def test_rstate():
"""Test the rstate helper."""
rng = rstate()
rng = rstate(rng)
rng1 = rstate(1)
rng2 = rstate(1)
nt.assert_equal(rng1.randint(5), rng2.randint(5))
nt.assert_raises(ValueError, rstate, 'foo')
def _sample_prior(model, priors, n, rng=None):
rng = rstate(rng)
# unpack priors
# TODO -- Bobak: This snippet is copied from learning/sampling.py
# and should probably be put into a Prior base class.
priors = dict(priors)
active = np.ones(model.nhyper, dtype=bool)
logged = np.ones(model.nhyper, dtype=bool)
for (key, block, log) in get_params(model):
inactive = (key in priors) and (priors[key] is None)
logged[block] = log
active[block] = not inactive
if inactive:
del priors[key]
else:
priors[key] = (block, log, priors[key])
priors = priors.values()
# sample hyperparameters from prior
hypers = np.tile(model.get_hyper(), (n, 1))
for (block, log, prior) in priors:
hypers[:, block] = (np.log(prior.sample(n, rng=rng)) if log else
prior.sample(n, rng=rng))
return hypers
def sample(self, X, m=None, latent=True, rng=None):
"""
Sample values from the posterior at points `X`. Given an `(n,d)`-array
`X` this will return an `n`-vector corresponding to the resulting
sample.
If `m` is not `None` an `(m,n)`-array will be returned instead,
corresponding to `m` such samples. If `latent` is `False` the sample
will instead be returned corrupted by the observation noise. Finally
`rng` can be used to seed the randomness.
"""
X = self._kernel.transform(X)
# this boolean indicates whether we'll flatten the sample to return a
# vector, or if we'll return a set of samples as an array.
flatten = (m is None)
# get the relevant sizes.
m = 1 if flatten else m
n = len(X)
# if a seed or instantiated RandomState is given use that, otherwise
# use the global object.
rng = rstate(rng)
# add a tiny amount to the diagonal to make the cholesky of Sigma
# stable and then add this correlated noise onto mu to get the sample.
mu, Sigma = self._full_posterior(X)
Sigma += 1e-10 * np.eye(n)
f = mu[None] + np.dot(rng.normal(size=(m, n)), sla.cholesky(Sigma))
if not latent:
f = self._likelihood.sample(f.ravel(), rng).reshape(m, n)
return f.ravel() if flatten else f
def sample(self, X, m=None, latent=True, rng=None):
"""
Sample values from the posterior at points `X`. Given an `(n,d)`-array
`X` this will return an `n`-vector corresponding to the resulting
sample.
If `m` is not `None` an `(m,n)`-array will be returned instead,
corresponding to `m` such samples. If `latent` is `False` the sample
will instead be returned corrupted by the observation noise. Finally
`rng` can be used to seed the randomness.
"""
X = self._kernel.transform(X)
# this boolean indicates whether we'll flatten the sample to return a
# vector, or if we'll return a set of samples as an array.
flatten = (m is None)
# get the relevant sizes.
m = 1 if flatten else m
n = len(X)
# if a seed or instantiated RandomState is given use that, otherwise
# use the global object.
rng = rstate(rng)
# add a tiny amount to the diagonal to make the cholesky of Sigma
# stable and then add this correlated noise onto mu to get the sample.
mu, Sigma = self._full_posterior(X)
Sigma += 1e-10 * np.eye(n)
f = mu[None] + np.dot(rng.normal(size=(m, n)), sla.cholesky(Sigma))
if not latent:
f = self._likelihood.sample(f.ravel(), rng).reshape(m, n)
return f.ravel() if flatten else f
def sample(self, size=1, rng=None):
rng = rstate(rng)
if self._std.ndim == 1:
sample = self._mu + self._std * rng.randn(size, self.ndim)
elif self._s2.ndim == 2:
sample = self._mu + np.dot(rng.randn(size, self.ndim), self._std)
return sample
def sample_spectrum(self, N, rng=None):
rng = rstate(rng)
sf2 = np.exp(self._logsf*2)
ell = np.exp(self._logell)
a = self._d / 2.
g = np.tile(rng.gamma(a, 1/a, N), (self.ndim, 1)).T
W = (rng.randn(N, self.ndim) / ell) / np.sqrt(g)
return W, sf2
def sample(self, size=1, rng=None):
rng = rstate(rng)
if self._std.ndim == 1:
sample = self._mu + self._std * rng.randn(size, self.ndim)
elif self._s2.ndim == 2:
sample = self._mu + np.dot(rng.randn(size, self.ndim), self._std)
return sample
def sample_spectrum(self, N, rng=None):
rng = rstate(rng)
sf2 = np.exp(self._logsf * 2)
ell = np.exp(self._logell)
a = self._d / 2.
