol ** 2
log_prob = (np.sum((means ** 2 * precisions), 1) -
2. * np.dot(X, (means * precisions).T) +
np.dot(X ** 2, precisions.T))
elif covariance_type == 'spherical':
precisions = precisions_chol ** 2
log_prob = (np.sum(means ** 2, 1) * precisions -
2 * np.dot(X, means.T * precisions) +
np.outer(row_norms(X, squared=True), precisions))
return -.5 * (n_features * np.log(2 * np.pi) + log_prob) + log_det
def test_minibatch_update_consistency():
# Check that dense and sparse minibatch update give the same results
rng = np.random.RandomState(42)
centers_old = centers + rng.normal(size=centers.shape)
centers_old_csr = centers_old.copy()
centers_new = np.zeros_like(centers_old)
centers_new_csr = np.zeros_like(centers_old_csr)
weight_sums = np.zeros(centers_old.shape[0], dtype=X.dtype)
weight_sums_csr = np.zeros(centers_old.shape[0], dtype=X.dtype)
x_squared_norms = (X**2).sum(axis=1)
x_squared_norms_csr = row_norms(X_csr, squared=True)
sample_weight = np.ones(X.shape[0], dtype=X.dtype)
# extract a small minibatch
X_mb = X[:10]
X_mb_csr = X_csr[:10]
x_mb_squared_norms = x_squared_norms[:10]
x_mb_squared_norms_csr = x_squared_norms_csr[:10]
sample_weight_mb = sample_weight[:10]
# step 1: compute the dense minibatch update
old_inertia = _mini_batch_step(
X_mb,
x_mb_squared_norms,
sample_weight_mb,
centers_old,
centers_new,
weight_sums,
np.random.RandomState(0),
random_reassign=False,
)
assert old_inertia > 0.0
# compute the new inertia on the same batch to check that it decreased
labels, new_inertia = _labels_inertia(
X_mb, sample_weight_mb, x_mb_squared_norms, centers_new
)
assert new_inertia > 0.0
assert new_inertia < old_inertia
# step 2: compute the sparse minibatch update
old_inertia_csr = _mini_batch_step(
X_mb_csr,
x_mb_squared_norms_csr,
sample_weight_mb,
centers_old_csr,
centers_new_csr,
weight_sums_csr,
np.random.RandomState(0),
random_reassign=False,
)
assert old_inertia_csr > 0.0
# compute the new inertia on the same batch to check that it decreased
labels_csr, new_inertia_csr = _labels_inertia(
X_mb_csr, sample_weight_mb, x_mb_squared_norms_csr, centers_new_csr
)
assert new_inertia_csr > 0.0
assert new_inertia_csr < old_inertia_csr
# step 3: check that sparse and dense updates lead to the same results
assert_array_equal(labels, labels_csr)
assert_allclose(centers_new, centers_new_csr)
assert_allclose(old_inertia, old_inertia_csr)
assert_allclose(new_inertia, new_inertia_csr)
def _daal4py_check(self, X, y_, check_input):
#conver to 2d format
X = make2d(X)
y = make2d(y_)
#convet from list type
if isinstance(X, list):
X = np.asarray(X, np.float64)
if isinstance(y, list):
y = np.asarray(y, np.float64)
_fptype = getFPType(X)
#check alpha
if self.alpha == 0:
warnings.warn(
"With alpha=0, this algorithm does not converge "
"well. You are advised to use the LinearRegression "
"estimator",
stacklevel=2)
#check precompute
if isinstance(self.precompute, np.ndarray):
if check_input:
check_array(self.precompute, dtype=_fptype)
self.precompute = make2d(self.precompute)
#only for compliance with Sklearn
if self.fit_intercept:
X_offset = np.average(X, axis=0, weights=None)
if self.normalize:
X_scale = row_norms(X)
if np.isscalar(X_scale):
if X_scale == .0:
X_scale = 1.
elif isinstance(X_scale, np.ndarray):
X_scale[X_scale == 0.0] = 1.0
else:
X_scale = np.ones(X.shape[1], dtype=_fptype)
else:
X_offset = np.zeros(X.shape[1], dtype=_fptype)
X_scale = np.ones(X.shape[1], dtype=_fptype)
if (self.fit_intercept
and not np.allclose(X_offset, np.zeros(X.shape[1]))
or self.normalize
and not np.allclose(X_scale, np.ones(X.shape[1]))):
warnings.warn(
"Gram matrix was provided but X was centered"
" to fit intercept, "
"or X was normalized : recomputing Gram matrix.", UserWarning)
else:
if self.precompute not in [False, True, 'auto']:
raise ValueError("precompute should be one of True, False, "
"'auto' or array-like. Got %r" % self.precompute)
#check X and y
if check_input:
X, y = check_X_y(X,
y,
dtype=[np.float64, np.float32],
multi_output=True,
y_numeric=True)
else:
#only for compliance with Sklearn, this assert is not required for DAAL
if (X.flags['F_CONTIGUOUS'] == False):
raise ValueError("ndarray is not Fortran contiguous")
#check selection
if self.selection not in ['random', 'cyclic']:
raise ValueError("selection should be either random or cyclic.")
