English | 한국어

EZVitsDataset

Description

EZVitsDataset processes video datasets. This library creates a dataset by downloading a YouTube video or loading the video from a specified path.

Installation

This project can be run only if ffmpeg is installed on your computer.

git clone https://github.com/hopoduck/EZVitsDataset.git
cd ez-vits-dataset 

conda create --name dataset python=3.10

conda activate dataset

pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install "audio-separator[gpu]>=0.15.2" # or pip install "audio-separator[cpu]>=0.15.2"
pip install -r requirements.txt

How to use

Import the EZVitsDataset and EZVitsDatasetParams modules and initialize EZVitsDataset. For initialization use: For options that can be set, see the EZVitsDatasetParams object.
Then call the main function to run download or local mode. The mode parameter selects download or local, and the path_or_url parameter selects a YouTube URL or specifies the path where the video file is located.

from ezvitsdataset.EZVitsDataset import EZVitsDataset, EZVitsDatasetParams

dataset = EZVitsDataset(
    EZVitsDatasetParams(
        device="cuda",
        language="en",
    )
)

dataset.main(
    mode="download",
    path_or_url="youtube video or playlist url",
)

After that, run main.py.

python main.py

EZVitsDatasetParams parameters

• language: Video file language (automatically detects whisper if not set, but speed may be slow)
• skip_download: Skip YouTube download (default: False)
• youtube_dl_option: yt-dlp options
• sampling_rate: Output wav file sampling rate (default: 44100)
• skip_min_time: Do not process the dataset if it is shorter than the specified number of seconds (default: 10)
• skip_max_time: Do not process the dataset if it is longer than the specified number of seconds (default: 10)
• remove_original_file: Delete the original file after processing the dataset (default: False)
• whisper_model: whisper model name (default: 'large-v2')
• device: device to run cpu/cuda (default: 'cuda')
• compute_type: Change depending on the execution device (default: 'float16')
• overwrite_data: Whether to overwrite in case of duplication during data processing (default: False)
• audio_separator_model: uvr model name (default: 'UVR_MDXNET_KARA_2.onnx')
  • You can check available models at here.
• batch_size: Amount of data to be processed at once when processing whisper model, can be lowered when VRAM is insufficient (default value: 8)
• chunk_size: Length of one sentence when processing whisper model (default: 6)
• remove_tmp_file: Whether to delete temporary files after task completion (default: False)
• download_path: Path to download video
• uvr_path: Path to save audio file after uvr processing
• output_path: Path to save the audio file after all processing.
• filelist_path: filelist.txt, train.txt, val.txt creation path

License

EZVitsDataset is released under the MIT license. For more information, see the LICENSE file in the repository.