File size: 8,921 Bytes
2f5f13b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
import os
import sys
import glob
import time
import tqdm
import torch
import torchcrepe
import numpy as np
import concurrent.futures
import multiprocessing as mp
import json

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:
    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, audio_array, method, hop_length):
        if method == "crepe":
            return self._get_crepe(audio_array, hop_length, type="full")
        elif method == "crepe-tiny":
            return self._get_crepe(audio_array, hop_length, type="tiny")
        elif method == "rmvpe":
            return self.model_rmvpe.infer_from_audio(audio_array, thred=0.03)

    def _get_crepe(self, x, hop_length, type):
        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,
            type,
            batch_size=hop_length * 2,
            device=audio.device,
            pad=True,
        )
        source = pitch.squeeze(0).cpu().float().numpy()
        source[source < 0.001] = np.nan
        return np.nan_to_num(
            np.interp(
                np.arange(0, len(source) * (x.size // self.hop), len(source))
                / (x.size // self.hop),
                np.arange(0, len(source)),
                source,
            )
        )

    def coarse_f0(self, f0):
        f0_mel = 1127.0 * np.log(1.0 + f0 / 700.0)
        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):
        inp_path, opt_path_coarse, opt_path_full, _ = file_info
        if os.path.exists(opt_path_coarse) and os.path.exists(opt_path_full):
            return

        try:
            np_arr = load_audio(inp_path, self.fs)
            feature_pit = self.compute_f0(np_arr, f0_method, hop_length)
            np.save(opt_path_full, feature_pit, allow_pickle=False)
            coarse_pit = self.coarse_f0(feature_pit)
            np.save(opt_path_coarse, 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, threads):
        self.device = device
        if f0_method == "rmvpe":
            self.model_rmvpe = RMVPE0Predictor(
                os.path.join("rvc", "models", "predictors", "rmvpe.pt"),
                device=device,
            )

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

        with tqdm.tqdm(total=len(files), leave=True) as pbar:
            with concurrent.futures.ThreadPoolExecutor(max_workers=threads) as executor:
                futures = [executor.submit(worker, f) for f in files]
                for _ in concurrent.futures.as_completed(futures):
                    pbar.update(1)


def run_pitch_extraction(files, devices, f0_method, hop_length, threads):
    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()
    with concurrent.futures.ProcessPoolExecutor(max_workers=len(devices)) as executor:
        tasks = [
            executor.submit(
                fe.process_files,
                files[i :: len(devices)],
                f0_method,
                hop_length,
                devices[i],
                threads // len(devices),
            )
            for i in range(len(devices))
        ]
        concurrent.futures.wait(tasks)

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


def process_file_embedding(

    files, embedder_model, embedder_model_custom, device_num, device, n_threads

):
    model = load_embedding(embedder_model, embedder_model_custom).to(device).float()
    model.eval()
    n_threads = max(1, n_threads)

    def worker(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(device).float()
        feats = feats.view(1, -1)
        with torch.no_grad():
            result = model(feats)["last_hidden_state"]
        feats_out = result.squeeze(0).float().cpu().numpy()
        if not np.isnan(feats_out).any():
            np.save(out_file_path, feats_out, allow_pickle=False)
        else:
            print(f"{wav_file_path} produced NaN values; skipping.")

    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(worker, f) for f in files]
            for _ in concurrent.futures.as_completed(futures):
                pbar.update(1)


def run_embedding_extraction(

    files, devices, embedder_model, embedder_model_custom, threads

):
    devices_str = ", ".join(devices)
    print(
        f"Starting embedding extraction with {num_processes} cores on {devices_str}..."
    )
    start_time = time.time()
    with concurrent.futures.ProcessPoolExecutor(max_workers=len(devices)) as executor:
        tasks = [
            executor.submit(
                process_file_embedding,
                files[i :: len(devices)],
                embedder_model,
                embedder_model_custom,
                i,
                devices[i],
                threads // len(devices),
            )
            for i in range(len(devices))
        ]
        concurrent.futures.wait(tasks)

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


if __name__ == "__main__":
    exp_dir = sys.argv[1]
    f0_method = sys.argv[2]
    hop_length = int(sys.argv[3])
    num_processes = int(sys.argv[4])
    gpus = sys.argv[5]
    sample_rate = sys.argv[6]
    embedder_model = sys.argv[7]
    embedder_model_custom = sys.argv[8] if len(sys.argv) > 8 else None
    include_mutes = int(sys.argv[9]) if len(sys.argv) > 9 else 2

    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, "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["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,
            os.path.join(exp_dir, "f0", file_name + ".npy"),
            os.path.join(exp_dir, "f0_voiced", file_name + ".npy"),
            os.path.join(exp_dir, "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, embedder_model, embedder_model_custom, num_processes
    )

    generate_config(sample_rate, exp_dir)
    generate_filelist(exp_dir, sample_rate, include_mutes)