-
Notifications
You must be signed in to change notification settings - Fork 2
/
test_unbiased_vs_shapiq.py
156 lines (123 loc) · 5.46 KB
/
test_unbiased_vs_shapiq.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
import copy
import itertools
import random
import numpy as np
from approximators.unbiased import CovertRegression, calculate_uksh_from_samples, get_weights
from approximators.shapiq import SHAPIQEstimator
from games import NLPLookupGame
class GameWrapper:
def __init__(self, game):
self.game = game
self.players = game.n
def __call__(self, S):
result = np.empty(S.shape[0])
for i in range(S.shape[0]):
sample = S[i]
s_set = set()
for j, value in enumerate(sample):
if value == 1:
s_set.add(j)
result[i] = self.game.set_call(s_set)
return result
def grand(self):
"""Get grand coalition value."""
return self.__call__(np.ones((1, self.players), dtype=bool))[0]
def null(self):
"""Get null coalition value."""
return self.__call__(np.zeros((1, self.players), dtype=bool))[0]
def get_S_and_game(budget, num_players, weight_vector, N, pairing, game_fun):
complete_subsets = []
incomplete_subsets = list(N)
incomplete_subsets.remove(0)
all_subsets_to_sample = []
for complete_subset in complete_subsets:
combinations = itertools.combinations(N, complete_subset)
for subset in combinations:
subset = set(subset)
all_subsets_to_sample.append(subset)
remaining_weight = weight_vector[incomplete_subsets] / sum(weight_vector[incomplete_subsets])
if len(incomplete_subsets) > 0:
sampled_subsets = []
while len(sampled_subsets) < budget:
subset_size = random.choices(incomplete_subsets, remaining_weight, k=1)
ids = np.random.choice(num_players, size=subset_size, replace=False)
sampled_subsets.append(tuple(sorted(ids)))
if pairing:
if len(sampled_subsets) < budget:
sampled_subsets.append(tuple(N - set(ids)))
for subset in sampled_subsets:
all_subsets_to_sample.append(set(subset))
game_values = [game_fun(subset) for subset in all_subsets_to_sample]
return all_subsets_to_sample, game_values
def compare_unbiasedksh_and_shapx(
game,
budget: int,
pairing: bool = True,
u_ksh_sample_size: int = 5_000
):
num_players = game.n
game_fun = game.set_call
N = set(range(num_players))
empty_value = game_fun({})
full_value = game_fun(N)
weights = get_weights(num_players)
weight_vector = (np.asarray([0] + [*weights] + [0])) / sum(weights)
all_subsets_to_sample, game_values = get_S_and_game(
budget, num_players, weight_vector, N, pairing, game_fun)
# SII
interaction_estimator = SHAPIQEstimator(N=N, order=1, interaction_type='SII')
values_shapx_FSI = interaction_estimator.compute_from_samples(
S_list=all_subsets_to_sample, game_values=game_values,
val_empty=empty_value, val_full=full_value)
values_shapx_sii = copy.deepcopy(values_shapx_FSI[1])
# STI
interaction_estimator = SHAPIQEstimator(N=N, order=1, interaction_type='STI')
values_shapx_FSI = interaction_estimator.compute_from_samples(
S_list=all_subsets_to_sample, game_values=game_values,
val_empty=empty_value, val_full=full_value)
values_shapx_sti = copy.deepcopy(values_shapx_FSI[1])
# FSI
interaction_estimator = SHAPIQEstimator(N=N, order=1, interaction_type='FSI')
values_shapx_FSI = interaction_estimator.compute_from_samples(
S_list=all_subsets_to_sample, game_values=game_values,
val_empty=empty_value, val_full=full_value)
values_shapx_FSI = copy.deepcopy(values_shapx_FSI[1])
values_ksh = calculate_uksh_from_samples(
game=GameWrapper(game),
game_values=game_values,
S_list=all_subsets_to_sample
)
# Original Unbiased Kernel SHAP
u_ksh_covert, _, _, _ = CovertRegression(
game=GameWrapper(game),
batch_size=1,
detect_convergence=False,
n_samples=u_ksh_sample_size,
paired_sampling=pairing
)
values_shapx_sii = [round(value, 5) for value in values_shapx_sii]
values_shapx_sti = [round(value, 5) for value in values_shapx_sti]
values_shapx_FSI = [round(value, 5) for value in values_shapx_FSI]
values_ksh = [round(value, 5) for value in values_ksh]
u_ksh_covert = [round(value, 5) for value in u_ksh_covert]
print(f"shapx-defined-samples (sii): {values_shapx_sii} (n: {budget})\n"
f"shapx-defined-samples (sti): {values_shapx_sti} (n: {budget})\n"
f"shapx-defined-samples (FSI): {values_shapx_FSI} (n: {budget})\n"
f"u-ksh-defined-samples: {values_ksh} (n: {budget})\n"
f"u-ksh-sampling: {u_ksh_covert} (n: {u_ksh_sample_size})")
return values_ksh, values_shapx_sii, values_shapx_sti, values_shapx_FSI, u_ksh_covert
if __name__ == "__main__":
n = 14
N = set(range(n))
game = NLPLookupGame(n=n, sentence_id=172, set_zero=True)
game_fun = game.set_call
shap = SHAPIQEstimator(N, 1, "SII")
exact_values = shap.compute_interactions_complete(game_fun)[1]
result = compare_unbiasedksh_and_shapx(
game=game, budget=500, pairing=False, u_ksh_sample_size=2000)
values_ksh, values_shapx_sii, values_shapx_sti, values_shapx_FSI, u_ksh_covert = result
feature_names = game.input_sentence.split(" ")
print(feature_names)
print(np.sum((exact_values - values_ksh) ** 2))
print(np.sum((exact_values - values_shapx_sii) ** 2))
assert values_ksh == values_shapx_sii