g = np.tile(rng.gamma(a, 1 / a, N), (self.ndim, 1)).T
W = (rng.randn(N, self.ndim) / ell) / np.sqrt(g)
return W, sf2
def __init__(self, bounds, gp, N=None, rng=None):
self.bounds = np.array(bounds, dtype=float, ndmin=2)
self._gp = gp.copy()
self._rng = random.rstate(rng)
# generate some sampled observations.
N = N if (N is not None) else 100 * len(self.bounds)
X = random.latin(bounds, N, self._rng)
y = self._gp.sample(X, latent=False, rng=self._rng)
# add them back to get a new "posterior".
self._gp.add_data(X, y)
def _slice_sample(logprob,
x0,
sigma=1.0,
step_out=True,
max_steps_out=1000,
rng=None):
"""
Implementation of slice sampling taken almost directly from Snoek's
spearmint package (with a few minor modifications).
"""
rng = rstate(rng)
def direction_slice(direction, x0):
def dir_logprob(z):
return logprob(direction * z + x0)
upper = sigma * rng.rand()
lower = upper - sigma
llh_s = np.log(rng.rand()) + dir_logprob(0.0)
l_steps_out = 0
u_steps_out = 0
if step_out:
while dir_logprob(lower) > llh_s and l_steps_out < max_steps_out:
l_steps_out += 1
lower -= sigma
while dir_logprob(upper) > llh_s and u_steps_out < max_steps_out:
u_steps_out += 1
upper += sigma
while True:
new_z = (upper - lower) * rng.rand() + lower
new_llh = dir_logprob(new_z)
if np.isnan(new_llh):
raise Exception("Slice sampler got a NaN")
if new_llh > llh_s:
break
elif new_z < 0:
lower = new_z
elif new_z > 0:
upper = new_z
else:
raise Exception("Slice sampler shrank to zero!")
return new_z * direction + x0
# FIXME: I've removed how blocks work because I want to rewrite that bit.
# so right now this samples everything as one big block.
direction = rng.randn(x0.shape[0])
direction = direction / np.sqrt(np.sum(direction**2))
return direction_slice(direction, x0)
def _slice_sample(logprob,
x0,
sigma=1.0,
step_out=True,
max_steps_out=1000,
rng=None):
"""
Implementation of slice sampling taken almost directly from Snoek's
spearmint package (with a few minor modifications).
"""
rng = rstate(rng)
def direction_slice(direction, x0):
def dir_logprob(z):
return logprob(direction*z + x0)
upper = sigma*rng.rand()
lower = upper - sigma
llh_s = np.log(rng.rand()) + dir_logprob(0.0)
l_steps_out = 0
u_steps_out = 0
if step_out:
while dir_logprob(lower) > llh_s and l_steps_out < max_steps_out:
l_steps_out += 1
lower -= sigma
while dir_logprob(upper) > llh_s and u_steps_out < max_steps_out:
u_steps_out += 1
upper += sigma
while True:
new_z = (upper - lower)*rng.rand() + lower
new_llh = dir_logprob(new_z)
if np.isnan(new_llh):
raise Exception("Slice sampler got a NaN")
if new_llh > llh_s:
break
elif new_z < 0:
lower = new_z
elif new_z > 0:
upper = new_z
else:
raise Exception("Slice sampler shrank to zero!")
return new_z*direction + x0
# FIXME: I've removed how blocks work because I want to rewrite that bit.
# so right now this samples everything as one big block.
direction = rng.randn(x0.shape[0])
direction = direction / np.sqrt(np.sum(direction**2))
return direction_slice(direction, x0)
def __init__(self, model, prior, n=100, rng=None):
self._prior = prior
self._n = n
self._rng = rstate(rng)
# we won't add any data unless the model already has it.
data = None
if model.ndata > 0:
data = model.data
model = model.copy()
model.reset()
self._samples = [model.copy(h)
for h in _sample_prior(model, prior, n, rng=self._rng)]
self._logweights = np.zeros(n) - np.log(n)
self._loglikes = np.zeros(n)
if data is not None:
self.add_data(data[0], data[1])
def __init__(self, model, prior, n=100, burn=100, rng=None):
self._model = model.copy()
self._prior = prior
self._samples = []
self._n = n
self._burn = burn
self._rng = rstate(rng)
if self._model.ndata > 0:
if self._burn > 0:
sample(self._model, self._prior, self._burn, rng=self._rng)
self._samples = sample(self._model,
self._prior,
self._n,
raw=False,
rng=self._rng)
else:
