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)], [], [])
def test_get_internal_candidate_evaluations():
"""we do not test the case with no evaluations, as it is assumed
that there will be always some evaluations generated in the beginning
of the BO loop."""
candidates = [
CandidateEvaluation((2, 3.3, 'X'), dictionarize_objective(5.3)),
CandidateEvaluation((1, 9.9, 'Y'), dictionarize_objective(10.9)),
CandidateEvaluation((7, 6.1, 'X'), dictionarize_objective(13.1)),
]
state = TuningJobState(
hp_ranges=HyperparameterRanges_Impl(
HyperparameterRangeInteger('integer', 0, 10, LinearScaling()),
HyperparameterRangeContinuous('real', 0, 10, LinearScaling()),
HyperparameterRangeCategorical('categorical', ('X', 'Y')),
),
candidate_evaluations=candidates,
failed_candidates=[candidates[0].candidate
], # these should be ignored by the model
pending_evaluations=[])
result = get_internal_candidate_evaluations(state,
DEFAULT_METRIC,
normalize_targets=True,
num_fantasize_samples=20)
assert len(result.X.shape) == 2, "Input should be a matrix"
assert len(result.y.shape) == 2, "Output should be a matrix"
assert result.X.shape[0] == len(candidates)
assert result.y.shape[
-1] == 1, "Only single output value per row is suppored"
assert np.abs(np.mean(
result.y)) < 1e-8, "Mean of the normalized outputs is not 0.0"
assert np.abs(np.std(result.y) -
1.0) < 1e-8, "Std. of the normalized outputs is not 1.0"
np.testing.assert_almost_equal(result.mean, 9.766666666666666)
np.testing.assert_almost_equal(result.std, 3.283629428273267)
def default_models() -> List[GaussProcSurrogateModel]:
X = [
(0.0, 0.0),
(1.0, 0.0),
(0.0, 1.0),
(1.0, 1.0),
(0.0, 0.0
), # same evals are added multiple times to force GP to unlearn prior
(1.0, 0.0),
(0.0, 1.0),
(1.0, 1.0),
(0.0, 0.0),
(1.0, 0.0),
(0.0, 1.0),
(1.0, 1.0),
]
Y = [dictionarize_objective(np.sum(x) * 10.0) for x in X]
state = TuningJobState(
HyperparameterRanges_Impl(
HyperparameterRangeContinuous('x', 0.0, 1.0, LinearScaling()),
HyperparameterRangeContinuous('y', 0.0, 1.0, LinearScaling()),
),
[CandidateEvaluation(x, y) for x, y in zip(X, Y)],
[],
[],
)
random_seed = 0
gpmodel = default_gpmodel(state,
random_seed=random_seed,
optimization_config=DEFAULT_OPTIMIZATION_CONFIG)
gpmodel_mcmc = default_gpmodel_mcmc(state,
random_seed=random_seed,
mcmc_config=DEFAULT_MCMC_CONFIG)
return [
GaussProcSurrogateModel(state,
DEFAULT_METRIC,
random_seed,
gpmodel,
fit_parameters=True,
num_fantasy_samples=20),
GaussProcSurrogateModel(state,
DEFAULT_METRIC,
random_seed,
gpmodel_mcmc,
fit_parameters=True,
num_fantasy_samples=20)
]
def tuning_job_state() -> TuningJobState:
X = [
(0.0, 0.0),
(1.0, 0.0),
(0.0, 1.0),
(1.0, 1.0),
]
Y = [dictionarize_objective(np.sum(x) * 10.0) for x in X]
return TuningJobState(
HyperparameterRanges_Impl(
HyperparameterRangeContinuous('x', 0.0, 1.0, LinearScaling()),
HyperparameterRangeContinuous('y', 0.0, 1.0, LinearScaling()),
), [CandidateEvaluation(x, y) for x, y in zip(X, Y)], [], [])
def _candidate_evaluations(num):
return [
CandidateEvaluation(
candidate=(i,),
metrics=dictionarize_objective(float(i)))
for i in range(num)]
def test_gp_fantasizing():
"""
Compare whether acquisition function evaluations (values, gradients) with
fantasizing are the same as averaging them by hand.
"""
random_seed = 4567
_set_seeds(random_seed)
num_fantasy_samples = 10
num_pending = 5
hp_ranges = HyperparameterRanges_Impl(
HyperparameterRangeContinuous('x', 0.0, 1.0, LinearScaling()),
HyperparameterRangeContinuous('y', 0.0, 1.0, LinearScaling()))
X = [
(0.0, 0.0),
(1.0, 0.0),
(0.0, 1.0),
(1.0, 1.0),
]
num_data = len(X)
Y = [
dictionarize_objective(np.random.randn(1, 1)) for _ in range(num_data)
]
# Draw fantasies. This is done for a number of fixed pending candidates
# The model parameters are fit in the first iteration, when there are
# no pending candidates
# Note: It is important to not normalize targets, because this would be
# done on the observed targets only, not the fantasized ones, so it
# would be hard to compare below.
pending_evaluations = []
for _ in range(num_pending):
pending_cand = tuple(np.random.rand(2, ))
pending_evaluations.append(PendingEvaluation(pending_cand))
state = TuningJobState(hp_ranges,
[CandidateEvaluation(x, y) for x, y in zip(X, Y)],
failed_candidates=[],
pending_evaluations=pending_evaluations)
gpmodel = default_gpmodel(state,
random_seed,
optimization_config=DEFAULT_OPTIMIZATION_CONFIG)
model = GaussProcSurrogateModel(state,
DEFAULT_METRIC,
random_seed,
gpmodel,
fit_parameters=True,
num_fantasy_samples=num_fantasy_samples,
normalize_targets=False)
fantasy_samples = model.fantasy_samples
# Evaluate acquisition function and gradients with fantasizing
num_test = 50
X_test = [
hp_ranges.to_ndarray(tuple(np.random.rand(2, )))
for _ in range(num_test)
]
acq_func = EIAcquisitionFunction(model)
fvals, grads = _compute_acq_with_gradient_many(acq_func, X_test)
# Do the same computation by averaging by hand
fvals_cmp = np.empty((num_fantasy_samples, ) + fvals.shape)
grads_cmp = np.empty((num_fantasy_samples, ) + grads.shape)
X_full = X + state.pending_candidates
for it in range(num_fantasy_samples):
Y_full = Y + [
dictionarize_objective(eval.fantasies[DEFAULT_METRIC][:, it])
for eval in fantasy_samples
]
state2 = TuningJobState(
hp_ranges,
[CandidateEvaluation(x, y) for x, y in zip(X_full, Y_full)],
failed_candidates=[],
pending_evaluations=[])
# We have to skip parameter optimization here
model2 = GaussProcSurrogateModel(
state2,
DEFAULT_METRIC,
random_seed,
gpmodel,
fit_parameters=False,
num_fantasy_samples=num_fantasy_samples,
normalize_targets=False)
acq_func2 = EIAcquisitionFunction(model2)
fvals_, grads_ = _compute_acq_with_gradient_many(acq_func2, X_test)
fvals_cmp[it, :] = fvals_
grads_cmp[it, :] = grads_
# Comparison
fvals2 = np.mean(fvals_cmp, axis=0)
grads2 = np.mean(grads_cmp, axis=0)
assert np.allclose(fvals, fvals2)
assert np.allclose(grads, grads2)