process/tweet_emoji.py

import json
import os
from random import shuffle, seed
import numpy as np
import pandas as pd
from datasets import load_dataset

test = load_dataset("cardiffnlp/super_tweeteval", "tweet_emoji", split="test").shuffle(seed=42)
test = list(test.to_pandas().T.to_dict().values())
train = load_dataset("cardiffnlp/super_tweeteval", "tweet_emoji", split="train").shuffle(seed=42)
train = list(train.to_pandas().T.to_dict().values())
validation = load_dataset("cardiffnlp/super_tweeteval", "tweet_emoji", split="validation").shuffle(seed=42)
validation = list(validation.to_pandas().T.to_dict().values())
full = train + test + validation
df = pd.DataFrame(full)
df["date_dt"] = pd.to_datetime(df.date)
df = df.sort_values(by="date_dt")

# remove emoji 68 (face holding back tear) as it's not in the training set (so the total number is 99, not 100)
df = df[df["gold_label"] != 68]
df["gold_label"] = [i if i < 68 else i - 1 for i in df["gold_label"]]

dist_date = df.groupby("date_dt").size()
total_n = len(df)
n = 0
while True:
    n += 1
    if dist_date[:n].sum() > total_n/2:
        break
split_date = dist_date.index[n]
print(split_date)

train = df[df["date_dt"] <= split_date]
test = df[df["date_dt"] > split_date]
print(train.date_dt.min(), train.date_dt.max())
print(test.date_dt.min(), test.date_dt.max())

train.pop("date_dt")
test.pop("date_dt")
train = list(train.T.to_dict().values())
test = list(test.T.to_dict().values())

seed(42)
shuffle(train)
shuffle(test)
valid = train[:int(len(train)*0.2)]
train = train[len(valid):]
assert np.unique([i["gold_label"] for i in train], return_counts=True)[1].shape[0] == 99
assert np.unique([i["gold_label"] for i in test], return_counts=True)[1].shape[0] == 99
assert np.unique([i["gold_label"] for i in valid], return_counts=True)[1].shape[0] == 99

n_test = int(len(test)/4)
n_train = len(train)
n_validation = len(valid)
test_1 = test[:n_test]
test_2 = test[n_test:n_test*2]
test_3 = test[n_test*2:n_test*3]
test_4 = test[n_test*3:]
os.makedirs("data/tweet_emoji", exist_ok=True)
with open("data/tweet_emoji/test.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in test]))
with open("data/tweet_emoji/test_1.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in test_1]))
with open("data/tweet_emoji/test_2.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in test_2]))
with open("data/tweet_emoji/test_3.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in test_3]))
with open("data/tweet_emoji/test_4.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in test_4]))
with open("data/tweet_emoji/train.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in train]))
with open("data/tweet_emoji/validation.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in valid]))


def sampler(dataset_test, r_seed):
    seed(r_seed)
    shuffle(dataset_test)
    shuffle(train)
    shuffle(valid)
    test_tr = dataset_test[:int(n_train / 2)]
    test_vl = dataset_test[int(n_train / 2): int(n_train / 2) + int(n_validation / 2)]
    new_train = test_tr + train[:n_train - len(test_tr)]
    new_validation = test_vl + valid[:n_validation - len(test_vl)]
    print(np.unique([i["gold_label"] for i in new_train], return_counts=True)[1].shape[0])
    print(np.unique([i["gold_label"] for i in new_validation], return_counts=True)[1].shape[0])
    assert np.unique([i["gold_label"] for i in new_train], return_counts=True)[1].shape[0] == 99
    assert np.unique([i["gold_label"] for i in new_validation], return_counts=True)[1].shape[0] == 99
    return new_train, new_validation

id2test = {n: t for n, t in enumerate([test_1, test_2, test_3, test_4])}
for n, _test in enumerate([
        test_4 + test_2 + test_3,
        test_1 + test_4 + test_3,
        test_1 + test_2 + test_4,
        test_1 + test_2 + test_3]):
    for s in range(3):
        os.makedirs(f"data/tweet_emoji_test{n}_seed{s}", exist_ok=True)
        _train, _valid = sampler(_test, s)
        with open(f"data/tweet_emoji_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_emoji_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_emoji_test{n}_seed{s}/test.jsonl", "w") as f:
            f.write("\n".join([json.dumps(i) for i in id2test[n]]))