| # 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) |