-
Notifications
You must be signed in to change notification settings - Fork 0
/
experiment.py
180 lines (150 loc) · 4.7 KB
/
experiment.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
from collections import defaultdict
import numpy as np
import pandas as pd
from data import load_adult, preprocess_adult
from metrics import eval_model
from train import gen_decaf, train_decaf
# Define DAG for Adult dataset
DAG = {
"age": [
"occupation",
"hours-per-week",
"income",
"workclass",
"marital-status",
"education",
"relationship",
],
"education": [
"occupation",
"hours-per-week",
"income",
"workclass",
"relationship",
],
"hours-per-week": ["income"],
"marital-status": [
"occupation",
"hours-per-week",
"income",
"workclass",
"relationship",
"education",
],
"native-country": [
"marital-status",
"hours-per-week",
"education",
"workclass",
"income",
"relationship",
],
"occupation": ["income"],
"race": ["occupation", "income", "hours-per-week", "education", "marital-status"],
"relationship": ["income"],
"sex": [
"occupation",
"marital-status",
"income",
"workclass",
"education",
"relationship",
],
"workclass": ["income"],
}
def dag_to_idx(df, dag):
"""Convert columns in a DAG to the corresponding indices."""
dag_idx = []
for node, children in dag.items():
for child in children:
dag_idx.append([df.columns.get_loc(node), df.columns.get_loc(child)])
return dag_idx
def create_bias_dict(df, edge_map):
"""
Convert the given edge tuples to a bias dict used for generating
debiased synthetic data.
"""
bias_dict = {}
for key, val in edge_map.items():
bias_dict[df.columns.get_loc(key)] = [df.columns.get_loc(f) for f in val]
return bias_dict
def train_models(num_runs=10):
dataset_train = preprocess_adult(load_adult())
dataset_test = preprocess_adult(load_adult(load_test=True))
dag_seed = dag_to_idx(dataset_train, DAG)
bias_dicts = {
"nd": {},
"ftu": create_bias_dict(dataset_train, {"income": ["sex"]}),
"cf": create_bias_dict(
dataset_train, {"income": ["marital-status", "sex", "relationship"]}
),
"dp": create_bias_dict(
dataset_train,
{
"income": [
"occupation",
"hours-per-week",
"marital-status",
"education",
"sex",
"workclass",
"relationship",
]
},
),
"cf-y": create_bias_dict(
dataset_train,
{
"income": ["sex", "marital-status"],
"relationship": ["sex"],
},
),
"dp-y": create_bias_dict(
dataset_train,
{
"income": ["sex"],
"occupation": ["sex"],
"marital-status": ["sex"],
"workclass": ["sex"],
"education": ["sex"],
"relationship": ["sex"],
},
),
}
results = defaultdict(dict)
results["original"] = defaultdict(list)
for ver in bias_dicts.keys():
results[f"decaf_{ver}"] = defaultdict(list)
for run in range(num_runs):
model = train_decaf(
dataset_train, model_name=f"decaf_run_{run+1}", dag_seed=dag_seed, epochs=50
)
for ver, bias_edges in bias_dicts.items():
synth_data = gen_decaf(model, dataset_train, bias_edges)
if ver.endswith("-y"):
synth_X = gen_decaf(model, dataset_train, bias_dicts[ver[:-2]])
synth_X["income"] = synth_data["income"]
synth_data = synth_X
model_results = eval_model(synth_data, dataset_test)
for key, value in model_results.items():
results[f"decaf_{ver}"][key].append(value)
# Original
model_results = eval_model(dataset_train, dataset_test)
for key, value in model_results.items():
results["original"][key].append(value)
return results
def results_df(results):
def formatter(results):
return f"{np.mean(results):.3f}±{np.std(results):.3f}"
cols = ["model"] + list(results["original"].keys())
rows = list(results.keys())
df = pd.DataFrame(np.zeros((len(rows), len(cols))), columns=cols)
df["model"] = rows
for model, model_results in results.items():
for col in cols[1:]:
df.loc[df["model"] == model, col] = formatter(model_results[col])
return df
if __name__ == "__main__":
results = train_models(num_runs=10)
df = results_df(results)
print(df)