def test_exponential_moving_average(self): for dtype in (torch.float, torch.double): tkwargs = {"device": self.device, "dtype": dtype} # test max iter sc = ExpMAStoppingCriterion(maxiter=2) self.assertEqual(sc.maxiter, 2) self.assertEqual(sc.n_window, 10) self.assertEqual(sc.rel_tol, 1e-5) self.assertFalse(sc.evaluate(fvals=torch.ones(1, **tkwargs))) self.assertTrue(sc.evaluate(fvals=torch.zeros(1, **tkwargs))) # test convergence n_window = 4 for minimize in (True, False): # test basic sc = ExpMAStoppingCriterion(minimize=minimize, n_window=n_window, rel_tol=0.0375) self.assertEqual(sc.rel_tol, 0.0375) self.assertIsNone(sc._prev_fvals) weights_exp = torch.tensor([0.1416, 0.1976, 0.2758, 0.3849]) self.assertTrue( torch.allclose(sc.weights, weights_exp, atol=1e-4)) f_vals = 1 + torch.linspace(1, 0, 25, **tkwargs)**2 if not minimize: f_vals = -f_vals for i, fval in enumerate(f_vals): if sc.evaluate(fval): self.assertEqual(i, 10) break # test multiple components sc = ExpMAStoppingCriterion(minimize=minimize, n_window=n_window, rel_tol=0.0375) df = torch.linspace(0, 0.1, 25, **tkwargs) if not minimize: df = -df f_vals = torch.stack([f_vals, f_vals + df], dim=-1) for i, fval in enumerate(f_vals): if sc.evaluate(fval): self.assertEqual(i, 10) break
def optimize_acqf_cyclic( acq_function: AcquisitionFunction, bounds: Tensor, q: int, num_restarts: int, raw_samples: int, options: Optional[Dict[str, Union[bool, float, int, str]]] = None, inequality_constraints: Optional[List[Tuple[Tensor, Tensor, float]]] = None, equality_constraints: Optional[List[Tuple[Tensor, Tensor, float]]] = None, fixed_features: Optional[Dict[int, float]] = None, post_processing_func: Optional[Callable[[Tensor], Tensor]] = None, batch_initial_conditions: Optional[Tensor] = None, cyclic_options: Optional[Dict[str, Union[bool, float, int, str]]] = None, ) -> float: r"""Generate a set of `q` candidates via cyclic optimization. Args: acq_function: An AcquisitionFunction bounds: A `2 x d` tensor of lower and upper bounds for each column of `X`. q: The number of candidates. num_restarts: Number of starting points for multistart acquisition function optimization. raw_samples: Number of samples for initialization options: Options for candidate generation. inequality constraints: A list of tuples (indices, coefficients, rhs), with each tuple encoding an inequality constraint of the form `\sum_i (X[indices[i]] * coefficients[i]) >= rhs` equality constraints: A list of tuples (indices, coefficients, rhs), with each tuple encoding an inequality constraint of the form `\sum_i (X[indices[i]] * coefficients[i]) = rhs` fixed_features: A map `{feature_index: value}` for features that should be fixed to a particular value during generation. post_processing_func: A function that post-processes an optimization result appropriately (i.e., according to `round-trip` transformations). batch_initial_conditions: A tensor to specify the initial conditions. If no initial conditions are provided, the default initialization will be used. cyclic_options: Options for stopping criterion for outer cyclic optimization. Returns: A two-element tuple containing - a `q x d`-dim tensor of generated candidates. - a `q`-dim tensor of expected acquisition values, where the `i`th value is the acquistion value conditional on having observed all candidates except candidate `i`. Example: >>> # generate `q=3` candidates cyclically using 15 random restarts >>> # 256 raw samples, and 4 cycles >>> >>> qEI = qExpectedImprovement(model, best_f=0.2) >>> bounds = torch.tensor([[0.], [1.]]) >>> candidates, acq_value_list = optimize_acqf_cyclic( >>> qEI, bounds, 3, 15, 256, cyclic_options={"maxiter": 4} >>> ) """ # for the first cycle, optimize the q candidates sequentially candidates, acq_vals = optimize_acqf( acq_function=acq_function, bounds=bounds, q=q, num_restarts=num_restarts, raw_samples=raw_samples, options=options, inequality_constraints=inequality_constraints, equality_constraints=equality_constraints, fixed_features=fixed_features, post_processing_func=post_processing_func, batch_initial_conditions=batch_initial_conditions, return_best_only=True, sequential=True, ) if q > 1: cyclic_options = cyclic_options or {} stopping_criterion = ExpMAStoppingCriterion(**cyclic_options) stop = stopping_criterion.evaluate(fvals=acq_vals) base_X_pending = acq_function.X_pending idxr = torch.ones(q, dtype=torch.bool, device=bounds.device) while not stop: for i in range(q): # optimize only candidate i idxr[i] = 0 acq_function.set_X_pending( torch.cat([base_X_pending, candidates[idxr]], dim=-2) if base_X_pending is not None else candidates[idxr] ) candidate_i, acq_val_i = optimize_acqf( acq_function=acq_function, bounds=bounds, q=1, num_restarts=num_restarts, raw_samples=raw_samples, options=options, inequality_constraints=inequality_constraints, equality_constraints=equality_constraints, fixed_features=fixed_features, post_processing_func=post_processing_func, batch_initial_conditions=candidates[i].unsqueeze(0), return_best_only=True, sequential=True, ) candidates[i] = candidate_i acq_vals[i] = acq_val_i idxr[i] = 1 stop = stopping_criterion.evaluate(fvals=acq_vals) acq_function.set_X_pending(base_X_pending) return candidates, acq_vals
def fit_gpytorch_torch( mll: MarginalLogLikelihood, bounds: Optional[ParameterBounds] = None, optimizer_cls: Optimizer = Adam, options: Optional[Dict[str, Any]] = None, track_iterations: bool = True, approx_mll: bool = True, ) -> Tuple[MarginalLogLikelihood, Dict[str, Union[float, List[OptimizationIteration]]]]: r"""Fit a gpytorch model by maximizing MLL with a torch optimizer. The model and likelihood in mll must already be in train mode. Note: this method requires that the model has `train_inputs` and `train_targets`. Args: mll: MarginalLogLikelihood to be maximized. bounds: A ParameterBounds dictionary mapping parameter names to tuples of lower and upper bounds. Bounds specified here take precedence over bounds on the same parameters specified in the constraints registered with the module. optimizer_cls: Torch optimizer to use. Must not require a closure. options: options for model fitting. Relevant options will be passed to the `optimizer_cls`. Additionally, options can include: "disp" to specify whether to display model fitting diagnostics and "maxiter" to specify the maximum number of iterations. track_iterations: Track the function values and wall time for each iteration. approx_mll: If True, use gpytorch's approximate MLL computation ( according to the gpytorch defaults based on the training at size). Unlike for the deterministic algorithms used in fit_gpytorch_scipy, this is not an issue for stochastic optimizers. Returns: 2-element tuple containing - mll with parameters optimized in-place. - Dictionary with the following key/values: "fopt": Best mll value. "wall_time": Wall time of fitting. "iterations": List of OptimizationIteration objects with information on each iteration. If track_iterations is False, will be empty. Example: >>> gp = SingleTaskGP(train_X, train_Y) >>> mll = ExactMarginalLogLikelihood(gp.likelihood, gp) >>> mll.train() >>> fit_gpytorch_torch(mll) >>> mll.eval() """ optim_options = {"maxiter": 100, "disp": True, "lr": 0.05} optim_options.update(options or {}) exclude = optim_options.pop("exclude", None) if exclude is not None: mll_params = [ t for p_name, t in mll.named_parameters() if p_name not in exclude ] else: mll_params = list(mll.parameters()) optimizer = optimizer_cls( params=[{"params": mll_params}], **_filter_kwargs(optimizer_cls, **optim_options), ) # get bounds specified in model (if any) bounds_: ParameterBounds = {} if hasattr(mll, "named_parameters_and_constraints"): for param_name, _, constraint in