def classify_speaker(htk_feats, labels, ubm_params):
# For some bizzare reason this breaks...???
#means = adapt_UBM(ubm_params, htk_feats)
gmm = GMM(M, D, means=np.array(ubm_params['means']), covars=np.array(ubm_params['covars']),\
weights=np.array(ubm_params['weights']), cvtype='diag')
gmm.train(htk_feats, max_em_iters=1)
means = gmm.components.means
means = means.reshape(1, M*D)
result = speaker_svm.classify(means, labels)
return result
def classify_speaker(htk_feats, labels, ubm_params):
# For some bizzare reason this breaks...???
#means = adapt_UBM(ubm_params, htk_feats)
gmm = GMM(M, D, means=np.array(ubm_params['means']), covars=np.array(ubm_params['covars']),\
weights=np.array(ubm_params['weights']), cvtype='diag')
gmm.train(htk_feats, max_em_iters=1)
means = gmm.components.means
means = means.reshape(1, M * D)
result = speaker_svm.classify(means, labels)
return result
def test_training_once(self):
gmm0 = GMM(self.M, self.D, cvtype='diag')
likelihood0 = gmm0.train(self.X)
means0 = gmm0.components.means.flatten()
covars0 = gmm0.components.covars.flatten()
gmm1 = GMM(self.M, self.D, cvtype='diag')
likelihood1 = gmm1.train(self.X)
means1 = gmm1.components.means.flatten()
covars1 = gmm1.components.covars.flatten()
self.assertAlmostEqual(likelihood0, likelihood1, delta=1)
for a,b in zip(means0, means1): self.assertAlmostEqual(a,b,places=2)
for a,b in zip(covars0, covars1): self.assertAlmostEqual(a,b,places=2)
def test_prediction_full(self):
print "test prediction full"
gmm0 = GMM(self.M, self.D, cvtype='full')
likelihood0 = gmm0.train(self.X)
Y0 = gmm0.predict(self.X)
gmm1 = GMM(self.M, self.D, cvtype='full')
likelihood1 = gmm1.train(self.X)
Y1 = gmm1.predict(self.X)
for a,b in zip(Y0, Y1): self.assertAlmostEqual(a,b, places=3)
self.assertTrue(len(set(Y0)) > 1)
def adapt_UBM(ubm_params, data):
updated_means = np.array(ubm_params['means'], dtype=np.float32)
for it in range(1): # adaptation loop
gmm = GMM(M, D, means=updated_means, covars=np.array(ubm_params['covars']),\
weights=np.array(ubm_params['weights']), cvtype='diag')
gmm.train(data, max_em_iters=1)
# ==== Actual adaptation code =====
new_means = gmm.components.means
new_weights = gmm.components.weights
T = data.shape[0]
updated_means = adapt_means(ubm_params['means'], ubm_params['covars'],\
ubm_params['weights'], new_means, new_weights, T).flatten('C')
# ==== Demo adaptation code =====
#updated_means = gmm.components.means
return updated_means
def adapt_UBM(ubm_params, data):
updated_means = np.array(ubm_params['means'], dtype=np.float32)
for it in range(1): # adaptation loop
gmm = GMM(M, D, means=updated_means, covars=np.array(ubm_params['covars']),\
weights=np.array(ubm_params['weights']), cvtype='diag')
gmm.train(data, max_em_iters=1)
# ==== Actual adaptation code =====
new_means = gmm.components.means
new_weights = gmm.components.weights
T = data.shape[0]
updated_means = adapt_means(ubm_params['means'], ubm_params['covars'],\
ubm_params['weights'], new_means, new_weights, T).flatten('C')
# ==== Demo adaptation code =====
#updated_means = gmm.components.means
return updated_means
def train_UBM(M, D, all_feats, total_num_feats):
ubm_feats = all_feats[0]
