-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathrun_exp_bq.py
290 lines (244 loc) · 12.2 KB
/
run_exp_bq.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
import os
import sys
import pickle
import torch
import numpy as np
from pprint import pprint
from tqdm import tqdm
from utils.logger import setup_logging
from utils.arg_helper import parse_arguments, get_bo_config
from ahgp.gp.gp_helper import standardize
from ahgp.nn import *
from utils.bo_functions import *
from utils.bo_bq_model import Emukit_BO_BQ_GP_Model, BaseGaussianProcessCustomModel, QuadratureKernelCustom
from emukit.quadrature.methods import VanillaBayesianQuadrature
from emukit.test_functions.quadrature.baselines import univariate_approximate_ground_truth_integral, bivariate_approximate_ground_truth_integral
from emukit.test_functions.quadrature import hennig1D, sombrero2D, hennig2D, circular_gaussian
from emukit.core.optimization import LocalSearchAcquisitionOptimizer
from emukit.core.parameter_space import ParameterSpace
from emukit.quadrature.acquisitions import IntegralVarianceReduction
import matplotlib.pyplot as plt
import time
### Figure config
LEGEND_SIZE = 15
FIGURE_SIZE = (12, 8)
torch.set_printoptions(profile='full')
torch.set_printoptions(precision=4,linewidth=200)
np.set_printoptions(precision=4,linewidth=150)
def bq_loop(config, ai_model=None):
user_function, integral_bounds = eval(config.name)()
lb = integral_bounds[0][0] # lower bound
ub = integral_bounds[0][1] # upper bound
data_dim = config.data_dim
init_num_data = config.init_num_data
interval_std = config.interval_std
interval = np.zeros((1,data_dim))
std = np.zeros((1,data_dim))
mean = np.zeros((1,data_dim))
integral_bounds_scaled = integral_bounds.copy()
for ii in range(data_dim):
interval[0,ii] = integral_bounds[ii][1] - integral_bounds[ii][0]
std[0,ii] = interval[0,ii]/interval_std
mean[0,ii] = (integral_bounds[ii][1] + integral_bounds[ii][0])/2
integral_bounds_scaled[ii] = ((integral_bounds[ii] - mean[0,ii])/std[0,ii]).tolist()
lb_scaled = integral_bounds_scaled[0][0] # lower bound
ub_scaled = integral_bounds_scaled[0][1] # upper bound
lb = integral_bounds[0][0] # lower bound
ub = integral_bounds[0][1] # upper bound
results_list = [None] * config.repeated_runs
npr = np.random.RandomState(config.seed)
for kk in tqdm(range(config.repeated_runs)):
integral_mean_list = np.zeros(config.bq_iter+1)
integral_std_list = np.zeros(config.bq_iter+1)
#initialize data points
X_init = (npr.rand(init_num_data,data_dim)-0.5)*interval + mean
Y_init = user_function.f(X_init)
X_init_norm = (X_init - mean)/std
Y_init_norm, mean_Y, std_Y = standardize(Y_init)
X = X_init_norm
Y = Y_init_norm
X[np.abs(X)<1.0e-5] = 1.0e-5
#normalized function
function_norm= lambda x: (user_function.f(x*std+mean)- mean_Y)/std_Y
if data_dim == 1:
ground_truth = univariate_approximate_ground_truth_integral(function_norm, (lb_scaled, ub_scaled))[0]
elif data_dim ==2:
ground_truth = bivariate_approximate_ground_truth_integral(function_norm, integral_bounds_scaled)[0]
#Set up Emukit_BO_BQ_GP_Model
emukit_gp_model = Emukit_BO_BQ_GP_Model(X_init_norm, Y_init_norm, config, ai_model)
emukit_gp_model.optimize()
emukit_gp_model.set_kernel()
emukit_quad_kern = QuadratureKernelCustom(emukit_gp_model, integral_bounds_scaled)
emukit_model = BaseGaussianProcessCustomModel(kern=emukit_quad_kern, gp_model=emukit_gp_model)
emukit_method = VanillaBayesianQuadrature(base_gp=emukit_model, X=X, Y=Y)
#set up Bayesian quadrature
if config.plot:
x_plot = np.linspace(integral_bounds_scaled[0][0], integral_bounds_scaled[0][1], 300)[:, None]
y_plot = function_norm(x_plot)
mu_plot, var_plot = emukit_method.predict(x_plot)
plt.figure(figsize=FIGURE_SIZE)
plt.plot(X_init_norm, Y_init_norm, "ro", markersize=10, label="Observations")
plt.plot(x_plot, y_plot, "k", label="The Integrand")
plt.plot(x_plot, mu_plot, "C0", label="Model")
plt.fill_between(x_plot[:, 0],
mu_plot[:, 0] + np.sqrt(var_plot)[:, 0],
mu_plot[:, 0] - np.sqrt(var_plot)[:, 0], color="C0", alpha=0.6)
plt.fill_between(x_plot[:, 0],
mu_plot[:, 0] + 2 * np.sqrt(var_plot)[:, 0],
mu_plot[:, 0] - 2 * np.sqrt(var_plot)[:, 0], color="C0", alpha=0.4)
plt.fill_between(x_plot[:, 0],
