def generate_gmm(image_features_list):
concatenated_descriptors = np.concatenate(image_features_list)
gmm_filename = 'gmm.pkl'
N, D = concatenated_descriptors.shape
K = 128
print("The sizes are {0} and {1}".format(N, D))
if (N > 3000000):
batch_size = 3000000
else:
batch_size = N
ggmm.init(batch_size * D)
gmm = ggmm.GMM(K, D)
thresh = 1e-3 # convergence threshold
n_iter = 500 # maximum number of EM iterations
init_params = 'wmc' # initialize weights, means, and covariances
# train GMM
converged = gmm.fit(concatenated_descriptors[:batch_size],
thresh,
n_iter,
init_params=init_params)
print("GMM converged? ... {0}".format(converged))
pickle.dump((gmm.get_weights(), gmm.get_means(), gmm.get_covars()),
open(gmm_filename, 'wb'))
return gmm
def EMEngine(X, thresh_, n_iter_, timer, K_, init_='wmc'):
N, D = X.shape
K = K_
start = time.time()
# train gmm
ggmm.init()
gmm = ggmm.GMM(K,D)
gmm.fit(X, thresh = thresh_, n_iter = n_iter_, init_params=init_, verbose=False, iterations=timer)
# retrieve parameters
weights = gmm.get_weights()
means = gmm.get_means()
covars = gmm.get_covars()
posteriors = gmm.compute_posteriors(X)
end = time.time() - start
if timer:
print 'Max Iterations:',n_iter_
print 'Convergence Threshold:',thresh_
print 'Exectution Time:', end
print 'Number of Gaussians:', K
print '==========================='
return (means, weights, covars)
def generate_gmm(image_features_list):
concatenated_descriptors = np.concatenate(image_features_list)
gmm_filename = 'gmm.pkl'
N, D = concatenated_descriptors.shape
K=128
print("The sizes are {0} and {1}".format(N,D))
if(N > 3000000):
batch_size = 3000000
else:
batch_size = N
ggmm.init(batch_size * D)
gmm = ggmm.GMM(K,D)
thresh = 1e-3 # convergence threshold
n_iter = 500 # maximum number of EM iterations
init_params = 'wmc' # initialize weights, means, and covariances
# train GMM
converged = gmm.fit(concatenated_descriptors[:batch_size], thresh, n_iter, init_params=init_params)
print("GMM converged? ... {0}".format(converged))
pickle.dump((gmm.get_weights(), gmm.get_means(), gmm.get_covars()), open(gmm_filename,'wb'))
return gmm
def load_gmm():
wmc = pickle.load(open("gmm.pkl", "rb"))
ggmm.init(wmc[0].shape[0] * 64)
gmm = ggmm.GMM(wmc[0].shape[0], 64)
gmm.set_weights(wmc[0])
gmm.set_means(wmc[1])
gmm.set_covars(wmc[2])
print("Loaded GMM Info!")
return gmm
def load_gmm():
wmc = pickle.load(open("gmm.pkl","rb"))
ggmm.init(wmc[0].shape[0] * 64)
gmm = ggmm.GMM(wmc[0].shape[0],64)
gmm.set_weights(wmc[0])
gmm.set_means(wmc[1])
gmm.set_covars(wmc[2])
print("Loaded GMM Info!")
return gmm
def setup():
ggmm.init() # activates cublas