def test_direct(self):
maximizer = Direct(self.acquisition_func, self.X_lower, self.X_upper)
x = maximizer.maximize()
assert x.shape[0] == 1
assert x.shape[1] == self.dims
assert np.all(x[:, 0] >= self.X_lower[0])
assert np.all(x[:, 0] <= self.X_upper[0])
def test_direct(self):
maximizer = Direct(self.objective_function, self.lower, self.upper)
x = maximizer.maximize()
assert x.shape[0] == 1
assert len(x.shape) == 1
assert np.all(x >= self.lower)
assert np.all(x <= self.upper)
def test_direct(self):
maximizer = Direct(self.objective_function, self.lower, self.upper)
x = maximizer.maximize()
assert x.shape[0] == 2
assert len(x.shape) == 1
assert np.all(x >= self.lower)
assert np.all(x <= self.upper)
class Bohamiann(Optimizer):
def __init__(self, config_space, burnin=3000, n_iters=10000):
super(Bohamiann,
self).__init__(sacred_space_to_configspace(config_space))
self.rng = np.random.RandomState(np.random.seed())
self.n_dims = len(self.config_space.get_hyperparameters())
# All inputs are mapped to be in [0, 1]^D
self.lower = np.zeros([self.n_dims])
self.upper = np.ones([self.n_dims])
self.incumbents = []
self.X = None
self.y = None
self.model = BayesianNeuralNetwork(sampling_method="sghmc",
l_rate=np.sqrt(1e-4),
mdecay=0.05,
burn_in=burnin,
n_iters=n_iters,
precondition=True,
normalize_input=True,
normalize_output=True)
self.acquisition_func = LogEI(self.model)
self.maximizer = Direct(self.acquisition_func,
self.lower,
self.upper,
verbose=False)
def suggest_configuration(self):
if self.X is None and self.y is None:
# No data points yet to train a model, just return a random configuration instead
new_x = init_random_uniform(self.lower,
self.upper,
n_points=1,
rng=self.rng)[0, :]
else:
# Train the model on all finished runs
self.model.train(self.X, self.y)
self.acquisition_func.update(self.model)
# Maximize the acquisition function
new_x = self.maximizer.maximize()
# Maps from [0, 1]^D space back to original space
next_config = Configuration(self.config_space, vector=new_x)
# Transform to sacred configuration
result = configspace_config_to_sacred(next_config)
return result
def suggest_configuration(self):
if self.X is None and self.Y is None:
new_x = init_random_uniform(self.X_lower,
self.X_upper,
N=1,
rng=self.rng)
elif self.X.shape[0] == 1:
# We need at least 2 data points to train a GP
Xopt = init_random_uniform(self.X_lower,
self.X_upper,
N=1,
rng=self.rng)
else:
prior = DNGOPrior()
model = DNGO(batch_size=100,
num_epochs=20000,
learning_rate=0.1,
momentum=0.9,
l2=1e-16,
adapt_epoch=5000,
n_hypers=20,
prior=prior,
do_optimize=True,
do_mcmc=True)
#acquisition_func = EI(model, task.X_lower, task.X_upper)
lo = np.ones([model.n_units_3]) * -1
up = np.ones([model.n_units_3])
ei = LogEI(model, lo, up)
acquisition_func = IntegratedAcquisition(model, ei, self.X_lower,
self.X_upper)
maximizer = Direct(acquisition_func, self.X_lower, self.X_upper)
model.train(self.X, self.Y)
acquisition_func.update(model)
new_x = maximizer.maximize()
# Map from [0, 1]^D space back to original space
next_config = Configuration(self.config_space, vector=new_x[0, :])
# Transform to sacred configuration
result = configspace_config_to_sacred(next_config)
return result
def suggest_configuration(self):
if self.X is None and self.y is None:
new_x = init_random_uniform(self.lower, self.upper,
n_points=1, rng=self.rng)[0, :]
elif self.X.shape[0] == 1:
# We need at least 2 data points to train a GP
new_x = init_random_uniform(self.lower, self.upper,
n_points=1, rng=self.rng)[0, :]
else:
cov_amp = 1
n_dims = self.lower.shape[0]
initial_ls = np.ones([n_dims])
exp_kernel = george.kernels.Matern52Kernel(initial_ls,
ndim=n_dims)
kernel = cov_amp * exp_kernel
prior = DefaultPrior(len(kernel) + 1)
model = GaussianProcessMCMC(kernel, prior=prior,
n_hypers=self.n_hypers,
chain_length=self.chain_length,
burnin_steps=self.burnin,
normalize_input=False,
normalize_output=True,
rng=self.rng,
lower=self.lower,
upper=self.upper)
a = LogEI(model)
acquisition_func = MarginalizationGPMCMC(a)
max_func = Direct(acquisition_func, self.lower, self.upper, verbose=False)
model.train(self.X, self.y)
acquisition_func.update(model)
new_x = max_func.maximize()
next_config = Configuration(self.config_space, vector=new_x)
# Transform to sacred configuration
result = configspace_config_to_sacred(next_config)
return result
def mtbo(objective_function,
lower,
upper,
n_tasks=2,
n_init=2,
num_iterations=30,
burnin=100,
chain_length=200,
n_hypers=20,
output_path=None,
rng=None):
"""
Interface to MTBO[1] which uses an auxiliary cheaper task to speed up the optimization
of a more expensive but similar task.
