print " number of components to retain: %d" % num_components
P = eigv.astype('float32') * np.sqrt(1.0/(eigs + bias)) # PCA whitening
P = P[:, -num_components:] # truncate transformation matrix
tock()
# kmeans
print "Learn kmeans"
k = settings['num_means']
print " whiten patches"
patches_subset_whitened = np.dot(patches_subset_centered, P)
print " run kmeans"
centroids = kmeans.spherical_kmeans(patches_subset_whitened, k, num_iterations=20, batch_size=10000)
tock()
# feature extraction and summarisation
print "Feature extraction and summarisation"
threshold = settings['threshold']
PC = np.dot(P, centroids.T)
# def summarise_features(features):
# features = features.reshape(features.shape[0], -1, features.shape[3]) # merge time and frequency axes
# return np.hstack([features.mean(1), features.std(1), features.min(1), features.max(1)]) # summarize over time axis
# # return features.mean(1)
# # return features.max(1)
print(" number of components to retain: %d" % num_components)
P = eigv.astype('float32') * np.sqrt(1.0 / (eigs + bias)) # PCA whitening
P = P[:, -num_components:] # truncate transformation matrix
tock()
# kmeans
print("Learn kmeans")
k = settings['num_means']
print(" whiten patches")
patches_subset_whitened = np.dot(patches_subset_centered, P)
print(" run kmeans")
centroids = kmeans.spherical_kmeans(patches_subset_whitened,
k,
num_iterations=20,
batch_size=10000)
tock()
# feature extraction and summarisation
print("Feature extraction and summarisation")
threshold = settings['threshold']
PC = np.dot(P, centroids.T)
# def summarise_features(features):
# features = features.reshape(features.shape[0], -1, features.shape[3]) # merge time and frequency axes
# return np.hstack([features.mean(1), features.std(1), features.min(1), features.max(1)]) # summarize over time axis
# # return features.mean(1)
# # return features.max(1)