return X, y
def test_minibatch_update_consistency():
# Check that dense and sparse minibatch update give the same results
rng = np.random.RandomState(42)
old_centers = centers + rng.normal(size=centers.shape)
new_centers = old_centers.copy()
new_centers_csr = old_centers.copy()
counts = np.zeros(new_centers.shape[0], dtype=np.int32)
counts_csr = np.zeros(new_centers.shape[0], dtype=np.int32)
x_squared_norms = (X**2).sum(axis=1)
x_squared_norms_csr = row_norms(X_csr, squared=True)
buffer = np.zeros(centers.shape[1], dtype=np.double)
buffer_csr = np.zeros(centers.shape[1], dtype=np.double)
# extract a small minibatch
X_mb = X[:10]
X_mb_csr = X_csr[:10]
x_mb_squared_norms = x_squared_norms[:10]
x_mb_squared_norms_csr = x_squared_norms_csr[:10]
# step 1: compute the dense minibatch update
old_inertia, incremental_diff = _mini_batch_step(X_mb,
x_mb_squared_norms,
new_centers,
counts,
buffer,
1,
None,
random_reassign=False)
assert_greater(old_inertia, 0.0)
# compute the new inertia on the same batch to check that it decreased
labels, new_inertia = _labels_inertia(X_mb, x_mb_squared_norms,
new_centers)
assert_greater(new_inertia, 0.0)
assert_less(new_inertia, old_inertia)
# check that the incremental difference computation is matching the
# final observed value
effective_diff = np.sum((new_centers - old_centers)**2)
assert_almost_equal(incremental_diff, effective_diff)
# step 2: compute the sparse minibatch update
old_inertia_csr, incremental_diff_csr = _mini_batch_step(
X_mb_csr,
x_mb_squared_norms_csr,
new_centers_csr,
counts_csr,
buffer_csr,
1,
None,
random_reassign=False)
assert_greater(old_inertia_csr, 0.0)
# compute the new inertia on the same batch to check that it decreased
labels_csr, new_inertia_csr = _labels_inertia(X_mb_csr,
x_mb_squared_norms_csr,
new_centers_csr)
assert_greater(new_inertia_csr, 0.0)
assert_less(new_inertia_csr, old_inertia_csr)
# check that the incremental difference computation is matching the
# final observed value
effective_diff = np.sum((new_centers_csr - old_centers)**2)
assert_almost_equal(incremental_diff_csr, effective_diff)
# step 3: check that sparse and dense updates lead to the same results
assert_array_equal(labels, labels_csr)
assert_array_almost_equal(new_centers, new_centers_csr)
assert_almost_equal(incremental_diff, incremental_diff_csr)
assert_almost_equal(old_inertia, old_inertia_csr)
assert_almost_equal(new_inertia, new_inertia_csr)
def get_data(dataname, verbose=True):
assert dataname in data_bank, 'Dataset name not recognized!'
# meta information about the data
meta_dict = {
'isSparse': False,
'n_true_classes': None,
'xsn_train': None,
'xsn_test': None
}
train_lb = None
if dataname == 'rcv1':
print('Fetching RCV1 data from sklearn')
train = fetch_rcv1(subset='test')
train = train.data
test = fetch_rcv1(subset='train')
test = test.data
meta_dict['n_true_classes'] = 103
elif dataname == 'mnist':
print('Fetching MNIST data from sklearn')
mnist = fetch_mldata('MNIST original')
data_ind = range(mnist.data.shape[0])
random.shuffle(data_ind)
train_ind = data_ind[:60000]
test_ind = data_ind[-10000:]
train = mnist.data[train_ind, :]
test = mnist.data[test_ind, :]
meta_dict['n_true_classes'] = 10
elif dataname == 'gauss':
""" Synthetic Gaussian """
print('Generating Gaussian blurbs datasets')
#centers = [[2, 2], [-2, -2], [2, -2]]
#centers = np.asarray(centers)
n_s = 7000
gauss, _ = make_blobs(n_samples=n_s,
n_features=10,
centers=50,
center_box=(-30, 30),
cluster_std=10.0)
data_ind = range(gauss.shape[0])
random.shuffle(data_ind)
train_ind = data_ind[:6 * n_s / 7]
test_ind = data_ind[-n_s / 7:]
train = gauss[train_ind, :]
test = gauss[test_ind, :]
meta_dict['n_true_classes'] = 50
elif dataname == 'covtype':
""" Forest covertype """
print('Fetching forest covertype datasets')
cov = fetch_covtype()
train = cov.data[:500000]
train_lb = cov.target[:500000]
test = cov.data[500000:]
test_lb = cov.target[500000:]
meta_dict['n_true_classes'] = 7
elif dataname == 'cifar10_raw':
"""
This will make over 4 million data points for training, each with
dim 8*8*3
"""
# take user input to know which batch to preprocess
usr_input = raw_input('Which batch to preprocess: ')
# get data home
data_home = '/Users/tangch/scikit_learn_data/cifar10' # make this portable in the future
os.chdir(data_home)
prefix = 'data_batch_'
# combine training batches
if usr_input != 'test':
curr_batch = prefix + str(usr_input)
fname = curr_batch
dataname = dataname + 'batch_' + str(usr_input)
else:
fname = data_home + 'test_batch'
dataname = dataname + 'test_batch'
#print 'Opening '+fname + 'in directory' + os.getcwd()
with open(fname, 'rb') as f:
dict = pickle.load(f)
train = dict['data']
train_lb = dict['labels']
print 'processing batch %s' % fname
print '%d by %d training data with %d label' %(train.shape[0],\
train.shape[1], len(train_lb))
os.chdir(
'/Users/tangch/Documents/Python_projects/myprojects/mbkm_2016/mbkm'
)
# Reduce the dimension of data by random sampling
train = train.reshape(train.shape[0], 32, 32, 3)
width = 8
height = width
# subsampling patches u.a.r.
train, train_lb = convsubsample(train,
1,
width,
height,
labels=train_lb)
# random shuffle training data
#ind = range(train.shape[0])
print 'shuffling dataset randomly'
#ind = random.shuffle(ind)
#pdb.set_trace()
#train = train[ind]
#train_lb = train_lb[ind]
train = np.random.sample(train, train.shape[0])
train = np.array(train)
train_lb = np.random.sample(train_lb, train_lb.shape[0])
train_lb = np.array(train_lb)
test = None
# get test data/labels
#fname = data_home + 'test_batch'
#with open(fname,'rb') as f:
# dict = pickle.load(f)
#test = dict['data']
#test_lb = dict['labels']
