def test_mean_var_interface_returns_same_as_cpp(self):
"""Test that the /gp/mean_var endpoint does the same thing as the C++ interface."""
tolerance = 1.0e-11
for test_case in self.gp_test_environments:
python_domain, python_gp = test_case
python_cov, historical_data = python_gp.get_core_data_copy()
cpp_cov = SquareExponential(python_cov.hyperparameters)
cpp_gp = GaussianProcess(cpp_cov, historical_data)
points_to_evaluate = python_domain.generate_uniform_random_points_in_domain(
10)
# mean and var from C++
cpp_mean = cpp_gp.compute_mean_of_points(points_to_evaluate)
cpp_var = cpp_gp.compute_variance_of_points(points_to_evaluate)
# mean and var from REST
json_payload = self._build_json_payload(
python_domain, python_cov, historical_data,
points_to_evaluate.tolist())
resp = self.testapp.post(self.endpoint, json_payload)
resp_schema = GpMeanVarResponse()
resp_dict = resp_schema.deserialize(json.loads(resp.body))
rest_mean = numpy.asarray(resp_dict.get('mean'))
rest_var = numpy.asarray(resp_dict.get('var'))
self.assert_vector_within_relative(rest_mean, cpp_mean, tolerance)
self.assert_vector_within_relative(rest_var, cpp_var, tolerance)
def optimize(self, do_optimize=True, **kwargs):
if self.prior is None:
self.p0 = numpy.random.rand(1 + self.dim + self._num_derivatives +
1)
else:
self.p0 = self.prior.sample_from_prior(1)
self.hypers = [optimize.minimize(self.nll, self.p0.ravel()).x]
self.is_trained = True
self._models = []
hypers_list = []
noises_list = []
for sample in self.hypers:
print sample
if numpy.any((-20 > sample) + (sample > 20)):
continue
sample = numpy.exp(sample)
# Instantiate a GP for each hyperparameter configuration
cov_hyps = sample[:(self.dim + 1)]
hypers_list.append(cov_hyps)
se = SquareExponential(cov_hyps)
if self.noisy:
noise = sample[(self.dim + 1):]
else:
noise = numpy.array((1 + self._num_derivatives) * [1.e-8])
noises_list.append(noise)
model = GaussianProcess(se, noise, self._historical_data,
self.derivatives)
self._models.append(model)
self._gaussian_process_mcmc = GaussianProcessMCMC(
numpy.array(hypers_list), numpy.array(noises_list),
self._historical_data, self.derivatives)
def test_python_and_cpp_return_same_cholesky_variance_and_gradient(self):
"""Compare chol_var/grad chol_var results from Python & C++, checking seeral random points per test case."""
num_tests_per_case = 2
var_tolerance = 3.0e-12
# TODO(GH-240): set RNG seed for this case and restore toleranace to 3.0e-12 or better
grad_var_tolerance = 3.0e-10
for test_case in self.gp_test_environments:
domain, python_gp = test_case
python_cov, historical_data = python_gp.get_core_data_copy()
cpp_cov = SquareExponential(python_cov.hyperparameters)
cpp_gp = GaussianProcess(cpp_cov, historical_data)
for num_to_sample in self.num_to_sample_list:
for _ in xrange(num_tests_per_case):
points_to_sample = domain.generate_uniform_random_points_in_domain(num_to_sample)
cpp_var = cpp_gp.compute_cholesky_variance_of_points(points_to_sample)
python_var = python_gp.compute_cholesky_variance_of_points(points_to_sample)
self.assert_vector_within_relative(python_var, cpp_var, var_tolerance)
cpp_grad_var = cpp_gp.compute_grad_cholesky_variance_of_points(points_to_sample)
python_grad_var = python_gp.compute_grad_cholesky_variance_of_points(points_to_sample)
self.assert_vector_within_relative(python_grad_var, cpp_grad_var, grad_var_tolerance)
def test_python_and_cpp_return_same_mu_and_gradient(self):
"""Compare mu/grad mu results from Python & C++, checking seeral random points per test case."""
