def IMRun_EGO(nbrs_list_i, incident_edge_list_i, edges_i,
J, h, S_ref, S_in, h_effective_in,
beta_true, gammas, SAW_length_min, SAW_length_max,
nmoves, nadaptations, nruns_per_param_update,
policy_size, strategy, exp_temp, MH_rate,
S_reset = None, h_effective_reset = None):
assert(not(S_reset is None and h_effective_reset is not None))
print 'strategy', strategy
if S_reset is None:
S_reset = S_in
h_effective_reset = h_effective_in
num_nodes_c, num_edges_c,\
edges_i_c, \
nbrs_list_ptrs, nbrs_list_ptrs_c, \
nbrs_list_lens, nbrs_list_lens_c, \
incident_edge_list_ptrs, incident_edge_list_ptrs_c, \
incident_edge_list_lens, incident_edge_list_lens_c \
= grph_c.make_ctypes_for_graph_repr \
(nbrs_list_i, incident_edge_list_i, edges_i)
J_c, h_c, h_effective_in_c, h_effective_reset_c = \
map(grph_c.make_ctypes_for_graph_weights,
[J, h, h_effective_in, h_effective_reset])
S_end = zeros(S_in.shape, dtype=int32)
S_ref_c, S_in_c, S_reset_c, S_end_c = \
map(grph_c.make_ctypes_for_node_activations,
[S_ref, S_in, S_reset, S_end])
beta_true_c, nmoves_c, nruns_per_param_update_c, \
gammas_c, len_gammas_c, \
SAW_length_min_c, SAW_length_max_c \
= make_ctypes_for_simulation_parameters\
(beta_true, nmoves, nruns_per_param_update, \
gammas, SAW_length_min, SAW_length_max)
nadaptations_c = c.c_int32(nadaptations)
nactions = (SAW_length_max - SAW_length_min + 1) * len(gammas)
E_samples, E_proposeds, log_f_fwds, log_f_revs, \
MH_ratios, SAW_lengths, used_gammas, rewards \
= make_output_buffers(nmoves, nactions)
E_samples_c, E_proposeds_c, log_f_fwds_c, log_f_revs_c, \
MH_ratios_c, SAW_lengths_c, used_gammas_c, rewards_c \
= make_ctypes_for_output_buffers\
(E_samples, E_proposeds, log_f_fwds, log_f_revs,
MH_ratios, SAW_lengths, used_gammas, rewards)
policy_size_c = c.c_int32(policy_size)
policy = zeros([policy_size*2, 1], dtype=double)
policy_c = policy.ctypes.data_as(c.POINTER(c.c_double))
kernel_hyperparms = \
array([SAW_length_max - SAW_length_min + 1,
gammas[-1] - gammas[0]]) * 0.1
kernel = GaussianKernel_ard(kernel_hyperparms)
gp = GaussianProcess(kernel, noise=0.1)
# Bind the GP in closures. References to these closures need to be
# kept, so that they only get garbage-collected when the C code
# has finished running.
maximize_EI_lmbd = \
lambda bounds_ptr, max_x_ptr, x_size:\
ego_c.maximizeEI_wrap(gp, bounds_ptr, max_x_ptr, x_size)
add_data_to_gp_lmbd = \
lambda x_ptr, x_size, y:\
ego_c.add_data_to_gp(gp, x_ptr, x_size, y)
maximize_gp_lmbd = \
lambda x_ptr, x_size, y_ptr:\
ego_c.maximize_gp(gp, x_ptr, x_size, y_ptr)
gp_posterior_mean_lmbd = \
lambda coords_ptr, coords_size:\
ego_c.gp_posterior_mean(gp, coords_ptr, coords_size)
maximize_EI_local_lmbd = \
lambda bounds_ptr, max_x_ptr, x_size:\
ego_c.maximizeEI_local_wrap(gp, bounds_ptr, max_x_ptr, x_size, strategy)
sample_gp_posterior_mean_lmbd = \
lambda coords_ptr, coords_size, bounds_ptr, sample_x_ptr, num_samples:\
ego_c.sample_gp_posterior_mean(gp, coords_ptr, coords_size, bounds_ptr, sample_x_ptr,
num_samples, None, exp_temp, MH_rate)
# Create the callbacks.
maximize_EI_c = ego_c.ACQ_MAXIMIZER_FUNC(maximize_EI_lmbd)
add_data_to_gp_c = ego_c.GP_ADD_DATA_FUNC(add_data_to_gp_lmbd)
maximize_gp_c = ego_c.GP_MAXIMIZER_FUNC(maximize_gp_lmbd)
gp_posterior_mean_c = ego_c.GP_POSTERIOR_MEAN_FUNC(gp_posterior_mean_lmbd)
maximize_EI_local_c = ego_c.LOCAL_ACQ_MAXIMIZER_FUNC(maximize_EI_local_lmbd)
sample_gp_posterior_mean_c = ego_c.SAMPLE_GP_POSTERIOR_MEAN_FUNC(sample_gp_posterior_mean_lmbd)
