def recall(self, idlist):
"""Construct an IdxsValsList representation of the elements of idlist.
The result will not be renumbered.
"""
if idlist:
idset = set(idlist)
if len(idset) != len(idlist):
raise NotImplementedError('dups in idlist')
# for each variable in the bandit (each idxs, vals pair)
# extract the database elements and put them into a new (idxs, vals)
# pair that we can return.
rval_idxs = []
rval_vals = []
for idxs, vals in zip(self.db_idxs, self.db_vals):
assert len(idxs) == len(vals)
ii_vv = [(ii, vv)
for (ii, vv) in zip(idxs, vals) if ii in idset]
rval_idxs.append([iv[0] for iv in ii_vv])
rval_vals.append([iv[1] for iv in ii_vv])
else:
rval_idxs = [[] for s in self.s_idxs]
rval_vals = [[] for s in self.s_idxs]
return IdxsValsList.fromlists(rval_idxs, rval_vals)
def suggest_from_prior(self, trials, results, N):
logger.info('suggest_from_prior')
if not hasattr(self, '_prior_sampler'):
self._prior_sampler = theano.function([self.s_N],
self.s_prior.flatten(),
mode=self.mode)
rvals = self._prior_sampler(N)
return IdxsValsList.fromflattened(rvals)
def suggest_from_prior(self, trials, results, N):
logger.info('suggest_from_prior')
if not hasattr(self, '_prior_sampler'):
self._prior_sampler = theano.function(
[self.s_N],
self.s_prior.flatten(),
mode=self.mode)
rvals = self._prior_sampler(N)
return IdxsValsList.fromflattened(rvals)
def suggest_from_prior(self, N):
try:
prior_sampler = self._prior_sampler
except AttributeError:
prior_sampler = self._prior_sampler = theano.function(
[self.helper_locals['n_to_draw']],
self.helper_locals['s_prior'].flatten(),
mode=self.mode)
rvals = prior_sampler(N)
return IdxsValsList.fromflattened(rvals)
def suggest_from_prior(self, N):
try:
prior_sampler = self._prior_sampler
except AttributeError:
prior_sampler = self._prior_sampler = theano.function(
[self.helper_locals['n_to_draw']],
self.helper_locals['s_prior'].flatten(),
mode=self.mode)
rvals = prior_sampler(N)
return IdxsValsList.fromflattened(rvals)
def suggest(self, trials, results, N):
# normally a TheanoBanditAlto.suggest would start with this call:
## ivls = self.idxs_vals_by_status(trials, results)
rvals = self._sampler(N)
# A TheanoBanditAlgo.suggest implementation should usually
# return suggest_ivl(...).
return self.suggest_ivl(
IdxsValsList.fromlists(
rvals[:len(rvals)/2],
rvals[len(rvals)/2:]))
def posterior(self, priors, observations, s_rng):
"""
priors - an IdxsValsList of random variables
observations - an IdxsValsList of corresponding samples
returns - an IdxsValsList of posterior random variables
"""
assert len(priors) == len(observations)
# observation.idxs could be invalid.
# They could be invalid, in the sense of describing samples that
# could not have been drawn from the prior.
# This code does not try to detect that situation.
post_vals = [
self.s_posterior_helper(p, o, s_rng)
for p, o in zip(priors, observations)
]
# At this point, each post_vals[i] is connected to the original graph
# formed of prior nodes.
#
# We want each post_vals[i] to be connected instead to the other
# post_vals just created, corresponding to the posterior values of other
# random vars.
#
# The way to get all the new_s_val nodes hooked up to one another is to
# use the clone_keep_replacements function.
# XXX: this clones everything. It should be possible to do a more
# selective clone of just the pieces that change.
inputs = theano.gof.graph.inputs(priors.flatten() + post_vals)
env = ienv.std_interactive_env(inputs,
priors.flatten() + post_vals,
clone_inputs_and_orphans=False)
env.prefer_replace(zip(priors.valslist(), post_vals),
reason='IndependentNodeTreeEstimator.posterior')
# raise an exception if we created cycles
env.toposort()
# extract the cloned results from the env
rval = IdxsValsList.fromlists(
[env.newest(v) for v in priors.idxslist()],
[env.newest(v) for v in post_vals])
# remove all references in the variables to the env. Prepare them
# to be inserted into another env if necessary.
env.disown()
return rval
def posterior(self, priors, observations, s_rng):
"""
priors - an IdxsValsList of random variables
observations - an IdxsValsList of corresponding samples
returns - an IdxsValsList of posterior random variables
"""
assert len(priors) == len(observations)
