def _init_params(self, x):
"""Initialize GMM parameters by K-means.
:param x: (n_samples, n_features) features.
:param n_components: the number of components.
:return: Initialized GMM parameters:
pi: (n_components,) mixing coefficients
mean: (n_components, n_features) means
cov: (n_components, n_features, n_features) covariances
"""
n_samples, n_features = x.shape
k_means = KMeans(self.n_components)
assigned_indices = k_means.fit_predict(x)
mean_init = k_means.centers
pi_init = np.zeros(self.n_components)
cov_init = np.zeros((self.n_components, n_features, n_features))
for k in range(self.n_components):
cond = assigned_indices == k
d_k = x[cond] - mean_init[k]
pi_init[k] = np.sum(cond) / n_samples
cov_init[k] = np.dot(d_k.T, d_k) / np.sum(cond)
return pi_init, mean_init, cov_init
def main():
iris_url = 'http://archive.ics.uci.edu/ml/machine-learning-databases/' \
'iris/iris.data'
x_col_names = ['sepal_length', 'sepal_width', 'petal_length', 'petal_width']
y_col_name = 'label'
iris_df = pd.read_csv(iris_url, names=x_col_names + [y_col_name])
x_data = np.array(iris_df[x_col_names])
# perform k-means clustering
k_means = KMeans(n_centers=3, init='k-means++',
random_state=np.random.RandomState(0))
y_pred = k_means.fit_predict(x_data)
centers = k_means.centers
# plot
plot_colors = ['r', 'g', 'b']
for ci in range(k_means.n_centers):
plt.scatter(x_data[y_pred == ci, 0], x_data[y_pred == ci, 1],
c=plot_colors[ci])
plt.scatter(centers[:, 0], centers[:, 1], c='y', label='centers')
plt.title('k-means example on the iris dataset')
plt.xlabel(x_col_names[0])
plt.ylabel(x_col_names[1])
plt.legend()
plt.show()
def _init_params(self, x):
"""Initialize GMM parameters by K-means.
:param x: (n_samples, n_features) features.
:param n_components: the number of components.
:return: Initialized GMM parameters:
pi: (n_components,) mixing coefficients
mean: (n_components, n_features) means
cov: (n_components, n_features, n_features) covariances
"""
n_samples, n_features = x.shape
k_means = KMeans(self.n_components)
assigned_indices = k_means.fit_predict(x)
mean_init = k_means.centers
pi_init = np.zeros(self.n_components)
cov_init = np.zeros((self.n_components, n_features, n_features))
for k in range(self.n_components):
cond = assigned_indices == k
d_k = x[cond] - mean_init[k]
pi_init[k] = np.sum(cond) / n_samples
cov_init[k] = np.dot(d_k.T, d_k) / np.sum(cond)
return pi_init, mean_init, cov_init
