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 __init__(self, bandit, good_estimator, bad_estimator):
TheanoBanditAlgo.__init__(self, bandit)
self.good_estimator = good_estimator
self.bad_estimator = bad_estimator
self.build_helpers()
def __init__(self, bandit, good_estimator, bad_estimator):
TheanoBanditAlgo.__init__(self, bandit)
self.good_estimator = good_estimator
self.bad_estimator = bad_estimator
self.build_helpers()