# subsampling patches u.a.r.
#test, test_lb = convsubsample(test, 1, width,height, labels = test_lb)
meta_dict['n_true_classes'] = 10
print 'finished preprocessing current batch'
elif dataname == 'cifar10_norm':
"""
Only works if we have cifar10_raw
"""
usr_input = raw_input('Which batch to preprocess: ')
### load cifar10_raw
fname = 'cifar10_raw'
if usr_input != 'test':
fname = fname + 'batch_' + str(usr_input)
dataname = dataname + 'batch_' + str(usr_input)
else:
fname = fname + 'test_batch'
dataname = dataname + 'test_batch'
try:
with open(fname, 'rb') as f:
train, test, meta = pickle.load(f)
except Exception as e:
print('Cannot open file ' + fname)
### Normalization
train = normalize(train, norm='l2')
test = normalize(test, norm='l2')
### meta info extraction
meta_dict['n_true_classes'] = meta['n_true_classes']
train_lb = meta['train_lb']
test_lb = meta['test_lb']
elif dataname == 'cifar10_white_norm':
"""
Only works if we have cifar10_norm
"""
### load cifar10_norm
fname = 'cifar10_norm'
try:
with open(fname, 'rb') as f:
train_old, test_old, meta = pickle.load(f)
except Exception as e:
print('Cannot find file ' + fname)
train_test = np.vstack((train_old, test_old))
### Whitening
pca = RandomizedPCA(whiten=True) #use approx PCA to save computation
train_test = pca.fit_transform(train_test)
train = train_test[:train_old.shape[0]]
test = train_test[train_old.shape[0]:]
### Extract meta info
meta_dict['n_true_classes'] = meta['n_true_classes']
train_lb = meta['train_lb']
test_lb = meta['test_lb']
else:
print 'nothing'
meta_dict['dataname'] = dataname
# add true labels if exists
if not train_lb is None:
meta_dict['train_lb'] = train_lb
#meta_dict['test_lb'] = test_lb
# Check if data is sparse
if sp.issparse(train):
meta_dict['isSparse'] = True
print('The %s data is sparse' % dataname)
print('%d training data' % train.shape[0])
print('data dimension is %d' % train.shape[1])
if len(train.shape) == 3:
print('data has %d channels' % train.shape[2])
if test is None:
print 'No test data'
else:
print('%d test data' % test.shape[0])
print('The number of true classes is %d' % meta_dict['n_true_classes'])
# What does this do?
train = as_float_array(train, copy=True)
if not test is None:
test = as_float_array(test, copy=True)
# precompute squared norms for faster computation
x_squared_norms_tr = row_norms(train, squared=True)
meta_dict['xsn_train'] = x_squared_norms_tr
if not test is None:
x_squared_norms_tt = row_norms(test, squared=True)
meta_dict['xsn_test'] = x_squared_norms_tt
return train, test, meta_dict
def constraint_kmeans(
X,
labels,
sample_weight,
centers,
inertia,
iter,
max_iter, # pylint: disable=W0622
strategy='gain',
verbose=0,
state=None,
learning_rate=1.,
history=False,
fLOG=None):
"""
Completes the constraint :epkg:`k-means`.
@param X features
@param labels initialized labels (unused)
@param sample_weight sample weight
@param centers initialized centers
@param inertia initialized inertia (unused)
@param iter number of iteration already done
@param max_iter maximum of number of iteration
@param strategy strategy used to sort observations before
mapping them to clusters
@param verbose verbose
@param state random state
@param learning_rate used by strategy `'weights'`
@param history return list of centers accross iterations
@param fLOG logging function (needs to be specified otherwise
verbose has no effects)
@return tuple (best_labels, best_centers, best_inertia,
iter, all_centers)
"""
if labels.dtype != numpy.int32:
raise TypeError("Labels must be an array of int not '{0}'".format(
labels.dtype))
if strategy == 'weights':
return _constraint_kmeans_weights(X,
labels,
sample_weight,
centers,
inertia,
iter,
max_iter,
verbose=verbose,
state=state,
learning_rate=learning_rate,
history=history,
fLOG=fLOG)
else:
if isinstance(X, DataFrame):
X = X.values
x_squared_norms = row_norms(X, squared=True)
counters = numpy.empty((centers.shape[0], ), dtype=numpy.int32)
limit = X.shape[0] // centers.shape[0]
leftover = X.shape[0] - limit * centers.shape[0]
leftclose = numpy.empty((centers.shape[0], ), dtype=numpy.int32)
n_clusters = centers.shape[0]
distances_close = numpy.empty((X.shape[0], ), dtype=X.dtype)
best_inertia = None
best_iter = None
all_centers = []
# association
_constraint_association(leftover,
counters,
labels,
leftclose,
distances_close,
centers,
X,
x_squared_norms,
limit,
strategy,
state=state)
if sample_weight is None:
sw = numpy.ones((X.shape[0], ))
else:
sw = sample_weight
if scipy.sparse.issparse(X):
_centers_fct = _centers_sparse
else:
_centers_fct = _centers_dense
while iter < max_iter:
# compute new clusters
centers = _centers_fct(X, sw, labels, n_clusters, distances_close)
if history:
all_centers.append(centers)
# association
_constraint_association(leftover,
counters,
labels,
leftclose,
distances_close,
centers,
X,
x_squared_norms,
limit,
strategy,
state=state)
# inertia
_, inertia = _labels_inertia_skl(X=X,
sample_weight=sw,
x_squared_norms=x_squared_norms,
centers=centers,
distances=distances_close)
iter += 1
if verbose and fLOG:
fLOG("CKMeans %d/%d inertia=%f" % (iter, max_iter, inertia))
# best option so far?
if best_inertia is None or inertia < best_inertia:
best_inertia = inertia
best_centers = centers.copy()
best_labels = labels.copy()
best_iter = iter
# early stop
if (best_inertia is not None and inertia >= best_inertia
and iter > best_iter + 5):
break
return (best_labels, best_centers, best_inertia, None, iter,
all_centers)
def spherical_k_means(X, n_clusters, init='k-means++', n_init=10,
max_iter=300, verbose=False, tol=1e-4, random_state=None,
copy_x=True, n_jobs=1, algorithm="auto", return_n_iter=False):
"""Modified from sklearn.cluster.k_means_.k_means.
"""
if n_init <= 0:
raise ValueError("Invalid number of initializations."