mll.named_parameters_and_constraints(): if constraint is not None and not constraint.enforced: bounds_[param_name] = constraint.lower_bound, constraint.upper_bound # update with user-supplied bounds (overwrites if already exists) if bounds is not None: bounds_.update(bounds) iterations = [] t1 = time.time() param_trajectory: Dict[str, List[Tensor]] = { name: [] for name, param in mll.named_parameters() } loss_trajectory: List[float] = [] i = 0 stop = False stopping_criterion = ExpMAStoppingCriterion( **_filter_kwargs(ExpMAStoppingCriterion, **optim_options) ) train_inputs, train_targets = mll.model.train_inputs, mll.model.train_targets while not stop: optimizer.zero_grad() with gpt_settings.fast_computations(log_prob=approx_mll): output = mll.model(*train_inputs) # we sum here to support batch mode args = [output, train_targets] + _get_extra_mll_args(mll) loss = -mll(*args).sum() loss.backward() loss_trajectory.append(loss.item()) for name, param in mll.named_parameters(): param_trajectory[name].append(param.detach().clone()) if optim_options["disp"] and ( (i + 1) % 10 == 0 or i == (optim_options["maxiter"] - 1) ): print(f"Iter {i + 1}/{optim_options['maxiter']}: {loss.item()}") if track_iterations: iterations.append(OptimizationIteration(i, loss.item(), time.time() - t1)) optimizer.step() # project onto bounds: if bounds_: for pname, param in mll.named_parameters(): if pname in bounds_: param.data = param.data.clamp(*bounds_[pname]) i += 1 stop = stopping_criterion.evaluate(fvals=loss.detach()) info_dict = { "fopt": loss_trajectory[-1], "wall_time": time.time() - t1, "iterations": iterations, } return mll, info_dict
def gen_candidates_torch( initial_conditions: Tensor, acquisition_function: Callable, lower_bounds: Optional[Union[float, Tensor]] = None, upper_bounds: Optional[Union[float, Tensor]] = None, optimizer: Type[Optimizer] = torch.optim.Adam, options: Optional[Dict[str, Union[float, str]]] = None, verbose: bool = True, fixed_features: Optional[Dict[int, Optional[float]]] = None, ) -> Iterable[Any]: # -> Tuple[Tensor, Any, Optional[Tensor]]: r"""Generate a set of candidates using a `torch.optim` optimizer. Optimizes an acquisition function starting from a set of initial candidates using an optimizer from `torch.optim`. Args: initial_conditions: Starting points for optimization. acquisition_function: Acquisition function to be used. lower_bounds: Minimum values for each column of initial_conditions. upper_bounds: Maximum values for each column of initial_conditions. optimizer (Optimizer): The pytorch optimizer to use to perform candidate search. options: Options used to control the optimization. Includes maxiter: Maximum number of iterations verbose: If True, provide verbose output. fixed_features: This is a dictionary of feature indices to values, where all generated candidates will have features fixed to these values. If the dictionary value is None, then that feature will just be fixed to the clamped value and not optimized. Assumes values to be compatible with lower_bounds and upper_bounds! Returns: 2-element tuple containing - The set of generated candidates. - The acquisition value for each t-batch. """ options = options or {} _jitter = options.get('jitter', 0.) clamped_candidates = columnwise_clamp( X=initial_conditions, lower=lower_bounds, upper=upper_bounds ).requires_grad_(True) candidates = fix_features(clamped_candidates, fixed_features) bayes_optimizer = optimizer( params=[clamped_candidates], lr=options.get("lr", 0.025) ) i = 0 stop = False stopping_criterion = ExpMAStoppingCriterion( **_filter_kwargs(ExpMAStoppingCriterion, **options) ) while not stop: i += 1 batch_loss = acquisition_function(candidates) loss = -batch_loss.sum() if verbose: print("Iter: {} - Value: {:.3f}".format(i, -(loss.item()))) if torch.isnan(loss): print('loss is nan, exiting optimization of the acquisition