def __init__(
self,
lb,
ub,
restarts=10,
samps=1000,
dim=1,
acqf=None,
extra_acqf_args=None,
model=None,
):
if extra_acqf_args is None:
extra_acqf_args = {}
super().__init__(lb=lb,
ub=ub,
dim=dim,
acqf=acqf,
extra_acqf_args=extra_acqf_args)
self.restarts = restarts
self.samps = samps
self.likelihood = gpytorch.likelihoods.BernoulliLikelihood()
if model is None:
self.model = GPClassificationModel(inducing_min=self.lb,
inducing_max=self.ub)
else:
self.model = model
def test_1d_classification(self):
"""
Just see if we memorize the training set
"""
X, y = self.X, self.y
model = GPClassificationModel(
torch.Tensor([-3]), torch.Tensor([3]), inducing_size=10
)
model.fit(X[:50], y[:50])
# pspace
pm, _ = model.predict(X[:50], probability_space=True)
pred = (pm > 0.5).numpy()
npt.assert_allclose(pred, y[:50])
# fspace
pm, _ = model.predict(X[:50], probability_space=False)
pred = (pm > 0).numpy()
npt.assert_allclose(pred, y[:50])
# smoke test update
model.update(X, y)
# pspace
pm, _ = model.predict(X, probability_space=True)
pred = (pm > 0.5).numpy()
npt.assert_allclose(pred, y)
# fspace
pm, _ = model.predict(X, probability_space=False)
pred = (pm > 0).numpy()
npt.assert_allclose(pred, y)
def test_dim_grid_model_size(self):
lb = -4.0
ub = 4.0
dim = 1
gridsize = 10
mb = GPClassificationModel(lb=lb, ub=ub, dim=dim)
grid = GPClassificationModel.dim_grid(mb, gridsize=gridsize)
self.assertEqual(grid.shape, torch.Size([10, 1]))
def test_reset_variational_strategy(self):
model = GPClassificationModel(lb=[-3], ub=[3], inducing_size=20)
variational_params_before = [
v.clone().detach().numpy() for v in model.variational_parameters()
]
induc_before = model.variational_strategy.inducing_points
model.fit(torch.Tensor(self.X), torch.Tensor(self.y))
variational_params_after = [
v.clone().detach().numpy() for v in model.variational_parameters()
]
induc_after = model.variational_strategy.inducing_points
model._reset_variational_strategy()
variational_params_reset = [
v.clone().detach().numpy() for v in model.variational_parameters()
]
induc_reset = model.variational_strategy.inducing_points
# before should be different from after and after should be different
# from reset
self.assertFalse(np.allclose(induc_before, induc_after))
self.assertFalse(np.allclose(induc_after, induc_reset))
for before, after in zip(variational_params_before, variational_params_after):
self.assertFalse(np.allclose(before, after))
for after, reset in zip(variational_params_after, variational_params_reset):
self.assertFalse(np.allclose(after, reset))
def test_time_limits(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
X, y = make_classification(
n_samples=100,
n_features=8,
n_redundant=3,
n_informative=5,
random_state=1,
n_clusters_per_class=4,
)
X, y = torch.Tensor(X), torch.Tensor(y)
model = GPClassificationModel(
lb=-3 * torch.ones(8),
ub=3 * torch.ones(8),
max_fit_time=0.5,
inducing_size=10,
)
model.fit(X, y)
generator = OptimizeAcqfGenerator(acqf=MCLevelSetEstimation,
acqf_kwargs={
"beta": 1.96,
"target": 0.5
})
start = time.time()
generator.gen(1, model)
end = time.time()
long = end - start
generator = OptimizeAcqfGenerator(
acqf=MCLevelSetEstimation,
acqf_kwargs={
"beta": 1.96,
"target": 0.5
},
max_gen_time=0.1,
)
start = time.time()
generator.gen(1, model)
end = time.time()
short = end - start
# very loose test because fit time is only approximately computed
self.assertTrue(long > short)
def test_mi_acqf(self):
mean = ConstantMean().initialize(constant=1.2)
covar = LinearKernel().initialize(variance=1.0)
model = GPClassificationModel(
inducing_min=torch.Tensor([0]),
inducing_max=torch.Tensor([1]),
