def test_standardize(self, cuda=False): tkwargs = {"device": torch.device("cuda" if cuda else "cpu")} for dtype in (torch.float, torch.double): tkwargs["dtype"] = dtype X = torch.tensor([0.0, 0.0], **tkwargs) self.assertTrue(torch.equal(X, standardize(X))) X2 = torch.tensor([0.0, 1.0, 1.0, 1.0], **tkwargs) expected_X2_stdized = torch.tensor([-1.5, 0.5, 0.5, 0.5], **tkwargs) self.assertTrue(torch.equal(expected_X2_stdized, standardize(X2))) X3 = torch.tensor([[0.0, 1.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]], **tkwargs).transpose(1, 0) X3_stdized = standardize(X3) self.assertTrue(torch.equal(X3_stdized[:, 0], expected_X2_stdized)) self.assertTrue( torch.equal(X3_stdized[:, 1], torch.zeros(4, **tkwargs)))
def forward(self, X: Tensor, num_samples: int = 1) -> Tensor: r"""Sample from a tempered value of the acquisition function value. Args: X: A `batch_shape x N x d`-dim Tensor from which to sample (in the `N` dimension) according to the maximum posterior value under the objective. Note that if a batched model is used in the underlying acquisition function, then its batch shape must be broadcastable to `batch_shape`. num_samples: The number of samples to draw. Returns: A `batch_shape x num_samples x d`-dim Tensor of samples from `X`, where `X[..., i, :]` is the `i`-th sample. """ # TODO: Can we get the model batch shape property from the model? # we move the `N` dimension to the front for evaluating the acquisition function # so that X_eval has shape `N x batch_shape x 1 x d` X_eval = X.permute(-2, *range(X.ndim - 2), -1).unsqueeze(-2) acqval = self.acq_func(X_eval) # N x batch_shape # now move the `N` dimension back (this is the number of categories) acqval = acqval.permute(*range(1, X.ndim - 1), 0) # batch_shape x N weights = torch.exp(self.eta * standardize(acqval)) # batch_shape x N idcs = batched_multinomial(weights=weights, num_samples=num_samples, replacement=self.replacement) # now do some gathering acrobatics to select the right elements from X return torch.gather(X, -2, idcs.unsqueeze(-1).expand(*idcs.shape, X.size(-1)))
def standardize_obs(self, observations, new_y): '''Takes a tensor of observations, extracts the y values and spits out the standardised value of a new observation ''' y_vals = np.array([obs[1] for obs in observations]) t_y_vals = torch.tensor(y_vals).double() augmented_obs = torch.cat( (t_y_vals, torch.tensor([new_y]).double())) standardized_y = standardize(augmented_obs)[-1] return (standardized_y)
def test_standardize(self): for dtype in (torch.float, torch.double): tkwargs = {"device": self.device, "dtype": dtype} Y = torch.tensor([0.0, 0.0], **tkwargs) self.assertTrue(torch.equal(Y, standardize(Y))) Y2 = torch.tensor([0.0, 1.0, 1.0, 1.0], **tkwargs) expected_Y2_stdized = torch.tensor([-1.5, 0.5, 0.5, 0.5], **tkwargs) self.assertTrue(torch.equal(expected_Y2_stdized, standardize(Y2))) Y3 = torch.tensor( [[0.0, 1.0, 1.0, 1.0], [0.0, 0.0, 0.0, 0.0]], **tkwargs ).transpose(1, 0) Y3_stdized = standardize(Y3) self.assertTrue(torch.equal(Y3_stdized[:, 0], expected_Y2_stdized)) self.assertTrue(torch.equal(Y3_stdized[:, 1], torch.zeros(4, **tkwargs))) Y4 = torch.cat([Y3, Y2.unsqueeze(-1)], dim=-1) Y4_stdized = standardize(Y4) self.assertTrue(torch.equal(Y4_stdized[:, 0], expected_Y2_stdized)) self.assertTrue(torch.equal(Y4_stdized[:, 1], torch.zeros(4, **tkwargs))) self.assertTrue(torch.equal(Y4_stdized[:, 2], expected_Y2_stdized))
