-
Notifications
You must be signed in to change notification settings - Fork 3
/
pbo.py
196 lines (165 loc) · 6.27 KB
/
pbo.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
import random
import time
import warnings
from copy import deepcopy
from itertools import permutations
import numpy as np
import torch
from botorch.acquisition.analytic import PosteriorMean
from botorch.acquisition.monte_carlo import qExpectedImprovement
from botorch.optim.optimize import optimize_acqf
from botorch.utils.gp_sampling import get_gp_samples
from scipy.optimize import minimize
from acquisition_functions import ExpectedUtility
from constants import * # noqa: F403, F401
from sim_helpers import fit_pref_model, organize_comparisons
warnings.filterwarnings("ignore", message="Could not update `train_inputs` with transformed inputs")
def pref2rff(pref_model, n_samples):
# assume pref_model on cpu
pref_model = pref_model.eval().double()
# force the model to infer utility
pref_model.posterior(pref_model.datapoints)
modified_pref_model = deepcopy(pref_model)
class LikelihoodForRFF:
noise = torch.tensor(1.0).double()
modified_pref_model.likelihood = LikelihoodForRFF()
modified_pref_model.train_targets = pref_model.utility
modified_pref_model.input_transforms = None
gp_samples = get_gp_samples(
model=modified_pref_model,
num_outputs=1,
n_samples=n_samples,
num_rff_features=500,
)
gp_samples.input_transform = deepcopy(pref_model.input_transform)
# gp_samples = gp_samples.to(device=device)
return gp_samples
def get_pbo_pe_comparisons(
outcome_X,
train_comps,
problem,
utils,
init_round,
total_training_round,
comp_noise_type,
comp_noise,
pe_strategy,
):
"""
Generate TS-based comparisons on previously observed points
Args:
outcome_X ([type]): [description]
train_comps ([type]): [description]
problem ([type]): [description]
utils ([type]): [description]
init_round ([type]): [description]
total_training_round ([type]): [description]
comp_noise_type ([type]): [description]
comp_noise ([type]): [description]
pe_strategy (str): being either "random", "ts" or "eubo"
Returns:
[type]: [description]
"""
all_pairs = torch.combinations(torch.tensor(range(outcome_X.shape[-2])), r=2).to(train_comps)
for i in range(total_training_round):
pbo_pe_start_time = time.time()
if (
(pe_strategy != "random")
and (train_comps is not None)
and (train_comps.shape[-2] >= init_round)
):
pbo_pref_model = fit_pref_model(
outcome_X,
train_comps,
kernel="default",
transform_input=True,
Y_bounds=problem.bounds,
)
if (
(pe_strategy == "random")
or (train_comps is None)
or (train_comps.shape[-2] < init_round)
):
cand_comps = all_pairs[
torch.randint(high=all_pairs.shape[-2], size=(1,)),
]
elif pe_strategy == "ts":
cand_comps = None
# use TS to draw comparisons
comp1 = pbo_pref_model.posterior(outcome_X).sample().argmax(dim=-2)
# exclude the first sample
sample2 = pbo_pref_model.posterior(outcome_X).sample()
sample2[:, comp1.squeeze(), :] = -float("Inf")
comp2 = sample2.argmax(dim=-2)
# Create candidate comparisons
cand_comps = torch.cat((comp1, comp2), dim=-1)
elif pe_strategy == "eubo":
eubo_acqf = ExpectedUtility(
preference_model=pbo_pref_model,
outcome_model=None,
previous_winner=None,
search_space_type="y",
)
cand_comps = None
max_eubo_val = -np.inf
for j in range(all_pairs.shape[-2]):
X_pair = outcome_X[all_pairs[j, :]]
eubo_val = eubo_acqf(X_pair).item()
if eubo_val > max_eubo_val:
max_eubo_val = eubo_val
cand_comps = all_pairs[[j], :]
else:
raise ValueError("Unsupported PE strategy for PBO")
cand_comps = organize_comparisons(utils, cand_comps, comp_noise_type, comp_noise)
pbo_pe_time = time.time() - pbo_pe_start_time
train_comps = cand_comps if train_comps is None else torch.cat((train_comps, cand_comps))
print(
f"PBO with PE strategy {pe_strategy} gen time: {pbo_pe_time:.2f}s, train_comps shape: {train_comps.shape}"
)
return train_comps
def gen_pbo_candidates(outcome_X, train_comps, q, problem, pbo_gen_method):
"""generate pbo candidates
Args:
outcome_X (_type_): _description_
train_comps (_type_): _description_
q (_type_): _description_
problem (_type_): _description_
pbo_gen_method (_type_): _description_
"""
if pbo_gen_method == "ts":
problem_cpu = deepcopy(problem).cpu()
pref_model = fit_pref_model(
outcome_X, train_comps, kernel="default", transform_input=True, Y_bounds=problem.bounds
)
outcome_cand_X = []
for _ in range(q):
gp_samples = pref2rff(pref_model.cpu(), n_samples=1)
acqf = PosteriorMean(gp_samples)
single_outcome_cand_X, _ = optimize_acqf(
acqf,
bounds=problem_cpu.bounds,
q=1,
num_restarts=NUM_RESTARTS,
raw_samples=RAW_SAMPLES,
options={"batch_limit": 1},
)
outcome_cand_X.append(single_outcome_cand_X)
outcome_cand_X = torch.cat(outcome_cand_X).to(outcome_X)
elif pbo_gen_method == "ei":
pref_model = fit_pref_model(
outcome_X, train_comps, kernel="default", transform_input=True, Y_bounds=problem.bounds
)
# to fill in utility values
pref_model.posterior(pref_model.datapoints)
acqf = qExpectedImprovement(model=pref_model, best_f=pref_model.utility.max().item())
outcome_cand_X, _ = optimize_acqf(
acqf,
bounds=problem.bounds,
q=q,
num_restarts=NUM_RESTARTS,
raw_samples=RAW_SAMPLES,
sequential=True,
)
else:
raise ValueError("Unsupported gen_method for PBO")
return outcome_cand_X