# Make the call.
libimsampler.sampleIM_ego_ext\
(num_nodes_c, num_edges_c,
nbrs_list_ptrs_c, nbrs_list_lens_c,
incident_edge_list_ptrs_c, incident_edge_list_lens_c,
edges_i_c,
J_c, h_c,
S_ref_c, S_in_c, S_reset_c, S_end_c,
h_effective_in_c, h_effective_reset_c,
beta_true_c,
len_gammas_c, gammas_c,
SAW_length_min_c, SAW_length_max_c,
nmoves_c, nadaptations_c, nruns_per_param_update_c,
maximize_EI_c, add_data_to_gp_c, maximize_gp_c, gp_posterior_mean_c,
maximize_EI_local_c, sample_gp_posterior_mean_c,
strategy, E_samples_c, E_proposeds_c, log_f_fwds_c, log_f_revs_c,
MH_ratios_c, SAW_lengths_c, used_gammas_c, rewards_c,
policy_c, policy_size_c)
gamma_min = gammas[0]
gamma_max = gammas[-1]
# Return the answers.
return S_end, E_samples, E_proposeds, log_f_fwds, log_f_revs, \
MH_ratios, SAW_lengths, used_gammas, rewards, gp, \
gamma_min, gamma_max, policy
def SARDRun_EGO(nbrs_list_i, incident_edge_list_i, edges_i,
J, h, S_in, h_effective_in,
beta_true, gamma_hi_bounds, gamma_lo_bounds,
iters_per_move_bounds,
max_SAW_length_bounds, min_SAW_length_bounds,
nmoves, nadaptations, nruns_per_param_update,
policy_size, strategy, exp_temp, MH_rate, return_samples,
S_reset = None, h_effective_reset = None):
assert(not(S_reset is None and h_effective_reset is not None))
nparams = 8
nmoves = array([nmoves], dtype='int')
nadaptations = array([nadaptations], dtype='int')
nruns_per_param_update = \
array([nruns_per_param_update], dtype='int')
policy_size = array([policy_size], dtype='int')
if S_reset is None:
S_reset = S_in
h_effective_reset = h_effective_in
num_nodes_c, num_edges_c,\
edges_i_c, \
nbrs_list_ptrs, nbrs_list_ptrs_c, \
nbrs_list_lens, nbrs_list_lens_c, \
incident_edge_list_ptrs, incident_edge_list_ptrs_c, \
incident_edge_list_lens, incident_edge_list_lens_c \
= grph_c.make_ctypes_for_graph_repr \
(nbrs_list_i, incident_edge_list_i, edges_i)
J_c, h_c, h_effective_in_c, h_effective_reset_c = \
map(grph_c.make_ctypes_for_graph_weights,
[J, h, h_effective_in, h_effective_reset])
S_end = zeros(S_in.shape, dtype=int32)
S_in_c, S_reset_c, S_end_c = \
map(grph_c.make_ctypes_for_node_activations,
[S_in, S_reset, S_end])
beta_true_c, \
gamma_hi_max_c, gamma_hi_min_c, gamma_lo_max_c, gamma_lo_min_c, \
max_SAW_length_bound_max_c, max_SAW_length_bound_min_c, \
min_SAW_length_bound_max_c, min_SAW_length_bound_min_c, \
iters_per_move_max_c, iters_per_move_min_c, \
nmoves_c, nadaptations_c, nruns_per_param_update_c \
= map(make_ctype_for_var,
[beta_true,
gamma_hi_bounds[1], gamma_hi_bounds[0],
gamma_lo_bounds[1], gamma_lo_bounds[0],
max_SAW_length_bounds[1], max_SAW_length_bounds[0],
min_SAW_length_bounds[1], min_SAW_length_bounds[0],
iters_per_move_bounds[1], iters_per_move_bounds[0],
nmoves, nadaptations, nruns_per_param_update])
iters_per_move_max = iters_per_move_bounds[1]
E_samples, E_proposeds, log_f_fwds, log_f_revs, \
MH_ratios, move_types, used_sigma_lengths, used_rho_lengths, \
used_iters_per_move, used_gammas_his, used_gammas_los, \
used_P_LL, used_P_LH, used_P_HL, rewards \
= make_output_buffers(nmoves, iters_per_move_max)
E_samples_c, E_proposeds_c, log_f_fwds_c, log_f_revs_c, \
MH_ratios_c, move_types_c, \
used_sigma_lengths_c, used_rho_lengths_c, used_iters_per_move_c, \
used_gammas_his_c, used_gammas_los_c, \
used_P_LL_c, used_P_LH_c, used_P_HL_c, rewards_c \
= map(make_ctype_for_buffer,
[E_samples, E_proposeds, log_f_fwds, log_f_revs,
MH_ratios, move_types,
used_sigma_lengths, used_rho_lengths, used_iters_per_move,
used_gammas_his, used_gammas_los,
used_P_LL, used_P_LH, used_P_HL, rewards])
return_samples_c = make_ctype_for_var(return_samples)
policy_size_c = make_ctype_for_var(policy_size)
policy = zeros([policy_size * nparams, 1], dtype=double)
policy_c = make_ctype_for_buffer(policy)
kernel_hyperparms = \
array(
[iters_per_move_bounds[1] - iters_per_move_bounds[0] + 1,
min_SAW_length_bounds[1] - min_SAW_length_bounds[0] + 1,
max_SAW_length_bounds[1] - max_SAW_length_bounds[0] + 1,
1.0, 1.0, 1.0,
gamma_hi_bounds[1] - gamma_hi_bounds[0],
gamma_lo_bounds[1] - gamma_lo_bounds[0]])
kernel = GaussianKernel_ard(kernel_hyperparms)
gp = GaussianProcess(kernel, noise=0.1)
