kendex/prep_push_to_hf.py

from os.path import basename, join
from pathlib import Path

import librosa
import numpy as np
import pandas as pd
from datasets import Audio, Dataset, DatasetDict
from transformers import AutoTokenizer

from leviticus import normalize

MAX_DURATION_IN_SECONDS = 10.0
MIN_DURATION_IN_SECONDS = 1.0
MAX_LEN = 50
MIN_LEN = 5
SR = 16_000
TOKENIZER_CHECKPOINT = "distilbert-base-uncased-finetuned-sst-2-english"

tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_CHECKPOINT)


def duration_filter(item):
    return MIN_DURATION_IN_SECONDS < item < MAX_DURATION_IN_SECONDS


def text_filter(item):
    return MIN_LEN < len([i for i in item.split(" ") if len(i) > 0]) < MAX_LEN


def text_mapper(item):
    text = item["text"]
    item["normalized_text"] = normalize(text)
    return item


def create_datasets(df):
    def create_dataset(df_slice):
        audio_column = "audio"
        text_column = "text"
        duration_column = "duration"

        dataset = Dataset.from_pandas(df_slice)
        dataset = dataset.cast_column(audio_column, Audio(sampling_rate=SR))
        dataset = dataset.filter(text_filter, input_columns=[text_column])
        dataset = dataset.filter(duration_filter, input_columns=[duration_column])
        dataset = dataset.map(text_mapper, batched=False)

        return dataset

    train, test = np.split(df, [int(0.9 * len(df))])
    train_dataset = create_dataset(train)
    test_dataset = create_dataset(test)

    return train_dataset, test_dataset


def main():
    repo_dir = Path(__file__).resolve().parent.parent
    data_dir = join(repo_dir, "data")
    kendex_dir = join(data_dir, "Kendex")
    audio_dir = join(kendex_dir, "wavs")

    metadata = pd.read_csv(join(kendex_dir, "metadata.csv"), delimiter="|", header=None)
    wavs = pd.Series([join(audio_dir, f"{f}.wav") for f in metadata[0]])
    data = {
        "audio": wavs,
        "file": [basename(w) for w in wavs],
        "text": metadata[1],
        "duration": [librosa.get_duration(path=w) for w in wavs],
    }

    df = pd.DataFrame(data).sample(frac=1, random_state=666).reset_index(drop=True)

    train, test = create_datasets(df)

    full_dataset = DatasetDict({"train": train, "test": test})
    full_dataset.push_to_hub("michaelnetbiz/Kendex")


if __name__ == "__main__":
    main()