[1] Multi-Task Bayesian Optimization
K. Swersky and J. Snoek and R. Adams
Proceedings of the 27th International Conference on Advances in Neural Information Processing Systems (NIPS'13)
Parameters
----------
objective_function: function
Objective function that will be optimized
lower: np.array(D,)
Lower bound of the input space
upper: np.array(D,)
Upper bound of the input space
n_tasks: int
Number of task
n_init: int
Number of initial design points
num_iterations: int
Number of iterations
chain_length : int
The length of the MCMC chain for each walker.
burnin : int
The number of burnin steps before the actual MCMC sampling starts.
output_path: string
Specifies the path where the intermediate output after each iteration will be saved.
If None no output will be saved to disk.
rng: numpy.random.RandomState
Random number generator
Returns
-------
dict with all results
"""
assert n_init <= num_iterations, "Number of initial design point has to be <= than the number of iterations"
assert lower.shape[0] == upper.shape[
0], "Dimension miss match between upper and lower bound"
time_start = time.time()
if rng is None:
rng = np.random.RandomState(np.random.randint(0, 10000))
n_dims = lower.shape[0]
# Bookkeeping
time_func_eval = []
time_overhead = []
incumbents = []
runtime = []
X = []
y = []
c = []
# Define model for the objective function
cov_amp = 1 # Covariance amplitude
kernel = cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims + 1,
dim=d)
task_kernel = george.kernels.TaskKernel(n_dims + 1, n_dims, n_tasks)
kernel *= task_kernel
# Take 3 times more samples than we have hyperparameters
if n_hypers < 2 * len(kernel):
n_hypers = 3 * len(kernel)
if n_hypers % 2 == 1:
n_hypers += 1
prior = MTBOPrior(len(kernel) + 1,
n_ls=n_dims,
n_kt=len(task_kernel),
rng=rng)
model_objective = MTBOGPMCMC(kernel,
prior=prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
lower=lower,
upper=upper,
rng=rng)
# Define model for the cost function
cost_cov_amp = 1
cost_kernel = cost_cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
cost_kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims + 1,
dim=d)
cost_task_kernel = george.kernels.TaskKernel(n_dims + 1, n_dims, n_tasks)
cost_kernel *= cost_task_kernel
cost_prior = MTBOPrior(len(cost_kernel) + 1,
n_ls=n_dims,
n_kt=len(task_kernel),
rng=rng)
model_cost = MTBOGPMCMC(cost_kernel,
prior=cost_prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
lower=lower,
upper=upper,
rng=rng)
# Extend input space by task variable
extend_lower = np.append(lower, 0)
extend_upper = np.append(upper, n_tasks - 1)
is_env = np.zeros(extend_lower.shape[0])
is_env[-1] = 1
# Define acquisition function and maximizer
ig = InformationGainPerUnitCost(model_objective,
model_cost,
extend_lower,
extend_upper,
sampling_acquisition=EI,
is_env_variable=is_env,
n_representer=50)
acquisition_func = MarginalizationGPMCMC(ig)
maximizer = Direct(acquisition_func,
extend_lower,
extend_upper,
n_func_evals=200)
# Initial Design
logger.info("Initial Design")
for it in range(n_init):
start_time_overhead = time.time()
# Draw random configuration and evaluate it just on the auxiliary task
task = 0
x = init_random_uniform(lower, upper, 1, rng)[0]
logger.info("Evaluate candidate %s", str(x))
st = time.time()
func_val, cost = objective_function(x, task)
time_func_eval.append(time.time() - st)
logger.info("Configuration achieved a performance of %f with cost %f",
func_val, cost)
logger.info("Evaluation of this configuration took %f seconds",
time_func_eval[-1])
# Bookkeeping
config = np.append(x, task)
X.append(config)
y.append(np.log(