for d in all_feats[1:]:
ubm_feats = np.concatenate((ubm_feats, d))
print "total feats: ", total_num_feats
ubm_feats = ubm_feats.reshape(total_num_feats, D)
# === TRAIN UBM ===
ubm = GMM(M, D, cvtype='diag')
ubm.train(ubm_feats, max_em_iters=5)
ubm_params = {}
ubm_params['means'] = ubm.components.means
ubm_params['covars'] = ubm.components.covars
ubm_params['weights'] = ubm.components.weights
return ubm_params
def train_UBM(M, D, all_feats, total_num_feats):
ubm_feats = all_feats[0]
for d in all_feats[1:]:
ubm_feats = np.concatenate((ubm_feats, d))
print "total feats: ", total_num_feats
ubm_feats = ubm_feats.reshape(total_num_feats, D)
# === TRAIN UBM ===
ubm = GMM(M, D, cvtype='diag')
ubm.train(ubm_feats, max_em_iters=5)
ubm_params = {}
ubm_params['means'] = ubm.components.means
ubm_params['covars'] = ubm.components.covars
ubm_params['weights'] = ubm.components.weights
return ubm_params
def test_prediction_once(self):
gmm0 = GMM(self.M, self.D, cvtype='diag')
likelihood0 = gmm0.train(self.X)
Y0 = gmm0.predict(self.X)
gmm1 = GMM(self.M, self.D, cvtype='diag')
likelihood1 = gmm1.train(self.X)
Y1 = gmm1.predict(self.X)
for a, b in zip(Y0, Y1):
self.assertAlmostEqual(a, b)
self.assertTrue(len(set(Y0)) > 1)
def test_training_repeat(self):
gmm0 = GMM(self.M, self.D, cvtype='diag')
likelihood0 = gmm0.train(self.X)
likelihood0 = gmm0.train(self.X)
likelihood0 = gmm0.train(self.X)
likelihood0 = gmm0.train(self.X)
likelihood0 = gmm0.train(self.X)
means0 = gmm0.components.means.flatten()
covars0 = gmm0.components.covars.flatten()
gmm1 = GMM(self.M, self.D, cvtype='diag')
likelihood1 = gmm1.train(self.X)
likelihood1 = gmm1.train(self.X)
likelihood1 = gmm1.train(self.X)
likelihood1 = gmm1.train(self.X)
likelihood1 = gmm1.train(self.X)
means1 = gmm1.components.means.flatten()
covars1 = gmm1.components.covars.flatten()
self.assertAlmostEqual(likelihood0, likelihood1, delta=1)
for a, b in zip(means0, means1):
self.assertAlmostEqual(a, b, places=2)
for a, b in zip(covars0, covars1):
self.assertAlmostEqual(a, b, places=2)
def test_getter_methods(self):
gmm0 = GMM(self.M, self.D, cvtype="diag")
likelihood0 = gmm0.train(self.X)
all_weights0 = gmm0.get_all_component_weights()
all_means0 = gmm0.get_all_component_means()
all_diag_covars0 = gmm0.get_all_component_diag_covariance()
gmm1 = GMM(self.M, self.D, cvtype="diag")
likelihood1 = gmm1.train(self.X)
all_weights1 = gmm1.get_all_component_weights()
all_means1 = gmm1.get_all_component_means()
all_diag_covars1 = gmm1.get_all_component_diag_covariance()
self.assertAlmostEqual(likelihood0, likelihood1, places=3)
for (a, b) in zip(all_weights0, all_weights1):
self.assertAlmostEqual(a, b, places=3)
for m in range(gmm0.M):
for a, b in zip(all_means0[m], all_means1[m]):
self.assertAlmostEqual(a, b, places=3)
for a, b in zip(all_diag_covars0, all_diag_covars1):
self.assertAlmostEqual(a, b, places=3)
def test_prediction_full(self):
gmm0 = GMM(self.M, self.D, cvtype='full')
likelihood0 = gmm0.train(self.X)
Y0 = gmm0.predict(self.X)
gmm1 = GMM(self.M, self.D, cvtype='full')
likelihood1 = gmm1.train(self.X)
Y1 = gmm1.predict(self.X)
for a,b in zip(Y0, Y1): self.assertAlmostEqual(a,b)
self.assertTrue(len(set(Y0)) > 1)