def generate_outer_restart_points(self, acqf: OneShotrhoKG, w_samples: Tensor = None) -> Tensor: """ Generates the restart points for acqf optimization. :param acqf: The acquisition function being optimized :param w_samples: the list of w samples to use :return: restart points """ X = draw_constrained_sobol( bounds=self.outer_bounds, n=self.raw_samples, q=self.q, inequality_constraints=self.inequality_constraints, ).to(dtype=self.dtype, device=self.device) # get the optimizers of the inner problem if w_samples is None: w_samples = (acqf.fixed_samples if acqf.fixed_samples is not None else torch.rand(acqf.num_samples, acqf.dim_w, dtype=self.dtype, device=self.device)) inner_rho = InnerRho( model=acqf.model, w_samples=w_samples, alpha=acqf.alpha, dim_x=acqf.dim_x, num_repetitions=acqf.num_repetitions, inner_seed=acqf.inner_seed, CVaR=acqf.CVaR, expectation=acqf.expectation, weights=getattr(acqf, "weights", None), ) inner_solutions, inner_values = super().optimize_inner( inner_rho, False) # sample from the optimizers n_value = int((1 - self.random_frac) * self.num_fantasies) weights = torch.exp(self.eta * standardize(inner_values)) idx = torch.multinomial(weights, self.raw_samples * n_value, replacement=True) # set the respective raw samples to the sampled optimizers X[..., -n_value * self.dim_x:] = inner_solutions[idx, 0].view( self.raw_samples, 1, -1) if w_samples is not None: w_ind = torch.randint(w_samples.shape[0], (self.raw_samples, self.q)) if self.q > 1: raise NotImplementedError("This does not support q>1!") X[..., self.dim_x:self.dim] = w_samples[w_ind, :] return self.generate_restart_points_from_samples(X, acqf)
def gen_one_shot_kg_initial_conditions(acq_function, bounds, q, num_restarts, raw_samples, options=None): r"""[Copy of original botorch function] Generate a batch of smart initializations for qKnowledgeGradient. This function generates initial conditions for optimizing one-shot KG using the maximizer of the posterior objective. Intutively, the maximizer of the fantasized posterior will often be close to a maximizer of the current posterior. This function uses that fact to generate the initital conditions for the fantasy points. Specifically, a fraction of `1 - frac_random` (see options) is generated by sampling from the set of maximizers of the posterior objective (obtained via random restart optimization) according to a softmax transformation of their respective values. This means that this initialization strategy internally solves an acquisition function maximization problem. The remaining `frac_random` fantasy points as well as all `q` candidate points are chosen according to the standard initialization strategy in `gen_batch_initial_conditions`. Args: acq_function: The qKnowledgeGradient instance to be optimized. bounds: A `2 x d` tensor of lower and upper bounds for each column of task features. q: The number of candidates to consider. num_restarts: The number of starting points for multistart acquisition function optimization. raw_samples: The number of raw samples to consider in the initialization heuristic. options: Options for initial condition generation. These contain all settings for the standard heuristic initialization from `gen_batch_initial_conditions`. In addition, they contain `frac_random` (the fraction of fully random fantasy points), `num_inner_restarts` and `raw_inner_samples` (the number of random restarts and raw samples for solving the posterior objective maximization problem, respectively) and `eta` (temperature parameter for sampling heuristic from posterior objective maximizers). Returns: A `num_restarts x q' x d` tensor that can be used as initial conditions for `optimize_acqf()`. Here `q' = q + num_fantasies` is the total number of points (candidate points plus fantasy points). Example: >>> qKG = qKnowledgeGradient(model, num_fantasies=64) >>> bounds = torch.tensor([[0., 0.], [1., 1.]]) >>> Xinit = gen_one_shot_kg_initial_conditions( >>> qKG, bounds, q=3, num_restarts=10, raw_samples=512, >>> options={"frac_random": 0.25}, >>> ) """ options = options or {} frac_random: float = options.get("frac_random", 0.1) if not 0 < frac_random < 1: raise ValueError( f"frac_random must take on values in (0,1). Value: {frac_random}") q_aug = acq_function.get_augmented_q_batch_size(q=q) # TODO: Avoid unnecessary computation by not generating all candidates ics = gen_batch_initial_conditions( acq_function=acq_function, bounds=bounds, q=q_aug, num_restarts=num_restarts, raw_samples=raw_samples, options=options, ) # compute maximizer of the value function value_function = _get_value_function( model=acq_function.model, objective=acq_function.objective, sampler=acq_function.inner_sampler, ) fantasy_cands, fantasy_vals = optimize_acqf( acq_function=value_function, bounds=bounds, q=1, num_restarts=options.get("num_inner_restarts", 20), raw_samples=options.get("raw_inner_samples", 1024), return_best_only=False, ) # sampling from the optimizers n_value = int((1 - frac_random) * (q_aug - q)) # number of non-random ICs eta = options.get("eta", 2.0) weights = torch.exp(eta * transforms.standardize(fantasy_vals)) idx = torch.multinomial(weights, num_restarts * n_value, replacement=True) # set the respective initial conditions to the sampled optimizers ics[..., -n_value:, :] = fantasy_cands[idx, 0].view(num_restarts, n_value, -1) return ics
def test_cache_root(self): sample_cached_path = ( "botorch.acquisition.cached_cholesky.sample_cached_cholesky") raw_state_dict = { "likelihood.noise_covar.raw_noise": torch.tensor([[0.0895], [0.2594]], dtype=torch.float64), "mean_module.constant": torch.tensor([[-0.4545], [-0.1285]], dtype=torch.float64), "covar_module.raw_outputscale": torch.tensor([1.4876, 1.4897], dtype=torch.float64), "covar_module.base_kernel.raw_lengthscale": torch.tensor([[[-0.7202, -0.2868]], [[-0.8794, -1.2877]]], dtype=torch.float64), } # test batched models (e.g. for MCMC) for train_batch_shape, m, dtype in product( (torch.Size([]), torch.Size([3])), (1, 2), (torch.float, torch.double)): state_dict = deepcopy(raw_state_dict) for k, v in state_dict.items(): if m == 1: v = v[0] if len(train_batch_shape) > 0: v = v.unsqueeze(0).expand(*train_batch_shape, *v.shape) state_dict[k] = v tkwargs = {"device": self.device, "dtype": dtype} if m == 2: objective = GenericMCObjective(lambda Y, X: Y.sum(dim=-1)) else: objective = None for k, v in state_dict.items(): state_dict[k] = v.to(**tkwargs) all_close_kwargs = ({ "atol": 1e-1, "rtol": 0.0, } if dtype == torch.float else { "atol": 1e-4, "rtol": 0.0 }) torch.manual_seed(1234) train_X = torch.rand(*train_batch_shape, 3, 2, **tkwargs) train_Y = ( torch.sin(train_X * 2 * pi) + torch.randn(*train_batch_shape, 3, 2, **tkwargs))[..., :m] train_Y = standardize(train_Y) model = SingleTaskGP( train_X, train_Y, ) if len(train_batch_shape) > 0: X_baseline = train_X[0] else: X_baseline = train_X model.load_state_dict(state_dict, strict=False) # test sampler with collapse_batch_dims=False sampler = IIDNormalSampler(5, seed=0, collapse_batch_dims=False) with self.assertRaises(UnsupportedError): qNoisyExpectedImprovement( model=model, X_baseline=X_baseline, sampler=sampler, objective=objective, prune_baseline=False, cache_root=True, ) sampler = IIDNormalSampler(5, seed=0) torch.manual_seed(0) acqf = qNoisyExpectedImprovement( model=model, X_baseline=X_baseline, sampler=sampler, objective=objective, prune_baseline=False, cache_root=True, ) orig_base_samples = acqf.base_sampler.base_samples.detach().clone() sampler2 = IIDNormalSampler(5, seed=0) sampler2.base_samples = orig_base_samples torch.manual_seed(0) acqf_no_cache = qNoisyExpectedImprovement( model=model, X_baseline=X_baseline, sampler=sampler2, objective=objective, prune_baseline=False, cache_root=False, ) for q, batch_shape in product( (1, 3), (torch.Size([]), torch.Size([3]), torch.Size([4, 3]))): test_X = (0.3 + 0.05 * torch.randn(*batch_shape, q, 2, **tkwargs) ).requires_grad_(True) with mock.patch( sample_cached_path, wraps=sample_cached_cholesky) as mock_sample_cached: torch.manual_seed(0) val = acqf(test_X) mock_sample_cached.assert_called_once() val.sum().backward() base_samples = acqf.sampler.base_samples.detach().clone() X_grad = test_X.grad.clone() test_X2 = test_X.detach().clone().requires_grad_(True) acqf_no_cache.sampler.base_samples = base_samples