# NOTE: This method is no longer useful. Take it out!
def check_constraint(nx_pt, bounds, coords_size):
if nx_pt[1] <= nx_pt[2]:
return False
else:
return False
# Samples to be returned
samples = []
# Bind the GP in closures. References to these closures need to be
# kept, so that they only get garbage-collected when the C code
# has finished running.
maximize_EI_lmbd = \
lambda bounds_ptr, max_x_ptr, x_size:\
ego_c.maximizeEI_wrap(gp, bounds_ptr, max_x_ptr, x_size)
add_data_to_gp_lmbd = \
lambda x_ptr, x_size, y:\
ego_c.add_data_to_gp(gp, x_ptr, x_size, y)
maximize_gp_lmbd = \
lambda x_ptr, x_size, y_ptr:\
ego_c.maximize_gp(gp, x_ptr, x_size, y_ptr)
gp_posterior_mean_lmbd = \
lambda coords_ptr, coords_size:\
ego_c.gp_posterior_mean(gp, coords_ptr, coords_size)
maximize_EI_local_lmbd = \
lambda bounds_ptr, max_x_ptr, x_size:\
ego_c.maximizeEI_local_wrap(gp, bounds_ptr, max_x_ptr, x_size, strategy)
sample_gp_posterior_mean_lmbd = \
lambda coords_ptr, coords_size, bounds_ptr, sample_x_ptr, num_samples:\
ego_c.sample_gp_posterior_mean(gp, coords_ptr, coords_size, bounds_ptr, sample_x_ptr,
num_samples, check_constraint, exp_temp, MH_rate)
return_sample_lmbd = \
lambda sample_x_ptr, coords_size:\
ego_c.return_sample(samples, sample_x_ptr, coords_size)
# Create the callbacks.
maximize_EI_c = ego_c.ACQ_MAXIMIZER_FUNC(maximize_EI_lmbd)
add_data_to_gp_c = ego_c.GP_ADD_DATA_FUNC(add_data_to_gp_lmbd)
maximize_gp_c = ego_c.GP_MAXIMIZER_FUNC(maximize_gp_lmbd)
gp_posterior_mean_c = ego_c.GP_POSTERIOR_MEAN_FUNC(gp_posterior_mean_lmbd)
maximize_EI_local_c = ego_c.LOCAL_ACQ_MAXIMIZER_FUNC(maximize_EI_local_lmbd)
sample_gp_posterior_mean_c = ego_c.SAMPLE_GP_POSTERIOR_MEAN_FUNC(sample_gp_posterior_mean_lmbd)
return_sample_c = ego_c.RETURN_SAMPLE(return_sample_lmbd)
# Make the call.
libsardsampler.sampleSARD_ego_ext\
(num_nodes_c, num_edges_c,
nbrs_list_ptrs_c, nbrs_list_lens_c,
incident_edge_list_ptrs_c, incident_edge_list_lens_c,
edges_i_c,
J_c, h_c, S_in_c, S_reset_c, S_end_c,
h_effective_in_c, h_effective_reset_c,
beta_true_c,
gamma_hi_max_c, gamma_hi_min_c,
gamma_lo_max_c, gamma_lo_min_c,
max_SAW_length_bound_max_c, max_SAW_length_bound_min_c,
min_SAW_length_bound_max_c, min_SAW_length_bound_min_c,
iters_per_move_min_c, iters_per_move_max_c,
nmoves_c, nadaptations_c, nruns_per_param_update_c,
maximize_EI_c, add_data_to_gp_c,
maximize_gp_c, gp_posterior_mean_c, maximize_EI_local_c,
sample_gp_posterior_mean_c, return_sample_c, return_samples_c,
strategy, E_samples_c, E_proposeds_c, log_f_fwds_c, log_f_revs_c,
MH_ratios_c, move_types_c,
used_sigma_lengths_c, used_rho_lengths_c,
used_iters_per_move_c,
used_gammas_his_c, used_gammas_los_c,
used_P_LL_c, used_P_LH_c, used_P_HL_c,
rewards_c, policy_c, policy_size_c)
print len(samples)
return S_end, E_samples, E_proposeds, log_f_fwds, log_f_revs, \
MH_ratios, move_types, \
used_sigma_lengths, used_rho_lengths, used_iters_per_move, \
used_gammas_his, used_gammas_los, \
used_P_LL, used_P_LH, used_P_HL, \
rewards, gp, policy, samples