func_val)) # Model the target function on a logarithmic scale
c.append(np.log(cost)) # Model the cost on a logarithmic scale
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y)
incumbents.append(X[best_idx][:-1])
time_overhead.append(time.time() - start_time_overhead)
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(
data,
open(os.path.join(output_path, "mtbo_iter_%d.json" % it), "w"))
X = np.array(X)
y = np.array(y)
c = np.array(c)
for it in range(n_init, num_iterations):
logger.info("Start iteration %d ... ", it)
start_time = time.time()
# Train models
model_objective.train(X, y, do_optimize=True)
model_cost.train(X, c, do_optimize=True)
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y)
incumbent = X[best_idx][:-1]
incumbent = np.append(incumbent, 1)
incumbent_value = y[best_idx]
incumbents.append(incumbent[:-1])
logger.info("Current incumbent %s with estimated performance %f",
str(incumbent), incumbent_value)
# Maximize acquisition function
acquisition_func.update(model_objective, model_cost)
new_x = maximizer.maximize()
new_x[-1] = np.rint(
new_x[-1]) # Map float value to discrete task variable
time_overhead.append(time.time() - start_time)
logger.info("Optimization overhead was %f seconds", time_overhead[-1])
# Evaluate the chosen configuration
logger.info("Evaluate candidate %s", str(new_x))
start_time = time.time()
new_y, new_c = objective_function(new_x[:-1], new_x[-1])
time_func_eval.append(time.time() - start_time)
logger.info("Configuration achieved a performance of %f with cost %f",
new_y, new_c)
logger.info("Evaluation of this configuration took %f seconds",
time_func_eval[-1])
# Add new observation to the data
X = np.concatenate((X, new_x[None, :]), axis=0)
y = np.concatenate(
(y, np.log(np.array([new_y]))),
axis=0) # Model the target function on a logarithmic scale
c = np.concatenate(
(c, np.log(np.array([new_c]))),
axis=0) # Model the cost function on a logarithmic scale
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(
data,
open(os.path.join(output_path, "mtbo_iter_%d.json" % it), "w"))
# Estimate the final incumbent
model_objective.train(X, y)
incumbent, incumbent_value = projected_incumbent_estimation(
model_objective, X[:, :-1], proj_value=n_tasks - 1)
logger.info("Final incumbent %s with estimated performance %f",
str(incumbent), incumbent_value)
results = dict()
results["x_opt"] = incumbent[:-1].tolist()
results["incumbents"] = [inc.tolist() for inc in incumbents]
results["runtime"] = runtime
results["overhead"] = time_overhead
results["time_func_eval"] = time_func_eval
results["X"] = X
results["y"] = y
results["c"] = c
return results
def fabolas(objective_function,
lower,
upper,
s_min,
s_max,
n_init=40,
num_iterations=100,
subsets=[256, 128, 64],
inc_estimation="mean",
burnin=100,
chain_length=100,
n_hypers=12,
output_path=None,
rng=None):
"""
Fast Bayesian Optimization of Machine Learning Hyperparameters
on Large Datasets
Fast Bayesian Optimization of Machine Learning Hyperparameters on Large Datasets
A. Klein and S. Falkner and S. Bartels and P. Hennig and F. Hutter
http://arxiv.org/abs/1605.07079
Parameters
----------
objective_function: function
Objective function that will be optimized
lower: np.array(D,)
Lower bound of the input space
upper: np.array(D,)
Upper bound of the input space
s_min: int
Minimum number of data points for the training data set
s_max: int
Maximum number of data points for the training data set
n_init: int
Number of initial design points
n_hypers: int
Number of hyperparameter samples for the GP
subsets: list
The ratio of the subsets size of the initial design.
