extract.py

import os
import sys
import glob
import time
import tqdm
import torch
import numpy as np
import concurrent.futures
import multiprocessing as mp
import json
import shutil
import argparse
import torchcrepe
import resampy
import penn

now_dir = os.getcwd()
sys.path.append(os.path.join(now_dir))

# Zluda hijack
import rvc.lib.zluda

from rvc.lib.utils import load_audio, load_embedding
from rvc.train.extract.preparing_files import generate_config, generate_filelist
from rvc.lib.predictors.RMVPE import RMVPE0Predictor
from rvc.configs.config import Config

# Load config
config = Config()

mp.set_start_method("spawn", force=True)


class FeatureInput:
    """Class for F0 extraction."""

    def __init__(self, sample_rate=16000, hop_size=160, device="cpu"):
        self.fs = sample_rate
        self.hop = hop_size
        self.f0_bin = 256
        self.f0_max = 1100.0
        self.f0_min = 50.0
        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
        self.device = device
        self.model_rmvpe = None

    def compute_f0(self, np_arr, f0_method, hop_length):
        """Extract F0 using the specified method."""
        if f0_method == "crepe":
            return self.get_crepe(np_arr, hop_length)
        elif f0_method == "rmvpe":
            # Ensure model is loaded if needed (handled in process_files)
            if self.model_rmvpe is None:
                 raise RuntimeError("RMVPE model not initialized. Call process_files first.")
            return self.model_rmvpe.infer_from_audio(np_arr, thred=0.03)
        elif f0_method == "fcnf0":
            return self.get_fcnf0(np_arr)
        else:
            raise ValueError(f"Unknown F0 method: {f0_method}")

    def get_crepe(self, x, hop_length):
        """Extract F0 using CREPE."""
        audio = torch.from_numpy(x.astype(np.float32)).to(self.device)
        audio /= torch.quantile(torch.abs(audio), 0.999)
        audio = audio.unsqueeze(0)
        pitch = torchcrepe.predict(
            audio,
            self.fs,
            hop_length,
            self.f0_min,
            self.f0_max,
            "full",
            batch_size=hop_length * 2,
            device=audio.device,
            pad=True,
        )
        source = pitch.squeeze(0).cpu().float().numpy()
        source[source < 0.001] = np.nan
        target = np.interp(
            np.arange(0, len(source) * (x.size // self.hop), len(source))
            / (x.size // self.hop),
            np.arange(0, len(source)),
            source,
        )
        return np.nan_to_num(target)

    def get_fcnf0(self, x):
        """Extract F0 using FCNF0++"""
        device_obj = torch.device(self.device)

        # FCNF0++ uses 8kHz sample rate per paper for increased accuracy
        audio_8k = resampy.resample(x, self.fs, 8000, filter='kaiser_best') 
        audio_tensor = torch.from_numpy(audio_8k.astype(np.float32)).to(device_obj) 
        audio_tensor = audio_tensor.unsqueeze(0)

        gpu_index = device_obj.index if device_obj.type == 'cuda' else None

        # These settings are based on both paper and authors examples
        pitch, periodicity = penn.from_audio(
            audio=audio_tensor,
            sample_rate=8000,        
            hopsize=0.01, # 10 ms   
            fmin=30,                      
            fmax=1600,                  
            checkpoint=None,  # Defaults stock FCNF0++ ckpt     
            batch_size=2048,            
            center='half-hop',         
            interp_unvoiced_at=0.065, 
            gpu=gpu_index
        )

        source = pitch.squeeze().cpu().float().numpy()


        time_original = np.arange(x.size // self.hop) * (self.hop / self.fs)
        time_fcnf0 = np.arange(len(source)) * 0.01 # Time points for penn output