# observation.idxs could be invalid.
# They could be invalid, in the sense of describing samples that
# could not have been drawn from the prior.
# This code does not try to detect that situation.
post_vals = [self.s_posterior_helper(p, o, s_rng)
for p, o in zip(priors, observations)]
# At this point, each post_vals[i] is connected to the original graph
# formed of prior nodes.
#
# We want each post_vals[i] to be connected instead to the other
# post_vals just created, corresponding to the posterior values of other
# random vars.
#
# The way to get all the new_s_val nodes hooked up to one another is to
# use the clone_keep_replacements function.
# XXX: this clones everything. It should be possible to do a more
# selective clone of just the pieces that change.
inputs = theano.gof.graph.inputs(priors.flatten() + post_vals)
env = ienv.std_interactive_env(inputs, priors.flatten() + post_vals,
clone_inputs_and_orphans=False)
env.prefer_replace(
zip(priors.valslist(), post_vals),
reason='IndependentNodeTreeEstimator.posterior')
# raise an exception if we created cycles
env.toposort()
# extract the cloned results from the env
rval = IdxsValsList.fromlists(
[env.newest(v) for v in priors.idxslist()],
[env.newest(v) for v in post_vals])
# remove all references in the variables to the env. Prepare them
# to be inserted into another env if necessary.
env.disown()
return rval
def __init__(self, bandit):
TheanoBanditAlgo.__init__(self, bandit)
self.numpy_rng = numpy.random.RandomState(234)
self.s_prior = IdxsValsList.fromlists(self.s_idxs, self.s_vals)
self.s_n_train = tensor.lscalar('n_train')
self.s_n_test = tensor.lscalar('n_test')
self.y_obs = tensor.vector('y_obs')
self.y_obs_var = tensor.vector('y_obs_var')
self.x_obs_IVL = self.s_prior.new_like_self()
self.cand_x = self.s_prior.new_like_self()
self.cand_EI_thresh = tensor.scalar()
self.init_kernels()
self.init_gram_weights()
self.params.extend(self.convex_coefficient_params)
self.param_bounds.extend(self.convex_coefficient_params_bounds)
self.s_big_param_vec = tensor.vector()
### assumes all variables are refinable
### assumes all variables are vectors
n_elements_used = 0
for k, iv in zip(self.kernels, self.cand_x):
if self.is_refinable[k]:
n_elements_in_v = iv.idxs.shape[0]
start = n_elements_used
stop = n_elements_used + n_elements_in_v
iv.vals = self.s_big_param_vec[start:stop]
n_elements_used += n_elements_in_v
self.gprmath = GPR_math(self.x_obs_IVL,
self.y_obs,
self.y_obs_var,
picklable_instancemethod(self, 'K_fn'),
N=self.s_n_train,
min_variance=self.y_minvar)
self.nll_obs = self.gprmath.s_nll()
self.cand_EI = tensor.log(self.gprmath.s_expectation_lt_thresh(
self.cand_x,
self.cand_EI_thresh))