inducing_size=10,
mean_module=mean,
covar_module=covar,
)
x = torch.rand(size=(10, 1))
acqf = BernoulliMCMutualInformation(model=model, objective=ProbitObjective())
acq_pytorch = acqf(x)
samps_numpy = norm.cdf(
multivariate_normal.rvs(mean=np.ones(10) * 1.2, cov=x @ x.T, size=10000)
)
samp_entropies = bernoulli(samps_numpy).entropy()
mean_entropy = bernoulli(samps_numpy.mean(axis=0)).entropy()
acq_numpy = mean_entropy - samp_entropies.mean(axis=0)
# this assertion fails, not sure why, these should be equal to numerical
# precision
# self.assertTrue(np.allclose(acq_numpy, acq_pytorch.detach().numpy().flatten()))
# this one succeeds
self.assertTrue(
pearsonr(acq_numpy, acq_pytorch.detach().numpy().flatten())[0] > (1 - 1e-5)
)
def test_1d_query(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 150
n_opt = 1
lb = -4.0
ub = 4.0
target = 0.5
def obj(x):
return -((Normal(0, 1).cdf(x[..., 0]) - target) ** 2)
# Test sine function with period 4
def test_fun(x):
return np.sin(np.pi * x / 4)
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound,
acqf_kwargs={"beta": 1.96, "objective": GenericMCObjective(obj)},
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(norm.cdf(test_fun(next_x)))])
# We expect the global max to be at (2, 1), the min at (-2, -1)
fmax, argmax = strat.get_max()
self.assertTrue(np.abs(fmax - 1) < 0.5)
self.assertTrue(np.abs(argmax[0] - 2) < 0.5)
fmin, argmin = strat.get_min()
self.assertTrue(np.abs(fmin + 1) < 0.5)
self.assertTrue(np.abs(argmin[0] + 2) < 0.5)
# Query at x=2 should be f=1
self.assertTrue(np.abs(strat.predict(torch.tensor([2]))[0] - 1) < 0.5)
# Inverse query at val 1 should return (1,[2])
val, loc = strat.inv_query(1.0, constraints={})
self.assertTrue(np.abs(val - 1) < 0.5)
self.assertTrue(np.abs(loc[0] - 2) < 0.5)
def test_1d_classification_different_scales(self):
"""
Just see if we memorize the training set
"""
np.random.seed(1)
torch.manual_seed(1)
X, y = make_classification(
n_features=2,
n_redundant=0,
n_informative=1,
random_state=1,
n_clusters_per_class=1,
)
X, y = torch.Tensor(X), torch.Tensor(y)
X[:, 0] = X[:, 0] * 1000
X[:, 1] = X[:, 1] / 1000
lb = [-3000, -0.003]
ub = [3000, 0.003]
model = GPClassificationModel(lb=lb, ub=ub, inducing_size=20)
model.fit(X[:50], y[:50])
# pspace
pm, _ = model.predict(X[:50], probability_space=True)
pred = (pm > 0.5).numpy()
npt.assert_allclose(pred, y[:50])
# fspace
pm, _ = model.predict(X[:50], probability_space=False)
pred = (pm > 0).numpy()
npt.assert_allclose(pred, y[:50])
# smoke test update
model.update(X, y)
# pspace
pm, _ = model.predict(X, probability_space=True)
pred = (pm > 0.5).numpy()
npt.assert_allclose(pred, y)
# fspace
pm, _ = model.predict(X, probability_space=False)
pred = (pm > 0).numpy()
npt.assert_allclose(pred, y)
def test_1d_jnd(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 150
n_opt = 1
lb = -4.0
ub = 4.0
target = 0.5
def obj(x):
return -((Normal(0, 1).cdf(x[..., 0]) - target) ** 2)
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(norm.cdf(next_x / 1.5))])
x = torch.linspace(-4, 4, 100)
zhat, _ = strat.predict(x)
# we expect jnd close to the target to be close to the correct
# jnd (1.5), and since this is linear model this should be true
# for both definitions of JND
jnd_step = strat.get_jnd(grid=x[:, None], method="step")
est_jnd_step = jnd_step[50]
# looser test because step-jnd is hurt more by reverting to the mean
self.assertTrue(np.abs(est_jnd_step - 1.5) < 0.5)
jnd_taylor = strat.get_jnd(grid=x[:, None], method="taylor")
est_jnd_taylor = jnd_taylor[50]
self.assertTrue(np.abs(est_jnd_taylor - 1.5) < 0.25)
def test_hyperparam_consistency(self):