        # Handle edge case where source might be empty or have only one value
        if len(source) < 2:
             # If empty or single value, return constant array of that value (or NaN)
             fill_value = source[0] if len(source) == 1 else np.nan
             target = np.full(x.size // self.hop, fill_value)
        else:
             target = np.interp(time_original, time_fcnf0, source, left=source[0], right=source[-1])

        return np.nan_to_num(target)

    def coarse_f0(self, f0):
        """Convert F0 to coarse F0."""
        f0_mel = 1127 * np.log(1 + f0 / 700)
        f0_mel = np.clip(
            (f0_mel - self.f0_mel_min)
            * (self.f0_bin - 2)
            / (self.f0_mel_max - self.f0_mel_min)
            + 1,
            1,
            self.f0_bin - 1,
        )
        return np.rint(f0_mel).astype(int)

    def process_file(self, file_info, f0_method, hop_length):
        """Process a single audio file for F0 extraction."""
        inp_path, opt_path1, opt_path2, _ = file_info

        if os.path.exists(opt_path1) and os.path.exists(opt_path2):
            return

        try:
            np_arr = load_audio(inp_path, 16000)
            feature_pit = self.compute_f0(np_arr, f0_method, hop_length)
            np.save(opt_path2, feature_pit, allow_pickle=False)
            coarse_pit = self.coarse_f0(feature_pit)
            np.save(opt_path1, coarse_pit, allow_pickle=False)
        except Exception as error:
            print(
                f"An error occurred extracting file {inp_path} on {self.device}: {error}"
            )

    def process_files(
        self, files, f0_method, hop_length, device_num, device, n_threads
    ):
        """Process multiple files."""
        self.device = device
        if f0_method == "rmvpe":
            self.model_rmvpe = RMVPE0Predictor(
                os.path.join("rvc", "models", "predictors", "rmvpe.pt"),
                is_half=False,
                device=device,
            )
        elif f0_method == "fcnf0":
            # Penn lib handles it
            pass 
        else: 
            n_threads = 1

        n_threads = 1 if n_threads == 0 else n_threads

        def process_file_wrapper(file_info):
            self.process_file(file_info, f0_method, hop_length)

        with tqdm.tqdm(total=len(files), leave=True, position=device_num) as pbar:
            # using multi-threading
            with concurrent.futures.ThreadPoolExecutor(
                max_workers=n_threads
            ) as executor:
                futures = [
                    executor.submit(process_file_wrapper, file_info)
                    for file_info in files
                ]
                for future in concurrent.futures.as_completed(futures):
                    pbar.update(1)


def run_pitch_extraction(files, devices, f0_method, hop_length, num_processes):
    devices_str = ", ".join(devices)
    print(
        f"Starting pitch extraction with {num_processes} cores on {devices_str} using {f0_method}..."
    )
    start_time = time.time()
    fe = FeatureInput()
    ps = []
    num_devices = len(devices)
    for i, device in enumerate(devices):
        p = mp.Process(
            target=fe.process_files,
            args=(
                files[i::num_devices],
                f0_method,
                hop_length,
                i,
                device,
                num_processes // num_devices,
            ),
        )
        ps.append(p)
        p.start()
    for i, device in enumerate(devices):
        ps[i].join()

    elapsed_time = time.time() - start_time
    print(f"Pitch extraction completed in {elapsed_time:.2f} seconds.")


def process_file_embedding(
    files, version, embedder_model, embedder_model_custom, device_num, device, n_threads
):
    dtype = torch.float32
    model = load_embedding(embedder_model, embedder_model_custom).to(dtype).to(device)
    n_threads = 1 if n_threads == 0 else n_threads

    def process_file_embedding_wrapper(file_info):
        wav_file_path, _, _, out_file_path = file_info
        if os.path.exists(out_file_path):
            return
        feats = torch.from_numpy(load_audio(wav_file_path, 16000)).to(dtype).to(device)
        feats = feats.view(1, -1)
        with torch.no_grad():
            feats = model(feats)["last_hidden_state"]
            feats = (
                model.final_proj(feats[0]).unsqueeze(0) if version == "v1" else feats
            )
        feats = feats.squeeze(0).float().cpu().numpy()
        if not np.isnan(feats).any():
            np.save(out_file_path, feats, allow_pickle=False)
        else:
            print(f"{file} contains NaN values and will be skipped.")

