def gen_mxfusion_model(self,
dtype,
D,
Z,
noise_var,
lengthscale,
variance,
rand_gen=None):
m = Model()
m.N = Variable()
m.X = Variable(shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(),
initial_value=mx.nd.array(noise_var,
dtype=dtype))
kernel = RBF(input_dim=3,
ARD=True,
variance=mx.nd.array(variance, dtype=dtype),
lengthscale=mx.nd.array(lengthscale, dtype=dtype),
dtype=dtype)
m.Y = SVGPRegression.define_variable(X=m.X,
kernel=kernel,
noise_var=m.noise_var,
inducing_inputs=m.Z,
shape=(m.N, D),
dtype=dtype)
gp = m.Y.factor
m.Y.factor.svgp_log_pdf.jitter = 1e-8
return m, gp
def test_prediction(self):
np.random.seed(0)
np.random.seed(0)
X = np.random.rand(10, 3)
Y = np.random.rand(10, 1)
Z = np.random.rand(3, 3)
qU_mean = np.random.rand(3, 1)
qU_cov_W = np.random.rand(3, 3)
qU_cov_diag = np.random.rand(3,)
noise_var = np.random.rand(1)
lengthscale = np.random.rand(3)
variance = np.random.rand(1)
qU_chol = np.linalg.cholesky(qU_cov_W.dot(qU_cov_W.T)+np.diag(qU_cov_diag))[None,:,:]
Xt = np.random.rand(5, 3)
m_gpy = GPy.core.SVGP(X=X, Y=Y, Z=Z, kernel=GPy.kern.RBF(3, ARD=True, lengthscale=lengthscale, variance=variance), likelihood=GPy.likelihoods.Gaussian(variance=noise_var))
m_gpy.q_u_mean = qU_mean
m_gpy.q_u_chol = GPy.util.choleskies.triang_to_flat(qU_chol)
dtype = 'float64'
m = Model()
m.N = Variable()
m.X = Variable(shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(), initial_value=mx.nd.array(noise_var, dtype=dtype))
kernel = RBF(input_dim=3, ARD=True, variance=mx.nd.array(variance, dtype=dtype), lengthscale=mx.nd.array(lengthscale, dtype=dtype), dtype=dtype)
m.Y = SVGPRegression.define_variable(X=m.X, kernel=kernel, noise_var=m.noise_var, inducing_inputs=m.Z, shape=(m.N, 1), dtype=dtype)
gp = m.Y.factor
observed = [m.X, m.Y]
infr = Inference(MAP(model=m, observed=observed), dtype=dtype)
infr.initialize(X=X.shape, Y=Y.shape)
infr.params[gp._extra_graphs[0].qU_mean] = mx.nd.array(qU_mean, dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_W] = mx.nd.array(qU_cov_W, dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_diag] = mx.nd.array(qU_cov_diag, dtype=dtype)
loss, _ = infr.run(X=mx.nd.array(X, dtype=dtype), Y=mx.nd.array(Y, dtype=dtype))
# noise_free, diagonal
mu_gpy, var_gpy = m_gpy.predict_noiseless(Xt)
infr2 = TransferInference(ModulePredictionAlgorithm(m, observed=[m.X], target_variables=[m.Y]), infr_params=infr.params, dtype=np.float64)
res = infr2.run(X=mx.nd.array(Xt, dtype=dtype))[0]
mu_mf, var_mf = res[0].asnumpy()[0], res[1].asnumpy()[0]
assert np.allclose(mu_gpy, mu_mf), (mu_gpy, mu_mf)
assert np.allclose(var_gpy[:,0], var_mf), (var_gpy[:,0], var_mf)
# noisy, diagonal
mu_gpy, var_gpy = m_gpy.predict(Xt)
infr2 = TransferInference(ModulePredictionAlgorithm(m, observed=[m.X], target_variables=[m.Y]), infr_params=infr.params, dtype=np.float64)
infr2.inference_algorithm.model.Y.factor.svgp_predict.noise_free = False
res = infr2.run(X=mx.nd.array(Xt, dtype=dtype))[0]
mu_mf, var_mf = res[0].asnumpy()[0], res[1].asnumpy()[0]
assert np.allclose(mu_gpy, mu_mf), (mu_gpy, mu_mf)
assert np.allclose(var_gpy[:,0], var_mf), (var_gpy[:,0], var_mf)
def test_sampling_prediction(self):
np.random.seed(0)
np.random.seed(0)
X = np.random.rand(10, 3)
Y = np.random.rand(10, 1)
Z = np.random.rand(3, 3)
qU_mean = np.random.rand(3, 1)
qU_cov_W = np.random.rand(3, 3)
qU_cov_diag = np.random.rand(3,)
noise_var = np.random.rand(1)
lengthscale = np.random.rand(3)
variance = np.random.rand(1)
qU_chol = np.linalg.cholesky(qU_cov_W.dot(qU_cov_W.T)+np.diag(qU_cov_diag))[None,:,:]
Xt = np.random.rand(5, 3)
