Spaces:
Running
Running
Update src/main.py
Browse files- src/main.py +22 -18
src/main.py
CHANGED
|
@@ -4,42 +4,40 @@ import os
|
|
| 4 |
import shlex
|
| 5 |
import subprocess
|
| 6 |
import librosa
|
|
|
|
| 7 |
import numpy as np
|
| 8 |
import soundfile as sf
|
| 9 |
import gradio as gr
|
| 10 |
from rvc import Config, load_hubert, get_vc, rvc_infer
|
| 11 |
|
| 12 |
BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
| 13 |
-
|
| 14 |
-
|
| 15 |
|
| 16 |
def get_rvc_model(voice_model):
|
| 17 |
-
model_dir = os.path.join(
|
| 18 |
rvc_model_path = next((os.path.join(model_dir, f) for f in os.listdir(model_dir) if f.endswith('.pth')), None)
|
| 19 |
rvc_index_path = next((os.path.join(model_dir, f) for f in os.listdir(model_dir) if f.endswith('.index')), None)
|
| 20 |
|
| 21 |
if rvc_model_path is None:
|
| 22 |
-
|
| 23 |
-
raise Exception(error_msg)
|
| 24 |
|
| 25 |
return rvc_model_path, rvc_index_path
|
| 26 |
|
| 27 |
def convert_to_stereo(audio_path):
|
| 28 |
wave, sr = librosa.load(audio_path, mono=False, sr=44100)
|
| 29 |
if type(wave[0]) != np.ndarray:
|
| 30 |
-
stereo_path =
|
| 31 |
command = shlex.split(f'ffmpeg -y -loglevel error -i "{audio_path}" -ac 2 -f wav "{stereo_path}"')
|
| 32 |
subprocess.run(command)
|
| 33 |
return stereo_path
|
| 34 |
-
|
| 35 |
-
return audio_path
|
| 36 |
|
| 37 |
def get_hash(filepath):
|
|
|
|
| 38 |
with open(filepath, 'rb') as f:
|
| 39 |
-
file_hash = hashlib.blake2b()
|
| 40 |
while chunk := f.read(8192):
|
| 41 |
file_hash.update(chunk)
|
| 42 |
-
|
| 43 |
return file_hash.hexdigest()[:11]
|
| 44 |
|
| 45 |
def display_progress(percent, message, progress=gr.Progress()):
|
|
@@ -47,30 +45,36 @@ def display_progress(percent, message, progress=gr.Progress()):
|
|
| 47 |
|
| 48 |
def voice_change(voice_model, vocals_path, output_path, pitch_change, f0_method, index_rate, filter_radius, rms_mix_rate, protect, crepe_hop_length):
|
| 49 |
rvc_model_path, rvc_index_path = get_rvc_model(voice_model)
|
| 50 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
config = Config(device, True)
|
| 52 |
-
hubert_model = load_hubert(device, config.is_half, os.path.join(
|
| 53 |
cpt, version, net_g, tgt_sr, vc = get_vc(device, config.is_half, config, rvc_model_path)
|
| 54 |
|
| 55 |
rvc_infer(rvc_index_path, index_rate, vocals_path, output_path, pitch_change, f0_method, cpt, version, net_g,
|
| 56 |
filter_radius, tgt_sr, rms_mix_rate, protect, crepe_hop_length, vc, hubert_model)
|
| 57 |
-
|
|
|
|
| 58 |
gc.collect()
|
|
|
|
| 59 |
|
| 60 |
def song_cover_pipeline(uploaded_file, voice_model, pitch_change, index_rate=0.5, filter_radius=3, rms_mix_rate=0.25, f0_method='rmvpe',
|
| 61 |
crepe_hop_length=128, protect=0.33, output_format='mp3', progress=gr.Progress()):
|
| 62 |
|
| 63 |
if not uploaded_file or not voice_model:
|
| 64 |
-
raise
|
| 65 |
|
| 66 |
display_progress(0, '[~] Запуск конвейера генерации AI-кавера...', progress)
|
| 67 |
|
| 68 |
if not os.path.exists(uploaded_file):
|
| 69 |
-
|
| 70 |
-
raise Exception(error_msg)
|
| 71 |
|
| 72 |
song_id = get_hash(uploaded_file)
|
| 73 |
-
song_dir = os.path.join(
|
| 74 |
os.makedirs(song_dir, exist_ok=True)
|
| 75 |
|
| 76 |
orig_song_path = convert_to_stereo(uploaded_file)
|
|
@@ -83,4 +87,4 @@ def song_cover_pipeline(uploaded_file, voice_model, pitch_change, index_rate=0.5
|
|
| 83 |
voice_change(voice_model, orig_song_path, ai_cover_path, pitch_change, f0_method, index_rate,
|
| 84 |
filter_radius, rms_mix_rate, protect, crepe_hop_length)
|
| 85 |
|
| 86 |
-
return ai_cover_path
|
|
|
|
| 4 |
import shlex
|
| 5 |
import subprocess
|
| 6 |
import librosa
|
| 7 |
+
import torch
|
| 8 |
import numpy as np
|
| 9 |
import soundfile as sf
|
| 10 |
import gradio as gr
|
| 11 |
from rvc import Config, load_hubert, get_vc, rvc_infer
|
| 12 |
|
| 13 |
BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
| 14 |
+
RVC_MODELS_DIR = os.path.join(BASE_DIR, 'rvc_models')
|
| 15 |
+
OUTPUT_DIR = os.path.join(BASE_DIR, 'song_output')
|
| 16 |
|
| 17 |
def get_rvc_model(voice_model):
|
| 18 |
+
model_dir = os.path.join(RVC_MODELS_DIR, voice_model)
|
| 19 |
rvc_model_path = next((os.path.join(model_dir, f) for f in os.listdir(model_dir) if f.endswith('.pth')), None)
|
| 20 |
rvc_index_path = next((os.path.join(model_dir, f) for f in os.listdir(model_dir) if f.endswith('.index')), None)
|
| 21 |
|
| 22 |
if rvc_model_path is None:
|
| 23 |
+
raise FileNotFoundError(f'В каталоге {model_dir} отсутствует файл модели.')