" n_init=%d must be bigger than zero." % n_init)
random_state = check_random_state(random_state)
if max_iter <= 0:
raise ValueError('Number of iterations should be a positive number,'
' got %d instead' % max_iter)
best_inertia = np.infty
X = as_float_array(X, copy=copy_x)
tol = _tolerance(X, tol)
if hasattr(init, '__array__'):
init = check_array(init, dtype=X.dtype.type, copy=True)
_validate_center_shape(X, n_clusters, init)
if n_init != 1:
warnings.warn(
'Explicit initial center position passed: '
'performing only one init in k-means instead of n_init=%d'
% n_init, RuntimeWarning, stacklevel=2)
n_init = 1
# precompute squared norms of data points
x_squared_norms = row_norms(X, squared=True)
if n_jobs == 1:
# For a single thread, less memory is needed if we just store one set
# of the best results (as opposed to one set per run per thread).
for it in range(n_init):
# run a k-means once
labels, inertia, centers, n_iter_ = _spherical_kmeans_single_lloyd(
X, n_clusters, max_iter=max_iter, init=init, verbose=verbose,
tol=tol, x_squared_norms=x_squared_norms,
random_state=random_state)
# determine if these results are the best so far
if best_inertia is None or inertia < best_inertia:
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
best_n_iter = n_iter_
else:
# parallelisation of k-means runs
seeds = random_state.randint(np.iinfo(np.int32).max, size=n_init)
results = Parallel(n_jobs=n_jobs, verbose=0)(
delayed(_spherical_kmeans_single_lloyd)(X, n_clusters,
max_iter=max_iter, init=init,
verbose=verbose, tol=tol,
x_squared_norms=x_squared_norms,
# Change seed to ensure variety
random_state=seed)
for seed in seeds)
# Get results with the lowest inertia
labels, inertia, centers, n_iters = zip(*results)
best = np.argmin(inertia)
best_labels = labels[best]
best_inertia = inertia[best]
best_centers = centers[best]
best_n_iter = n_iters[best]
if return_n_iter:
return best_centers, best_labels, best_inertia, best_n_iter
else:
return best_centers, best_labels, best_inertia
def calc_sampling_distribution(self):
x_squared_norms = row_norms(self.X, squared=True)
centers = _init_centroids(self.X, self.n_clusters, self.init, random_state=self.random_state,
x_squared_norms=x_squared_norms)
sens = sensitivity.kmeans_sensitivity(self.X, self.w, centers, max(np.log(self.n_clusters), 1))
self.p = sens / np.sum(sens)
def func(dat_matrix):
x_squared_norms = row_norms(dat_matrix, squared=True)
inertias = _labels_inertia(dat_matrix, x_squared_norms,
km.cluster_centers_)[1]
return inertias
def _k_init(X, n_clusters, random_state, n_local_trials=None):
"""Init n_clusters seeds according to k-means++
Parameters
-----------
X: array or sparse matrix, shape (n_samples, n_features)
The data to pick seeds for. To avoid memory copy, the input data
should be double precision (dtype=np.float64).
n_clusters: integer
The number of seeds to choose
x_squared_norms: array, shape (n_samples,)
Squared Euclidean norm of each data point.
random_state: numpy.RandomState
The generator used to initialize the centers.
n_local_trials: integer, optional
The number of seeding trials for each center (except the first),
of which the one reducing inertia the most is greedily chosen.
Set to None to make the number of trials depend logarithmically
on the number of seeds (2+log(k)); this is the default.
Notes
-----
Selects initial cluster centers for k-mean clustering in a smart way
to speed up convergence. see: Arthur, D. and Vassilvitskii, S.
"k-means++: the advantages of careful seeding". ACM-SIAM symposium
on Discrete algorithms. 2007
Version ported from http://www.stanford.edu/~darthur/kMeansppTest.zip,
which is the implementation used in the aforementioned paper.
"""
n_samples, n_features = X.shape
centers = np.empty((n_clusters, n_features), dtype=X.dtype)
# Modified !!!!
# assert x_squared_norms is not None, 'x_squared_norms None in _k_init'
x_squared_norms = row_norms(X, squared=True)
# Set the number of local seeding trials if none is given
if n_local_trials is None:
# This is what Arthur/Vassilvitskii tried, but did not report
# specific results for other than mentioning in the conclusion
# that it helped.
n_local_trials = 2 + int(np.log(n_clusters))
# Pick first center randomly
center_id = random_state.randint(n_samples)
if sp.issparse(X):
centers[0] = X[center_id].toarray()
else:
centers[0] = X[center_id]
# Initialize list of closest distances and calculate current potential
closest_dist_sq = euclidean_distances(centers[0, np.newaxis],
X,
Y_norm_squared=x_squared_norms,
squared=True)
current_pot = closest_dist_sq.sum()
# Pick the remaining n_clusters-1 points
for c in range(1, n_clusters):
# Choose center candidates by sampling with probability proportional
# to the squared distance to the closest existing center
rand_vals = random_state.random_sample(n_local_trials) * current_pot
candidate_ids = np.searchsorted(closest_dist_sq.cumsum(), rand_vals)
# Compute distances to center candidates
distance_to_candidates = euclidean_distances(
X[candidate_ids], X, Y_norm_squared=x_squared_norms, squared=True)
# Decide which candidate is the best
best_candidate = None
best_pot = None
best_dist_sq = None
for trial in range(n_local_trials):
# Compute potential when including center candidate
new_dist_sq = np.minimum(closest_dist_sq,
distance_to_candidates[trial])
new_pot = new_dist_sq.sum()
# Store result if it is the best local trial so far
if (best_candidate is None) or (new_pot < best_pot):
best_candidate = candidate_ids[trial]
best_pot = new_pot
best_dist_sq = new_dist_sq
# Permanently add best center candidate found in local tries
if sp.issparse(X):
centers[c] = X[best_candidate].toarray()
else:
centers[c] = X[best_candidate]
current_pot = best_pot
closest_dist_sq = best_dist_sq
return centers
def convert_sklearn_kmeans(scope, operator, container):
"""
Computation graph of distances to all centroids for a batch of examples.
Note that a centriod is just the center of a cluster. We use ``[]`` to
denote the dimension of a variable; for example, ``X[3, 2]`` means that
*X* is a *3-by-2* tensor. In addition, for a matrix *X*, $X'$ denotes its
transpose.
Symbols:
* *l*: # of examples.
* *n*: # of features per input example.
* *X*: input examples, l-by-n tensor.
* *C*: centroids, k-by-n tensor.
* :math:`C^2`: 2-norm of all centriod vectors, its shape is ``[k]``.
