class tuSOTOptimizer(AbstractOptimizer): primary_import = "pysot" def __init__(self, api_config): """Build wrapper class to use an optimizer in benchmark. Parameters ---------- api_config : dict-like of dict-like Configuration of the optimization variables. See API description. """ AbstractOptimizer.__init__(self, api_config) self.space_x = JointSpace(api_config) self.bounds = self.space_x.get_bounds() self.create_opt_prob() # Sets up the optimization problem (needs self.bounds) self.max_evals = np.iinfo(np.int32).max # NOTE: Largest possible int self.turbo_batch_size = None self.pysot_batch_size = None self.history = [] self.proposals = [] self.lb, self.ub = self.bounds[:, 0], self.bounds[:, 1] self.dim = len(self.bounds) self.turbo = Turbo1( f=None, lb=self.bounds[:, 0], ub=self.bounds[:, 1], n_init=2 * self.dim + 1, max_evals=self.max_evals, batch_size=4, # We need to update this later verbose=False, ) self.meta_init = self.get_init_hp() def get_init_hp(self): api_config = self.api_config feature = get_api_config_feature(api_config) result, values = [], [] for i in range(9): model = joblib.load( './lgb_meta_model_{}.pkl'.format( i)) value = model.predict(np.array(feature).reshape(1, -1))[-1] values.append(value) result.append(math.e ** value) result_dict = {} for i, (k, v) in enumerate(api_config.items()): if v['type'] == "real": lower, upper = v["range"][0], v["range"][1] if lower < result[i] < upper: out = result[i] else: out = np.random.uniform(lower, upper) elif v['type'] == "int": lower, upper = v["range"][0], v["range"][1] if lower < result[i] < upper: out = int(result[i]) else: out = np.random.randint(lower, upper) elif v['type'] == "bool": out = True if int(result[i]) >= 1 else False else: out = api_config[k]["values"][0] result_dict[k] = out return result[:len(api_config)], values[:len(api_config)], result_dict def restart(self): self.turbo._restart() self.turbo._X = np.zeros((0, self.turbo.dim)) self.turbo._fX = np.zeros((0, 1)) X_init = latin_hypercube(self.turbo.n_init, self.dim) self.X_init = from_unit_cube(X_init, self.lb, self.ub) def create_opt_prob(self): """Create an optimization problem object.""" opt = OptimizationProblem() opt.lb = self.bounds[:, 0] # In warped space opt.ub = self.bounds[:, 1] # In warped space opt.dim = len(self.bounds) opt.cont_var = np.arange(len(self.bounds)) opt.int_var = [] assert len(opt.cont_var) + len(opt.int_var) == opt.dim opt.objfun = None self.opt = opt def start(self): """Starts a new pySOT run.""" self.history = [] self.proposals = [] # Symmetric Latin hypercube design des_pts = max([self.pysot_batch_size, 2 * (self.opt.dim + 1)]) slhd = SymmetricLatinHypercube(dim=self.opt.dim, num_pts=des_pts) # Warped RBF interpolant rbf = RBFInterpolant(dim=self.opt.dim, kernel=CubicKernel(), tail=LinearTail(self.opt.dim), eta=1e-4) rbf = SurrogateUnitBox(rbf, lb=self.opt.lb, ub=self.opt.ub) # Optimization strategy self.strategy = SRBFStrategy( max_evals=self.max_evals, opt_prob=self.opt, exp_design=slhd, surrogate=rbf, asynchronous=True, batch_size=1, use_restarts=True, ) def pysot_suggest(self, n_suggestions=1): if self.pysot_batch_size is None: # First call to suggest self.pysot_batch_size = n_suggestions self.start() # Set the tolerances pretending like we are running batch d, p = float(self.opt.dim), float(n_suggestions) self.strategy.failtol = p * int(max(np.ceil(d / p), np.ceil(4 / p))) # Now we can make suggestions x_w = [] self.proposals = [] for _ in range(n_suggestions): proposal = self.strategy.propose_action() record = EvalRecord(proposal.args, status="pending") proposal.record = record proposal.accept() # This triggers all the callbacks # It is possible that pySOT proposes a previously evaluated point # when all variables are integers, so we just abort in this case # since we have likely converged anyway. See PySOT issue #30. x = list(proposal.record.params) # From tuple to list x_unwarped, = self.space_x.unwarp(x) if