with mock.patch( sample_cached_path, wraps=sample_cached_cholesky) as mock_sample_cached: torch.manual_seed(0) val2 = acqf_no_cache(test_X2) mock_sample_cached.assert_not_called() self.assertTrue(torch.allclose(val, val2, **all_close_kwargs)) val2.sum().backward() self.assertTrue( torch.allclose(X_grad, test_X2.grad, **all_close_kwargs)) # test we fall back to standard sampling for # ill-conditioned covariances acqf._baseline_L = torch.zeros_like(acqf._baseline_L) with warnings.catch_warnings( record=True) as ws, settings.debug(True): with torch.no_grad(): acqf(test_X) self.assertEqual(len(ws), 1) self.assertTrue(issubclass(ws[-1].category, BotorchWarning))
def train_loop(self): from botorch.models import SingleTaskGP from botorch.fit import fit_gpytorch_model from gpytorch.mlls import ExactMarginalLogLikelihood from botorch.optim import optimize_acqf from botorch.acquisition.monte_carlo import qExpectedImprovement from botorch.sampling.samplers import SobolQMCNormalSampler seed = 1 torch.manual_seed(seed) dt, d = torch.float32, 3 lb, ub = [1e-4, 0.1, 0.1], [3e-3, 1 - 1e-3, 1 - 1e-3] bounds = torch.tensor([lb, ub], dtype=dt) def gen_initial_data(): # auto # x = unnormalize(torch.rand(1, 3, dtype=dt), bounds=bounds) # manual x = torch.tensor([[1e-3, 0.9, 0.999]]) print('BO Initialization: \n') print('Initial Hyper-parameter: ' + str(x)) obj = self.train(x.view(-1)) print('Initial Error: ' + str(obj)) return x, obj.unsqueeze(1) def get_fitted_model(x, obj, state_dict=None): # initialize and fit model fitted_model = SingleTaskGP(train_X=x, train_Y=obj) if state_dict is not None: fitted_model.load_state_dict(state_dict) mll = ExactMarginalLogLikelihood(fitted_model.likelihood, fitted_model) mll.to(x) fit_gpytorch_model(mll) return fitted_model def optimize_acqf_and_get_observation(acq_func): """Optimizes the acquisition function, and returns a new candidate and a noisy observation""" candidates, _ = optimize_acqf( acq_function=acq_func, bounds=torch.stack([ torch.zeros(d, dtype=dt), torch.ones(d, dtype=dt), ]), q=1, num_restarts=10, raw_samples=200, ) x = unnormalize(candidates.detach(), bounds=bounds) print('Hyper-parameter: ' + str(x)) obj = self.train(x.view(-1)).unsqueeze(-1) print(print('Error: ' + str(obj))) return x, obj N_BATCH = 500 MC_SAMPLES = 2000 best_observed = [] train_x, train_obj = gen_initial_data() # (1,3), (1,1) best_observed.append(train_obj.view(-1)) print(f"\nRunning BO......\n ", end='') state_dict = None for iteration in range(N_BATCH): # fit the model model = get_fitted_model( normalize(train_x, bounds=bounds), standardize(train_obj), state_dict=state_dict, ) # define the qNEI acquisition module using a QMC sampler qmc_sampler = SobolQMCNormalSampler(num_samples=MC_SAMPLES, seed=seed) qEI = qExpectedImprovement(model=model, sampler=qmc_sampler, best_f=standardize(train_obj).max()) # optimize and get new observation new_x, new_obj = optimize_acqf_and_get_observation(qEI) # update training points train_x = torch.cat((train_x, new_x)) train_obj = torch.cat((train_obj, new_obj)) # update progress best_value = train_obj.max().item() best_observed.append(best_value) state_dict = model.state_dict() print(".", end='') print(best_observed)
def gen_value_function_initial_conditions( acq_function: AcquisitionFunction, bounds: Tensor, num_restarts: int, raw_samples: int, current_model: Model, options: Optional[Dict[str, Union[bool, float, int]]] = None, ) -> Tensor: r"""Generate a batch of smart initializations for optimizing the value function of qKnowledgeGradient. This function generates initial conditions for optimizing the inner problem of KG, i.e. its value function, using the maximizer of the posterior objective. Intutively, the maximizer of the fantasized posterior will often be close to a maximizer of the current posterior. This function uses that fact to generate the initital conditions for the fantasy points. Specifically, a