# verify that creating the model `from_config` or with `__init__` has the same hyperparams
m1 = GPClassificationModel(lb=[1, 2], ub=[3, 4])
m2 = GPClassificationModel.from_config(
config=Config(config_dict={"common": {"lb": "[1,2]", "ub": "[3,4]"}})
)
self.assertTrue(isinstance(m1.covar_module, type(m2.covar_module)))
self.assertTrue(
isinstance(m1.covar_module.base_kernel, type(m2.covar_module.base_kernel))
)
self.assertTrue(isinstance(m1.mean_module, type(m2.mean_module)))
m1priors = list(m1.covar_module.named_priors())
m2priors = list(m2.covar_module.named_priors())
for p1, p2 in zip(m1priors, m2priors):
name1, parent1, prior1, paramtransforms1, priortransforms1 = p1
name2, parent2, prior2, paramtransforms2, priortransforms2 = p2
self.assertTrue(name1 == name2)
self.assertTrue(isinstance(parent1, type(parent2)))
self.assertTrue(isinstance(prior1, type(prior2)))
def test_1d_classification(self):
"""
Just see if we memorize the training set
"""
np.random.seed(1)
torch.manual_seed(1)
X, y = make_classification(
n_features=1,
n_redundant=0,
n_informative=1,
random_state=1,
n_clusters_per_class=1,
)
X, y = torch.Tensor(X), torch.Tensor(y)
model = GPClassificationModel(torch.Tensor([-3]), torch.Tensor([3]))
model.set_train_data(X, y)
likelihood = gpytorch.likelihoods.BernoulliLikelihood()
mll = gpytorch.mlls.VariationalELBO(likelihood, model, 100)
fit_gpytorch_model(mll)
pred = (torch.sigmoid(model.posterior(X).mean) > 0.5).numpy()
npt.assert_allclose(pred[:, 0], y)
def test_1d_single_targeting(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 50
n_opt = 1
lb = -4.0
ub = 4.0
target = 0.75
def obj(x):
return -((Normal(0, 1).cdf(x[..., 0]) - target) ** 2)
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(norm.cdf(next_x))])
x = torch.linspace(-4, 4, 100)
zhat, _ = strat.predict(x)
# since target is 0.75, find the point at which f_est is 0.75
est_max = x[np.argmin((norm.cdf(zhat.detach().numpy()) - 0.75) ** 2)]
# since true z is just x, the true max is where phi(x)=0.75,
self.assertTrue(np.abs(est_max - norm.ppf(0.75)) < 0.5)
def test_1d_single_probit_pure_exploration(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 50
n_opt = 1
lb = -4.0
ub = 4.0
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(norm.cdf(next_x))])
x = torch.linspace(-4, 4, 100)
zhat, _ = strat.predict(x)
# f(x) = x so we're just looking at corr between cdf(zhat) and cdf(x)
self.assertTrue(
pearsonr(norm.cdf(zhat.detach().numpy()).flatten(), norm.cdf(x).flatten())[
0
]
> 0.95
)
def test_1d_single_lse(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 50
n_opt = 1
lb = -4.0
ub = 4.0
# target is in z space not phi(z) space, maybe that's
# weird
extra_acqf_args = {"target": 0.75, "beta": 1.96}
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
n_trials=n_opt,
generator=OptimizeAcqfGenerator(
MCLevelSetEstimation, acqf_kwargs=extra_acqf_args
),
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(norm.cdf(next_x))])
x = torch.linspace(-4, 4, 100)
zhat, _ = strat.predict(x)
# since target is 0.75, find the point at which f_est is 0.75
est_max = x[np.argmin((norm.cdf(zhat.detach().numpy()) - 0.75) ** 2)]
# since true z is just x, the true max is where phi(x)=0.75,
self.assertTrue(np.abs(est_max - norm.ppf(0.75)) < 0.5)
def from_config(cls, config):
classname = cls.__name__
model = GPClassificationModel.from_config(config)
lb = config.gettensor(classname, "lb")
ub = config.gettensor(classname, "ub")
restarts = config.getint(classname, "restarts", fallback=10)
samps = config.getint(classname, "samps", fallback=1000)
assert lb.shape[0] == ub.shape[0], "bounds are of different shapes!"