For example if subsets=[256, 128, 64] then the first random point from the
initial design is evaluated on s_max/256 of the data, the second point on
s_max/256 of the data and so on.
num_iterations: int
Number of iterations
chain_length : int
The length of the MCMC chain for each walker.
burnin : int
The number of burnin steps before the actual MCMC sampling starts.
output_path: string
Specifies the path where the intermediate output after each iteration will be saved.
If None no output will be saved to disk.
rng: numpy.random.RandomState
Random number generator
Returns
-------
dict
"""
assert n_init <= num_iterations, "Number of initial design point has to be <= than the number of iterations"
assert lower.shape[0] == upper.shape[
0], "Dimension miss match between upper and lower bound"
time_start = time.time()
if rng is None:
rng = np.random.RandomState(np.random.randint(0, 10000))
n_dims = lower.shape[0]
# Bookkeeping
time_func_eval = []
time_overhead = []
incumbents = []
runtime = []
X = []
y = []
c = []
# Define model for the objective function
cov_amp = 1 # Covariance amplitude
kernel = cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims + 1,
dim=d)
# Kernel for the environmental variable
# We use (1-s)**2 as basis function for the Bayesian linear kernel
degree = 1
env_kernel = george.kernels.BayesianLinearRegressionKernel(n_dims + 1,
dim=n_dims,
degree=degree)
env_kernel[:] = np.ones([degree + 1]) * 0.1
kernel *= env_kernel
# Take 3 times more samples than we have hyperparameters
if n_hypers < 2 * len(kernel):
n_hypers = 3 * len(kernel)
if n_hypers % 2 == 1:
n_hypers += 1
prior = EnvPrior(len(kernel) + 1, n_ls=n_dims, n_lr=(degree + 1), rng=rng)
quadratic_bf = lambda x: (1 - x)**2
linear_bf = lambda x: x
model_objective = FabolasGPMCMC(kernel,
prior=prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
normalize_output=False,
basis_func=quadratic_bf,
lower=lower,
upper=upper,
rng=rng)
# Define model for the cost function
cost_cov_amp = 1
cost_kernel = cost_cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
cost_kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims + 1,
dim=d)
cost_degree = 1
cost_env_kernel = george.kernels.BayesianLinearRegressionKernel(
n_dims + 1, dim=n_dims, degree=cost_degree)
cost_env_kernel[:] = np.ones([cost_degree + 1]) * 0.1
cost_kernel *= cost_env_kernel
cost_prior = EnvPrior(len(cost_kernel) + 1,
n_ls=n_dims,
n_lr=(cost_degree + 1),
rng=rng)
model_cost = FabolasGPMCMC(cost_kernel,
prior=cost_prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
basis_func=linear_bf,
normalize_output=False,
lower=lower,
upper=upper,
rng=rng)
# Extend input space by task variable
extend_lower = np.append(lower, 0)
extend_upper = np.append(upper, 1)
is_env = np.zeros(extend_lower.shape[0])
is_env[-1] = 1
# Define acquisition function and maximizer
ig = InformationGainPerUnitCost(model_objective,
model_cost,
extend_lower,
extend_upper,
sampling_acquisition=EI,
is_env_variable=is_env,
n_representer=50)
acquisition_func = MarginalizationGPMCMC(ig)
maximizer = Direct(acquisition_func,
extend_lower,
extend_upper,
verbose=True,
n_func_evals=200)
# Initial Design
logger.info("Initial Design")
for it in range(n_init):
start_time_overhead = time.time()
# Draw random configuration
s = int(s_max / float(subsets[it]))
x = init_random_uniform(lower, upper, 1, rng)[0]
logger.info("Evaluate %s on subset size %d", str(x), s)
st = time.time()
func_val, cost = objective_function(x, s)
time_func_eval.append(time.time() - st)
logger.info("Configuration achieved a performance of %f with cost %f",
func_val, cost)
logger.info("Evaluation of this configuration took %f seconds",
time_func_eval[-1])
# Bookkeeping
config = np.append(x, transform(s, s_min, s_max))
X.append(config)
y.append(np.log(
func_val)) # Model the target function on a logarithmic scale
c.append(np.log(cost)) # Model the cost on a logarithmic scale
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y)
incumbents.append(X[best_idx][:-1]) # Incumbent is always on s=s_max
time_overhead.append(time.time() - start_time_overhead)
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(
