Skip to content

fold_weights

Logic for calculating fold weights proportional to individual LB scores.

kelp.nn.inference.fold_weights.main

Main entrypoint for calculating fold weights.

Source code in kelp/nn/inference/fold_weights.py 
 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
def main() -> None:
    """Main entrypoint for calculating fold weights."""
    fold_scores = [
        ("fold=0", 0.7110),
        ("fold=1", 0.7086),
        ("fold=2", 0.7110),
        ("fold=3", 0.7139),
        ("fold=4", 0.7106),
        ("fold=5", 0.7100),
        ("fold=6", 0.7119),
        ("fold=7", 0.7105),
        ("fold=8", 0.7155),
        ("fold=9", 0.7047),
        ("fold=0v2", 0.7135),
        ("fold=1v2", 0.7114),
        ("fold=2v2", 0.7094),
        ("fold=3v2", 0.7133),
        ("fold=4v2", 0.7106),
    ]
    df = pd.DataFrame(fold_scores, columns=["fold", "score"])

    scaler = MinMaxScaler(feature_range=(0.2, 1.0))
    norm_scores = scaler.fit_transform(df[["score"]].values)
    df["score_norm"] = norm_scores
    df.to_parquet("scores.parquet")