dim = lb.shape[0]
acqf = config.getobj("experiment",
"acqf",
fallback=MCLevelSetEstimation)
acqf_name = acqf.__name__
default_extra_acqf_args = {
"beta": 3.98,
"target": 0.75,
"objective": ProbitObjective,
}
extra_acqf_args = {
k: config.getobj(acqf_name,
k,
fallback_type=float,
fallback=v,
warn=False)
for k, v in default_extra_acqf_args.items()
}
extra_acqf_args = _prune_extra_acqf_args(acqf, extra_acqf_args)
if ("objective" in extra_acqf_args.keys()
and extra_acqf_args["objective"] is not None):
extra_acqf_args["objective"] = extra_acqf_args["objective"]()
return cls(
lb=lb,
ub=ub,
restarts=restarts,
samps=samps,
dim=dim,
acqf=acqf,
model=model,
extra_acqf_args=extra_acqf_args,
)
def test_1d_single_probit(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 15
n_opt = 20
lb = -4.0
ub = 4.0
acqf = BernoulliMCMutualInformation
extra_acqf_args = {"objective": ProbitObjective()}
model_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, dim=1, inducing_size=10),
generator=OptimizeAcqfGenerator(acqf, extra_acqf_args),
n_trials=n_opt,
),
]
strat = SequentialStrategy(model_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(f_1d(next_x))])
x = torch.linspace(-4, 4, 100)
zhat, _ = strat.predict(x)
true = f_1d(x.detach().numpy())
est = zhat.detach().numpy()
# close enough!
self.assertTrue((((norm.cdf(est) - true) ** 2).mean()) < 0.25)
def test_2d_single_probit_pure_exploration(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 50
n_opt = 1
lb = [-1, -1]
ub = [1, 1]
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(cdf_new_novel_det(next_x))])
xy = np.mgrid[-1:1:30j, -1:1:30j].reshape(2, -1).T
post_mean, _ = strat.predict(torch.Tensor(xy))
phi_post_mean = norm.cdf(post_mean.reshape(30, 30).detach().numpy())
phi_post_true = cdf_new_novel_det(xy)
self.assertTrue(
pearsonr(phi_post_mean.flatten(), phi_post_true.flatten())[0] > 0.9
)
def test_1d_single_probit_new_interface(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 50
n_opt = 1
lb = -4.0
ub = 4.0
model_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(model_list)
while not strat.finished:
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(f_1d(next_x))])
self.assertTrue(strat.y.shape[0] == n_init + n_opt)
x = torch.linspace(-4, 4, 100)
zhat, _ = strat.predict(x)
# true max is 0, very loose test
self.assertTrue(np.abs(x[np.argmax(zhat.detach().numpy())]) < 0.5)
def test_2d_single_probit(self):
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 150
n_opt = 1
lb = [-1, -1]
ub = [1, 1]
strat_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=20),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(strat_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(f_2d(next_x[None, :]))])
xy = np.mgrid[-1:1:30j, -1:1:30j].reshape(2, -1).T
zhat, _ = strat.predict(torch.Tensor(xy))
self.assertTrue(np.all(np.abs(xy[np.argmax(zhat.detach().numpy())]) < 0.5))
def test_extra_ask_warns(self):
# test that when we ask more times than we have models, we warn but keep going
seed = 1
torch.manual_seed(seed)
np.random.seed(seed)
n_init = 3
n_opt = 1
lb = -4.0
ub = 4.0
model_list = [
Strategy(
lb=lb,
ub=ub,
n_trials=n_init,
generator=SobolGenerator(lb=lb, ub=ub, seed=seed),
),
Strategy(
lb=lb,
ub=ub,
model=GPClassificationModel(lb=lb, ub=ub, inducing_size=10),
generator=OptimizeAcqfGenerator(
qUpperConfidenceBound, acqf_kwargs={"beta": 1.96}
),
n_trials=n_opt,
),
]
strat = SequentialStrategy(model_list)
for _i in range(n_init + n_opt):
next_x = strat.gen()
strat.add_data(next_x, [bernoulli.rvs(norm.cdf(f_1d(next_x)))])
with self.assertWarns(RuntimeWarning):
strat.gen()
def test_predict_p(self):
"""
Verify analytic p-space mean and var is correct.