* *Y*: 2-norm of difference between examples and centroids,
*l-by-k* tensor. The value at i-th row and k-th column row,
``Y[i,k]``,is the distance from example *i* to centroid *k*.
* *L*: the id of the nearest centroid for each input example,
its shape is ``[l]``.
::
.------------------------------------------------------.
| |
| v
X [l, n] --> ReduceSumSquare -> X^2 [l] Gemm (alpha=-2, transB=1)
| | |- C [k, n]
| |
| v
`------> Add <-- -2XC' [l, k]
|
v
C^2 [k] --------> Add <----- Z [l, k]
|
v
L [l] <-- ArgMin <-- Y2 [l, k] --> Sqrt --> Y2 [l, k]
*scikit-learn* code:
::
X = data
Y = model.cluster_centers_
XX = row_norms(X, squared=True)
YY = row_norms(Y, squared=True)
distances = safe_sparse_dot(X, Y.T, dense_output=True)
distances *= -2
distances += XX[:, numpy.newaxis]
distances += YY[numpy.newaxis, :]
numpy.sqrt(distances, out=distances)
"""
X = operator.inputs[0]
out = operator.outputs
op = operator.raw_operator
opv = container.target_opset
C = op.cluster_centers_
input_name = X
dtype = guess_numpy_type(X.type)
if dtype != np.float64:
dtype = np.float32
if type(X.type) == Int64TensorType:
x_cast = OnnxCast(X, to=onnx_proto.TensorProto.FLOAT, op_version=opv)
input_name = x_cast
C2 = row_norms(C, squared=True).astype(dtype)
C = C.astype(dtype)
rs = OnnxReduceSumSquare(input_name, axes=[1], keepdims=1, op_version=opv)
N = X.type.shape[0]
if isinstance(N, int):
zeros = np.zeros((N, ), dtype=dtype)
else:
zeros = OnnxMul(rs, np.array([0], dtype=dtype),
op_version=opv)
z = OnnxAdd(rs, OnnxGemm(input_name, C, zeros, alpha=-2.,
transB=1, op_version=opv),
op_version=opv)
y2 = OnnxAdd(C2, z, op_version=opv)
ll = OnnxArgMin(y2, axis=1, keepdims=0, output_names=out[:1],
op_version=opv)
y2s = OnnxSqrt(y2, output_names=out[1:], op_version=opv)
ll.add_to(scope, container)
y2s.add_to(scope, container)
def normalize(self, X, norm='l2', axis=1, copy=True):
"""Normalize a dataset along any axis
Parameters
----------
X : array or scipy.sparse matrix with shape [n_samples, n_features]
The data to normalize, element by element.
scipy.sparse matrices should be in CSR format to avoid an
un-necessary copy.
norm : 'l1' or 'l2', optional ('l2' by default)
The norm to use to normalize each non zero sample (or each non-zero
feature if axis is 0).
axis : 0 or 1, optional (1 by default)
axis used to normalize the data along. If 1, independently normalize
each sample, otherwise (if 0) normalize each feature.
copy : boolean, optional, default is True
set to False to perform inplace row normalization and avoid a
copy (if the input is already a numpy array or a scipy.sparse
CSR matrix and if axis is 1).
See also
--------
:class:`sklearn.preprocessing.Normalizer` to perform normalization
using the ``Transformer`` API (e.g. as part of a preprocessing
:class:`sklearn.pipeline.Pipeline`)
"""
if norm not in ('l1', 'l2'):
raise ValueError("'%s' is not a supported norm" % norm)
if axis == 0:
sparse_format = 'csc'
elif axis == 1:
sparse_format = 'csr'
else:
raise ValueError("'%d' is not a supported axis" % axis)
X = check_arrays(X, sparse_format=sparse_format, copy=copy)[0]
warn_if_not_float(X, 'The normalize function')
if axis == 0:
X = X.T
if sparse.issparse(X):
if norm == 'l1':
inplace_csr_row_normalize_l1(X)
elif norm == 'l2':
inplace_csr_row_normalize_l2(X)
else:
if norm == 'l1':
norms = np.abs(X).sum(axis=1)
norms[norms == 0.0] = 1.0
elif norm == 'l2':
norms = row_norms(X)
norms[norms == 0.0] = 1.0
X /= norms[:, np.newaxis]
if axis == 0:
X = X.T
return X
def normalize(X, norm='l2', axis=1, copy=True):
"""Scale input vectors individually to unit norm (vector length).
Parameters
----------
X : array or scipy.sparse matrix with shape [n_samples, n_features]
The data to normalize, element by element.
scipy.sparse matrices should be in CSR format to avoid an
un-necessary copy.
norm : 'l1' or 'l2', optional ('l2' by default)
The norm to use to normalize each non zero sample (or each non-zero
feature if axis is 0).
axis : 0 or 1, optional (1 by default)
axis used to normalize the data along. If 1, independently normalize
each sample, otherwise (if 0) normalize each feature.
copy : boolean, optional, default True
set to False to perform inplace row normalization and avoid a
copy (if the input is already a numpy array or a scipy.sparse
CSR matrix and if axis is 1).