    with tqdm.tqdm(total=len(files), leave=True, position=device_num) as pbar:
        with concurrent.futures.ThreadPoolExecutor(max_workers=n_threads) as executor:
            futures = [
                executor.submit(process_file_embedding_wrapper, file_info)
                for file_info in files
            ]
            for future in concurrent.futures.as_completed(futures):
                pbar.update(1)


def run_embedding_extraction(
    files, devices, version, embedder_model, embedder_model_custom, num_processes # Add num_processes here
):
    start_time = time.time()
    devices_str = ", ".join(devices)

    print(
        f"Starting embedding extraction with {num_processes} cores on {devices_str}..."
    )
    ps = []
    num_devices = len(devices)
    for i, device in enumerate(devices):
        p = mp.Process(
            target=process_file_embedding,
            args=(
                files[i::num_devices],
                version,
                embedder_model,
                embedder_model_custom,
                i,
                device,
                num_processes // num_devices,
            ),
        )
        ps.append(p)
        p.start()
    for i, device in enumerate(devices):
        ps[i].join()
    elapsed_time = time.time() - start_time
    print(f"Embedding extraction completed in {elapsed_time:.2f} seconds.")


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Extract features for RVC training.")
    parser.add_argument("exp_dir", type=str, help="Experiment directory (e.g., logs/my_model).")
    parser.add_argument("f0_method", type=str, choices=["crepe", "rmvpe", "fcnf0"], help="F0 extraction method.")
    parser.add_argument("hop_length", type=int, help="Hop length for F0 extraction.")
    parser.add_argument("num_processes", type=int, help="Number of parallel processes.")
    parser.add_argument("gpus", type=str, help="GPU IDs to use, separated by '-', or '-' for CPU.")
    parser.add_argument("version", type=str, choices=["v1", "v2"], help="RVC model version.")
    parser.add_argument("sample_rate", type=str, choices=["32000", "40000", "48000"], help="Target sample rate.")
    parser.add_argument("embedder_model", type=str, help="Pretrained embedder model name or 'custom'.")
    parser.add_argument("embedder_model_custom", type=str, nargs='?', default=None, help="Path to custom embedder model (if embedder_model is 'custom').")
    parser.add_argument("--val", action="store_true", help="Generate filelist for validation (skips adding mute files).")

    args = parser.parse_args()

    exp_dir = args.exp_dir
    f0_method = args.f0_method
    hop_length = args.hop_length
    num_processes = args.num_processes
    gpus = args.gpus
    version = args.version
    sample_rate = args.sample_rate
    embedder_model = args.embedder_model
    embedder_model_custom = args.embedder_model_custom
    is_validation = args.val


    wav_path = os.path.join(exp_dir, "sliced_audios_16k")
    os.makedirs(os.path.join(exp_dir, "f0"), exist_ok=True)
    os.makedirs(os.path.join(exp_dir, "f0_voiced"), exist_ok=True)
    os.makedirs(os.path.join(exp_dir, version + "_extracted"), exist_ok=True)

    chosen_embedder_model = (
        embedder_model_custom if embedder_model == "custom" else embedder_model
    )

    file_path = os.path.join(exp_dir, "model_info.json")
    if os.path.exists(file_path):
        with open(file_path, "r") as f:
            data = json.load(f)
    else:
        data = {}
    data.update(
        {
            "embedder_model": chosen_embedder_model,
        }
    )
    with open(file_path, "w") as f:
        json.dump(data, f, indent=4)

    files = []
    for file in glob.glob(os.path.join(wav_path, "*.wav")):
        file_name = os.path.basename(file)
        file_info = [
            file,  # full path to sliced 16k wav
            os.path.join(exp_dir, "f0", file_name + ".npy"),
            os.path.join(exp_dir, "f0_voiced", file_name + ".npy"),
            os.path.join(
                exp_dir, version + "_extracted", file_name.replace("wav", "npy")
            ),
        ]
        files.append(file_info)

    devices = ["cpu"] if gpus == "-" else [f"cuda:{idx}" for idx in gpus.split("-")]

    run_pitch_extraction(files, devices, f0_method, hop_length, num_processes)


    run_embedding_extraction(
        files, devices, version, embedder_model, embedder_model_custom, num_processes # Pass num_processes here
    )


    generate_config(version, sample_rate, exp_dir)
    generate_filelist(exp_dir, version, sample_rate, is_validation_set=is_validation)