"""
X, y = self.X, self.y
model = GPClassificationModel(
torch.Tensor([-3]), torch.Tensor([3]), inducing_size=10
)
model.fit(X, y)
pmean_analytic, pvar_analytic = model.predict(X, probability_space=True)
fsamps = model.sample(X, 150000)
psamps = norm.cdf(fsamps)
pmean_samp = psamps.mean(0)
pvar_samp = psamps.var(0)
# TODO these tolerances are a bit loose, verify this is right.
self.assertTrue(np.allclose(pmean_analytic, pmean_samp, atol=0.001))
self.assertTrue(np.allclose(pvar_analytic, pvar_samp, atol=0.001))
def test_reset_hyperparams(self):
model = GPClassificationModel(lb=[-3], ub=[3], inducing_size=20)
os_before = model.covar_module.outputscale.clone().detach().numpy()
ls_before = model.covar_module.base_kernel.lengthscale.clone().detach().numpy()
model.fit(torch.Tensor(self.X), torch.Tensor(self.y))
os_after = model.covar_module.outputscale.clone().detach().numpy()
ls_after = model.covar_module.base_kernel.lengthscale.clone().detach().numpy()
model._reset_hyperparameters()
os_reset = model.covar_module.outputscale.clone().detach().numpy()
ls_reset = model.covar_module.base_kernel.lengthscale.clone().detach().numpy()
# before should be different from after and after should be different
# from reset but before and reset should be same
self.assertFalse(np.allclose(os_before, os_after))
self.assertFalse(np.allclose(os_after, os_reset))
self.assertTrue(np.allclose(os_before, os_reset))
self.assertFalse(np.allclose(ls_before, ls_after))
self.assertFalse(np.allclose(ls_after, ls_reset))
self.assertTrue(np.allclose(ls_before, ls_reset))
class SingleProbitModelbridge(ModelBridge):
outcome_type = "single_probit"
def __init__(
self,
lb,
ub,
restarts=10,
samps=1000,
dim=1,
acqf=None,
extra_acqf_args=None,
model=None,
):
if extra_acqf_args is None:
extra_acqf_args = {}
super().__init__(lb=lb,
ub=ub,
dim=dim,
acqf=acqf,
extra_acqf_args=extra_acqf_args)
self.restarts = restarts
self.samps = samps
self.likelihood = gpytorch.likelihoods.BernoulliLikelihood()
if model is None:
self.model = GPClassificationModel(inducing_min=self.lb,
inducing_max=self.ub)
else:
self.model = model
def fit(self, train_x, train_y):
n = train_y.shape[0]
self.mll = gpytorch.mlls.VariationalELBO(self.likelihood, self.model,
n)
self.model.train()
self.model.set_train_data(train_x, train_y)
fit_gpytorch_model(self.mll)
def gen(self, num_points=1, **kwargs):
self.model.eval()
train_x = self.model.train_inputs[0]
acq = self._get_acquisition_fn()
new_candidate, batch_acq_values = optimize_acqf(
acq_function=acq,
bounds=torch.tensor(np.c_[self.lb, self.ub]).T.to(train_x),
q=num_points,
num_restarts=self.restarts,
raw_samples=self.samps,
)
return new_candidate.numpy()
def predict(self, x):
post = self.model.posterior(x)
return post.mean.squeeze(), post.variance.squeeze()
def sample(self, x, num_samples):
return self.model(x).rsample(torch.Size([num_samples]))
@classmethod
def from_config(cls, config):
classname = cls.__name__
model = GPClassificationModel.from_config(config)
lb = config.gettensor(classname, "lb")
ub = config.gettensor(classname, "ub")
restarts = config.getint(classname, "restarts", fallback=10)
samps = config.getint(classname, "samps", fallback=1000)
assert lb.shape[0] == ub.shape[0], "bounds are of different shapes!"
dim = lb.shape[0]
acqf = config.getobj("experiment",
"acqf",
fallback=MCLevelSetEstimation)
acqf_name = acqf.__name__
default_extra_acqf_args = {
"beta": 3.98,
"target": 0.75,
"objective": ProbitObjective,
}
extra_acqf_args = {
k: config.getobj(acqf_name,
k,
fallback_type=float,
fallback=v,
warn=False)
for k, v in default_extra_acqf_args.items()
}
extra_acqf_args = _prune_extra_acqf_args(acqf, extra_acqf_args)
if ("objective" in extra_acqf_args.keys()
and extra_acqf_args["objective"] is not None):
extra_acqf_args["objective"] = extra_acqf_args["objective"]()
return cls(
lb=lb,
ub=ub,
restarts=restarts,
samps=samps,
dim=dim,
acqf=acqf,
model=model,
extra_acqf_args=extra_acqf_args,
)
def test_select_inducing_points(self):
"""Verify that when we have n_induc > data size, we use data as inducing,
and otherwise we correctly select inducing points."""