x_unwarped in self.history: warnings.warn("pySOT proposed the same point twice") self.start() return self.suggest(n_suggestions=n_suggestions) # NOTE: Append unwarped to avoid rounding issues self.history.append(copy(x_unwarped)) self.proposals.append(proposal) x_w.append(copy(x_unwarped)) return x_w def turbo_suggest(self, n_suggestions=1): if self.turbo_batch_size is None: # Remember the batch size on the first call to suggest self.turbo_batch_size = n_suggestions self.turbo.batch_size = n_suggestions self.turbo.failtol = np.ceil(np.max([4.0 / self.turbo_batch_size, self.dim / self.turbo_batch_size])) self.turbo.n_init = max([self.turbo.n_init, self.turbo_batch_size]) self.restart() X_next = np.zeros((n_suggestions, self.dim)) # Pick from the initial points n_init = min(len(self.X_init), n_suggestions) if n_init > 0: X_next[:n_init] = deepcopy(self.X_init[:n_init, :]) self.X_init = self.X_init[n_init:, :] # Remove these pending points # Get remaining points from TuRBO n_adapt = n_suggestions - n_init if n_adapt > 0: if len(self.turbo._X) > 0: # Use random points if we can't fit a GP X = to_unit_cube(deepcopy(self.turbo._X), self.lb, self.ub) fX = copula_standardize(deepcopy(self.turbo._fX).ravel()) # Use Copula X_cand, y_cand, _ = self.turbo._create_candidates( X, fX, length=self.turbo.length, n_training_steps=100, hypers={} ) X_next[-n_adapt:, :] = self.turbo._select_candidates(X_cand, y_cand)[:n_adapt, :] X_next[-n_adapt:, :] = from_unit_cube(X_next[-n_adapt:, :], self.lb, self.ub) # Unwarp the suggestions suggestions = self.space_x.unwarp(X_next) suggestions[-1] = self.meta_init[2] return suggestions def suggest(self, n_suggestions=1): if n_suggestions == 1: return self.turbo_suggest(n_suggestions) else: suggestion = n_suggestions // 2 return self.turbo_suggest(suggestion) + self.pysot_suggest(n_suggestions - suggestion) def _observe(self, x, y): # Find the matching proposal and execute its callbacks idx = [x == xx for xx in self.history] if np.any(idx): i = np.argwhere(idx)[0].item() # Pick the first index if there are ties proposal = self.proposals[i] proposal.record.complete(y) self.proposals.pop(i) self.history.pop(i) def observe(self, X, y): """Send an observation of a suggestion back to the optimizer. Parameters ---------- X : list of dict-like Places where the objective function has already been evaluated. Each suggestion is a dictionary where each key corresponds to a parameter being optimized. y : array-like, shape (n,) Corresponding values where objective has been evaluated """ assert len(X) == len(y) for x_, y_ in zip(X, y): # Just ignore, any inf observations we got, unclear if right thing if np.isfinite(y_): self._observe(x_, y_) XX, yy = self.space_x.warp(X), np.array(y)[:, None] if len(self.turbo._fX) >= self.turbo.n_init: self.turbo._adjust_length(yy) self.turbo.n_evals += self.turbo_batch_size self.turbo._X = np.vstack((self.turbo._X, deepcopy(XX))) self.turbo._fX = np.vstack((self.turbo._fX, deepcopy(yy))) self.turbo.X = np.vstack((self.turbo.X, deepcopy(XX))) self.turbo.fX = np.vstack((self.turbo.fX, deepcopy(yy))) # Check for a restart if self.turbo.length < self.turbo.length_min: self.restart()
class PySOTOptimizer(AbstractOptimizer): primary_import = "pysot" def __init__(self, api_config): """Build wrapper class to use an optimizer in benchmark. Parameters ---------- api_config : dict-like of dict-like Configuration of the optimization variables. See API description. """ AbstractOptimizer.