def test_gp_mcmc_fit(tuning_job_state):
def tuning_job_state_mcmc(X, Y) -> TuningJobState:
Y = [dictionarize_objective(y) for y in Y]
return TuningJobState(
HyperparameterRanges_Impl(
HyperparameterRangeContinuous('x', -4., 4., LinearScaling())),
[CandidateEvaluation(x, y) for x, y in zip(X, Y)], [], [])
_set_seeds(0)
def f(x):
return 0.1 * np.power(x, 3)
X = np.concatenate((np.random.uniform(-4., -1.,
10), np.random.uniform(1., 4., 10)))
Y = f(X)
X_test = np.sort(np.random.uniform(-1., 1., 10))
X = [(x, ) for x in X]
X_test = [(x, ) for x in X_test]
tuning_job_state = tuning_job_state_mcmc(X, Y)
# checks if fitting is running
random_seed = 0
gpmodel = default_gpmodel_mcmc(tuning_job_state,
random_seed,
mcmc_config=DEFAULT_MCMC_CONFIG)
model = GPMXNetModel(tuning_job_state,
DEFAULT_METRIC,
random_seed,
gpmodel,
fit_parameters=True,
num_fantasy_samples=20)
X = [tuning_job_state.hp_ranges.to_ndarray(x) for x in X]
predictions = model.predict(np.array(X))
Y_std_list = [stds for means, stds in predictions]
Y_mean_list = [means for means, stds in predictions]
Y_mean = np.mean(Y_mean_list, axis=0)
Y_std = np.mean(Y_std_list, axis=0)
assert np.all(np.abs(Y_mean - Y) < 1e-1), \
"in a noiseless setting, mean of GP should coincide closely with outputs at training points"
X_test = [tuning_job_state.hp_ranges.to_ndarray(x) for x in X_test]
predictions_test = model.predict(np.array(X_test))
Y_std_test_list = [stds for means, stds in predictions_test]
Y_std_test = np.mean(Y_std_test_list, axis=0)
assert np.max(Y_std) < np.min(Y_std_test), \
"Standard deviation on un-observed points should be greater than at observed ones"
def test_gp_fit(tuning_job_state):
_set_seeds(0)
X = [
(0.0, 0.0),
(1.0, 0.0),
(0.0, 1.0),
(1.0, 1.0),
]
Y = [np.sum(x) * 10.0 for x in X]
# checks if fitting is running
random_seed = 0
gpmodel = default_gpmodel(tuning_job_state,
random_seed,
optimization_config=DEFAULT_OPTIMIZATION_CONFIG)
model = GPMXNetModel(tuning_job_state,
DEFAULT_METRIC,
random_seed,
gpmodel,
fit_parameters=True,
num_fantasy_samples=20)
X = [tuning_job_state.hp_ranges.to_ndarray(x) for x in X]
Y_mean, Y_std = model.predict(np.array(X))[0]
assert np.all(np.abs(Y_mean - Y) < 1e-1), \
"in a noiseless setting, mean of GP should coincide closely with outputs at training points"
X_test = [
(0.2, 0.2),
(0.4, 0.2),
(0.1, 0.9),
(0.5, 0.5),
]
X_test = [tuning_job_state.hp_ranges.to_ndarray(x) for x in X_test]
Y_mean_test, Y_std_test = model.predict(np.array(X_test))[0]
assert np.min(Y_std) < np.min(Y_std_test), \
"Standard deviation on un-observed points should be greater than at observed ones"
def fit_predict_ours(data: dict,
random_seed: int,
optimization_config: OptimizationConfig,
test_intermediates: Optional[dict] = None) -> dict:
# Create surrogate model
num_dims = len(data['ss_limits'])
_gpmodel = GaussianProcessRegression(
kernel=Matern52(num_dims, ARD=True),
mean=ZeroMeanFunction(), # Instead of ScalarMeanFunction
optimization_config=optimization_config,
random_seed=random_seed,
test_intermediates=test_intermediates)
model = GPMXNetModel(data['state'],
DEFAULT_METRIC,
random_seed,
_gpmodel,
fit_parameters=True,
num_fantasy_samples=20)
model_params = model.get_params()
print('Hyperparameters: {}'.format(model_params))
# Prediction
means, stddevs = model.predict(data['test_inputs'])[0]
return {'means': means, 'stddevs': stddevs}