X, y = make_classification(
n_samples=100,
n_features=1,
n_redundant=0,
n_informative=1,
random_state=1,
n_clusters_per_class=1,
)
X, y = torch.Tensor(X), torch.Tensor(y)
model = GPClassificationModel(
torch.Tensor([-3]), torch.Tensor([3]), inducing_size=20
)
model.set_train_data(X[:10, ...], y[:10])
# (inducing point selection sorts the inputs so we sort X to verify)
self.assertTrue(
np.allclose(
model._select_inducing_points(method="auto"),
X[:10].sort(0).values,
)
)
model.set_train_data(X, y)
self.assertTrue(len(model._select_inducing_points(method="auto")) <= 20)
self.assertTrue(len(model._select_inducing_points(method="pivoted_chol")) <= 20)
self.assertEqual(len(model._select_inducing_points(method="kmeans++")), 20)
with self.assertRaises(AssertionError):
model._select_inducing_points(method="12345")
def plot_prior_samps_2d():
config = Config(
config_dict={
"common": {
"outcome_type": "single_probit",
"target": 0.75,
"lb": "[-3, -3]",
"ub": "[3, 3]",
},
"default_mean_covar_factory": {},
"song_mean_covar_factory": {},
"monotonic_mean_covar_factory": {"monotonic_idxs": "[1]"},
}
)
lb = torch.Tensor([-3, -3])
ub = torch.Tensor([3, 3])
nsamps = 5
gridsize = 30
grid = _dim_grid(lower=lb, upper=ub, dim=2, gridsize=gridsize)
np.random.seed(global_seed)
torch.random.manual_seed(global_seed)
with gpytorch.settings.prior_mode(True):
rbf_mean, rbf_covar = default_mean_covar_factory(config)
rbf_model = GPClassificationModel(
inducing_min=lb,
inducing_max=ub,
inducing_size=100,
mean_module=rbf_mean,
covar_module=rbf_covar,
)
# add just two samples at high and low
rbf_model.set_train_data(torch.Tensor([-3, -3])[:, None], torch.LongTensor([0]))
rbf_samps = rbf_model(grid).sample(torch.Size([nsamps]))
song_mean, song_covar = song_mean_covar_factory(config)
song_model = GPClassificationModel(
inducing_min=lb,
inducing_max=ub,
inducing_size=100,
mean_module=song_mean,
covar_module=song_covar,
)
song_model.set_train_data(
torch.Tensor([-3, -3])[:, None], torch.LongTensor([0])
)
song_samps = song_model(grid).sample(torch.Size([nsamps]))
mono_mean, mono_covar = monotonic_mean_covar_factory(config)
mono_model = MonotonicRejectionGP(
likelihood="probit-bernoulli",
monotonic_idxs=[1],
mean_module=mono_mean,
covar_module=mono_covar,
num_induc=1000,
)
bounds_ = torch.tensor([-3.0, -3.0, 3.0, 3.0]).reshape(2, -1)
# Select inducing points
mono_model.inducing_points = draw_sobol_samples(
bounds=bounds_, n=mono_model.num_induc, q=1
).squeeze(1)
inducing_points_aug = mono_model._augment_with_deriv_index(
mono_model.inducing_points, 0
)
scales = ub - lb
dummy_train_x = mono_model._augment_with_deriv_index(
torch.Tensor([-3, 3])[None, :], 0
)
mono_model.model = MixedDerivativeVariationalGP(
train_x=dummy_train_x,
train_y=torch.LongTensor([0]),
inducing_points=inducing_points_aug,
scales=scales,
fixed_prior_mean=torch.Tensor([0.75]),
covar_module=mono_covar,
mean_module=mono_mean,
)
mono_samps = mono_model.sample(grid, nsamps)
intensity_grid = np.linspace(-3, 3, gridsize)
fig, ax = plt.subplots(1, 3, figsize=(7.5, 3))
fig.tight_layout(rect=[0, 0.03, 1, 0.9])
fig.suptitle("Prior samples")
square_samps = np.array([s.reshape((gridsize,) * 2).numpy() for s in song_samps])