m_gpy = GPy.core.SVGP(X=X, Y=Y, Z=Z, kernel=GPy.kern.RBF(3, ARD=True, lengthscale=lengthscale, variance=variance), likelihood=GPy.likelihoods.Gaussian(variance=noise_var))
m_gpy.q_u_mean = qU_mean
m_gpy.q_u_chol = GPy.util.choleskies.triang_to_flat(qU_chol)
dtype = 'float64'
m = Model()
m.N = Variable()
m.X = Variable(shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(), initial_value=mx.nd.array(noise_var, dtype=dtype))
kernel = RBF(input_dim=3, ARD=True, variance=mx.nd.array(variance, dtype=dtype), lengthscale=mx.nd.array(lengthscale, dtype=dtype), dtype=dtype)
m.Y = SVGPRegression.define_variable(X=m.X, kernel=kernel, noise_var=m.noise_var, inducing_inputs=m.Z, shape=(m.N, 1), dtype=dtype)
gp = m.Y.factor
observed = [m.X, m.Y]
infr = Inference(MAP(model=m, observed=observed), dtype=dtype)
infr.initialize(X=X.shape, Y=Y.shape)
infr.params[gp._extra_graphs[0].qU_mean] = mx.nd.array(qU_mean, dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_W] = mx.nd.array(qU_cov_W, dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_diag] = mx.nd.array(qU_cov_diag, dtype=dtype)
loss, _ = infr.run(X=mx.nd.array(X, dtype=dtype), Y=mx.nd.array(Y, dtype=dtype))
# noise_free, diagonal
infr_pred = TransferInference(ModulePredictionAlgorithm(model=m, observed=[m.X], target_variables=[m.Y], num_samples=5),
infr_params=infr.params)
gp = m.Y.factor
gp.attach_prediction_algorithms(
targets=gp.output_names, conditionals=gp.input_names,
algorithm=SVGPRegressionSamplingPrediction(
gp._module_graph, gp._extra_graphs[0], [gp._module_graph.X]),
alg_name='svgp_predict')
gp.svgp_predict.diagonal_variance = False
gp.svgp_predict.jitter = 1e-6
y_samples = infr_pred.run(X=mx.nd.array(Xt, dtype=dtype))[0].asnumpy()
def test_with_samples(self):
from mxfusion.common import config
config.DEFAULT_DTYPE = 'float64'
dtype = 'float64'
D, X, Y, Z, noise_var, lengthscale, variance = self.gen_data()
m = Model()
m.N = Variable()
m.X = Normal.define_variable(mean=0, variance=1, shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(), initial_value=mx.nd.array(noise_var, dtype=dtype))
kernel = RBF(input_dim=3, ARD=True, variance=mx.nd.array(variance, dtype=dtype), lengthscale=mx.nd.array(lengthscale, dtype=dtype), dtype=dtype)
m.Y = SparseGPRegression.define_variable(X=m.X, kernel=kernel, noise_var=m.noise_var, inducing_inputs=m.Z, shape=(m.N, D), dtype=dtype)
m.Y.factor.sgp_log_pdf.jitter = 1e-8
q = create_Gaussian_meanfield(model=m, observed=[m.Y])
infr = GradBasedInference(
inference_algorithm=StochasticVariationalInference(
model=m, posterior=q, num_samples=10, observed=[m.Y]))
infr.run(Y=mx.nd.array(Y, dtype='float64'), max_iter=2,
learning_rate=0.1, verbose=True)
infr2 = Inference(ForwardSamplingAlgorithm(
model=m, observed=[m.X], num_samples=5))
infr2.run(X=mx.nd.array(X, dtype='float64'))
infr_pred = TransferInference(ModulePredictionAlgorithm(model=m, observed=[m.X], target_variables=[m.Y]), infr_params=infr.params)
xt = np.random.rand(13, 3)
res = infr_pred.run(X=mx.nd.array(xt, dtype=dtype))[0]
gp = m.Y.factor
gp.attach_prediction_algorithms(
targets=gp.output_names, conditionals=gp.input_names,
algorithm=SparseGPRegressionSamplingPrediction(
gp._module_graph, gp._extra_graphs[0], [gp._module_graph.X]),
alg_name='sgp_predict')
gp.sgp_predict.diagonal_variance = False
gp.sgp_predict.jitter = 1e-6
infr_pred2 = TransferInference(ModulePredictionAlgorithm(model=m, observed=[m.X], target_variables=[m.Y]), infr_params=infr.params)
xt = np.random.rand(13, 3)
res = infr_pred2.run(X=mx.nd.array(xt, dtype=dtype))[0]
def gen_mxfusion_model_w_mean(self, dtype, D, Z, noise_var, lengthscale,
variance, rand_gen=None):
net = nn.HybridSequential(prefix='nn_')