data,
open(os.path.join(output_path, "fabolas_iter_%d.json" % it),
"w"))
X = np.array(X)
y = np.array(y)
c = np.array(c)
for it in range(n_init, num_iterations):
logger.info("Start iteration %d ... ", it)
start_time = time.time()
# Train models
model_objective.train(X, y, do_optimize=True)
model_cost.train(X, c, do_optimize=True)
if inc_estimation == "last_seen":
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y)
incumbent = X[best_idx][:-1]
incumbent = np.append(incumbent, 1)
incumbent_value = y[best_idx]
else:
# Estimate incumbent by projecting all observed points to the task of interest and
# pick the point with the lowest mean prediction
incumbent, incumbent_value = projected_incumbent_estimation(
model_objective, X[:, :-1], proj_value=1)
incumbents.append(incumbent[:-1])
logger.info("Current incumbent %s with estimated performance %f",
str(incumbent), np.exp(incumbent_value))
# Maximize acquisition function
acquisition_func.update(model_objective, model_cost)
new_x = maximizer.maximize()
s = retransform(new_x[-1], s_min,
s_max) # Map s from log space to original linear space
time_overhead.append(time.time() - start_time)
logger.info("Optimization overhead was %f seconds", time_overhead[-1])
# Evaluate the chosen configuration
logger.info("Evaluate candidate %s on subset size %f", str(new_x[:-1]),
s)
start_time = time.time()
new_y, new_c = objective_function(new_x[:-1], s)
time_func_eval.append(time.time() - start_time)
logger.info("Configuration achieved a performance of %f with cost %f",
new_y, new_c)
logger.info("Evaluation of this configuration took %f seconds",
time_func_eval[-1])
# Add new observation to the data
X = np.concatenate((X, new_x[None, :]), axis=0)
y = np.concatenate(
(y, np.log(np.array([new_y]))),
axis=0) # Model the target function on a logarithmic scale
c = np.concatenate(
(c, np.log(np.array([new_c]))),
axis=0) # Model the cost function on a logarithmic scale
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(
data,
open(os.path.join(output_path, "fabolas_iter_%d.json" % it),
"w"))
# Estimate the final incumbent
model_objective.train(X, y, do_optimize=True)
incumbent, incumbent_value = projected_incumbent_estimation(
model_objective, X[:, :-1], proj_value=1)
logger.info("Final incumbent %s with estimated performance %f",
str(incumbent), incumbent_value)
results = dict()
results["x_opt"] = incumbent[:-1].tolist()
results["incumbents"] = [inc.tolist() for inc in incumbents]
results["runtime"] = runtime
results["overhead"] = time_overhead
results["time_func_eval"] = time_func_eval
results["X"] = X
results["y"] = y
results["c"] = c
return results
def warmstart_mtbo(objective_function,
lower,
upper,
observed_X,
observed_y,
n_tasks=2,
num_iterations=30,
target_task_id=1,
burnin=100,
chain_length=200,
n_hypers=20,
output_path=None,
rng=None):
"""
Interface to MTBO[1] which uses an auxiliary cheaper task to warm start the optimization on new but similar task.
Note here we only warmstart the optimization process, in case you want to speed up Bayesian optimization by
evaluating on auxiliary task during the optimization check out mtbo() or fabolas().
[1] Multi-Task Bayesian Optimization
K. Swersky and J. Snoek and R. Adams
Proceedings of the 27th International Conference on Advances in Neural Information Processing Systems (NIPS'13)
Parameters
----------
objective_function: function
Objective function that will be optimized
lower: np.array(D,)
Lower bound of the input space
upper: np.array(D,)
Upper bound of the input space
observed_X: np.array(N, D + 1)
observed point from the auxiliary task. Make sure that the last dimension identifies the auxiliary task
(default=0). We assume the main task to have the task id = 1
observed_y: np.array(N,)
corresponding target values
n_tasks: int
Number of task
target_task_id: int
the id of the target task
num_iterations: int
Number of iterations
chain_length : int
The length of the MCMC chain for each walker.
burnin : int
The number of burnin steps before the actual MCMC sampling starts.
output_path: string
Specifies the path where the intermediate output after each iteration will be saved.