fraction of `1 - frac_random` (see options) of raw samples is generated by sampling from the set of maximizers of the posterior objective (obtained via random restart optimization) according to a softmax transformation of their respective values. This means that this initialization strategy internally solves an acquisition function maximization problem. The remaining raw samples are generated using `draw_sobol_samples`. All raw samples are then evaluated, and the initial conditions are selected according to the standard initialization strategy in 'initialize_q_batch' individually for each inner problem. Args: acq_function: The value function instance to be optimized. bounds: A `2 x d` tensor of lower and upper bounds for each column of task features. num_restarts: The number of starting points for multistart acquisition function optimization. raw_samples: The number of raw samples to consider in the initialization heuristic. current_model: The model of the KG acquisition function that was used to generate the fantasy model of the value function. options: Options for initial condition generation. These contain all settings for the standard heuristic initialization from `gen_batch_initial_conditions`. In addition, they contain `frac_random` (the fraction of fully random fantasy points), `num_inner_restarts` and `raw_inner_samples` (the number of random restarts and raw samples for solving the posterior objective maximization problem, respectively) and `eta` (temperature parameter for sampling heuristic from posterior objective maximizers). Returns: A `num_restarts x batch_shape x q x d` tensor that can be used as initial conditions for `optimize_acqf()`. Here `batch_shape` is the batch shape of value function model. Example: >>> fant_X = torch.rand(5, 1, 2) >>> fantasy_model = model.fantasize(fant_X, SobolQMCNormalSampler(16)) >>> value_function = PosteriorMean(fantasy_model) >>> bounds = torch.tensor([[0., 0.], [1., 1.]]) >>> Xinit = gen_value_function_initial_conditions( >>> value_function, bounds, num_restarts=10, raw_samples=512, >>> options={"frac_random": 0.25}, >>> ) """ options = options or {} seed: Optional[int] = options.get("seed") frac_random: float = options.get("frac_random", 0.6) if not 0 < frac_random < 1: raise ValueError( f"frac_random must take on values in (0,1). Value: {frac_random}") # compute maximizer of the current value function value_function = _get_value_function( model=current_model, objective=acq_function.objective, sampler=getattr(acq_function, "sampler", None), project=getattr(acq_function, "project", None), ) from botorch.optim.optimize import optimize_acqf fantasy_cands, fantasy_vals = optimize_acqf( acq_function=value_function, bounds=bounds, q=1, num_restarts=options.get("num_inner_restarts", 20), raw_samples=options.get("raw_inner_samples", 1024), return_best_only=False, options={ k: v for k, v in options.items() if k not in ("frac_random", "num_inner_restarts", "raw_inner_samples", "eta") }, ) batch_shape = acq_function.model.batch_shape # sampling from the optimizers n_value = int((1 - frac_random) * raw_samples) # number of non-random ICs if n_value > 0: eta = options.get("eta", 2.0) weights = torch.exp(eta * standardize(fantasy_vals)) idx = batched_multinomial( weights=weights.expand(*batch_shape, -1), num_samples=n_value, replacement=True, ).permute(-1, *range(len(batch_shape))) resampled = fantasy_cands[idx] else: resampled = torch.empty(0, *batch_shape, 1, bounds.shape[-1], dtype=bounds.dtype) # add qMC samples randomized = draw_sobol_samples(bounds=bounds, n=raw_samples - n_value, q=1, batch_shape=batch_shape, seed=seed) # full set of raw samples X_rnd = torch.cat([resampled, randomized], dim=0) # evaluate the raw samples with torch.no_grad(): Y_rnd = acq_function(X_rnd) # select the restart points using the heuristic return initialize_q_batch(X=X_rnd, Y=Y_rnd, n=num_restarts, eta=options.get("eta", 2.0))