|
|
|
|
| 24 |
|
| 25 |
return rvc_model_path, rvc_index_path
|
| 26 |
|
| 27 |
def convert_to_stereo(audio_path):
|
| 28 |
wave, sr = librosa.load(audio_path, mono=False, sr=44100)
|
| 29 |
if type(wave[0]) != np.ndarray:
|
| 30 |
+
stereo_path = 'Voice_stereo.wav'
|
| 31 |
command = shlex.split(f'ffmpeg -y -loglevel error -i "{audio_path}" -ac 2 -f wav "{stereo_path}"')
|
| 32 |
subprocess.run(command)
|
| 33 |
return stereo_path
|
| 34 |
+
return audio_path
|
|
|
|
| 35 |
|
| 36 |
def get_hash(filepath):
|
| 37 |
+
file_hash = hashlib.blake2b()
|
| 38 |
with open(filepath, 'rb') as f:
|
|
|
|
| 39 |
while chunk := f.read(8192):
|
| 40 |
file_hash.update(chunk)
|
|
|
|
| 41 |
return file_hash.hexdigest()[:11]
|
| 42 |
|
| 43 |
def display_progress(percent, message, progress=gr.Progress()):
|
|
|
|
| 45 |
|
| 46 |
def voice_change(voice_model, vocals_path, output_path, pitch_change, f0_method, index_rate, filter_radius, rms_mix_rate, protect, crepe_hop_length):
|
| 47 |
rvc_model_path, rvc_index_path = get_rvc_model(voice_model)
|
| 48 |
+
|
| 49 |
+
if torch.cuda.is_available():
|
| 50 |
+
device = 'cuda:0'
|
| 51 |
+
else:
|
| 52 |
+
device = 'cpu'
|
| 53 |
+
|
| 54 |
config = Config(device, True)
|
| 55 |
+
hubert_model = load_hubert(device, config.is_half, os.path.join(RVC_MODELS_DIR, 'hubert_base.pt'))
|
| 56 |
cpt, version, net_g, tgt_sr, vc = get_vc(device, config.is_half, config, rvc_model_path)
|
| 57 |
|
| 58 |
rvc_infer(rvc_index_path, index_rate, vocals_path, output_path, pitch_change, f0_method, cpt, version, net_g,
|
| 59 |
filter_radius, tgt_sr, rms_mix_rate, protect, crepe_hop_length, vc, hubert_model)
|
| 60 |
+
|
| 61 |
+
del hubert_model, cpt, net_g, vc
|
| 62 |
gc.collect()
|
| 63 |
+
torch.cuda.empty_cache()
|
| 64 |
|
| 65 |
def song_cover_pipeline(uploaded_file, voice_model, pitch_change, index_rate=0.5, filter_radius=3, rms_mix_rate=0.25, f0_method='rmvpe',
|
| 66 |
crepe_hop_length=128, protect=0.33, output_format='mp3', progress=gr.Progress()):
|
| 67 |
|
| 68 |
if not uploaded_file or not voice_model:
|
| 69 |
+
raise ValueError('Убедитесь, что поле ввода песни и поле модели голоса заполнены.')
|
| 70 |
|
| 71 |
display_progress(0, '[~] Запуск конвейера генерации AI-кавера...', progress)
|
| 72 |
|
| 73 |
if not os.path.exists(uploaded_file):
|
| 74 |
+
raise FileNotFoundError(f'{uploaded_file} не существует.')
|
|
|
|
| 75 |
|
| 76 |
song_id = get_hash(uploaded_file)
|
| 77 |
+
song_dir = os.path.join(OUTPUT_DIR, song_id)
|
| 78 |
os.makedirs(song_dir, exist_ok=True)
|
| 79 |
|
| 80 |
orig_song_path = convert_to_stereo(uploaded_file)
|
|
|
|
| 87 |
voice_change(voice_model, orig_song_path, ai_cover_path, pitch_change, f0_method, index_rate,
|
| 88 |
filter_radius, rms_mix_rate, protect, crepe_hop_length)
|
| 89 |
|
| 90 |
+
return ai_cover_path
|