# FIXME: the likelihood won't play a role, so we can sample
# directly from the prior. This of course requires the prior to
# also be a well-defined distribution.
pass
def __init__(self, model, prior, n=100, burn=100, rng=None):
self._model = model.copy()
self._prior = prior
self._samples = []
self._n = n
self._burn = burn
self._rng = rstate(rng)
if self._model.ndata > 0:
if self._burn > 0:
sample(self._model, self._prior, self._burn, rng=self._rng)
self._samples = sample(self._model,
self._prior,
self._n,
raw=False,
rng=self._rng)
else:
# FIXME: the likelihood won't play a role, so we can sample
# directly from the prior. This of course requires the prior to
# also be a well-defined distribution.
pass
def __init__(self, N, likelihood, kernel, mean, X, y, rng=None):
# if given a seed or an instantiated RandomState make sure that we use
# it here, but also within the sample_spectrum code.
rng = rstate(rng)
if not isinstance(likelihood, Gaussian):
raise ValueError('Fourier samples only defined for Gaussian'
'likelihoods')
# this randomizes the feature.
W, alpha = kernel.sample_spectrum(N, rng)
self._W = W
self._b = rng.rand(N) * 2 * np.pi
self._a = np.sqrt(2 * alpha / N)
self._mean = mean
self._theta = None
if X is not None:
# evaluate the features
Z = np.dot(X, self._W.T) + self._b
Phi = np.cos(Z) * self._a
# get the components for regression
A = np.dot(Phi.T, Phi) + likelihood.s2 * np.eye(Phi.shape[1])
R = sla.cholesky(A)
r = y - mean
p = np.sqrt(likelihood.s2) * rng.randn(N)
# FIXME: we can do a smarter update here when the number of points
# is less than the number of features.
self._theta = sla.cho_solve((R, False), np.dot(Phi.T, r))
self._theta += sla.solve_triangular(R, p)
else:
self._theta = rng.randn(N)
def solve_bayesopt_dim(self,
objective,
bounds,
kernel="se",
niter=150,
dims=2,
init='sobol',
policy='ei',
solver='direct',
recommender='latent'):
"""
The following part, borrows some code from the pybo code.
Therefore, retains a similar structure and parameters.
From Pybo (with updates):
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
dims: number of dimensions per permutation
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
Nothing, everything is written to a file
"""
filename = self.outputdir + "_".join([policy[0], kernel, str(niter)])
bounds = np.array(bounds, dtype=float, ndmin=2)
# initialize the random number generator.
rng = rstate(None)
# get the model components.
init, policy, solver, recommender = \
self.get_components(init, policy, solver, recommender, rng)
# Create the bmodel with the data
model = self.initialmodel(objective, init, bounds, kernel)
with open(filename, "a+") as f:
for i, y in enumerate(model.data[1]):
f.write("0," + str(y) + ",0,0," + ",".join([str(a) for a in model.data[0][i]]) + "\n")
# Set up Dimension Scheduler Dictionaries
modelsDict = {}
dimDict = {}
bY = np.argmax(model.data[1])
bX = model.data[0][bY]
objective.set_initial(bX)
objective.set_best(model.data[1][bY])
dimensionsProb = [1.0/len(bounds)] * len(bounds)
self.console("Starting Bayesian Optimization with Dimensions Scheduler", 1)
self.console("Time,Objective,Mu,Var,X", 2)
for i in xrange(model.ndata, niter):
start = time.clock()
# Update the probabilities
if (i % 50 == 0):
mydata = model.data[0]
mu = mydata.mean(axis=0)
sigma = mydata.std(axis=0)
a = (mydata - mu) / sigma
C = np.cov(a.T)
evals, evecs = np.linalg.eig(C)
vars = evals / float(len(evals))
dimensionsProb = vars / vars.sum()
# generate new dimension and update the bounds and set the dimensions on the objective
d = self.getRandomDimensions(dimensionsProb, dims)
objective.set_dimensions(d)
boundslowerdim = objective.get_bounds()
# check if the model exists, else create a new one
if not modelsDict.has_key(str(d)):
tX ,tY= self.getInputsDimRed(model, d)
m = self.createNewModel(tX, tY, boundslowerdim, kernel)
try:
m.add_data(tX, tY)
except:
self.console("An error occured during modle creation during inputing data points into the GP. Try increasing GP:sn parameter ")
break
modelsDict[str(d)] = m
dimDict[str(d)] = d
# get model for current dimensions
modellowerdim = modelsDict.get(str(d))
# get the next point to evaluate.
index = policy(modellowerdim)
x, _ = solver(index, boundslowerdim)