If None no output will be saved to disk.
rng: numpy.random.RandomState
Random number generator
Returns
-------
dict with all results
"""
assert lower.shape[0] == upper.shape[
0], "Dimension miss match between upper and lower bound"
time_start = time.time()
if rng is None:
rng = np.random.RandomState(np.random.randint(0, 10000))
n_dims = lower.shape[0]
# Bookkeeping
time_func_eval = []
time_overhead = []
incumbents = []
incumbent_values = []
runtime = []
X = deepcopy(observed_X)
y = deepcopy(observed_y)
# Define model for the objective function
cov_amp = 1 # Covariance amplitude
kernel = cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims + 1,
dim=d)
task_kernel = george.kernels.TaskKernel(n_dims + 1, n_dims, n_tasks)
kernel *= task_kernel
# Take 3 times more samples than we have hyperparameters
if n_hypers < 2 * len(kernel):
n_hypers = 3 * len(kernel)
if n_hypers % 2 == 1:
n_hypers += 1
prior = MTBOPrior(len(kernel) + 1,
n_ls=n_dims,
n_kt=len(task_kernel),
rng=rng)
model_objective = MTBOGPMCMC(kernel,
prior=prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
lower=lower,
upper=upper,
rng=rng)
acquisition_func = LogEI(model_objective)
# Optimize acquisition function only on the main task
def wrapper(x):
x_ = np.append(x, np.ones([x.shape[0], 1]) * target_task_id, axis=1)
if y.shape[0] == init_points:
eta = 0
else:
eta = np.min(y[init_points:])
a = acquisition_func(x_, eta=eta)
return a
maximizer = Direct(wrapper, lower, upper, n_func_evals=200)
X = np.array(X)
y = np.array(y)
init_points = y.shape[0]
for it in range(num_iterations):
logger.info("Start iteration %d ... ", it)
start_time = time.time()
# Train models
model_objective.train(X, y, do_optimize=True)
# Maximize acquisition function
acquisition_func.update(model_objective)
new_x = maximizer.maximize()
new_x = np.append(new_x, np.array([target_task_id]))
time_overhead.append(time.time() - start_time)
logger.info("Optimization overhead was %f seconds", time_overhead[-1])
# Evaluate the chosen configuration
logger.info("Evaluate candidate %s", str(new_x))
start_time = time.time()
new_y = objective_function(new_x[:-1], int(new_x[-1]))
time_func_eval.append(time.time() - start_time)
logger.info("Configuration achieved a performance of %f", new_y)
logger.info("Evaluation of this configuration took %f seconds",
time_func_eval[-1])
# Add new observation to the data
X = np.concatenate((X, new_x[None, :]), axis=0)
y = np.concatenate(
(y, np.array([new_y])),
axis=0) # Model the target function on a logarithmic scale
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y[init_points:]) + init_points
incumbent = X[best_idx][:-1]
incumbent_value = y[best_idx]
incumbents.append(incumbent)
incumbent_values.append(incumbent_value)
logger.info("Current incumbent %s with estimated performance %f",
str(incumbent), incumbent_value)
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(
data,
open(os.path.join(output_path, "mtbo_iter_%d.json" % it), "w"))
logger.info("Final incumbent %s with estimated performance %f",
str(incumbent), incumbent_value)
results = dict()
results["x_opt"] = incumbent.tolist()
results["incumbents"] = [inc.tolist() for inc in incumbents]
results["runtime"] = runtime
results["overhead"] = time_overhead
results["time_func_eval"] = time_func_eval
results["incumbent_values"] = incumbent_values
results["X"] = X
results["y"] = y
return results
def fabolas(objective_function, lower, upper, s_min, s_max,
n_init=40, num_iterations=100, subsets=[256, 128, 64], inc_estimation="mean",
burnin=100, chain_length=100, n_hypers=12, output_path=None, rng=None):
"""
Fast Bayesian Optimization of Machine Learning Hyperparameters
on Large Datasets
Fast Bayesian Optimization of Machine Learning Hyperparameters on Large Datasets
A. Klein and S. Falkner and S. Bartels and P. Hennig and F. Hutter
http://arxiv.org/abs/1605.07079
Parameters
----------
objective_function: function
Objective function that will be optimized
lower: np.array(D,)
Lower bound of the input space
upper: np.array(D,)
Upper bound of the input space
s_min: int
Minimum number of data points for the training data set
s_max: int
Maximum number of data points for the training data set
n_init: int
Number of initial design points
n_hypers: int
Number of hyperparameter samples for the GP
subsets: list
The ratio of the subsets size of the initial design.
