def test_diag_fit_kmeans():
model1, _ = create_diag_pdf()
model1.initialize()
x = model1.sample(num_samples_train)
x_val = model1.sample(num_samples)
model2 = GMM(num_comp=4, x_dim=x_dim)
model2.initialize(x)
elbo = model2.fit(x, x_val=x_val)
model2_diag = GMMDiagCov(num_comp=4, x_dim=x_dim)
model2_diag.initialize(x)
elbo_diag = model2_diag.fit(x, x_val=x_val)
model2.plot2D(feat_idx=[0, 1], num_sigmas=1)
plt.savefig(output_dir + '/plot_fit_kmeans_init_D01.pdf')
plt.close()
model2.plot2D(feat_idx=[0, 2], num_sigmas=1)
plt.savefig(output_dir + '/plot_fit_kmeans_init_D02.pdf')
plt.close()
plt.figure()
plt.plot(np.repeat(model1.elbo(x) / x.shape[0], len(elbo[1])), 'b')
plt.plot(np.repeat(model1.elbo(x_val) / x.shape[0], len(elbo[1])), 'b--')
plt.plot(elbo[1], 'r')
plt.plot(elbo[3], 'r--')
plt.plot(elbo_diag[1], 'g')
plt.plot(elbo_diag[3], 'g--')
plt.savefig(output_dir + '/fit_kmeans_init_elbo.pdf')
plt.close()
def test_initialize_kmeans():
model1 = create_pdf()
x = model1.sample(num_samples=num_samples_init)
model2 = GMMDiagCov(num_comp=4, x_dim=x_dim)
model2.initialize(x)
def test_initialize_stdnormal():
model = GMMDiagCov(num_comp=1, x_dim=x_dim)
model.initialize()
assert (model.pi == 1)
assert_allclose(model.mu, np.zeros((1, x_dim)))
assert_allclose(model.Lambda, np.ones((1, x_dim)))
def test_initialize():
model1 = create_pdf()
model1.initialize()
model2 = GMMDiagCov(num_comp=model1.num_comp,
pi=model1.pi,
eta=model1.eta,
x_dim=model1.x_dim)
model2.initialize()
model3 = GMMDiagCov(num_comp=model2.num_comp,
pi=model2.pi,
mu=model2.mu,
Lambda=model2.Lambda,
x_dim=model1.x_dim)
model3.initialize()
assert_allclose(model1.eta, model2.eta)
assert_allclose(model1.eta, model3.eta)
assert_allclose(model1.A, model2.A)
assert_allclose(model1.A, model3.A)
assert_allclose(model1.mu, model2.mu)
assert_allclose(model1.mu, model3.mu)
assert_allclose(model1.Lambda, model2.Lambda)
assert_allclose(model1.Lambda, model3.Lambda)
def create_pdf():
model = GMMDiagCov(num_comp=len(pi1),
pi=pi1,
mu=mu1,
Lambda=1 / S1,
x_dim=x_dim)
return model
def test_fit_kmeans_split4():
model1 = create_pdf()
model1.initialize()
x = model1.sample(num_samples_train)
x_val = model1.sample(num_samples)
model2 = GMMDiagCov(num_comp=1, x_dim=x_dim)
model2.initialize()
elbo = model2.fit(x, x_val=x_val, epochs=1)
model2 = model2.split_comp(4)
elbo = model2.fit(x, x_val=x_val)
model1.plot2D(feat_idx=[0, 1], num_sigmas=1)
plt.savefig(output_dir + '/plot_fit_split4_init_D01.pdf')
plt.close()
model1.plot2D(feat_idx=[0, 2], num_sigmas=1)
plt.savefig(output_dir + '/plot_fit_split4_init_D02.pdf')
plt.close()
plt.figure()
plt.plot(np.repeat(model1.elbo(x) / x.shape[0], len(elbo[1])), 'b')
plt.plot(np.repeat(model1.elbo(x_val) / x.shape[0], len(elbo[1])), 'b--')
plt.plot(elbo[1], 'r')
plt.plot(elbo[3], 'r--')
plt.savefig(output_dir + '/fit_split4_init_elbo.pdf')
plt.close()