__init__(self, api_config) self.space_x = JointSpace(api_config) self.bounds = self.space_x.get_bounds() self.create_opt_prob() # Sets up the optimization problem (needs self.bounds) self.max_evals = np.iinfo(np.int32).max # NOTE: Largest possible int self.batch_size = None self.history = [] self.proposals = [] def create_opt_prob(self): """Create an optimization problem object.""" opt = OptimizationProblem() opt.lb = self.bounds[:, 0] # In warped space opt.ub = self.bounds[:, 1] # In warped space opt.dim = len(self.bounds) opt.cont_var = np.arange(len(self.bounds)) opt.int_var = [] assert len(opt.cont_var) + len(opt.int_var) == opt.dim opt.objfun = None self.opt = opt def start(self, max_evals): """Starts a new pySOT run.""" self.history = [] self.proposals = [] # Symmetric Latin hypercube design des_pts = max([self.batch_size, 2 * (self.opt.dim + 1)]) slhd = SymmetricLatinHypercube(dim=self.opt.dim, num_pts=des_pts) # Warped RBF interpolant rbf = RBFInterpolant(dim=self.opt.dim, kernel=CubicKernel(), tail=LinearTail(self.opt.dim), eta=1e-4) rbf = SurrogateUnitBox(rbf, lb=self.opt.lb, ub=self.opt.ub) # Optimization strategy self.strategy = SRBFStrategy( max_evals=self.max_evals, opt_prob=self.opt, exp_design=slhd, surrogate=rbf, asynchronous=True, batch_size=1, use_restarts=True, ) def suggest(self, n_suggestions=1): """Get a suggestion from the optimizer. Parameters ---------- n_suggestions : int Desired number of parallel suggestions in the output Returns ------- next_guess : list of dict List of `n_suggestions` suggestions to evaluate the objective function. Each suggestion is a dictionary where each key corresponds to a parameter being optimized. """ if self.batch_size is None: # First call to suggest self.batch_size = n_suggestions self.start(self.max_evals) # Set the tolerances pretending like we are running batch d, p = float(self.opt.dim), float(n_suggestions) self.strategy.failtol = p * int(max(np.ceil(d / p), np.ceil(4 / p))) # Now we can make suggestions x_w = [] self.proposals = [] for _ in range(n_suggestions): proposal = self.strategy.propose_action() record = EvalRecord(proposal.args, status="pending") proposal.record = record proposal.accept() # This triggers all the callbacks # It is possible that pySOT proposes a previously evaluated point # when all variables are integers, so we just abort in this case # since we have likely converged anyway. See PySOT issue #30. x = list(proposal.record.params) # From tuple to list x_unwarped, = self.space_x.unwarp(x) if x_unwarped in self.history: warnings.warn("pySOT proposed the same point twice") self.start(self.max_evals) return self.suggest(n_suggestions=n_suggestions) # NOTE: Append unwarped to avoid rounding issues self.history.append(copy(x_unwarped)) self.proposals.append(proposal) x_w.append(copy(x_unwarped)) return x_w def _observe(self, x, y): # Find the matching proposal and execute its callbacks idx = [x == xx for xx in self.history] i = np.argwhere(idx)[0].item() # Pick the first index if there are ties proposal = self.proposals[i] proposal.record.complete(y) self.proposals.pop(i) self.history.pop(i) def observe(self, X, y): """Send an observation of a suggestion back to the optimizer. Parameters ---------- X : list of dict-like Places where the objective function has already been evaluated. Each suggestion is a dictionary where each key corresponds to a parameter being optimized. y : array-like, shape (n,) Corresponding values where objective has been evaluated """ assert len(X) == len(y) for x_, y_ in zip(X, y): # Just ignore, any inf observations we got, unclear if right thing if np.isfinite(y_): self._observe(x_, y_)
class tuSOTOptimizer(AbstractOptimizer): primary_import = "pysot" def __init__(self, api_config): """Build wrapper class to use an optimizer in benchmark. Parameters ---------- api_config : dict-like of dict-like Configuration of the optimization variables. See API description. """ AbstractOptimizer.