function.') break bayes_optimizer.zero_grad() loss.backward() if options.get('clip_gradient', False): torch.nn.utils.clip_grad_value_(clamped_candidates, clip_value=options.get('clip_value', 10.)) bayes_optimizer.step() clamped_candidates.data = columnwise_clamp( clamped_candidates, lower_bounds + _jitter, upper_bounds - _jitter ) candidates = fix_features(clamped_candidates, fixed_features) stop = stopping_criterion.evaluate(fvals=loss.detach()) # clamped_candidates = columnwise_clamp( # X=candidates, lower=lower_bounds, upper=upper_bounds, raise_on_violation=True # ) with torch.no_grad(): batch_acquisition = acquisition_function(candidates) return candidates, batch_acquisition
def gen_candidates_torch( initial_conditions: Tensor, acquisition_function: AcquisitionFunction, lower_bounds: Optional[Union[float, Tensor]] = None, upper_bounds: Optional[Union[float, Tensor]] = None, optimizer: Type[Optimizer] = torch.optim.Adam, options: Optional[Dict[str, Union[float, str]]] = None, callback: Optional[Callable[[int, Tensor, Tensor], NoReturn]] = None, fixed_features: Optional[Dict[int, Optional[float]]] = None, ) -> Tuple[Tensor, Tensor]: r"""Generate a set of candidates using a `torch.optim` optimizer. Optimizes an acquisition function starting from a set of initial candidates using an optimizer from `torch.optim`. Args: initial_conditions: Starting points for optimization. acquisition_function: Acquisition function to be used. lower_bounds: Minimum values for each column of initial_conditions. upper_bounds: Maximum values for each column of initial_conditions. optimizer (Optimizer): The pytorch optimizer to use to perform candidate search. options: Options used to control the optimization. Includes maxiter: Maximum number of iterations callback: A callback function accepting the current iteration, loss, and gradients as arguments. This function is executed after computing the loss and gradients, but before calling the optimizer. fixed_features: This is a dictionary of feature indices to values, where all generated candidates will have features fixed to these values. If the dictionary value is None, then that feature will just be fixed to the clamped value and not optimized. Assumes values to be compatible with lower_bounds and upper_bounds! Returns: 2-element tuple containing - The set of generated candidates. - The acquisition value for each t-batch. Example: >>> qEI = qExpectedImprovement(model, best_f=0.2) >>> bounds = torch.tensor([[0., 0.], [1., 2.]]) >>> Xinit = gen_batch_initial_conditions( >>> qEI, bounds, q=3, num_restarts=25, raw_samples=500 >>> ) >>> batch_candidates, batch_acq_values = gen_candidates_torch( initial_conditions=Xinit, acquisition_function=qEI, lower_bounds=bounds[0], upper_bounds=bounds[1], ) """ options = options or {} # if there are fixed features we may optimize over a domain of lower dimension if fixed_features: subproblem = _remove_fixed_features_from_optimization( fixed_features=fixed_features, acquisition_function=acquisition_function, initial_conditions=initial_conditions, lower_bounds=lower_bounds, upper_bounds=upper_bounds, inequality_constraints=None, equality_constraints=None, ) # call the routine with no fixed_features clamped_candidates, batch_acquisition = gen_candidates_torch( initial_conditions=subproblem.initial_conditions, acquisition_function=subproblem.acquisition_function, lower_bounds=subproblem.lower_bounds, upper_bounds=subproblem.upper_bounds, optimizer=optimizer, options=options, callback=callback, fixed_features=None, ) clamped_candidates = subproblem.acquisition_function._construct_X_full( clamped_candidates ) return clamped_candidates, batch_acquisition _clamp = partial(columnwise_clamp, lower=lower_bounds, upper=upper_bounds) clamped_candidates = _clamp(initial_conditions).requires_grad_(True) _optimizer = optimizer(params=[clamped_candidates], lr=options.get("lr", 