plotsamps = norm.cdf(square_samps[:, ::5, :]).T.reshape(gridsize, -1)
ax[0].plot(intensity_grid, plotsamps, "b")
ax[0].set_title("Linear kernel model")
square_samps = np.array([s.reshape((gridsize,) * 2).numpy() for s in rbf_samps])
plotsamps = norm.cdf(square_samps[:, ::5, :]).T.reshape(gridsize, -1)
ax[1].plot(intensity_grid, plotsamps, "b")
ax[1].set_title("Nonmonotonic RBF kernel model")
square_samps = np.array([s.reshape((gridsize,) * 2).numpy() for s in mono_samps])
plotsamps = norm.cdf(square_samps[:, ::5, :]).T.reshape(gridsize, -1)
ax[2].plot(intensity_grid, plotsamps, "b")
ax[2].set_title("Monotonic RBF kernel model")
return fig
def plot_prior_samps_1d():
config = Config(
config_dict={
"common": {
"outcome_type": "single_probit",
"target": 0.75,
"lb": "[-3]",
"ub": "[3]",
},
"default_mean_covar_factory": {},
"song_mean_covar_factory": {},
"monotonic_mean_covar_factory": {"monotonic_idxs": "[0]"},
}
)
lb = torch.Tensor([-3])
ub = torch.Tensor([3])
nsamps = 10
gridsize = 50
grid = _dim_grid(lower=lb, upper=ub, dim=1, gridsize=gridsize)
np.random.seed(global_seed)
torch.random.manual_seed(global_seed)
with gpytorch.settings.prior_mode(True):
rbf_mean, rbf_covar = default_mean_covar_factory(config)
rbf_model = GPClassificationModel(
inducing_min=lb,
inducing_max=ub,
inducing_size=100,
mean_module=rbf_mean,
covar_module=rbf_covar,
)
# add just two samples at high and low
rbf_model.set_train_data(
torch.Tensor([-3, 3])[:, None], torch.LongTensor([0, 1])
)
rbf_samps = rbf_model(grid).sample(torch.Size([nsamps]))
song_mean, song_covar = song_mean_covar_factory(config)
song_model = GPClassificationModel(
inducing_min=lb,
inducing_max=ub,
inducing_size=100,
mean_module=song_mean,
covar_module=song_covar,
)
song_model.set_train_data(
torch.Tensor([-3, 3])[:, None], torch.LongTensor([0, 1])
)
song_samps = song_model(grid).sample(torch.Size([nsamps]))
mono_mean, mono_covar = monotonic_mean_covar_factory(config)
mono_model = MonotonicRejectionGP(
likelihood="probit-bernoulli",
monotonic_idxs=[0],
mean_module=mono_mean,
covar_module=mono_covar,
)
bounds_ = torch.tensor([-3.0, 3.0])[:, None]
# Select inducing points
mono_model.inducing_points = draw_sobol_samples(
bounds=bounds_, n=mono_model.num_induc, q=1
).squeeze(1)
inducing_points_aug = mono_model._augment_with_deriv_index(
mono_model.inducing_points, 0
)
scales = ub - lb
dummy_train_x = mono_model._augment_with_deriv_index(
torch.Tensor([-3, 3])[:, None], 0
)
mono_model.model = MixedDerivativeVariationalGP(
train_x=dummy_train_x,
train_y=torch.LongTensor([0, 1]),
inducing_points=inducing_points_aug,
scales=scales,
fixed_prior_mean=torch.Tensor([0.75]),
covar_module=mono_covar,
mean_module=mono_mean,
)
mono_samps = mono_model.sample(grid, nsamps)
fig, ax = plt.subplots(1, 3, figsize=(7.5, 3))
fig.tight_layout(rect=[0.01, 0.03, 1, 0.9])
fig.suptitle("GP prior samples (probit-transformed)")
ax[0].plot(grid.squeeze(), norm.cdf(song_samps.T), "b")
ax[0].set_ylabel("Response Probability")
ax[0].set_title("Linear kernel")
ax[1].plot(grid.squeeze(), norm.cdf(rbf_samps.T), "b")
ax[1].set_xlabel("Intensity")
ax[1].set_title("RBF kernel (nonmonotonic)")
ax[2].plot(grid.squeeze(), norm.cdf(mono_samps.T), "b")
ax[2].set_title("RBF kernel (monotonic)")
return fig