See also
--------
:class:`sklearn.preprocessing.Normalizer` to perform normalization
using the ``Transformer`` API (e.g. as part of a preprocessing
:class:`sklearn.pipeline.Pipeline`)
"""
if norm not in ('l1', 'l2'):
raise ValueError("'%s' is not a supported norm" % norm)
if axis == 0:
sparse_format = 'csc'
elif axis == 1:
sparse_format = 'csr'
else:
raise ValueError("'%d' is not a supported axis" % axis)
X = check_array(X, sparse_format, copy=copy)
warn_if_not_float(X, 'The normalize function')
if axis == 0:
X = X.T
if sparse.issparse(X):
X = check_array(X, accept_sparse=sparse_format, dtype=np.float64)
if norm == 'l1':
inplace_csr_row_normalize_l1(X)
elif norm == 'l2':
inplace_csr_row_normalize_l2(X)
else:
if norm == 'l1':
norms = np.abs(X).sum(axis=1)
norms[norms == 0.0] = 1.0
elif norm == 'l2':
norms = row_norms(X)
norms[norms == 0.0] = 1.0
X /= norms[:, np.newaxis]
if axis == 0:
X = X.T
return X
def fit(self,X,y,sample_weight=None):
if not isinstance(self.C, numbers.Number) or self.C < 0:
raise ValueError("Penalty term must be positive; got (C=%r)"
% self.C)
if not isinstance(self.max_iter, numbers.Number) or self.max_iter < 0:
raise ValueError("Maximum number of iteration must be positive;"
" got (max_iter=%r)" % self.max_iter)
if not isinstance(self.tol, numbers.Number) or self.tol < 0:
raise ValueError("Tolerance for stopping criteria must be "
"positive; got (tol=%r)" % self.tol)
solver = _check_solver(self.solver, self.penalty, self.dual)
if solver in ['newton-cg']:
_dtype = [np.float64, np.float32]
else:
_dtype = np.float64
X, y = check_X_y(X, y, accept_sparse='csr', dtype=_dtype, order="C",
accept_large_sparse=solver != 'liblinear')
self.X_fit_ = X
X_k = self._get_kernel(X)
check_classification_targets(y)
self.classes_ = np.unique(y)
n_samples, n_features = X_k.shape
multi_class = _check_multi_class(self.multi_class, solver,
len(self.classes_))
if solver == 'liblinear':
if effective_n_jobs(self.n_jobs) != 1:
warnings.warn("'n_jobs' > 1 does not have any effect when"
" 'solver' is set to 'liblinear'. Got 'n_jobs'"
" = {}.".format(effective_n_jobs(self.n_jobs)))
self.coef_, self.intercept_, n_iter_ = _fit_liblinear(
X_k, y, self.C, self.fit_intercept, self.intercept_scaling,
self.class_weight, self.penalty, self.dual, self.verbose,
self.max_iter, self.tol, self.random_state,
sample_weight=sample_weight)
self.n_iter_ = np.array([n_iter_])
return self
if solver in ['sag', 'saga']:
max_squared_sum = row_norms(X_k, squared=True).max()
else:
max_squared_sum = None
n_classes = len(self.classes_)
classes_ = self.classes_
if n_classes < 2:
raise ValueError("This solver needs samples of at least 2 classes"
" in the data, but the data contains only one"
" class: %r" % classes_[0])
if len(self.classes_) == 2:
n_classes = 1
classes_ = classes_[1:]
if self.warm_start:
warm_start_coef = getattr(self, 'coef_', None)
else:
warm_start_coef = None
if warm_start_coef is not None and self.fit_intercept:
warm_start_coef = np.append(warm_start_coef,
self.intercept_[:, np.newaxis],
axis=1)
self.coef_ = list()
self.intercept_ = np.zeros(n_classes)
# Hack so that we iterate only once for the multinomial case.
if multi_class == 'multinomial':
classes_ = [None]
warm_start_coef = [warm_start_coef]
if warm_start_coef is None:
warm_start_coef = [None] * n_classes
path_func = delayed(logistic_regression_path)
# The SAG solver releases the GIL so it's more efficient to use
# threads for this solver.
if solver in ['sag', 'saga']:
prefer = 'threads'
else:
prefer = 'processes'
fold_coefs_ = Parallel(n_jobs=self.n_jobs, verbose=self.verbose,
**_joblib_parallel_args(prefer=prefer))(
path_func(X_k, y, pos_class=class_, Cs=[self.C],
fit_intercept=self.fit_intercept, tol=self.tol,
verbose=self.verbose, solver=solver,
multi_class=multi_class, max_iter=self.max_iter,
class_weight=self.class_weight, check_input=False,
random_state=self.random_state, coef=warm_start_coef_,
penalty=self.penalty,
max_squared_sum=max_squared_sum,
sample_weight=sample_weight)
for class_, warm_start_coef_ in zip(classes_, warm_start_coef))
fold_coefs_, _, n_iter_ = zip(*fold_coefs_)
self.n_iter_ = np.asarray(n_iter_, dtype=np.int32)[:, 0]
if multi_class == 'multinomial':
self.coef_ = fold_coefs_[0][0]
else:
self.coef_ = np.asarray(fold_coefs_)
self.coef_ = self.coef_.reshape(n_classes, n_features +
int(self.fit_intercept))
if self.fit_intercept:
self.intercept_ = self.coef_[:, -1]
self.coef_ = self.coef_[:, :-1]
return self
def fit(self, X, y=None, sample_weight=None):
"""Compute k-means clustering.
Parameters
----------
X : array-like or sparse matrix, shape=(n_samples, n_features)
Training instances to cluster. It must be noted that the data
will be converted to C ordering, which will cause a memory
copy if the given data is not C-contiguous.
y : Ignored
Not used, present here for API consistency by convention.
sample_weight : array-like, shape (n_samples,), optional
The weights for each observation in X. If None, all observations
are assigned equal weight (default: None).
Returns
-------
self
Fitted estimator.
"""
random_state = check_random_state(self.random_state)
n_init = self.n_init
if n_init <= 0:
raise ValueError("Invalid number of initializations."
" n_init=%d must be bigger than zero." % n_init)
if self.max_iter <= 0:
raise ValueError(
'Number of iterations should be a positive number,'
' got %d instead' % self.max_iter)
# avoid forcing order when copy_x=False
order = "C" if self.copy_x else None
X = check_array(X,
accept_sparse='csr',
dtype=[np.float64, np.float32],
order=order,
copy=self.copy_x)
# verify that the number of samples given is larger than k
if _num_samples(X) < self.n_clusters:
raise ValueError("n_samples=%d should be >= n_clusters=%d" %
(_num_samples(X), self.n_clusters))
tol = _tolerance(X, self.tol)
