File size: 6,832 Bytes
6f5bbf2
 
f82071f
6f5bbf2
 
 
f82071f
 
6f5bbf2
 
f82071f
5815543
 
f82071f
 
6f5bbf2
f82071f
6f5bbf2
 
f82071f
 
6f5bbf2
f82071f
 
6f5bbf2
f82071f
 
6f5bbf2
f82071f
 
5815543
f82071f
6f5bbf2
 
f82071f
 
 
 
6f5bbf2
f82071f
 
 
 
6f5bbf2
 
 
f82071f
6f5bbf2
 
f82071f
 
 
 
 
 
 
 
 
 
 
 
 
 
6f5bbf2
f82071f
 
 
 
 
 
 
 
 
 
 
 
6f5bbf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f82071f
5815543
 
 
 
 
 
 
 
 
 
6f5bbf2
5815543
6f5bbf2
5815543
6f5bbf2
 
 
 
 
 
5815543
6f5bbf2
 
f82071f
 
 
5815543
f82071f
 
 
6f5bbf2
 
 
 
5815543
6f5bbf2
 
 
 
f82071f
6f5bbf2
5815543
 
 
 
6f5bbf2
5815543
 
 
 
 
6f5bbf2
 
 
5815543
6f5bbf2
 
 
f82071f
 
5815543
f82071f
6f5bbf2
 
 
5815543
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
import argparse
import logging
import os
import random
from concurrent.futures import ProcessPoolExecutor
from glob import glob
from random import shuffle

import librosa
import numpy as np
import torch
import torch.multiprocessing as mp
from loguru import logger
from tqdm import tqdm

import diffusion.logger.utils as du
import utils
from diffusion.vocoder import Vocoder
from modules.mel_processing import spectrogram_torch

logging.getLogger("numba").setLevel(logging.WARNING)
logging.getLogger("matplotlib").setLevel(logging.WARNING)

hps = utils.get_hparams_from_file("configs/config.json")
dconfig = du.load_config("configs/diffusion.yaml")
sampling_rate = hps.data.sampling_rate
hop_length = hps.data.hop_length
speech_encoder = hps["model"]["speech_encoder"]


def process_one(filename, hmodel, f0p, device, diff=False, mel_extractor=None):
    wav, sr = librosa.load(filename, sr=sampling_rate)
    audio_norm = torch.FloatTensor(wav)
    audio_norm = audio_norm.unsqueeze(0)
    soft_path = filename + ".soft.pt"
    if not os.path.exists(soft_path):
        wav16k = librosa.resample(wav, orig_sr=sampling_rate, target_sr=16000)
        wav16k = torch.from_numpy(wav16k).to(device)
        c = hmodel.encoder(wav16k)
        torch.save(c.cpu(), soft_path)

    f0_path = filename + ".f0.npy"
    if not os.path.exists(f0_path):
        f0_predictor = utils.get_f0_predictor(f0p,sampling_rate=sampling_rate, hop_length=hop_length,device=None,threshold=0.05)
        f0,uv = f0_predictor.compute_f0_uv(
            wav
        )
        np.save(f0_path, np.asanyarray((f0,uv),dtype=object))


    spec_path = filename.replace(".wav", ".spec.pt")
    if not os.path.exists(spec_path):
        # Process spectrogram
        # The following code can't be replaced by torch.FloatTensor(wav)
        # because load_wav_to_torch return a tensor that need to be normalized

        if sr != hps.data.sampling_rate:
            raise ValueError(
                "{} SR doesn't match target {} SR".format(
                    sr, hps.data.sampling_rate
                )
            )