# make an observation and record it.
try:
y = objective(x)
yp = str(y)
curX = objective.get_prev_input()
model.add_data(curX, y)
modellowerdim.add_data(x, y)
if objective.get_best() is not None:
if (y > objective.get_best()):
objective.set_best(y)
objective.change_initial(x)
else:
objective.set_best(y)
objective.change_initial(x)
except:
curX = objective.get_prev_input()
yp = "Failed"
glomu, glovar = model.posterior(curX, grad=False)[:2]
# Write to teh file
interval = time.clock() - start
data = str(interval) + "," + yp + "," + str(glomu[0]) + "," + str(glovar[0]) + ",".join(
[str(a) for a in curX])
self.console(data, 2)
with open(filename, "a+") as f:
f.write(data + "\n")
# Get the recommender point from the model
recx = recommender(model, bounds)
recmu, recvar = model.posterior(recx, grad=False)[:2]
with open(filename, "a+") as f:
f.write("0,Rec.," + str(recmu) + "," + str(recvar) + "," + ",".join([str(s) for s in recx]) + "\n")
def sample_spectrum(self, N, rng=None):
rng = rstate(rng)
sf2 = np.exp(self._logsf*2)
ell = np.exp(self._logell)
W = rng.randn(N, self.ndim) / ell
return W, sf2
def solve_bayesopt(objective,
bounds,
grad,
niter=100,
init='middle',
policy='ei',
solver='lbfgs',
recommender='latent',
model=None,
noisefree=False,
ftrue=None,
rng=None,
callback=None,
n_grad=20):
"""
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
model: the Bayesian model instantiation.
noisefree: a boolean denoting that the model is noisefree; this only
applies if a default model is used (ie. it is ignored if the
model argument is used).
ftrue: a ground-truth function (for evaluation).
rng: either an RandomState object or an integer used to seed the state;
this will be fed to each component that requests randomness.
callback: a function to call on each iteration for visualization.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
A numpy record array containing a trace of the optimization process.
The fields of this array are `x`, `y`, and `xbest` corresponding to the
query locations, outputs, and recommendations at each iteration. If
ground-truth is known an additional field `fbest` will be included.
"""
# make sure the bounds are a 2d-array.
bounds = np.array(bounds, dtype=float, ndmin=2)
# see if the query object itself defines ground truth.
if (ftrue is None) and hasattr(objective, 'get_f'):
ftrue = objective.get_f
# initialize the random number generator.
rng = rstate(rng)
# get the model components.
init, policy, solver, recommender = \
get_components(init, policy, solver, recommender, rng)
# create a list of initial points to query.
X = init(bounds)
Y = [objective(x) for x in X]
if model is None:
# initialize parameters of a simple GP model.
sf = np.std(Y) if (len(Y) > 1) else 10.
mu = np.mean(Y)
ell = bounds[:, 1] - bounds[:, 0]
# FIXME: this may not be a great setting for the noise parameter
sn = 1e-5 if noisefree else 1e-3
# specify a hyperprior for the GP.
prior = {
'sn': (
None if noisefree else
pygp.priors.Horseshoe(scale=0.1, min=1e-5)),
'sf': pygp.priors.LogNormal(mu=np.log(sf), sigma=1., min=1e-6),
'ell': pygp.priors.Uniform(ell / 100, ell * 2),
'mu': pygp.priors.Gaussian(mu, sf)}
# create the GP model (with hyperprior).
model = pygp.BasicGP(sn, sf, ell, mu, kernel='matern5')
model = pygp.meta.MCMC(model, prior, n=10, burn=100, rng=rng)
# add any initial data to our model.
model.add_data(X, Y)
# allocate a datastructure containing "convergence" info.
info = np.zeros(niter, [('x', np.float, (len(bounds),)),
('y', np.float),
('xbest', np.float, (len(bounds),))])
# initialize the data.
info['x'][:len(X)] = X
info['y'][:len(Y)] = Y
info['xbest'][:len(Y)] = [X[np.argmax(Y[:i+1])] for i in xrange(len(Y))]
for i in xrange(model.ndata, niter):
# get the next point to evaluate.
index = policy(model)
x, _ = solver(index, bounds)
# deal with any visualization.
if callback is not None:
callback(model, bounds, info[:i], x, index, ftrue)
# make an observation and record it.
y = objective(x)
model.add_data(x, y)
# record everything.
info[i] = (x, y, recommender(model, bounds))
if ftrue is not None:
fbest = ftrue(info['xbest'])
info = append_fields(info, 'fbest', fbest, usemask=False)
return info
def solve_bayesopt(self,
objective,
bounds,
kernel="se",
niter=150,
init='sobol',
policy='ei',
solver='direct',
recommender='latent'):
"""
The following part, borrows some code from the pybo code.
Therefore, retains a similar structure and parameters.