# If the distances are precomputed every job will create a matrix of
# shape (n_clusters, n_samples). To stop KMeans from eating up memory
# we only activate this if the created matrix is guaranteed to be
# under 100MB. 12 million entries consume a little under 100MB if they
# are of type double.
precompute_distances = self.precompute_distances
if precompute_distances == 'auto':
n_samples = X.shape[0]
precompute_distances = (self.n_clusters * n_samples) < 12e6
elif isinstance(precompute_distances, bool):
pass
else:
raise ValueError(
"precompute_distances should be 'auto' or True/False"
", but a value of %r was passed" % precompute_distances)
# Validate init array
init = self.init
if hasattr(init, '__array__'):
init = check_array(init, dtype=X.dtype.type, copy=True)
_validate_center_shape(X, self.n_clusters, init)
if n_init != 1:
warnings.warn(
'Explicit initial center position passed: '
'performing only one init in k-means instead of n_init=%d'
% n_init,
RuntimeWarning,
stacklevel=2)
n_init = 1
# subtract of mean of x for more accurate distance computations
if not sp.issparse(X):
X_mean = X.mean(axis=0)
# The copy was already done above
X -= X_mean
if hasattr(init, '__array__'):
init -= X_mean
# precompute squared norms of data points
x_squared_norms = row_norms(X, squared=True)
best_labels, best_inertia, best_centers = None, None, None
algorithm = self.algorithm
if self.n_clusters == 1:
# elkan doesn't make sense for a single cluster, full will produce
# the right result.
algorithm = "full"
if algorithm == "auto":
algorithm = "full" if sp.issparse(X) else 'elkan'
if algorithm == "full":
kmeans_single = _fuzzykmeans_single_lloyd
elif algorithm == "elkan":
kmeans_single = _fuzzykmeans_single_elkan
else:
raise ValueError("Algorithm must be 'auto', 'full' or 'elkan', got"
" %s" % str(algorithm))
seeds = random_state.randint(np.iinfo(np.int32).max, size=n_init)
if effective_n_jobs(self.n_jobs) == 1:
# For a single thread, less memory is needed if we just store one
# set of the best results (as opposed to one set per run per
# thread).
for seed in seeds:
# run a k-means once
fuzzy_labels, labels, inertia, centers, n_iter_ = kmeans_single(
X,
self.m,
sample_weight,
self.n_clusters,
max_iter=self.max_iter,
init=init,
verbose=self.verbose,
precompute_distances=precompute_distances,
tol=tol,
x_squared_norms=x_squared_norms,
random_state=seed)
# determine if these results are the best so far
if best_inertia is None or inertia < best_inertia:
best_fuzzy_labels = fuzzy_labels.copy()
best_labels = labels.copy()
best_centers = centers.copy()
best_inertia = inertia
best_n_iter = n_iter_
else:
# parallelisation of k-means runs
results = Parallel(n_jobs=self.n_jobs, verbose=0)(
delayed(kmeans_single)(
X,
self.m,
sample_weight,
self.n_clusters,
max_iter=self.max_iter,
init=init,
verbose=self.verbose,
tol=tol,
precompute_distances=precompute_distances,
x_squared_norms=x_squared_norms,
# Change seed to ensure variety
random_state=seed) for seed in seeds)
# Get results with the lowest inertia
fuzzy_labels, labels, inertia, centers, n_iters = zip(*results)
best = np.argmin(inertia)
best_fuzzy_labels = fuzzy_labels[best]
best_labels = labels[best]
best_inertia = inertia[best]
best_centers = centers[best]
best_n_iter = n_iters[best]
if not sp.issparse(X):
if not self.copy_x:
X += X_mean
best_centers += X_mean
distinct_clusters = len(set(best_labels))
if distinct_clusters < self.n_clusters:
warnings.warn(
"Number of distinct clusters ({}) found smaller than "
"n_clusters ({}). Possibly due to duplicate points "
"in X.".format(distinct_clusters, self.n_clusters),
ConvergenceWarning,
stacklevel=2)
self.cluster_centers_ = best_centers
self.fuzzy_labels_ = best_fuzzy_labels
self.labels_ = best_labels
self.inertia_ = best_inertia
self.n_iter_ = best_n_iter
return self
assert indices.shape[0] == n_clusters
assert (indices >= 0).all()
assert (indices <= data.shape[0]).all()
# Check for the correct number of seeds and that they are bound by the data
assert centers.shape[0] == n_clusters
assert (centers.max(axis=0) <= data.max(axis=0)).all()
assert (centers.min(axis=0) >= data.min(axis=0)).all()
# Check that indices correspond to reported centers
# Use X for comparison rather than data, test still works against centers
# calculated with sparse data.
assert_allclose(X[indices].astype(dtype), centers)
@pytest.mark.parametrize("x_squared_norms", [row_norms(X, squared=True), None])
def test_kmeans_plusplus_norms(x_squared_norms):
# Check that defining x_squared_norms returns the same as default=None.
centers, indices = kmeans_plusplus(X,
n_clusters,
x_squared_norms=x_squared_norms)
assert_allclose(X[indices], centers)
def test_kmeans_plusplus_dataorder():
# Check that memory layout does not effect result
centers_c, _ = kmeans_plusplus(X, n_clusters, random_state=0)
X_fortran = np.asfortranarray(X)
def _estimate_log_gaussian_prob(x, means, precisions_chol, covariance_type):
"""Estimate the log Gaussian probability.
Parameters
----------
x : array-like or csr_matrix, shape (n_samples, n_features)
means : array-like, shape (n_components, n_features)
precisions_chol : array-like,
Cholesky decompositions of the precision matrices.