mu_plot[:, 0] + 3 * np.sqrt(var_plot)[:, 0],
mu_plot[:, 0] - 3 * np.sqrt(var_plot)[:, 0], color="C0", alpha=0.2)
plt.legend(loc=2, prop={'size': LEGEND_SIZE})
plt.xlabel(r"$x$")
plt.ylabel(r"$f(x)$")
plt.grid(True)
plt.xlim(lb_scaled, ub_scaled)
plt.show()
initial_integral_mean, initial_integral_variance = emukit_method.integrate()
integral_mean_list[0] = initial_integral_mean
integral_std_list[0] = np.sqrt(initial_integral_variance)
if config.plot:
x_plot_integral = np.linspace(initial_integral_mean-3*np.sqrt(initial_integral_variance),
initial_integral_mean+3*np.sqrt(initial_integral_variance), 200)
y_plot_integral_initial = 1/np.sqrt(initial_integral_variance * 2 * np.pi) * \
np.exp( - (x_plot_integral - initial_integral_mean)**2 / (2 * initial_integral_variance) )
plt.figure(figsize=FIGURE_SIZE)
plt.plot(x_plot_integral, y_plot_integral_initial, "k", label="initial integral density")
plt.axvline(initial_integral_mean, color="red", label="initial integral estimate", \
linestyle="--")
plt.axvline(ground_truth, color="blue", label="ground truth integral", \
linestyle="--")
plt.legend(loc=2, prop={'size': LEGEND_SIZE})
plt.xlabel(r"$F$")
plt.ylabel(r"$p(F)$")
plt.grid(True)
plt.xlim(np.min(x_plot_integral), np.max(x_plot_integral))
plt.show()
print('The initial estimated integral is: ', round(initial_integral_mean, 4))
print('with a credible interval: ', round(2*np.sqrt(initial_integral_variance), 4), '.')
print('The ground truth rounded to 2 digits for comparison is: ', round(ground_truth, 4), '.')
for ii in range(config.bq_iter):
time_count = 0
result = {}
ivr_acquisition = IntegralVarianceReduction(emukit_method)
space = ParameterSpace(emukit_method.reasonable_box_bounds.convert_to_list_of_continuous_parameters())
num_steps = 200
num_init_points = 5
optimizer = LocalSearchAcquisitionOptimizer(space,num_steps,num_init_points)
x_new,_ = optimizer.optimize(ivr_acquisition)
y_new = function_norm(x_new)
X = np.append(X, x_new, axis=0)
Y = np.append(Y, y_new, axis=0)
X[np.abs(X)<1.0e-5] = 1.0e-5
emukit_method.set_data(X, Y)
start_time = time.time()
emukit_model.optimize()
time_count = time_count + time.time() - start_time
integral_mean, integral_variance = emukit_method.integrate()
integral_mean_list[ii+1] = integral_mean
integral_std_list[ii+1] = np.sqrt(integral_variance)
if config.plot:
mu_plot_final, var_plot_final = emukit_method.predict(x_plot)
y_plot_integral = 1/np.sqrt(integral_variance * 2 * np.pi) * \
np.exp( - (x_plot_integral - integral_mean)**2 / (2 * integral_variance) )
plt.figure(figsize=FIGURE_SIZE)
plt.plot(x_plot_integral, y_plot_integral_initial, "gray", label="initial integral density")
plt.plot(x_plot_integral, y_plot_integral, "k", label="new integral density")
plt.axvline(initial_integral_mean, color="gray", label="initial integral estimate", linestyle="--")
plt.axvline(integral_mean, color="red", label="new integral estimate", linestyle="--")
plt.axvline(ground_truth, color="blue", label="ground truth integral", \
linestyle="--")
plt.legend(loc=2, prop={'size': LEGEND_SIZE})
plt.xlabel(r"$F$")
plt.ylabel(r"$p(F)$")
plt.grid(True)
plt.xlim(np.min(x_plot_integral), np.max(x_plot_integral))
plt.show()
plt.figure(figsize=FIGURE_SIZE)
plt.plot(emukit_model.X, emukit_model.Y, "ro", markersize=10, label="Observations")
plt.plot(x_plot, y_plot, "k", label="The Integrand")
plt.plot(x_plot, mu_plot_final, "C0", label="Model")
plt.fill_between(x_plot[:, 0],
mu_plot_final[:, 0] + np.sqrt(var_plot_final)[:, 0],
mu_plot_final[:, 0] - np.sqrt(var_plot_final)[:, 0], color="C0", alpha=0.6)
plt.fill_between(x_plot[:, 0],
mu_plot_final[:, 0] + 2 * np.sqrt(var_plot_final)[:, 0],
mu_plot_final[:, 0] - 2 * np.sqrt(var_plot_final)[:, 0], color="C0", alpha=0.4)
plt.fill_between(x_plot[:, 0],
mu_plot_final[:, 0] + 3 * np.sqrt(var_plot_final)[:, 0],
mu_plot_final[:, 0] - 3 * np.sqrt(var_plot_final)[:, 0], color="C0", alpha=0.2)
plt.legend(loc=2, prop={'size': LEGEND_SIZE})
plt.xlabel(r"$x$")
plt.ylabel(r"$f(x)$")
plt.grid(True)
plt.xlim(lb_scaled, ub_scaled)
plt.show()
print('The estimated integral is: ', round(integral_mean, 4))
print('with a credible interval: ', round(2*np.sqrt(integral_variance), 4), '.')
