process/tweet_hate_balance.py

import json
import os
import numpy as np
import pandas as pd

with open("data/tweet_hate/train.jsonl") as f:
    df_train = pd.DataFrame([json.loads(i) for i in f if len(i)])
with open("data/tweet_hate/validation.jsonl") as f:
    df_val = pd.DataFrame([json.loads(i) for i in f if len(i)])
label_train_dist = dict(zip(*np.unique(df_train["gold_label_binary"], return_counts=True)))
label_val_dist = dict(zip(*np.unique(df_val["gold_label_binary"], return_counts=True)))
df_train_positive = df_train[df_train["gold_label_binary"] == 1]
df_train_negative = df_train[df_train["gold_label_binary"] == 0]
df_val_positive = df_val[df_val["gold_label_binary"] == 1]
df_val_negative = df_val[df_val["gold_label_binary"] == 0]

df_test = {}
with open("data/tweet_hate/test_1.jsonl") as f:
    df_test[0] = pd.DataFrame([json.loads(i) for i in f if len(i)])
with open("data/tweet_hate/test_2.jsonl") as f:
    df_test[1] = pd.DataFrame([json.loads(i) for i in f if len(i)])
with open("data/tweet_hate/test_3.jsonl") as f:
    df_test[2] = pd.DataFrame([json.loads(i) for i in f if len(i)])
with open("data/tweet_hate/test_4.jsonl") as f:
    df_test[3] = pd.DataFrame([json.loads(i) for i in f if len(i)])


def sampler(chunk_index, r_seed):
    df_test_tmp = pd.concat([v for k, v in df_test.items() if k != chunk_index])
    df_test_tmp = df_test_tmp.sample(len(df_test_tmp), random_state=r_seed)
    ratio = len(df_train)/(len(df_train) + len(df_val))
    df_train_tmp = df_test_tmp.iloc[:int(ratio * len(df_test_tmp))]
    df_val_tmp = df_test_tmp.iloc[int(ratio * len(df_test_tmp)):]

    label_train_tmp_dist = dict(zip(*np.unique(df_train_tmp["gold_label_binary"], return_counts=True)))
    assert label_train_tmp_dist[0] < label_train_dist[0] and label_train_tmp_dist[1] < label_train_dist[1]
    df_train_tmp_p = df_train_positive.sample(label_train_dist[1] - label_train_tmp_dist[1], random_state=r_seed)
    df_train_tmp_n = df_train_negative.sample(label_train_dist[0] - label_train_tmp_dist[0], random_state=r_seed)
    df_train_tmp = pd.concat([df_train_tmp, df_train_tmp_p, df_train_tmp_n])
    df_train_tmp.index = range(len(df_train_tmp))
    assert dict(zip(*np.unique(df_train_tmp["gold_label_binary"], return_counts=True))) == label_train_dist
    print(dict(zip(*np.unique(df_train_tmp["gold_label_binary"], return_counts=True))))

    label_val_tmp_dist = dict(zip(*np.unique(df_val_tmp["gold_label_binary"], return_counts=True)))
    assert label_val_tmp_dist[0] < label_val_dist[0] and label_val_tmp_dist[1] < label_val_dist[1]
    df_val_tmp_p = df_val_positive.sample(label_val_dist[1] - label_val_tmp_dist[1], random_state=r_seed)
    df_val_tmp_n = df_val_negative.sample(label_val_dist[0] - label_val_tmp_dist[0], random_state=r_seed)
    df_val_tmp = pd.concat([df_val_tmp, df_val_tmp_p, df_val_tmp_n])
    df_val_tmp.index = range(len(df_val_tmp))
    assert dict(zip(*np.unique(df_val_tmp["gold_label_binary"], return_counts=True))) == label_val_dist
    return list(df_train_tmp.T.to_dict().values()), list(df_val_tmp.T.to_dict().values())

for n in range(4):
    for s in range(3):
        os.makedirs(f"data/tweet_hate_balance_test{n}_seed{s}", exist_ok=True)
        _train, _valid = sampler(n, s)
        with open(f"data/tweet_hate_balance_test{n}_seed{s}/train.jsonl", "w") as f:
            f.write("\n".join([json.dumps(i) for i in _train]))
        with open(f"data/tweet_hate_balance_test{n}_seed{s}/validation.jsonl", "w") as f:
            f.write("\n".join([json.dumps(i) for i in _valid]))
        with open(f"data/tweet_hate_balance_test{n}_seed{s}/test.jsonl", "w") as f:
            f.write("\n".join([json.dumps(i) for i in list(df_test[n].T.to_dict().values())]))