From Pybo:
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
A dictionary with the model, recommended values, and finished iteration
"""
filename = self.outputdir + "_".join([policy[0], kernel, str(niter)])
bounds = np.array(bounds, dtype=float, ndmin=2)
rng = rstate(None)
# get the model components.
init, policy, solver, recommender = \
self.get_components(init, policy, solver, recommender, rng)
# Create the model with the data
model = self.initialmodel(objective, init, bounds, kernel)
# Write the initial data to a file
with open(filename, "a+") as f:
for i, y in enumerate(model.data[1]):
f.write("0," + str(y) + ",0,0," + ",".join([str(a) for a in model.data[0][i]]) + "\n")
# Send data to the observer
self.pipeout.send({"inidata": {"x": model.data[0].tolist(), "y": model.data[1].tolist()}})
self.console("Starting Bayesian Optimization")
# Start BayesOpt
for i in xrange(model.ndata, niter):
# Record current time and iteration
start = time.clock()
curiter = i
# Check if Observer has terminated the process
if self.pipeout.poll():
o = self.pipeout.recv()
if o.has_key("stop"):
self.stop = o["stop"]
if self.stop:
break
# get the next point to evaluate.
index = policy(model)
x, _ = solver(index, bounds)
glomu, glovar = model.posterior(x, grad=False)[:2]
# make an observation and record it.
try:
y = objective(x)
model.add_data(x, y)
yp = str(y)
except:
yp = "Failed"
# Send out the data to the observer and write to the file
interval = time.clock() - start
with open(filename, "a+") as f:
f.write(str(interval) + "," + yp + "," + str(glomu[0]) + "," + str(glovar[0]) + ",".join(
[str(a) for a in x]) + "\n")
if yp != "Failed":
self.pipeout.send({"data": {"y": y, "x": x, "mu": glomu[0], "var": glovar[0], "time": interval}})
# Get the recommender point from the model
recx = recommender(model, bounds)
recmu, recvar = model.posterior(recx, grad=False)[:2]
with open(filename, "a+") as f:
f.write("0,Rec.," + str(recmu[0]) + "," + str(recvar[0]) + "," + ",".join([str(s) for s in recx]) + "\n")
self.console("Finished experiment")
return {"model": model, "iterfinish": curiter, "rec": {"x": recx, "mu": recmu, "var": recvar}}
def sample(self, f, rng=None):
rng = rstate(rng)
return f + rng.normal(size=len(f), scale=np.exp(self._logsigma))
def sample(gp, priors, n, raw=True, rng=None):
rng = rstate(rng)
priors = dict(priors)
active = np.ones(gp.nhyper, dtype=bool)
logged = np.ones(gp.nhyper, dtype=bool)
for (key, block, log) in get_params(gp):
inactive = (key in priors) and (priors[key] is None)
logged[block] = log
active[block] = not inactive
if inactive:
del priors[key]
else:
priors[key] = (block, log, priors[key])
# priors is now just a list of the form (block, log, prior).
priors = priors.values()
# get the initial hyperparameters and transform into the non-log space.
hyper0 = gp.get_hyper()
hyper0[logged] = np.exp(hyper0[logged])
def logprob(x):
# copy the initial hyperparameters and then assign the "active"
# parameters that come from x.
hyper = hyper0.copy()
hyper[active] = x
logprob = 0
# compute the prior probabilities. we do this first so that if there
# are any infs they'll be caught in the least expensive computations
# first.
for block, log, prior in priors:
logprob += prior.logprior(hyper[block])
if np.isinf(logprob):
break
# now compute the likelihood term. note that we'll have to take the log
# of any logspace parameters before calling set_hyper.
if not np.isinf(logprob):
hyper[logged] = np.log(hyper[logged])
gp.set_hyper(hyper)
logprob += gp.loglikelihood()
return logprob
# create a big list of the hyperparameters so that we can just assign to
# the components that are active. also get an initial sample x
# corresponding only to the active parts of hyper0.
hypers = np.tile(hyper0, (n, 1))
x = hyper0.copy()[active]
# do the sampling.
for i in xrange(n):
x = _slice_sample(logprob, x, rng=rng)
hypers[i][active] = x
# change the logspace components back into logspace.
hypers[:, logged] = np.log(hypers[:, logged])
# make sure the gp gets updated to the last sampled hyperparameter.
gp.set_hyper(hypers[-1])
if raw:
return hypers
else:
return [gp.copy(h) for h in hypers]
def sample(self, size=1, rng=None):
rng = rstate(rng)
return np.vstack(self._min + rng.gamma(k, s, size=size)
for k, s in zip(self._k, self._scale)).T
def __init__(self, sigma=0.0, rng=None):
self._sigma = sigma
self._rng = rstate(rng)
def solve_bayesopt(objective,
bounds,
niter=100,
init='middle',
policy='ei',
solver='lbfgs',
recommender='latent',
model=None,
noisefree=False,
ftrue=None,
rng=None,
callback=None):
"""
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
model: the Bayesian model instantiation.
noisefree: a boolean denoting that the model is noisefree; this only
applies if a default model is used (ie. it is ignored if the
model argument is used).
ftrue: a ground-truth function (for evaluation).
rng: either an RandomState object or an integer used to seed the state;
this will be fed to each component that requests randomness.
callback: a function to call on each iteration for visualization.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
A numpy record array containing a trace of the optimization process.