print('The ground truth rounded to 2 digits for comparison is: ', round(ground_truth, 4), '.')
integral_error_list = np.abs(integral_mean_list - ground_truth)
result['integral_error_list'] = integral_error_list
integral_error_list_scaledback = integral_error_list * std_Y.item()
for jj in range(data_dim):
integral_error_list_scaledback = integral_error_list_scaledback * std[0,jj]
result['integral_error_list_scaledback'] = integral_error_list_scaledback
result['integral_std_list'] = integral_std_list
result['time_elapsed'] = time_count
results_list[kk] = result
print(time_count)
if config.plot:
plt.figure(figsize=(12, 8))
plt.fill_between(np.arange(config.bq_iter+1)+1, integral_error_list-0.2*integral_std_list, integral_error_list+0.2*integral_std_list, color='red', alpha=0.15)
plt.plot(np.arange(config.bq_iter+1)+1, integral_error_list, 'or-', lw=2, label='Estimated integral')
plt.legend(loc=2, prop={'size': LEGEND_SIZE})
plt.xlabel(r"iteration")
plt.ylabel(r"$f(x)$")
plt.grid(True)
plt.show()
#end of one run
return results_list
def main():
args = parse_arguments()
config = get_bo_config(args.config_file)
torch.manual_seed(config.seed)
torch.cuda.manual_seed_all(config.seed)
config.use_gpu = config.use_gpu and torch.cuda.is_available()
device = torch.device('cuda' if config.use_gpu else 'cpu')
# log info
log_file = os.path.join(config.save_dir,
"log_exp_{}.txt".format(config.run_id))
logger = setup_logging(args.log_level, log_file)
logger.info("Writing log file to {}".format(log_file))
logger.info("Exp instance id = {}".format(config.run_id))
logger.info("Exp comment = {}".format(args.comment))
logger.info("Config =")
print(">" * 80)
pprint(config)
print("<" * 80)
#load model
model = eval(config.model.name)(config.model)
model_snapshot = torch.load(config.model.pretrained_model, map_location=device)
model.load_state_dict(model_snapshot["model"], strict=True)
model.to(device)
if config.use_gpu:
model = nn.DataParallel(model, device_ids=config.gpus).cuda()
# Run the experiment
results_list = bq_loop(config.bq, model)
if config.bq.is_GPY:
if config.bq.is_sparseGP:
pickle.dump(results_list,
open(os.path.join(config.bq.save_dir, config.bq.name + '_sparseGP_init_p' + str(config.bq.init_num_data) + '_inducing_p'+ str(config.bq.num_inducing_pts) + '_results.p'), 'wb'))
else:
pickle.dump(results_list,
open(os.path.join(config.bq.save_dir, config.bq.name + '_fullGP_init_p' + str(config.bq.init_num_data) + '_results.p'), 'wb'))
else:
if config.bq.is_ai:
pickle.dump(results_list,
open(os.path.join(config.bq.save_dir, config.bq.name + '_ai_init_p' + str(config.bq.init_num_data) + '_results.p'), 'wb'))
else:
pickle.dump(results_list,
open(os.path.join(config.bq.save_dir, config.bq.name + '_opt_init_p' + str(config.bq.init_num_data) + 'iter' +str(config.bq.opt_iter) + 'lr' + str(config.bq.opt_lr) + '_results.p'), 'wb'))
sys.exit(0)
if __name__ == "__main__":
main()