def _get_pairwise_affinities(self, X):
"""Computes pairwise affinities."""
affines = np.zeros((self.n_samples, self.n_samples), dtype=np.float32)
target_entropy = np.log(self.perplexity)
distances = l2_distance(X)
for i in range(self.n_samples):
affines[i, :] = self._binary_search(distances[i], target_entropy)
# Fill diagonal with near zero value
np.fill_diagonal(affines, 1.0e-12)
affines = affines.clip(min=1e-100)
affines = (affines + affines.T) / (2 * self.n_samples)
return affines
def _get_pairwise_affinities(self, X):
"""Computes pairwise affinities."""
affines = np.zeros((self.n_samples, self.n_samples), dtype=np.float32)
target_entropy = np.log(self.perplexity)
distances = l2_distance(X)
for i in range(self.n_samples):
affines[i, :] = self._binary_search(distances[i], target_entropy)
# Fill diagonal with near zero value
np.fill_diagonal(affines, 1.e-12)
affines = affines.clip(min=1e-100)
affines = (affines + affines.T) / (2 * self.n_samples)
return affines
def fit_transform(self, X, y=None):
self._setup_input(X, y)
Y = np.random.randn(self.n_samples, self.n_components)
velocity = np.zeros_like(Y)
gains = np.ones_like(Y)
P = self._get_pairwise_affinities(X)
iter_num = 0
while iter_num < self.max_iter:
iter_num += 1
D = l2_distance(Y)
Q = self._q_distribution(D)
# Normalizer q distribution
Q_n = Q / np.sum(Q)
# Early exaggeration & momentum
pmul = 4.0 if iter_num < 100 else 1.0
momentum = 0.5 if iter_num < 20 else 0.8
# Perform gradient step
grads = np.zeros(Y.shape)
for i in range(self.n_samples):
grad = 4 * np.dot((pmul * P[i] - Q_n[i]) * Q[i], Y[i] - Y)
grads[i] = grad
gains = (gains + 0.2) * ((grads > 0) !=
(velocity > 0)) + (gains * 0.8) * (
(grads > 0) == (velocity > 0))
gains = gains.clip(min=self.min_gain)
velocity = momentum * velocity - self.lr * (gains * grads)
Y += velocity
Y = Y - np.mean(Y, 0)
error = np.sum(P * np.log(P / Q_n))
logging.info("Iteration %s, error %s" % (iter_num, error))
return Y
def fit_transform(self, X, y=None):
self._setup_input(X, y)
Y = np.random.randn(self.n_samples, self.n_components)
velocity = np.zeros_like(Y)
gains = np.ones_like(Y)
P = self._get_pairwise_affinities(X)
iter_num = 0
while iter_num < self.max_iter:
iter_num += 1
D = l2_distance(Y)
Q = self._q_distribution(D)
# Normalizer q distribution
Q_n = Q / np.sum(Q)
# Early exaggeration & momentum
pmul = 4.0 if iter_num < 100 else 1.0
momentum = 0.5 if iter_num < 20 else 0.8
# Perform gradient step
grads = np.zeros(Y.shape)
for i in range(self.n_samples):
grad = 4 * np.dot((pmul * P[i] - Q_n[i]) * Q[i], Y[i] - Y)
grads[i] = grad
gains = (gains + 0.2) * ((grads > 0) != (velocity > 0)) + (gains * 0.8) * ((grads > 0) == (velocity > 0))
gains = gains.clip(min=self.min_gain)
velocity = momentum * velocity - self.lr * (gains * grads)
Y += velocity
Y = Y - np.mean(Y, 0)
error = np.sum(P * np.log(P / Q_n))
logging.info("Iteration %s, error %s" % (iter_num, error))
return Y