with net.name_scope():
net.add(nn.Dense(D, flatten=False, activation="tanh",
in_units=3, dtype=dtype))
net.initialize(mx.init.Xavier(magnitude=3))
m = Model()
m.N = Variable()
m.X = Variable(shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(), initial_value=mx.nd.array(noise_var, dtype=dtype))
kernel = RBF(input_dim=3, ARD=True, variance=mx.nd.array(variance, dtype=dtype), lengthscale=mx.nd.array(lengthscale, dtype=dtype), dtype=dtype)
m.mean_func = MXFusionGluonFunction(net, num_outputs=1,
broadcastable=True)
m.Y = SparseGPRegression.define_variable(X=m.X, kernel=kernel, noise_var=m.noise_var, mean=m.mean_func(m.X), inducing_inputs=m.Z, shape=(m.N, D), dtype=dtype)
m.Y.factor.sgp_log_pdf.jitter = 1e-8
return m, net
def test_log_pdf_w_samples_of_noise_var(self):
D, X, Y, Z, noise_var, lengthscale, variance, qU_mean, \
qU_cov_W, qU_cov_diag, qU_chol = self.gen_data()
dtype = 'float64'
D = 2
Y = np.random.rand(10, D)
qU_mean = np.random.rand(3, D)
m = Model()
m.N = Variable()
m.X = Variable(shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(),
shape=(m.N, D))
kernel = RBF(input_dim=3,
ARD=True,
variance=mx.nd.array(variance, dtype=dtype),
lengthscale=mx.nd.array(lengthscale, dtype=dtype),
dtype=dtype)
m.Y = SVGPRegression.define_variable(X=m.X,
kernel=kernel,
noise_var=m.noise_var,
inducing_inputs=m.Z,
shape=(m.N, D),
dtype=dtype)
gp = m.Y.factor
m.Y.factor.svgp_log_pdf.jitter = 1e-8
observed = [m.X, m.Y]
infr = Inference(MAP(model=m, observed=observed), dtype=dtype)
infr.initialize(X=X.shape, Y=Y.shape)
infr.params[gp._extra_graphs[0].qU_mean] = mx.nd.array(qU_mean,
dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_W] = mx.nd.array(qU_cov_W,
dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_diag] = mx.nd.array(qU_cov_diag,
dtype=dtype)
loss, _ = infr.run(X=mx.nd.array(X, dtype=dtype),
Y=mx.nd.array(Y, dtype=dtype),
max_iter=1)
def test_log_pdf(self):
np.random.seed(0)
D = 2
X = np.random.rand(10, 3)
Y = np.random.rand(10, D)
Z = np.random.rand(3, 3)
qU_mean = np.random.rand(3, D)
qU_cov_W = np.random.rand(3, 3)
qU_cov_diag = np.random.rand(3,)
noise_var = np.random.rand(1)
lengthscale = np.random.rand(3)
variance = np.random.rand(1)
qU_chol = np.linalg.cholesky(qU_cov_W.dot(qU_cov_W.T)+np.diag(qU_cov_diag))[None,:,:]
m_gpy = GPy.core.SVGP(X=X, Y=Y, Z=Z, kernel=GPy.kern.RBF(3, ARD=True, lengthscale=lengthscale, variance=variance), likelihood=GPy.likelihoods.Gaussian(variance=noise_var))
m_gpy.q_u_mean = qU_mean
m_gpy.q_u_chol = GPy.util.choleskies.triang_to_flat(qU_chol)
l_gpy = m_gpy.log_likelihood()
dtype = 'float64'
m = Model()
m.N = Variable()
m.X = Variable(shape=(m.N, 3))
m.Z = Variable(shape=(3, 3), initial_value=mx.nd.array(Z, dtype=dtype))
m.noise_var = Variable(transformation=PositiveTransformation(), initial_value=mx.nd.array(noise_var, dtype=dtype))
kernel = RBF(input_dim=3, ARD=True, variance=mx.nd.array(variance, dtype=dtype), lengthscale=mx.nd.array(lengthscale, dtype=dtype), dtype=dtype)
m.Y = SVGPRegression.define_variable(X=m.X, kernel=kernel, noise_var=m.noise_var, inducing_inputs=m.Z, shape=(m.N, D), dtype=dtype)
gp = m.Y.factor
observed = [m.X, m.Y]
infr = Inference(MAP(model=m, observed=observed), dtype=dtype)
infr.initialize(X=X.shape, Y=Y.shape)
infr.params[gp._extra_graphs[0].qU_mean] = mx.nd.array(qU_mean, dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_W] = mx.nd.array(qU_cov_W, dtype=dtype)
infr.params[gp._extra_graphs[0].qU_cov_diag] = mx.nd.array(qU_cov_diag, dtype=dtype)
loss, _ = infr.run(X=mx.nd.array(X, dtype=dtype), Y=mx.nd.array(Y, dtype=dtype))
l_mf = -loss
assert np.allclose(l_mf.asnumpy(), l_gpy)