The fields of this array are `x`, `y`, and `xbest` corresponding to the
query locations, outputs, and recommendations at each iteration. If
ground-truth is known an additional field `fbest` will be included.
"""
# make sure the bounds are a 2d-array.
bounds = np.array(bounds, dtype=float, ndmin=2)
# see if the query object itself defines ground truth.
if (ftrue is None) and hasattr(objective, 'get_f'):
ftrue = objective.get_f
# initialize the random number generator.
rng = rstate(rng)
# get the model components.
init, policy, solver, recommender = \
get_components(init, policy, solver, recommender, rng)
# create a list of initial points to query.
X = init(bounds)
Y = [objective(x) for x in X]
if model is None:
# initialize parameters of a simple GP model.
sf = np.std(Y) if (len(Y) > 1) else 10.
mu = np.mean(Y)
ell = bounds[:, 1] - bounds[:, 0]
# FIXME: this may not be a great setting for the noise parameter
sn = 1e-5 if noisefree else 1e-3
# specify a hyperprior for the GP.
prior = {
'sn': (None if noisefree else pygp.priors.Horseshoe(scale=0.1,
min=1e-5)),
'sf':
pygp.priors.LogNormal(mu=np.log(sf), sigma=1., min=1e-6),
'ell':
pygp.priors.Uniform(ell / 100, ell * 2),
'mu':
pygp.priors.Gaussian(mu, sf)
}
# create the GP model (with hyperprior).
model = pygp.BasicGP(sn, sf, ell, mu, kernel='matern5')
model = pygp.meta.MCMC(model, prior, n=10, burn=100, rng=rng)
# add any initial data to our model.
model.add_data(X, Y)
# allocate a datastructure containing "convergence" info.
info = np.zeros(niter, [('x', np.float, (len(bounds), )), ('y', np.float),
('xbest', np.float, (len(bounds), ))])
# initialize the data.
info['x'][:len(X)] = X
info['y'][:len(Y)] = Y
info['xbest'][:len(Y)] = [X[np.argmax(Y[:i + 1])] for i in xrange(len(Y))]
for i in xrange(model.ndata, niter):
# get the next point to evaluate.
index = policy(model)
x, _ = solver(index, bounds)
# deal with any visualization.
if callback is not None:
callback(model, bounds, info[:i], x, index, ftrue)
# make an observation and record it.
y = objective(x)
model.add_data(x, y)
# record everything.
info[i] = (x, y, recommender(model, bounds))
if ftrue is not None:
fbest = ftrue(info['xbest'])
info = append_fields(info, 'fbest', fbest, usemask=False)
return info
def sample(self, size=1, rng=None):
rng = rstate(rng)
return np.vstack(self._min + rng.lognormal(m, s, size=size)
for m, s in zip(self._mu, self._sigma)).T
def solve_hyperopt(objective,
bounds,
niter=100,
init='middle',
policy='ei',
solver='lbfgs',
recommender='latent',
model=None,
noisefree=False,
ftrue=None,
rng=None,
callback=None):
"""
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
model: the Bayesian model instantiation.
noisefree: a boolean denoting that the model is noisefree; this only
applies if a default model is used (ie. it is ignored if the
model argument is used).
ftrue: a ground-truth function (for evaluation).
rng: either an RandomState object or an integer used to seed the state;
this will be fed to each component that requests randomness.
callback: a function to call on each iteration for visualization.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
A numpy record array containing a trace of the optimization process.
The fields of this array are `x`, `y`, and `xbest` corresponding to the
query locations, outputs, and recommendations at each iteration. If
ground-truth is known an additional field `fbest` will be included.