__init__(self, api_config) self.space_x = JointSpace(api_config) self.bounds = self.space_x.get_bounds() self.create_opt_prob( ) # Sets up the optimization problem (needs self.bounds) self.max_evals = np.iinfo(np.int32).max # NOTE: Largest possible int self.turbo_batch_size = None self.pysot_batch_size = None self.history = [] self.proposals = [] self.lb, self.ub = self.bounds[:, 0], self.bounds[:, 1] self.dim = len(self.bounds) self.turbo = Turbo1( f=None, lb=self.bounds[:, 0], ub=self.bounds[:, 1], n_init=2 * self.dim + 1, max_evals=self.max_evals, batch_size=4, # We need to update this later verbose=False, ) # hyperopt self.random = np_random space, self.round_to_values = tuSOTOptimizer.get_hyperopt_dimensions( api_config) self.domain = Domain(dummy_f, space, pass_expr_memo_ctrl=None) self.trials = Trials() # Some book keeping like opentuner wrapper self.trial_id_lookup = {} # Store just for data validation self.param_set_chk = frozenset(api_config.keys()) def restart(self): self.turbo._restart() self.turbo._X = np.zeros((0, self.turbo.dim)) self.turbo._fX = np.zeros((0, 1)) X_init = latin_hypercube(self.turbo.n_init, self.dim) self.X_init = from_unit_cube(X_init, self.lb, self.ub) def create_opt_prob(self): """Create an optimization problem object.""" opt = OptimizationProblem() opt.lb = self.bounds[:, 0] # In warped space opt.ub = self.bounds[:, 1] # In warped space opt.dim = len(self.bounds) opt.cont_var = np.arange(len(self.bounds)) opt.int_var = [] assert len(opt.cont_var) + len(opt.int_var) == opt.dim opt.objfun = None self.opt = opt def start(self): """Starts a new pySOT run.""" self.history = [] self.proposals = [] # Symmetric Latin hypercube design des_pts = max([self.pysot_batch_size, 2 * (self.opt.dim + 1)]) slhd = SymmetricLatinHypercube(dim=self.opt.dim, num_pts=des_pts) # Warped RBF interpolant rbf = RBFInterpolant(dim=self.opt.dim, kernel=CubicKernel(), tail=LinearTail(self.opt.dim), eta=1e-4) rbf = SurrogateUnitBox(rbf, lb=self.opt.lb, ub=self.opt.ub) # Optimization strategy self.strategy = SRBFStrategy( max_evals=self.max_evals, opt_prob=self.opt, exp_design=slhd, surrogate=rbf, asynchronous=True, batch_size=1, use_restarts=True, ) @staticmethod def hashable_dict(d): hashable_object = frozenset(d.items()) return hashable_object @staticmethod def get_hyperopt_dimensions(api_config): param_list = sorted(api_config.keys()) space = {} round_to_values = {} for param_name in param_list: param_config = api_config[param_name] param_type = param_config["type"] param_space = param_config.get("space", None) param_range = param_config.get("range", None) param_values = param_config.get("values", None) # Some setup for case that whitelist of values is provided: values_only_type = param_type in ("cat", "ordinal") if (param_values is not None) and (not values_only_type): assert param_range is None param_values = np.unique(param_values) param_range = (param_values[0], param_values[-1]) round_to_values[param_name] = interp1d( param_values, param_values, kind="nearest", fill_value="extrapolate") if param_type == "int": low, high = param_range if param_space in ("log", "logit"): space[param_name] = hp.qloguniform(param_name, np.log(low), np.log(high), 1) else: space[param_name] = hp.quniform(param_name, low, high, 1) elif param_type == "bool": assert param_range is None assert param_values is None space[param_name] = hp.choice(param_name, (False, True)) elif param_type in ("cat", "ordinal"): assert param_range is None space[param_name] = hp.choice(param_name, param_values) elif param_type == "real": low, high = param_range if param_space in ("log", "logit"): space[param_name] = hp.loguniform(param_name, np.log(low), np.log(high)) else: space[param_name] = hp.uniform(param_name, low, high) else: assert False, "type %s not handled in API" % param_type return space, round_to_values def get_trial(self, trial_id): for trial in self.trials._dynamic_trials: if trial["tid"] == trial_id: assert isinstance(trial, dict) # Make sure right kind of dict assert "state" in trial and "result" in trial assert trial["state"] == JOB_STATE_NEW return trial assert False, "No matching trial ID" def cleanup_guess(self, x_guess): assert isinstance(x_guess, dict) # Also, check the keys are only the vars we are searching over: assert frozenset(x_guess.keys()) == self.param_set_chk # Do the rounding # Make a