'full' : shape of (n_components, n_features, n_features)
'tied' : shape of (n_features, n_features)
'diag' : shape of (n_components, n_features)
'spherical' : shape of (n_components,)
covariance_type : {'full', 'tied', 'diag', 'spherical'}
Returns
-------
log_prob : array, shape (n_samples, n_components)
"""
n_samples, n_features = x.shape
n_components, _ = means.shape
# det(precision_chol) is half of det(precision)
log_det = _compute_log_det_cholesky(precisions_chol, covariance_type,
n_features)
if covariance_type == 'full':
log_prob = np.empty((n_samples, n_components))
for k, (mu, prec_chol) in enumerate(zip(means, precisions_chol)):
if issparse(x):
y = x.dot(prec_chol) - np.dot(mu, prec_chol)
else:
y = np.matmul(x, prec_chol) - np.dot(mu, prec_chol)
log_prob[:, k] = np.sum(np.square(y), axis=1)
elif covariance_type == 'tied':
log_prob = np.empty((n_samples, n_components))
for k, mu in enumerate(means):
if issparse(x):
y = x.dot(precisions_chol) - np.dot(mu, precisions_chol)
else:
y = np.dot(x, precisions_chol) - np.dot(mu, precisions_chol)
log_prob[:, k] = np.sum(np.square(y), axis=1)
elif covariance_type == 'diag':
precisions = precisions_chol**2
if issparse(x):
log_prob = (np.sum((means**2 * precisions), 1) - 2. *
(x * (means * precisions).T) +
x.multiply(x).dot(precisions.T))
else:
log_prob = (np.sum((means**2 * precisions), 1) -
2. * np.dot(x, (means * precisions).T) +
np.dot(x**2, precisions.T))
elif covariance_type == 'spherical':
precisions = precisions_chol**2
if issparse(x):
log_prob = (np.sum(means**2, 1) * precisions - 2 *
(x * (means.T * precisions)) +
np.outer(row_norms(x, squared=True), precisions))
else:
log_prob = (np.sum(means**2, 1) * precisions -
2 * np.dot(x, means.T * precisions) +
np.outer(row_norms(x, squared=True), precisions))
else: # pragma: no cover
raise ValueError()
return -.5 * (n_features * np.log(2 * np.pi) + log_prob) + log_det
def test_minibatch_update_consistency():
# Check that dense and sparse minibatch update give the same results
rng = np.random.RandomState(42)
old_centers = centers + rng.normal(size=centers.shape)
new_centers = old_centers.copy()
new_centers_csr = old_centers.copy()
counts = np.zeros(new_centers.shape[0], dtype=np.int32)
counts_csr = np.zeros(new_centers.shape[0], dtype=np.int32)
x_squared_norms = (X ** 2).sum(axis=1)
x_squared_norms_csr = row_norms(X_csr, squared=True)
buffer = np.zeros(centers.shape[1], dtype=np.double)
buffer_csr = np.zeros(centers.shape[1], dtype=np.double)
# extract a small minibatch
X_mb = X[:10]
X_mb_csr = X_csr[:10]
x_mb_squared_norms = x_squared_norms[:10]
x_mb_squared_norms_csr = x_squared_norms_csr[:10]
# step 1: compute the dense minibatch update
old_inertia, incremental_diff = _mini_batch_step(
X_mb, x_mb_squared_norms, new_centers, counts,
buffer, 1, None, random_reassign=False)
assert_greater(old_inertia, 0.0)
# compute the new inertia on the same batch to check that it decreased
labels, new_inertia = _labels_inertia(
X_mb, x_mb_squared_norms, new_centers)
assert_greater(new_inertia, 0.0)
assert_less(new_inertia, old_inertia)
# check that the incremental difference computation is matching the
# final observed value
effective_diff = np.sum((new_centers - old_centers) ** 2)
assert_almost_equal(incremental_diff, effective_diff)
# step 2: compute the sparse minibatch update
old_inertia_csr, incremental_diff_csr = _mini_batch_step(
X_mb_csr, x_mb_squared_norms_csr, new_centers_csr, counts_csr,
buffer_csr, 1, None, random_reassign=False)
assert_greater(old_inertia_csr, 0.0)
# compute the new inertia on the same batch to check that it decreased
labels_csr, new_inertia_csr = _labels_inertia(
X_mb_csr, x_mb_squared_norms_csr, new_centers_csr)
assert_greater(new_inertia_csr, 0.0)
assert_less(new_inertia_csr, old_inertia_csr)
# check that the incremental difference computation is matching the
# final observed value
effective_diff = np.sum((new_centers_csr - old_centers) ** 2)
assert_almost_equal(incremental_diff_csr, effective_diff)
# step 3: check that sparse and dense updates lead to the same results
assert_array_equal(labels, labels_csr)
assert_array_almost_equal(new_centers, new_centers_csr)
assert_almost_equal(incremental_diff, incremental_diff_csr)
assert_almost_equal(old_inertia, old_inertia_csr)
assert_almost_equal(new_inertia, new_inertia_csr)
def _init_w(self, V, X):
"""
Initialize the topics W.
If self.init='k-means++', we use the init method of
sklearn.cluster.KMeans.
If self.init='random', topics are initialized with a Gamma
distribution.
If self.init='k-means', topics are initialized with a KMeans on the
n-grams counts.
"""
if self.init == 'k-means++':
if LooseVersion(sklearn_version) < LooseVersion('0.24'):
W = _k_init(V,
self.n_components,
x_squared_norms=row_norms(V, squared=True),
random_state=self.random_state,
n_local_trials=None) + .1
else:
W, _ = kmeans_plusplus(V,
self.n_components,
x_squared_norms=row_norms(V,
squared=True),
random_state=self.random_state,
n_local_trials=None)
W = W + .1 # To avoid restricting topics to few n-grams only
elif self.init == 'random':
W = self.random_state.gamma(shape=self.gamma_shape_prior,
scale=self.gamma_scale_prior,
size=(self.n_components, self.n_vocab))
elif self.init == 'k-means':
prototypes = get_kmeans_prototypes(X,
self.n_components,
analyzer=self.analyzer,
random_state=self.random_state)
W = self.ngrams_count_.transform(prototypes).A + .1
if self.add_words:
W2 = self.word_count_.transform(prototypes).A + .1
W = np.hstack((W, W2))
# if k-means doesn't find the exact number of prototypes
if W.shape[0] < self.n_components:
if LooseVersion(sklearn_version) < LooseVersion('0.24'):
W2 = _k_init(V,
self.n_components - W.shape[0],
x_squared_norms=row_norms(V, squared=True),
random_state=self.random_state,
n_local_trials=None) + .1
else:
W2, _ = kmeans_plusplus(V,
self.n_components - W.shape[0],
x_squared_norms=row_norms(
V, squared=True),
random_state=self.random_state,
n_local_trials=None)
W2 = W2 + .1
W = np.concatenate((W, W2), axis=0)
else:
raise AttributeError('Initialization method %s does not exist.' %
self.init)
W /= W.sum(axis=1, keepdims=True)
A = np.ones((self.n_components, self.n_vocab)) * 1e-10
B = A.copy()
return W, A, B