num_tests_per_case = 4
mu_tolerance = 3.0e-13
grad_mu_tolerance = 3.0e-12
for test_case in self.gp_test_environments:
domain, python_gp = test_case
python_cov, historical_data = python_gp.get_core_data_copy()
cpp_cov = SquareExponential(python_cov.hyperparameters)
cpp_gp = GaussianProcess(cpp_cov, historical_data)
for num_to_sample in self.num_to_sample_list:
for _ in xrange(num_tests_per_case):
points_to_sample = domain.generate_uniform_random_points_in_domain(
num_to_sample)
cpp_mu = cpp_gp.compute_mean_of_points(points_to_sample)
python_mu = python_gp.compute_mean_of_points(
points_to_sample)
self.assert_vector_within_relative(python_mu, cpp_mu,
mu_tolerance)
cpp_grad_mu = cpp_gp.compute_grad_mean_of_points(
points_to_sample)
python_grad_mu = python_gp.compute_grad_mean_of_points(
points_to_sample)
self.assert_vector_within_relative(python_grad_mu,
cpp_grad_mu,
grad_mu_tolerance)
def test_sample_point_from_gp(self):
"""Test that sampling points from the GP works."""
point_one = SamplePoint([0.0, 1.0], -1.0, 0.0)
point_two = SamplePoint([2.0, 2.5], 1.0, 0.1)
covariance = SquareExponential([1.0, 1.0, 1.0])
historical_data = HistoricalData(len(point_one.point),
[point_one, point_two])
gaussian_process = GaussianProcess(covariance, historical_data)
out_values = numpy.zeros(3)
for i in xrange(3):
out_values[i] = gaussian_process.sample_point_from_gp(
point_two.point, 0.001)
gaussian_process._gaussian_process.reset_to_most_recent_seed()
out_values_test = numpy.ones(3)
for i in xrange(3):
out_values_test[i] = gaussian_process.sample_point_from_gp(
point_two.point, 0.001)
# Exact match b/c we should've run over the exact same computations
self.assert_vector_within_relative(out_values_test, out_values, 0.0)
# Sampling from a historical point (that had 0 noise) should produce the same value associated w/that point
value = gaussian_process.sample_point_from_gp(point_one.point, 0.0)
self.assert_scalar_within_relative(value, point_one.value,
numpy.finfo(numpy.float64).eps)
def test_interface_returns_same_as_cpp(self):
"""Test that the /gp/ei endpoint does the same thing as the C++ interface."""
tolerance = 1.0e-11
for test_case in self.gp_test_environments:
python_domain, python_gp = test_case
python_cov, historical_data = python_gp.get_core_data_copy()
cpp_cov = SquareExponential(python_cov.hyperparameters)
cpp_gp = GaussianProcess(cpp_cov, historical_data)
points_to_evaluate = python_domain.generate_uniform_random_points_in_domain(
10)
# EI from C++
expected_improvement_evaluator = ExpectedImprovement(
cpp_gp,
None,
)
# TODO(GH-99): Change test case to have the right shape:
# (num_to_evaluate, num_to_sample, dim)
# Here we assume the shape is (num_to_evaluate, dim) so we insert an axis, making num_to_sample = 1.
# Also might be worth testing more num_to_sample values (will require manipulating C++ RNG state).
cpp_expected_improvement = expected_improvement_evaluator.evaluate_at_point_list(
points_to_evaluate[:, numpy.newaxis, :], )
# EI from REST
json_payload = self._build_json_payload(
python_domain, python_cov, historical_data,
points_to_evaluate.tolist())
resp = self.testapp.post(self.endpoint, json_payload)
resp_schema = GpEiResponse()
resp_dict = resp_schema.deserialize(json.loads(resp.body))
rest_expected_improvement = numpy.asarray(
resp_dict.get('expected_improvement'))
self.assert_vector_within_relative(
rest_expected_improvement,
cpp_expected_improvement,
tolerance,
)
def train(self, do_optimize=True, **kwargs):
"""
Performs MCMC sampling to sample hyperparameter configurations from the
likelihood and trains for each sample a GP on X and y
Parameters
----------
X: np.ndarray (N, D)
Input data points. The dimensionality of X is (N, D),
with N as the number of points and D is the number of features.
y: np.ndarray (N,)
The corresponding target values.
do_optimize: boolean
If set to true we perform MCMC sampling otherwise we just use the
hyperparameter specified in the kernel.