"""
# make sure the bounds are a 2d-array.
bounds = np.array(bounds, dtype=float, ndmin=2)
# see if the query object itself defines ground truth.
if (ftrue is None) and hasattr(objective, 'get_f'):
ftrue = objective.get_f
# initialize the random number generator.
rng = rstate(rng)
# get the model components.
init, policy, solver, recommender = \
get_components(init, policy, solver, recommender, rng)
# create a list of initial points to query.
X = init(bounds)
Y = [objective(x) for x in X]
# add any initial data to our model.
model.add_data(X, Y)
# allocate a datastructure containing "convergence" info.
info = np.zeros(niter, [('x', np.float, (len(bounds),)),
('y', np.float),
('xbest', np.float, (len(bounds),))])
# initialize the data.
info['x'][:len(X)] = X
info['y'][:len(Y)] = Y
info['xbest'][:len(Y)] = [X[np.argmax(Y[:i+1])] for i in xrange(len(Y))]
# BFGS callback
X_y = ([], [])
def callbackF(Xi):
global X_y
X_y[0].append(Xi)
X_y[1].append(objective(Xi))
for i in xrange(niter):
# take n_grad gradient steps
res = minimize(objective, x0, method='BFGS',
jac=grad,
options={'gtol': 1e-6, 'disp': True, 'maxiter': n_grad},
callback=callbackF)
# get the next point to evaluate.
index = policy(model)
x, _ = solver(index, bounds)
# deal with any visualization.
if callback is not None:
callback(model, bounds, info[:i], x, index, ftrue)
# make an observation and record it.
y = objective(x)
model.add_data(x, y)
# record everything.
info[i] = (x, y, recommender(model, bounds))
if ftrue is not None:
fbest = ftrue(info['xbest'])
info = append_fields(info, 'fbest', fbest, usemask=False)
return info
def sample(self, size=1, rng=None):
rng = rstate(rng)
return self._a + (self._b - self._a) * rng.rand(size, self.ndim)
def __init__(self, sigma=0.0, rng=None):
self._sigma = sigma
self._rng = rstate(rng)
def sample(self, size=1, rng=None):
rng = rstate(rng)
return np.vstack(self._min + rng.gamma(k, s, size=size)
for k, s in zip(self._k, self._scale)).T
def sample_spectrum(self, N, rng=None):
rng = rstate(rng)
sf2 = np.exp(self._logsf * 2)
ell = np.exp(self._logell)
W = rng.randn(N, self.ndim) / ell
return W, sf2
def sample(self, size=1, rng=None):
rng = rstate(rng)
return np.vstack(self._min + rng.lognormal(m, s, size=size)
for m, s in zip(self._mu, self._sigma)).T
def solve_bayesopt(self,
objective,
bounds,
kernel="se",
niter=150,
init='sobol',
policy='ei',
solver='direct',
recommender='latent'):
"""
The following part, borrows some code from the pybo code.
Therefore, retains a similar structure and parameters.
From Pybo:
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
Nothing, everything is written to a file
"""
filename = self.outputdir + "_".join([policy[0], kernel, str(niter)])
bounds = np.array(bounds, dtype=float, ndmin=2)
rng = rstate(None)
# get the model components.
init, policy, solver, recommender = \
self.get_components(init, policy, solver, recommender, rng)
# Create the model with the data
model = self.initialmodel(objective, init, bounds, kernel)
with open(filename, "a+") as f:
for i, y in enumerate(model.data[1]):
f.write("0," + str(y) + ",0,0," + ",".join([str(a) for a in model.data[0][i]]) + "\n")
# Update the console
self.console("Starting Bayesian Optimization", 1)
self.console("Time,Objective,Mu,Var,X", 2)
# Start BayesOpt
for i in xrange(model.ndata, niter):
# Record current time
start = time.clock()
# get the next point to evaluate.
index = policy(model)
x, _ = solver(index, bounds)
glomu, glovar = model.posterior(x, grad=False)[:2]
try:
# make an observation and record it.
y = objective(x)
model.add_data(x, y)
yp = str(y)
except:
yp = "Failed"
#Write to the file
interval = time.clock() - start
data = str(interval) + "," + yp + "," + str(glomu[0]) + "," + str(glovar[0]) + ",".join(
[str(a) for a in x])
self.console(data, 2)
with open(filename, "a+") as f:
f.write(data + "\n")
# Get the recommender point from the model
recx = recommender(model, bounds)
recmu, recvar = model.posterior(recx, grad=False)[:2]
with open(filename, "a+") as f:
f.write("0,Rec.," + str(recmu[0]) + "," + str(recvar[0]) + "," + ",".join([str(s) for s in recx]) + "\n")
def sample(self, size=1, rng=None):
rng = rstate(rng)
return self._a + (self._b - self._a) * rng.rand(size, self.ndim)
def solve_hyperopt(objective,
bounds,
niter=100,
init='middle',
policy='ei',
solver='lbfgs',
recommender='latent',
model=None,
noisefree=False,
ftrue=None,
rng=None,
callback=None):
"""
Maximize the given function using Bayesian Optimization.