# self.gm_algo is used to draw candidates for subsequent refinement
# It is also entirely responsible for choosing categorical variables.
self.gm_algo = AdaptiveParzenGM(self.bandit)
self.gm_algo.n_EI_candidates = self.n_candidates_to_draw_in_GM
def __init__(self, bandit):
TheanoBanditAlgo.__init__(self, bandit)
self.numpy_rng = numpy.random.RandomState(234)
self.s_prior = IdxsValsList.fromlists(self.s_idxs, self.s_vals)
self.s_n_train = tensor.lscalar('n_train')
self.s_n_test = tensor.lscalar('n_test')
self.y_obs = tensor.vector('y_obs')
self.y_obs_var = tensor.vector('y_obs_var')
self.x_obs_IVL = self.s_prior.new_like_self()
self.cand_x = self.s_prior.new_like_self()
self.cand_EI_thresh = tensor.scalar()
self.init_kernels()
self.init_gram_weights()
self.params.extend(self.convex_coefficient_params)
self.param_bounds.extend(self.convex_coefficient_params_bounds)
self.s_big_param_vec = tensor.vector()
### assumes all variables are refinable
### assumes all variables are vectors
n_elements_used = 0
for k, iv in zip(self.kernels, self.cand_x):
if self.is_refinable[k]:
n_elements_in_v = iv.idxs.shape[0]
start = n_elements_used
stop = n_elements_used + n_elements_in_v
iv.vals = self.s_big_param_vec[start:stop]
n_elements_used += n_elements_in_v
self.gprmath = GPR_math(self.x_obs_IVL,
self.y_obs,
self.y_obs_var,
picklable_instancemethod(self, 'K_fn'),
N=self.s_n_train,
min_variance=self.y_minvar)
self.nll_obs = self.gprmath.s_nll()
self.cand_EI = tensor.log(
self.gprmath.s_expectation_lt_thresh(self.cand_x,
self.cand_EI_thresh))
# self.gm_algo is used to draw candidates for subsequent refinement
# It is also entirely responsible for choosing categorical variables.
self.gm_algo = AdaptiveParzenGM(self.bandit)
self.gm_algo.n_EI_candidates = self.n_candidates_to_draw_in_GM
def setUp(self):
self.TE = IndependentNullEstimator()
self.bandit = NestedUniform()
self.experiment = SerialExperiment(
GM_BanditAlgo(self.bandit,
good_estimator=IndependentNullEstimator(),
bad_estimator=IndependentNullEstimator()))
self.s_rng = montetheano.RandomStreams(123)
prior_idxs, prior_vals, s_N = self.bandit.template.theano_sampler(self.s_rng)
#print prior_idxs
#print prior_vals
self.prior = IdxsValsList.fromlists(
[i for i in prior_idxs if i is not None],
[v for v in prior_vals if v is not None])
self.s_N = s_N
self.observations = self.prior.new_like_self()
for i, o in enumerate(self.observations):
o.idxs.name = 'Obs_idxs{%i}' % i
o.vals.name = 'Obs_vals{%i}' % i
def setUp(self):
self.TE = IndependentNullEstimator()
self.bandit = NestedUniform()
self.experiment = SerialExperiment(
GM_BanditAlgo(self.bandit,
good_estimator=IndependentNullEstimator(),
bad_estimator=IndependentNullEstimator()))
self.s_rng = montetheano.RandomStreams(123)
prior_idxs, prior_vals, s_N = self.bandit.template.theano_sampler(
self.s_rng)
#print prior_idxs
#print prior_vals
self.prior = IdxsValsList.fromlists(
[i for i in prior_idxs if i is not None],
[v for v in prior_vals if v is not None])
self.s_N = s_N
self.observations = self.prior.new_like_self()
for i, o in enumerate(self.observations):
o.idxs.name = 'Obs_idxs{%i}' % i
o.vals.name = 'Obs_vals{%i}' % i
def build_helpers(self):
s_prior = IdxsValsList.fromlists(self.s_idxs, self.s_vals)
s_obs = s_prior.new_like_self()
# y_thresh is the boundary between 'good' and 'bad' regions of the
# search space.
y_thresh = tensor.scalar()
yvals = tensor.vector()
n_to_draw = self.s_N
n_to_keep = tensor.iscalar()
s_rng = montetheano.RandomStreams(self.seed + 9)
GE = self.good_estimator
BE = self.bad_estimator
Gobs = s_obs.symbolic_take(where(yvals < y_thresh))
Bobs = s_obs.symbolic_take(where(yvals >= y_thresh))