copy to be safe, and also unpack singletons # We may also need to consider clip_chk at some point like opentuner x_guess = {k: only(x_guess[k]) for k in x_guess} for param_name, round_f in self.round_to_values.items(): x_guess[param_name] = round_f(x_guess[param_name]) # Also ensure this is correct dtype so sklearn is happy x_guess = { k: DTYPE_MAP[self.api_config[k]["type"]](x_guess[k]) for k in x_guess } return x_guess def pysot_suggest(self, n_suggestions=1): if self.pysot_batch_size is None: # First call to suggest self.pysot_batch_size = n_suggestions self.start() # Set the tolerances pretending like we are running batch d, p = float(self.opt.dim), float(n_suggestions) self.strategy.failtol = p * int(max(np.ceil(d / p), np.ceil(4 / p))) # Now we can make suggestions x_w = [] self.proposals = [] for _ in range(n_suggestions): proposal = self.strategy.propose_action() record = EvalRecord(proposal.args, status="pending") proposal.record = record proposal.accept() # This triggers all the callbacks # It is possible that pySOT proposes a previously evaluated point # when all variables are integers, so we just abort in this case # since we have likely converged anyway. See PySOT issue #30. x = list(proposal.record.params) # From tuple to list x_unwarped, = self.space_x.unwarp(x) if x_unwarped in self.history: warnings.warn("pySOT proposed the same point twice") self.start() return self.suggest(n_suggestions=n_suggestions) # NOTE: Append unwarped to avoid rounding issues self.history.append(copy(x_unwarped)) self.proposals.append(proposal) x_w.append(copy(x_unwarped)) return x_w def pysot_get_suggest(self, suggests): turbo_suggest_warps = self.space_x.warp(suggests) for i, warps in enumerate(turbo_suggest_warps): proposal = self.strategy.make_proposal(warps) proposal.add_callback(self.strategy.on_initial_proposal) record = EvalRecord(proposal.args, status="pending") proposal.record = record proposal.accept() self.history.append(copy(suggests[i])) self.proposals.append(proposal) def turbo_suggest(self, n_suggestions=1): if self.turbo_batch_size is None: # Remember the batch size on the first call to suggest self.turbo_batch_size = n_suggestions self.turbo.batch_size = n_suggestions self.turbo.failtol = np.ceil( np.max([ 4.0 / self.turbo_batch_size, self.dim / self.turbo_batch_size ])) self.turbo.n_init = max([self.turbo.n_init, self.turbo_batch_size]) self.restart() X_next = np.zeros((n_suggestions, self.dim)) # Pick from the initial points n_init = min(len(self.X_init), n_suggestions) if n_init > 0: X_next[:n_init] = deepcopy(self.X_init[:n_init, :]) self.X_init = self.X_init[ n_init:, :] # Remove these pending points # Get remaining points from TuRBO n_adapt = n_suggestions - n_init if n_adapt > 0: if len(self.turbo._X ) > 0: # Use random points if we can't fit a GP X = to_unit_cube(deepcopy(self.turbo._X), self.lb, self.ub) fX = copula_standardize(deepcopy( self.turbo._fX).ravel()) # Use Copula X_cand, y_cand, _ = self.turbo._create_candidates( X, fX, length=self.turbo.length, n_training_steps=100, hypers={}) X_next[-n_adapt:, :] = self.turbo._select_candidates( X_cand, y_cand)[:n_adapt, :] X_next[-n_adapt:, :] = from_unit_cube(X_next[-n_adapt:, :], self.lb, self.ub) # Unwarp the suggestions suggestions = self.space_x.unwarp(X_next) return suggestions def _hyperopt_suggest(self): new_ids = self.trials.new_trial_ids(1) assert len(new_ids) == 1 self.trials.refresh() seed = random_seed(self.random) new_trials = tpe.suggest(new_ids, self.domain, self.trials, seed) assert len(new_trials) == 1 self.trials.insert_trial_docs(new_trials) self.trials.refresh() new_trial, = new_trials # extract singleton return new_trial def _hyperopt_transform(self, x): new_id = self.trials.new_trial_ids(1)[0] domain = self.domain rng = np.random.RandomState(1) idxs, vals = pyll.rec_eval(domain.s_idxs_vals, memo={ domain.s_new_ids: [new_id], domain.s_rng: rng, }) rval_miscs = [dict(tid=new_id, cmd=domain.cmd, workdir=domain.workdir)] rval_results = domain.new_result() for (k, _) in vals.items(): vals[k][0] = x[k] miscs_update_idxs_vals(rval_miscs, idxs, vals) rval_docs = self.trials.new_trial_docs([new_id], [None], rval_results, rval_miscs) return rval_docs[0] def hyperopt_suggest(self, n_suggestions=1): assert n_suggestions >= 1, "invalid value for n_suggestions" # Get the new trials, it seems hyperopt either uses random search or # guesses one at a time anyway, so we might as welll call serially. new_trials = [self._hyperopt_suggest() for _ in range(n_suggestions)] X = [] for trial in new_trials: x_guess = self.cleanup_guess(trial["misc"]["vals"]) X.append(x_guess) # Build lookup to get original trial object x_guess_ = tuSOTOptimizer.hashable_dict(x_guess) assert x_guess_ not in self.trial_id_lookup, "the suggestions should not already be in the trial dict" self.trial_id_lookup[x_guess_] = trial["tid"] assert len(X) == n_suggestions return X def hyperopt_get_suggest(self, suggests): trials = [self._hyperopt_transform(x) for x in suggests] for trial in trials: x_guess = self.cleanup_guess(trial["misc"]["vals"]) x_guess_ = tuSOTOptimizer.hashable_dict(x_guess) assert x_guess_ not in self.trial_id_lookup, "the suggestions should not already be in the trial dict" self.trial_id_lookup[x_guess_] = trial["tid"] self.trials.insert_trial_docs(trials) self.trials.refresh() def suggest(self, n_suggestions=1): if n_suggestions == 1: return self.turbo_suggest(n_suggestions) else: t_suggestion = n_suggestions // 2 # p_suggestion = int((n_suggestions - t_suggestion) * 3/4) h_suggestion = n_suggestions - t_suggestion turbo_suggest = self.turbo_suggest(t_suggestion) # pysot_suggest = self.pysot_suggest(p_suggestion) hyperopt_suggest = self.hyperopt_suggest(h_suggestion) self.hyperopt_get_suggest(turbo_suggest) # self.pysot_get_suggest(turbo_suggest + hyperopt_suggest) return turbo_suggest + hyperopt_suggest def _observe(self, x, y): # Find the matching proposal and execute its callbacks idx = [x == xx for xx in self.history] if np.any(idx): i = np.argwhere( idx)[0].item() # Pick the first index if there are ties proposal = self.proposals[i] proposal.record.complete(y) self.proposals.pop(i) self.history.pop(i) def observe(self, X, y): """Send an observation of a suggestion back to the optimizer. Parameters ---------- X : list of dict-like Places where the objective function has already been evaluated. Each suggestion is a dictionary where each key corresponds to a parameter being optimized. y : array-like, shape (n,) Corresponding values where objective has been evaluated """ assert len(X) == len(y) # # pysot observe # for x_, y_ in zip(X, y): # # Just ignore, any inf observations we got, unclear if right thing # if np.isfinite(y_): # self._observe(x_, y_) # turbo observe XX, yy = self.space_x.warp(X), np.array(y)[:, None] if len(self.turbo._fX) >= self.turbo.n_init: self.turbo._adjust_length(yy) self.turbo.n_evals += self.turbo_batch_size self.turbo._X = np.vstack((self.turbo._X, deepcopy(XX))) self.turbo._fX = np.vstack((self.turbo._fX, deepcopy(yy))) self.turbo.X = np.vstack((self.turbo.X, deepcopy(XX))) self.turbo.fX = np.vstack((self.turbo.fX, deepcopy(yy))) # Check for a restart if self.turbo.length < self.turbo.length_min: self.restart() # hyperopt observe for x_guess, y_ in zip(X, y): x_guess_ = tuSOTOptimizer.hashable_dict(x_guess) assert x_guess_ in self.trial_id_lookup, "Appears to be guess that did not originate from suggest" assert x_guess_ in self.trial_id_lookup, "trial object not available in trial dict" trial_id = self.trial_id_lookup.pop(x_guess_) trial = self.get_trial(trial_id) assert self.cleanup_guess( trial["misc"]["vals"] ) == x_guess, "trial ID not consistent with x values stored" # Cast to float to ensure native type result = {"loss": float(y_), "status": STATUS_OK} trial["state"] = JOB_STATE_DONE trial["result"] = result self.trials.refresh()