0.025)) i = 0 stop = False stopping_criterion = ExpMAStoppingCriterion( **_filter_kwargs(ExpMAStoppingCriterion, **options) ) while not stop: i += 1 with torch.no_grad(): X = _clamp(clamped_candidates).requires_grad_(True) loss = -acquisition_function(X).sum() grad = torch.autograd.grad(loss, X)[0] if callback: callback(i, loss, grad) def assign_grad(): _optimizer.zero_grad() clamped_candidates.grad = grad return loss _optimizer.step(assign_grad) stop = stopping_criterion.evaluate(fvals=loss.detach()) clamped_candidates = _clamp(clamped_candidates) with torch.no_grad(): batch_acquisition = acquisition_function(clamped_candidates) return clamped_candidates, batch_acquisition
def gen_candidates_torch( initial_conditions: Tensor, acquisition_function: Callable, lower_bounds: Optional[Union[float, Tensor]] = None, upper_bounds: Optional[Union[float, Tensor]] = None, optimizer: Type[Optimizer] = torch.optim.Adam, options: Optional[Dict[str, Union[float, str]]] = None, verbose: bool = True, fixed_features: Optional[Dict[int, Optional[float]]] = None, ) -> Tuple[Tensor, Tensor]: r"""Generate a set of candidates using a `torch.optim` optimizer. Optimizes an acquisition function starting from a set of initial candidates using an optimizer from `torch.optim`. Args: initial_conditions: Starting points for optimization. acquisition_function: Acquisition function to be used. lower_bounds: Minimum values for each column of initial_conditions. upper_bounds: Maximum values for each column of initial_conditions. optimizer (Optimizer): The pytorch optimizer to use to perform candidate search. options: Options used to control the optimization. Includes maxiter: Maximum number of iterations verbose: If True, provide verbose output. fixed_features: This is a dictionary of feature indices to values, where all generated candidates will have features fixed to these values. If the dictionary value is None, then that feature will just be fixed to the clamped value and not optimized. Assumes values to be compatible with lower_bounds and upper_bounds! Returns: 2-element tuple containing - The set of generated candidates. - The acquisition value for each t-batch. Example: >>> qEI = qExpectedImprovement(model, best_f=0.2) >>> bounds = torch.tensor([[0., 0.], [1., 2.]]) >>> Xinit = gen_batch_initial_conditions( >>> qEI, bounds, q=3, num_restarts=25, raw_samples=500 >>> ) >>> batch_candidates, batch_acq_values = gen_candidates_torch( initial_conditions=Xinit, acquisition_function=qEI, lower_bounds=bounds[0], upper_bounds=bounds[1], ) """ options = options or {} clamped_candidates = columnwise_clamp( X=initial_conditions, lower=lower_bounds, upper=upper_bounds ).requires_grad_(True) candidates = fix_features(clamped_candidates, fixed_features) bayes_optimizer = optimizer( params=[clamped_candidates], lr=options.get("lr", 0.025) ) param_trajectory: Dict[str, List[Tensor]] = {"candidates": []} loss_trajectory: List[float] = [] i = 0 stop = False stopping_criterion = ExpMAStoppingCriterion( **_filter_kwargs(ExpMAStoppingCriterion, **options) ) while not stop: i += 1 loss = -acquisition_function(candidates).sum() if verbose: print("Iter: {} - Value: {:.3f}".format(i, -(loss.item()))) loss_trajectory.append(loss.item()) param_trajectory["candidates"].append(candidates.clone()) def closure(): bayes_optimizer.zero_grad() loss.backward() return loss bayes_optimizer.step(closure) clamped_candidates.data = columnwise_clamp( clamped_candidates, lower_bounds, upper_bounds ) candidates = fix_features(clamped_candidates, fixed_features) stop = stopping_criterion.evaluate(fvals=loss.detach()) clamped_candidates = columnwise_clamp( X=candidates, lower=lower_bounds, upper=upper_bounds, raise_on_violation=True ) with torch.no_grad(): batch_acquisition = acquisition_function(candidates) return candidates, batch_acquisition