# To "optimize" EI we just draw a pile of samples from the density
# of good points and then just take the best of those.
Gsamples = GE.posterior(s_prior, Gobs, s_rng)
Bsamples = BE.posterior(s_prior, Bobs, s_rng)
G_ll = GE.log_likelihood(Gsamples,
Gsamples,
llik=tensor.zeros((n_to_draw, )))
B_ll = BE.log_likelihood(Bsamples,
Gsamples,
llik=tensor.zeros((n_to_draw, )))
# subtract B_ll from G_ll
log_EI = G_ll - B_ll
keep_idxs = argsort(log_EI)[-n_to_keep:]
# store all these vars for the unittests
self.helper_locals = locals()
del self.helper_locals['self']
def build_helpers(self):
s_prior = IdxsValsList.fromlists(self.s_idxs, self.s_vals)
s_obs = s_prior.new_like_self()
# y_thresh is the boundary between 'good' and 'bad' regions of the
# search space.
y_thresh = tensor.scalar()
yvals = tensor.vector()
n_to_draw = self.s_N
n_to_keep = tensor.iscalar()
s_rng = montetheano.RandomStreams(self.seed + 9)
GE = self.good_estimator
BE = self.bad_estimator
Gobs = s_obs.symbolic_take(where(yvals < y_thresh))
Bobs = s_obs.symbolic_take(where(yvals >= y_thresh))
# To "optimize" EI we just draw a pile of samples from the density
# of good points and then just take the best of those.
Gsamples = GE.posterior(s_prior, Gobs, s_rng)
Bsamples = BE.posterior(s_prior, Bobs, s_rng)
G_ll = GE.log_likelihood(Gsamples, Gsamples,
llik = tensor.zeros((n_to_draw,)))
B_ll = BE.log_likelihood(Bsamples, Gsamples,
llik = tensor.zeros((n_to_draw,)))
# subtract B_ll from G_ll
log_EI = G_ll - B_ll
keep_idxs = argsort(log_EI)[-n_to_keep:]
# store all these vars for the unittests
self.helper_locals = locals()
del self.helper_locals['self']
def suggest_from_model(self, ivls, N):
helper = self._suggest_from_model_fn
ylist = numpy.asarray(sorted(ivls['losses']['ok'].vals), dtype='float')
y_thresh_idx = int(self.gamma * len(ylist))
y_thresh = ylist[y_thresh_idx : y_thresh_idx + 2].mean()
logger.info('GM_BanditAlgo splitting results at y_thresh = %f'
% y_thresh)
logger.info('GM_BanditAlgo keeping %i results as good'
% y_thresh_idx)
logger.info('GM_BanditAlgo keeping %i results as bad'
% (len(ylist) - y_thresh_idx))
logger.info('GM_BanditAlgo good scores: %s'
% str(ylist[:y_thresh_idx]))
x_all = ivls['x_IVLs']['ok'].as_list()
y_all_iv = ivls['losses']['ok'].as_list()
assert y_all_iv.idxset() == x_all.idxset(), (y_all_iv.idxset(),
x_all.idxset())
for pseudo_bad_status in 'new', 'running':
logger.info('GM_BanditAlgo assigning bad scores to %i new jobs'
% len(ivls['losses'][pseudo_bad_status].idxs))
x_all.stack(ivls['x_IVLs'][pseudo_bad_status])
y_all_iv.stack(IdxsVals(
ivls['losses'][pseudo_bad_status].idxs,
[y_thresh + 1] * len(ivls['losses'][pseudo_bad_status].idxs)))
assert y_all_iv.idxset() == x_all.idxset(), (y_all_iv.idxset(),
x_all.idxset())
# renumber the configurations in x_all to be 0 .. (n_train - 1)
idmap = y_all_iv.reindex()
idmap = x_all.reindex(idmap)
assert y_all_iv.idxset() == x_all.idxset(), (y_all_iv.idxset(),
x_all.idxset())
assert numpy.all(y_all_iv.idxs == numpy.arange(len(y_all_iv.idxs))), (
y_all_iv.idxs)
y_all = y_all_iv.as_numpy(vdtype=theano.config.floatX).vals
x_all = x_all.as_numpy_floatX()
logger.info('GM_BanditAlgo drawing %i candidates'