"""
if do_optimize:
# We have one walker for each hyperparameter configuration
sampler = emcee.EnsembleSampler(self.n_chains, 1 + self.dim + self._num_derivatives + 1,
self.compute_log_likelihood)
# Do a burn-in in the first iteration
if not self.burned:
# Initialize the walkers by sampling from the prior
if self.prior is None:
self.p0 = numpy.random.rand(self.n_chains, 1 + self.dim + self._num_derivatives + 1)
else:
self.p0 = self.prior.sample_from_prior(self.n_chains)
# Run MCMC sampling
self.p0, _, _ = sampler.run_mcmc(self.p0, self.burnin_steps,
rstate0=self.rng)
self.burned = True
# Start sampling
pos, _, _ = sampler.run_mcmc(self.p0, self.chain_length,
rstate0=self.rng)
# Save the current position, it will be the start point in
# the next iteration
self.p0 = pos
# Take the last samples from each walker
self.hypers = sampler.chain[numpy.random.choice(self.n_chains, self.n_hypers), -1]
self.is_trained = True
self._models = []
hypers_list = []
noises_list = []
for sample in self.hypers:
if numpy.any((-20 > sample) + (sample > 20)):
continue
sample = numpy.exp(sample)
# Instantiate a GP for each hyperparameter configuration
cov_hyps = sample[:(self.dim+1)]
hypers_list.append(cov_hyps)
se = SquareExponential(cov_hyps)
if self.noisy:
noise = sample[(self.dim+1):]
else:
noise = numpy.array((1+self._num_derivatives)*[1.e-8])
noises_list.append(noise)
model = GaussianProcess(se, noise,
self._historical_data,
self.derivatives)
self._models.append(model)
self._gaussian_process_mcmc = GaussianProcessMCMC(numpy.array(hypers_list), numpy.array(noises_list),
self._historical_data, self.derivatives)
}
for i in range(len(hist_data)):
result["data_{0}_points_sampled".format(i)] = hist_data[i].points_sampled
result["data_{0}_points_sampled_value".format(i)] = hist_data[i].points_sampled_value
for key in range(len(hist_data)):
# Setup prior for MAP
if key == 0:
prior_mean = np.concatenate(([max(0.01, np.var(hist_data[key].points_sampled_value) - np.mean(hist_data[key].points_sampled_noise_variance))], [(d[1]-d[0]) for d in problem.obj_func_min._search_domain]))
else:
prior_mean = np.concatenate(([max(0.01, np.var(hist_data[key].points_sampled_value) - np.mean(hist_data[key].points_sampled_noise_variance) - np.mean(hist_data[0].points_sampled_noise_variance))], [(d[1]-d[0]) for d in problem.obj_func_min._search_domain]))
prior_sig = np.diag(np.power(prior_mean/2., 2.0))
prior = NormalPrior(prior_mean, prior_sig)
hyper_bounds = [(0.001, prior_mean[i]+2.*np.sqrt(prior_sig[i,i])) for i in range(problem.obj_func_min.getDim()+1)]
hyperparam_search_domain = pythonTensorProductDomain([ClosedInterval(bound[0], bound[1]) for bound in hyper_bounds])
multistart_pts = hyperparam_search_domain.generate_uniform_random_points_in_domain(num_hyper_multistart)
best_f = np.inf
cov = SquareExponential(prior_mean)
for i in range(num_hyper_multistart):
hyper, f, output = hyper_opt(cov, data=hist_data[key], init_hyper=multistart_pts[i, :],
hyper_bounds=hyper_bounds, approx_grad=False, hyper_prior=prior)
# print output
if f < best_f:
best_hyper = hyper
best_f = f
result['prior_mean'].append(prior_mean)
result['prior_sig'].append(np.diag(prior_sig))
result['hyper_bounds'].append(hyper_bounds)
result['hyperparam'] = np.concatenate((result['hyperparam'], best_hyper))
result['loglikelihood'].append(-best_f)
send_data_to_s3(bucket, problem.hyper_path, result)
