# Balance dataset by oversampling using SMOTEN (ideal for categorical data)
# (https://imbalanced-learn.org/stable/references/generated/imblearn.over_sampling.SMOTEN.html)
from imblearn.over_sampling import SMOTEN

ros = SMOTEN(n_jobs=14)  # Resampled on a 16-core, 60GB RAM machine

df_balanced_text, df_balanced_labels = ros.fit_resample(train_df[["text_a", "text_b"]], train_df["labels"])
df_balanced_text["labels"] = df_balanced_labels

print(f"""Balanced dataset size multiplier:
{len(df_balanced_labels) / len(train_df)}""")  # 4.07

# Need to cache the balanced data
df_balanced_text.to_hdf("./balanced.h5", "data", complevel=9)