For example if subsets=[256, 128, 64] then the first random point from the
initial design is evaluated on s_max/256 of the data, the second point on
s_max/256 of the data and so on.
num_iterations: int
Number of iterations
chain_length : int
The length of the MCMC chain for each walker.
burnin : int
The number of burnin steps before the actual MCMC sampling starts.
output_path: string
Specifies the path where the intermediate output after each iteration will be saved.
If None no output will be saved to disk.
rng: numpy.random.RandomState
Random number generator
Returns
-------
dict
"""
assert n_init <= num_iterations, "Number of initial design point has to be <= than the number of iterations"
assert lower.shape[0] == upper.shape[0], "Dimension miss match between upper and lower bound"
time_start = time.time()
if rng is None:
rng = np.random.RandomState(np.random.randint(0, 10000))
n_dims = lower.shape[0]
# Bookkeeping
time_func_eval = []
time_overhead = []
incumbents = []
runtime = []
X = []
y = []
c = []
# Define model for the objective function
cov_amp = 1 # Covariance amplitude
kernel = cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims+1, dim=d)
# Kernel for the environmental variable
# We use (1-s)**2 as basis function for the Bayesian linear kernel
degree = 1
env_kernel = george.kernels.BayesianLinearRegressionKernel(n_dims+1,
dim=n_dims,
degree=degree)
env_kernel[:] = np.ones([degree + 1]) * 0.1
kernel *= env_kernel
# Take 3 times more samples than we have hyperparameters
if n_hypers < 2*len(kernel):
n_hypers = 3 * len(kernel)
if n_hypers % 2 == 1:
n_hypers += 1
prior = EnvPrior(len(kernel) + 1,
n_ls=n_dims,
n_lr=(degree + 1),
rng=rng)
quadratic_bf = lambda x: (1 - x) ** 2
linear_bf = lambda x: x
model_objective = FabolasGPMCMC(kernel,
prior=prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
normalize_output=False,
basis_func=quadratic_bf,
lower=lower,
upper=upper,
rng=rng)
# Define model for the cost function
cost_cov_amp = 1
cost_kernel = cost_cov_amp
# ARD Kernel for the configuration space
for d in range(n_dims):
cost_kernel *= george.kernels.Matern52Kernel(np.ones([1]) * 0.01,
ndim=n_dims+1, dim=d)
cost_degree = 1
cost_env_kernel = george.kernels.BayesianLinearRegressionKernel(n_dims+1,
dim=n_dims,
degree=cost_degree)
cost_env_kernel[:] = np.ones([cost_degree + 1]) * 0.1
cost_kernel *= cost_env_kernel
cost_prior = EnvPrior(len(cost_kernel) + 1,
n_ls=n_dims,
n_lr=(cost_degree + 1),
rng=rng)
model_cost = FabolasGPMCMC(cost_kernel,
prior=cost_prior,
burnin_steps=burnin,
chain_length=chain_length,
n_hypers=n_hypers,
basis_func=linear_bf,
normalize_output=False,
lower=lower,
upper=upper,
rng=rng)
# Extend input space by task variable
extend_lower = np.append(lower, 0)
extend_upper = np.append(upper, 1)
is_env = np.zeros(extend_lower.shape[0])
is_env[-1] = 1
# Define acquisition function and maximizer
ig = InformationGainPerUnitCost(model_objective,
model_cost,
extend_lower,
extend_upper,
sampling_acquisition=EI,
is_env_variable=is_env,
n_representer=50)
acquisition_func = MarginalizationGPMCMC(ig)
maximizer = Direct(acquisition_func, extend_lower, extend_upper, verbose=True, n_func_evals=200)
# Initial Design
logger.info("Initial Design")
for it in range(n_init):
start_time_overhead = time.time()
# Draw random configuration
s = int(s_max / float(subsets[it]))
x = init_random_uniform(lower, upper, 1, rng)[0]
logger.info("Evaluate %s on subset size %d", str(x), s)
st = time.time()
func_val, cost = objective_function(x, s)
