-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
125 lines (110 loc) · 4.7 KB
/
utils.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
import duckdb
from jinja2 import Template
import pandas as pd
from relbench.datasets import get_dataset
from relbench.tasks import get_task
from sklearn.feature_selection import mutual_info_classif, mutual_info_regression
DATASET_INFO = {
'rel-stack': {
'tables': ['users', 'posts', 'votes', 'badges', 'comments', 'postHistory'],
'tasks': ['user-engagement', 'user-badge', 'post-votes']
},
'rel-amazon': {
'tables': ['review', 'customer', 'product'],
'tasks': ['user-churn', 'user-ltv', 'product-ltv', 'product-churn']
},
'rel-hm': {
'tables': ['article', 'customer', 'transactions'],
'tasks': ['user-churn', 'item-sales'],
},
'rel-f1': {
'tables': ['races', 'circuits', 'drivers', 'results', 'standings', 'constructors',
'constructor_results', 'constructor_standings', 'qualifying'],
'tasks': ['driver-position', 'driver-dnf', 'driver-top3']
},
'rel-trial': {
'tables': ['studies', 'outcomes', 'outcome_analyses', 'drop_withdrawals',
'reported_event_totals', 'designs', 'eligibilities', 'interventions',
'conditions', 'facilities', 'sponsors', 'interventions_studies',
'conditions_studies', 'facilities_studies', 'sponsors_studies'],
'tasks': ['study-outcome', 'study-adverse', 'site-success']
},
'rel-event': {
'tables': ['users', 'events', 'event_attendees', 'event_interest', 'user_friends'],
'tasks': ['user-repeat', 'user-ignore', 'user-attendance']
}
}
def db_setup(dataset_name: str, db_filename: str):
""" Sets up a DuckDB database (at db_filename) with the tables from the specified dataset.
Args:
dataset_name (str): The name of the relbench dataset.
db_filename (str): Path to the DuckDB database file.
"""
conn = duckdb.connect(db_filename)
dataset = get_dataset(name=dataset_name, download=True) # noqa
tasks = DATASET_INFO[dataset_name]['tasks']
tables = DATASET_INFO[dataset_name]['tables']
for table_name in tables:
exec(f'{table_name} = dataset.get_db().table_dict["{table_name}"].df')
conn.sql(f'create table {table_name} as select * from {table_name}')
for task_name in tasks:
task = get_task(dataset_name, task_name, download=True)
train_table = task.get_table("train").df # noqa
val_table = task.get_table("val").df # noqa
test_table = task.get_table("test").df # noqa
task_name = task_name.replace('-', '_')
conn.sql(f'create table {task_name}_train as select * from train_table')
conn.sql(f'create table {task_name}_val as select * from val_table')
conn.sql(f'create table {task_name}_test as select * from test_table')
conn.close()
def render_jinja_sql(query: str, context: dict) -> str:
return Template(query).render(context)
def validate_feature_tables(
task: str, conn: duckdb.DuckDBPyConnection = None, db_filename: str = None
):
task = task.replace('-', '_')
if conn is None:
conn = duckdb.connect(db_filename)
error_count = 0
for s in ['train', 'val', 'test']:
table_name = f'{task}_{s}'
print(f'Validating {s}')
labels = conn.sql(f'select * from {table_name}').df()
feats = conn.sql(f'select * from {table_name}_feats').df()
print(f'{s} labels size: {len(labels):,} x {len(labels.columns):,}')
print(f'{s} feats size: {len(feats):,} x {len(feats.columns):,}')
# validate feats \subset labels
joined = labels.merge(feats, how='inner', on=labels.columns.tolist(), suffixes=('', '_r'))
if (diff := len(labels) - len(joined)) != 0:
print(f'⚠️ {diff:,} samples are missing from feats table!')
error_count += 1
print()
if error_count == 0:
print('✅ All tables are valid!')
else:
print(f'❌ {error_count} errors found!')
if db_filename is not None:
conn.close()
def feature_summary_df(df: pd.DataFrame, y_col: str, classification: bool = True):
y = df[y_col]
df = df.drop(y_col, axis=1)
invalid_cols = df.select_dtypes(exclude=['number', 'category']).columns
df = df.drop(invalid_cols, axis=1)
if classification:
mi = mutual_info_classif(df.fillna(-1).values, y)
else:
mi = mutual_info_regression(df.fillna(-1).values, y)
res = pd.DataFrame(
{
'Label Corr.': df.corrwith(y),
'Label MI': mi,
'NaN %': df.isna().mean(),
},
index=df.columns
)
return (
res
.sort_values(by='Label MI', ascending=False)
.style
.format({'Label Corr.': '{:.3f}', 'Label MI': '{:.3f}', 'NaN %': '{:.1%}'})
)