% self.n_EI_candidates)
helper_rval = helper(self.n_EI_candidates, N,
y_thresh, y_all, *x_all.flatten())
assert len(helper_rval) == 6 * len(x_all)
keep_flat = helper_rval[:2 * len(x_all)]
Gobs_flat = helper_rval[2 * len(x_all): 4 * len(x_all)]
Bobs_flat = helper_rval[4 * len(x_all):]
assert len(keep_flat) == len(Gobs_flat) == len(Bobs_flat)
Gobs = IdxsValsList.fromflattened(Gobs_flat)
Bobs = IdxsValsList.fromflattened(Bobs_flat)
# guard against book-keeping error
# ensure that all observations were counted as either good or bad
gis = Gobs.idxset()
bis = Bobs.idxset()
xis = x_all.idxset()
assert len(xis) == len(y_all)
assert gis.union(bis) == xis
assert gis.intersection(bis) == set()
rval = IdxsValsList.fromflattened(keep_flat)
# relabel the return values to be elements 0 ... N - 1
rval.reindex()
return rval
def suggest_from_model(self, ivls, N):
helper = self._suggest_from_model_fn
ylist = numpy.asarray(sorted(ivls['losses']['ok'].vals), dtype='float')
y_thresh_idx = int(self.gamma * len(ylist))
y_thresh = ylist[y_thresh_idx:y_thresh_idx + 2].mean()
logger.info('GM_BanditAlgo splitting results at y_thresh = %f' %
y_thresh)
logger.info('GM_BanditAlgo keeping %i results as good' % y_thresh_idx)
logger.info('GM_BanditAlgo keeping %i results as bad' %
(len(ylist) - y_thresh_idx))
logger.info('GM_BanditAlgo good scores: %s' %
str(ylist[:y_thresh_idx]))
x_all = ivls['x_IVLs']['ok'].as_list()
y_all_iv = ivls['losses']['ok'].as_list()
assert y_all_iv.idxset() == x_all.idxset(), (y_all_iv.idxset(),
x_all.idxset())
for pseudo_bad_status in 'new', 'running':
logger.info('GM_BanditAlgo assigning bad scores to %i new jobs' %
len(ivls['losses'][pseudo_bad_status].idxs))
x_all.stack(ivls['x_IVLs'][pseudo_bad_status])
y_all_iv.stack(
IdxsVals(ivls['losses'][pseudo_bad_status].idxs,
[y_thresh + 1] *
len(ivls['losses'][pseudo_bad_status].idxs)))
assert y_all_iv.idxset() == x_all.idxset(), (y_all_iv.idxset(),
x_all.idxset())
# renumber the configurations in x_all to be 0 .. (n_train - 1)
idmap = y_all_iv.reindex()
idmap = x_all.reindex(idmap)
assert y_all_iv.idxset() == x_all.idxset(), (y_all_iv.idxset(),
x_all.idxset())
assert numpy.all(
y_all_iv.idxs == numpy.arange(len(y_all_iv.idxs))), (y_all_iv.idxs)
y_all = y_all_iv.as_numpy(vdtype=theano.config.floatX).vals
x_all = x_all.as_numpy_floatX()
logger.info('GM_BanditAlgo drawing %i candidates' %
self.n_EI_candidates)
helper_rval = helper(self.n_EI_candidates, N, y_thresh, y_all,
*x_all.flatten())
assert len(helper_rval) == 6 * len(x_all)
keep_flat = helper_rval[:2 * len(x_all)]
Gobs_flat = helper_rval[2 * len(x_all):4 * len(x_all)]
Bobs_flat = helper_rval[4 * len(x_all):]
assert len(keep_flat) == len(Gobs_flat) == len(Bobs_flat)
Gobs = IdxsValsList.fromflattened(Gobs_flat)
Bobs = IdxsValsList.fromflattened(Bobs_flat)
# guard against book-keeping error
# ensure that all observations were counted as either good or bad
gis = Gobs.idxset()
bis = Bobs.idxset()
xis = x_all.idxset()
assert len(xis) == len(y_all)
assert gis.union(bis) == xis
assert gis.intersection(bis) == set()
rval = IdxsValsList.fromflattened(keep_flat)
# relabel the return values to be elements 0 ... N - 1
rval.reindex()
return rval