time_func_eval.append(time.time() - st)
logger.info("Configuration achieved a performance of %f with cost %f", func_val, cost)
logger.info("Evaluation of this configuration took %f seconds", time_func_eval[-1])
# Bookkeeping
config = np.append(x, transform(s, s_min, s_max))
X.append(config)
y.append(np.log(func_val)) # Model the target function on a logarithmic scale
c.append(np.log(cost)) # Model the cost on a logarithmic scale
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y)
incumbents.append(X[best_idx][:-1]) # Incumbent is always on s=s_max
time_overhead.append(time.time() - start_time_overhead)
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(data, open(os.path.join(output_path, "fabolas_iter_%d.json" % it), "w"))
X = np.array(X)
y = np.array(y)
c = np.array(c)
for it in range(n_init, num_iterations):
logger.info("Start iteration %d ... ", it)
start_time = time.time()
# Train models
model_objective.train(X, y, do_optimize=True)
model_cost.train(X, c, do_optimize=True)
if inc_estimation == "last_seen":
# Estimate incumbent as the best observed value so far
best_idx = np.argmin(y)
incumbent = X[best_idx][:-1]
incumbent = np.append(incumbent, 1)
incumbent_value = y[best_idx]
else:
# Estimate incumbent by projecting all observed points to the task of interest and
# pick the point with the lowest mean prediction
incumbent, incumbent_value = projected_incumbent_estimation(model_objective, X[:, :-1],
proj_value=1)
incumbents.append(incumbent[:-1])
logger.info("Current incumbent %s with estimated performance %f",
str(incumbent), np.exp(incumbent_value))
# Maximize acquisition function
acquisition_func.update(model_objective, model_cost)
new_x = maximizer.maximize()
s = retransform(new_x[-1], s_min, s_max) # Map s from log space to original linear space
time_overhead.append(time.time() - start_time)
logger.info("Optimization overhead was %f seconds", time_overhead[-1])
# Evaluate the chosen configuration
logger.info("Evaluate candidate %s on subset size %f", str(new_x[:-1]), s)
start_time = time.time()
new_y, new_c = objective_function(new_x[:-1], s)
time_func_eval.append(time.time() - start_time)
logger.info("Configuration achieved a performance of %f with cost %f", new_y, new_c)
logger.info("Evaluation of this configuration took %f seconds", time_func_eval[-1])
# Add new observation to the data
X = np.concatenate((X, new_x[None, :]), axis=0)
y = np.concatenate((y, np.log(np.array([new_y]))), axis=0) # Model the target function on a logarithmic scale
c = np.concatenate((c, np.log(np.array([new_c]))), axis=0) # Model the cost function on a logarithmic scale
runtime.append(time.time() - time_start)
if output_path is not None:
data = dict()
data["optimization_overhead"] = time_overhead[it]
data["runtime"] = runtime[it]
data["incumbent"] = incumbents[it].tolist()
data["time_func_eval"] = time_func_eval[it]
data["iteration"] = it
json.dump(data, open(os.path.join(output_path, "fabolas_iter_%d.json" % it), "w"))
# Estimate the final incumbent
model_objective.train(X, y, do_optimize=True)
incumbent, incumbent_value = projected_incumbent_estimation(model_objective, X[:, :-1],
proj_value=1)
logger.info("Final incumbent %s with estimated performance %f",
str(incumbent), incumbent_value)
results = dict()
results["x_opt"] = incumbent[:-1].tolist()
results["incumbents"] = [inc.tolist() for inc in incumbents]
results["runtime"] = runtime
results["overhead"] = time_overhead
results["time_func_eval"] = time_func_eval
results["X"] = X
results["y"] = y
results["c"] = c
return results