Args:
objective: function handle representing the objective function.
bounds: bounds of the search space as a (d,2)-array.
niter: horizon for optimization.
init: the initialization component.
policy: the acquisition component.
solver: the inner-loop solver component.
recommender: the recommendation component.
model: the Bayesian model instantiation.
noisefree: a boolean denoting that the model is noisefree; this only
applies if a default model is used (ie. it is ignored if the
model argument is used).
ftrue: a ground-truth function (for evaluation).
rng: either an RandomState object or an integer used to seed the state;
this will be fed to each component that requests randomness.
callback: a function to call on each iteration for visualization.
Note that the modular way in which this function has been written allows
one to also pass parameters directly to some of the components. This works
for the `init`, `policy`, `solver`, and `recommender` inputs. These
components can be passed as either a string, a function, or a 2-tuple where
the first item is a string/function and the second is a dictionary of
additional arguments to pass to the component.
Returns:
A numpy record array containing a trace of the optimization process.
The fields of this array are `x`, `y`, and `xbest` corresponding to the
query locations, outputs, and recommendations at each iteration. If
ground-truth is known an additional field `fbest` will be included.
"""
# make sure the bounds are a 2d-array.
bounds = np.array(bounds, dtype=float, ndmin=2)
# see if the query object itself defines ground truth.
if (ftrue is None) and hasattr(objective, 'get_f'):
ftrue = objective.get_f
# initialize the random number generator.
rng = rstate(rng)
# get the model components.
init, policy, solver, recommender = \
get_components(init, policy, solver, recommender, rng)
# create a list of initial points to query.
X = init(bounds)
Y = [objective(x) for x in X]
# add any initial data to our model.
model.add_data(X, Y)
# allocate a datastructure containing "convergence" info.
info = np.zeros(niter, [('x', np.float, (len(bounds), )), ('y', np.float),
('xbest', np.float, (len(bounds), ))])
# initialize the data.
info['x'][:len(X)] = X
info['y'][:len(Y)] = Y
info['xbest'][:len(Y)] = [X[np.argmax(Y[:i + 1])] for i in xrange(len(Y))]
# BFGS callback
X_y = ([], [])
def callbackF(Xi):
global X_y
X_y[0].append(Xi)
X_y[1].append(objective(Xi))
for i in xrange(niter):
# take n_grad gradient steps
res = minimize(objective,
x0,
method='BFGS',
jac=grad,
options={
'gtol': 1e-6,
'disp': True,
'maxiter': n_grad
},
callback=callbackF)
# get the next point to evaluate.
index = policy(model)
x, _ = solver(index, bounds)
# deal with any visualization.
if callback is not None:
callback(model, bounds, info[:i], x, index, ftrue)
# make an observation and record it.
y = objective(x)
model.add_data(x, y)
# record everything.
info[i] = (x, y, recommender(model, bounds))
if ftrue is not None:
fbest = ftrue(info['xbest'])
info = append_fields(info, 'fbest', fbest, usemask=False)
return info
def sample(gp, priors, n, raw=True, rng=None):
rng = rstate(rng)
priors = dict(priors)
active = np.ones(gp.nhyper, dtype=bool)
logged = np.ones(gp.nhyper, dtype=bool)
for (key, block, log) in get_params(gp):
inactive = (key in priors) and (priors[key] is None)
logged[block] = log
active[block] = not inactive
if inactive:
del priors[key]
else:
priors[key] = (block, log, priors[key])
# priors is now just a list of the form (block, log, prior).
priors = priors.values()
# get the initial hyperparameters and transform into the non-log space.
hyper0 = gp.get_hyper()
hyper0[logged] = np.exp(hyper0[logged])
def logprob(x):
# copy the initial hyperparameters and then assign the "active"
# parameters that come from x.
hyper = hyper0.copy()
hyper[active] = x
logprob = 0
# compute the prior probabilities. we do this first so that if there
# are any infs they'll be caught in the least expensive computations
# first.
for block, log, prior in priors:
logprob += prior.logprior(hyper[block])
if np.isinf(logprob):
break
# now compute the likelihood term. note that we'll have to take the log
# of any logspace parameters before calling set_hyper.
if not np.isinf(logprob):
hyper[logged] = np.log(hyper[logged])
gp.set_hyper(hyper)
logprob += gp.loglikelihood()
return logprob
# create a big list of the hyperparameters so that we can just assign to
# the components that are active. also get an initial sample x
# corresponding only to the active parts of hyper0.
hypers = np.tile(hyper0, (n, 1))
x = hyper0.copy()[active]
# do the sampling.
for i in xrange(n):
x = _slice_sample(logprob, x, rng=rng)
hypers[i][active] = x
# change the logspace components back into logspace.
hypers[:, logged] = np.log(hypers[:, logged])
# make sure the gp gets updated to the last sampled hyperparameter.
gp.set_hyper(hypers[-1])
if raw:
return hypers
else:
return [gp.copy(h) for h in hypers]
