inst.py

import hashlib
import os
import shutil
import sys
from datetime import datetime
from pathlib import Path

import click
import yaml
from ml_collections import ConfigDict

from UVR import ModelData, AudioTools, Ensembler, DENOISER_MODEL_PATH, DEVERBER_MODEL_PATH, MDX_MODELS_DIR, \
    MDX_MIXER_PATH, load_model_hash_data, MDX_MODEL_NAME_SELECT, model_hash_table, MDX_HASH_DIR, MDX_C_CONFIG_PATH
from args import mdx23c_8kfft_instvoc_hq_process_data, htdemucs_ft_process_data, uvr_mdx_net_voc_ft_process_data
from download import download_model, get_model_file
from gui_data.constants import VR_ARCH_TYPE, MDX_ARCH_TYPE, DEMUCS_ARCH_TYPE, ENSEMBLE_MODE, TIME_STRETCH, \
    MANUAL_ENSEMBLE, MATCH_INPUTS, ALIGN_INPUTS, ALL_STEMS, DEFAULT, VOCAL_STEM, MP3_BIT_RATES, WAV, DEMUCS_2_SOURCE, \
    DEMUCS_2_SOURCE_MAPPER, INST_STEM, CKPT, ONNX, MDX_POP_NFFT, secondary_stem, PRIMARY_STEM, SECONDARY_STEM
from lib_v5 import spec_utils
from separate import (
    SeperateDemucs, SeperateMDX, SeperateMDXC, SeperateVR,  # Model-related
    save_format, clear_gpu_cache,  # Utility functions
    cuda_available, mps_available,  # directml_available,
)


def run_ensemble_models(audio_path, export_path, format=WAV, clean=True):
    start = datetime.now()
    process_datas = [mdx23c_8kfft_instvoc_hq_process_data, uvr_mdx_net_voc_ft_process_data,
                     htdemucs_ft_process_data]
    # download models
    for process_data in process_datas:
        download_model(process_data['model_name'])

    # create folder
    os.makedirs(export_path, exist_ok=True)
    temp_export_path = os.path.join(export_path, 'uvr5_' + datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
    os.makedirs(temp_export_path, exist_ok=True)
    print(f'temp_export_path', temp_export_path)

    instrumental_export_paths = []
    vocals_export_paths = []

    for process_data in process_datas:
        current_model = process_data['model_data']
        audio_file_base = Path(audio_path).stem + '_' + current_model.model_basename
        process_data['export_path'] = temp_export_path
        process_data['audio_file_base'] = audio_file_base
        process_data['audio_file'] = audio_path

        if current_model.process_method == VR_ARCH_TYPE:
            seperator = SeperateVR(current_model, process_data)
        elif current_model.process_method == MDX_ARCH_TYPE:
            seperator = SeperateMDXC(current_model, process_data) if current_model.is_mdx_c else SeperateMDX(
                current_model, process_data)
        elif current_model.process_method == DEMUCS_ARCH_TYPE:
            seperator = SeperateDemucs(current_model, process_data, vocal_stem_path=(audio_path, audio_file_base))
        else:
            raise Exception(f'model not found')

        seperator.seperate()

        instrumental_path = Path(temp_export_path) / f"{audio_file_base}_(Instrumental).{format.lower()}"
        vocals_path = Path(temp_export_path) / f"{audio_file_base}_(Vocals).{format.lower()}"
        instrumental_export_paths.append(str(instrumental_path))
        vocals_export_paths.append(str(vocals_path))

    # merge each model outputs
    vocals_final_path = Path(export_path) / f"{Path(audio_path).stem}.vocal.{format.lower()}"
    instrumental_final_path = Path(export_path) / f"{Path(audio_path).stem}.instrumental.{format.lower()}"

    ensemble(vocals_export_paths, vocals_final_path)
    ensemble(instrumental_export_paths, instrumental_final_path)

    if clean:
        shutil.rmtree(temp_export_path)
    print(f'instrumental_final_path', instrumental_final_path)
    print(f'vocals_final_path', vocals_final_path)
    print(f'Finished in {datetime.now() - start}')
    return instrumental_final_path, vocals_final_path


def ensemble(stem_outputs, stem_save_path, format=WAV):
    algorithm = 'Average'
    is_normalization = True
    spec_utils.ensemble_inputs(stem_outputs, algorithm, is_normalization, 'PCM_16', stem_save_path, is_wave=True)
    save_format(stem_save_path, format, '320k')


# /Users/taoluo/Downloads/test/kimk_audio_MDX23C-8KFFT-InstVoc_HQ_(Instrumental).WAV
#
if __name__ == '__main__':
    audio_file = '/Users/taoluo/Downloads/assets/audio/kimk_audio.mp3'
    audio_file = sys.argv[1]
    if not os.path.isfile(audio_file):
        raise FileNotFoundError(f"File {audio_file} not exist.")

    output_dir = os.path.dirname(audio_file)
    print(output_dir)
    run_ensemble_models(audio_file, output_dir)