def test_lmvnpdftied_consistent_with_lmvnpdffull(self):
nstates = 4
ndim = 20
nobs = 200
mu = np.random.randint(10) * np.random.rand(nstates, ndim)
tiedcv = _generate_random_spd_matrix(ndim)
obs = np.random.randint(10) * np.random.rand(nobs, ndim)
cv = np.tile(tiedcv, (nstates, 1, 1))
reference = gmm.lmvnpdf(obs, mu, cv, "full")
lpr = gmm.lmvnpdf(obs, mu, tiedcv, "tied")
assert_array_almost_equal(lpr, reference)
def clustering(self,data,mode="kmeans",itermax=100,threshold=1.0e-2):
"""
clustering data
input
data : 2d-array, data to be clusterd
mode : strings, clustering method, "kmeans" or "gmm"
itermax : int, max cycle of iteration in gmm train
threshold : float, gmm iteration is converged if changes are below it
return
codes : array, codes[i] is the corresponding cluster of data[i]
score : array, in kmeans, score is the distance, in gmm probability
iter_log : array, log of iteration (only gmm)
aic : float, AIC of the model (only gmm)
bic : float, BIC of the model (only gmm)
"""
print "running %s mode" % mode
if mode == "kmeans": # run kmeans mode
self.sort_order = False # ascending order
# get centroids
centroids, var = kmeans(data,self.nclust)
# assign data to the nearest cluster
codes, distances = vq(data,centroids)
return codes,distances
elif mode == "gmm": # run gmm mode
self.sort_order = True # descending order
# construct a GMM instance
self.gm = gmm.GMM(self.nclust,cvtype="full")
# train gmm
iter_log = self.gm.fit(data,itermax,thresh=threshold)
if len(iter_log) == itermax :
stderr.write("warning!! EM step not converged\n")
# assign data to the nearest cluster
logprobs, codes = self.gm.decode(data)
# calc probability that each datum belongs to corresponding cluster
lpr = gmm.lmvnpdf(data,self.gm.means, \
self.gm._covars,self.gm._cvtype)
probs = np.array([lpr[i,codes[i]] for i in range(len(lpr))])
# calc AIC and BIC for model evaluation
aic = AIC(iter_log[-1],data,self.nclust)
bic = BIC(iter_log[-1],data,self.nclust)
return codes,probs,iter_log,aic,bic
else:
raise ClusteringModeError(mode)
def _test_lmvnpdfspherical(self, ndim, nstates, nobs=100):
mu = np.random.randint(10) * np.random.rand(nstates, ndim)
spherecv = np.random.rand(nstates, 1) ** 2 + 1
obs = np.random.randint(10) * np.random.rand(nobs, ndim)
cv = np.tile(spherecv, (ndim, 1))
reference = self._slow_lmvnpdfdiag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, spherecv, "spherical")
assert_array_almost_equal(lpr, reference)
def _test_lmvnpdftied_with_diagonal_covariance(self, ndim, nstates, nobs=100):
mu = np.random.randint(10) * np.random.rand(nstates, ndim)
tiedcv = (np.random.rand(ndim) + 1.0) ** 2
obs = np.random.randint(10) * np.random.rand(nobs, ndim)
cv = np.tile(tiedcv, (nstates, 1))
reference = self._slow_lmvnpdfdiag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, np.diag(tiedcv), "tied")
assert_array_almost_equal(lpr, reference)
def _test_lmvnpdffull_with_diagonal_covariance(self, ndim, nstates, nobs=100):
mu = np.random.randint(10) * np.random.rand(nstates, ndim)
cv = (np.random.rand(nstates, ndim) + 1.0) ** 2
obs = np.random.randint(10) * np.random.rand(nobs, ndim)
fullcv = np.array([np.diag(x) for x in cv])
reference = self._slow_lmvnpdfdiag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, fullcv, "full")
assert_array_almost_equal(lpr, reference)
def test_lmvnpdf_spherical():
n_dim, n_states, n_obs = 2, 3, 10
mu = np.random.randint(10) * np.random.rand(n_states, n_dim)
spherecv = np.random.rand(n_states, 1) ** 2 + 1
obs = np.random.randint(10) * np.random.rand(n_obs, n_dim)
cv = np.tile(spherecv, (n_dim, 1))
reference = _naive_lmvnpdf_diag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, spherecv, 'spherical')
assert_array_almost_equal(lpr, reference)
def _test_lmvnpdfdiag(self, ndim, nstates, nobs=100):
# test the slow and naive implementation of lmvnpdf and
# compare it to the vectorized version (gmm.lmvnpdf) to test
# for correctness
mu = np.random.randint(10) * np.random.rand(nstates, ndim)
cv = (np.random.rand(nstates, ndim) + 1.0) ** 2
obs = np.random.randint(10) * np.random.rand(nobs, ndim)
reference = self._slow_lmvnpdfdiag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, cv, "diag")
assert_array_almost_equal(lpr, reference)
def test_lmvnpdf_full():
n_dim, n_states, n_obs = 2, 3, 10
mu = np.random.randint(10) * np.random.rand(n_states, n_dim)
cv = (np.random.rand(n_states, n_dim) + 1.0) ** 2
obs = np.random.randint(10) * np.random.rand(n_obs, n_dim)
fullcv = np.array([np.diag(x) for x in cv])
reference = _naive_lmvnpdf_diag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, fullcv, 'full')
assert_array_almost_equal(lpr, reference)
def test_lmvnpdf_diag():
"""
test a slow and naive implementation of lmvnpdf and
compare it to the vectorized version (gmm.lmvnpdf) to test
for correctness
"""
n_dim, n_states, n_obs = 2, 3, 10
mu = np.random.randint(10) * np.random.rand(n_states, n_dim)
cv = (np.random.rand(n_states, n_dim) + 1.0) ** 2
obs = np.random.randint(10) * np.random.rand(n_obs, n_dim)
ref = _naive_lmvnpdf_diag(obs, mu, cv)
lpr = gmm.lmvnpdf(obs, mu, cv, 'diag')
assert_array_almost_equal(lpr, ref)