def test_prediction_full(self):
print "test prediction full"
gmm0 = GMM(self.M, self.D, cvtype="full")
likelihood0 = gmm0.train(self.X)
Y0 = gmm0.predict(self.X)
gmm1 = GMM(self.M, self.D, cvtype="full")
likelihood1 = gmm1.train(self.X)
Y1 = gmm1.predict(self.X)
for a, b in zip(Y0, Y1):
self.assertAlmostEqual(a, b, places=3)
self.assertTrue(len(set(Y0)) > 1)
def test_training_once(self):
print "test training once"
gmm0 = GMM(self.M, self.D, cvtype='diag')
likelihood0 = gmm0.train(self.X)
means0 = gmm0.components.means.flatten()
covars0 = gmm0.components.covars.flatten()
gmm1 = GMM(self.M, self.D, cvtype='diag')
likelihood1 = gmm1.train(self.X)
means1 = gmm1.components.means.flatten()
covars1 = gmm1.components.covars.flatten()
self.assertAlmostEqual(likelihood0, likelihood1, places=3)
for a,b in zip(means0, means1): self.assertAlmostEqual(a,b, places=3)
for a,b in zip(covars0, covars1): self.assertAlmostEqual(a,b, places=3)
def test_training_repeat(self):
print "test training repeat"
gmm0 = GMM(self.M, self.D, cvtype="diag")
likelihood0 = gmm0.train(self.X)
likelihood0 = gmm0.train(self.X)
likelihood0 = gmm0.train(self.X)
means0 = gmm0.components.means.flatten()
covars0 = gmm0.components.covars.flatten()
gmm1 = GMM(self.M, self.D, cvtype="diag")
likelihood1 = gmm1.train(self.X)
likelihood1 = gmm1.train(self.X)
likelihood1 = gmm1.train(self.X)
means1 = gmm1.components.means.flatten()
covars1 = gmm1.components.covars.flatten()
self.assertAlmostEqual(likelihood0, likelihood1, places=3)
for a, b in zip(means0, means1):
self.assertAlmostEqual(a, b, places=3)
for a, b in zip(covars0, covars1):
self.assertAlmostEqual(a, b, places=3)
def test_init(self):
gmm = GMM(3, 2, cvtype='diag')
self.assertIsNotNone(gmm)
def test_pure_python(self):
print "test pure python"
gmm = GMM(self.M, self.D, cvtype='diag')
means, covars = gmm.train_using_python(self.X)
Y = gmm.predict_using_python(self.X)
self.assertTrue(len(set(Y)) > 1)
def test_pure_python(self):
print "test pure python"
gmm = GMM(self.M, self.D, cvtype="diag")
means, covars = gmm.train_using_python(self.X)
Y = gmm.predict_using_python(self.X)
self.assertTrue(len(set(Y)) > 1)
def test_getter_methods(self):
gmm0 = GMM(self.M, self.D, cvtype='diag')
likelihood0 = gmm0.train(self.X)
all_weights0 = gmm0.get_all_component_weights()
all_means0 = gmm0.get_all_component_means()
all_diag_covars0 = gmm0.get_all_component_diag_covariance()
gmm1 = GMM(self.M, self.D, cvtype='diag')
likelihood1 = gmm1.train(self.X)
all_weights1 = gmm1.get_all_component_weights()
all_means1 = gmm1.get_all_component_means()
all_diag_covars1 = gmm1.get_all_component_diag_covariance()
self.assertAlmostEqual(likelihood0, likelihood1, places=3)
for (a, b) in zip(all_weights0, all_weights1):
self.assertAlmostEqual(a, b, places=3)
for m in range(gmm0.M):
for a, b in zip(all_means0[m], all_means1[m]):
self.assertAlmostEqual(a, b, places=3)
for a, b in zip(all_diag_covars0, all_diag_covars1):
self.assertAlmostEqual(a, b, places=3)
def test_pure_python(self):
gmm = GMM(self.M, self.D, cvtype='diag')
means, covars = gmm.train_using_python(self.X)
Y = gmm.predict_using_python(self.X)
self.assertTrue(len(set(Y)) > 1)