        #audio_norm = audio / hps.data.max_wav_value

        spec = spectrogram_torch(
            audio_norm,
            hps.data.filter_length,
            hps.data.sampling_rate,
            hps.data.hop_length,
            hps.data.win_length,
            center=False,
        )
        spec = torch.squeeze(spec, 0)
        torch.save(spec, spec_path)

    if diff or hps.model.vol_embedding:
        volume_path = filename + ".vol.npy"
        volume_extractor = utils.Volume_Extractor(hop_length)
        if not os.path.exists(volume_path):
            volume = volume_extractor.extract(audio_norm)
            np.save(volume_path, volume.to('cpu').numpy())

    if diff:
        mel_path = filename + ".mel.npy"
        if not os.path.exists(mel_path) and mel_extractor is not None:
            mel_t = mel_extractor.extract(audio_norm.to(device), sampling_rate)
            mel = mel_t.squeeze().to('cpu').numpy()
            np.save(mel_path, mel)
        aug_mel_path = filename + ".aug_mel.npy"
        aug_vol_path = filename + ".aug_vol.npy"
        max_amp = float(torch.max(torch.abs(audio_norm))) + 1e-5
        max_shift = min(1, np.log10(1/max_amp))
        log10_vol_shift = random.uniform(-1, max_shift)
        keyshift = random.uniform(-5, 5)
        if mel_extractor is not None:
            aug_mel_t = mel_extractor.extract(audio_norm * (10 ** log10_vol_shift), sampling_rate, keyshift = keyshift)
        aug_mel = aug_mel_t.squeeze().to('cpu').numpy()
        aug_vol = volume_extractor.extract(audio_norm * (10 ** log10_vol_shift))
        if not os.path.exists(aug_mel_path):
            np.save(aug_mel_path,np.asanyarray((aug_mel,keyshift),dtype=object))
        if not os.path.exists(aug_vol_path):
            np.save(aug_vol_path,aug_vol.to('cpu').numpy())


def process_batch(file_chunk, f0p, diff=False, mel_extractor=None, device="cpu"):
    logger.info("Loading speech encoder for content...")
    rank = mp.current_process()._identity
    rank = rank[0] if len(rank) > 0 else 0
    if torch.cuda.is_available():
        gpu_id = rank % torch.cuda.device_count()
        device = torch.device(f"cuda:{gpu_id}")
    logger.info(f"Rank {rank} uses device {device}")
    hmodel = utils.get_speech_encoder(speech_encoder, device=device)
    logger.info(f"Loaded speech encoder for rank {rank}")
    for filename in tqdm(file_chunk):
        process_one(filename, hmodel, f0p, device, diff, mel_extractor)

def parallel_process(filenames, num_processes, f0p, diff, mel_extractor, device):
    with ProcessPoolExecutor(max_workers=num_processes) as executor:
        tasks = []
        for i in range(num_processes):
            start = int(i * len(filenames) / num_processes)
            end = int((i + 1) * len(filenames) / num_processes)
            file_chunk = filenames[start:end]
            tasks.append(executor.submit(process_batch, file_chunk, f0p, diff, mel_extractor, device=device))
        for task in tqdm(tasks):
            task.result()

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('-d', '--device', type=str, default=None)
    parser.add_argument(
        "--in_dir", type=str, default="dataset/44k", help="path to input dir"
    )
    parser.add_argument(
        '--use_diff',action='store_true', help='Whether to use the diffusion model'
    )
    parser.add_argument(
        '--f0_predictor', type=str, default="dio", help='Select F0 predictor, can select crepe,pm,dio,harvest,rmvpe,fcpe|default: pm(note: crepe is original F0 using mean filter)'
    )
    parser.add_argument(
        '--num_processes', type=int, default=1, help='You are advised to set the number of processes to the same as the number of CPU cores'
    )
    args = parser.parse_args()
    f0p = args.f0_predictor
    device = args.device
    if device is None:
        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    print(speech_encoder)
    logger.info("Using device: ", device)
    logger.info("Using SpeechEncoder: " + speech_encoder)
    logger.info("Using extractor: " + f0p)
    logger.info("Using diff Mode: " + str( args.use_diff))

    if args.use_diff:
        print("use_diff")
        print("Loading Mel Extractor...")
        mel_extractor = Vocoder(dconfig.vocoder.type, dconfig.vocoder.ckpt, device=device)
        print("Loaded Mel Extractor.")
    else:
        mel_extractor = None
    filenames = glob(f"{args.in_dir}/*/*.wav", recursive=True)  # [:10]
    shuffle(filenames)
    mp.set_start_method("spawn", force=True)

    num_processes = args.num_processes
    if num_processes == 0:
        num_processes = os.cpu_count()

    parallel_process(filenames, num_processes, f0p, args.use_diff, mel_extractor, device)