|
| 1 | +import traceback |
| 2 | +import logging |
| 3 | + |
| 4 | +logger = logging.getLogger(__name__) |
| 5 | + |
| 6 | +import numpy as np |
| 7 | +import soundfile as sf |
| 8 | +import torch |
| 9 | +from io import BytesIO |
| 10 | + |
| 11 | +from infer.lib.audio import load_audio, wav2 |
| 12 | +from infer.lib.infer_pack.models import ( |
| 13 | + SynthesizerTrnMs256NSFsid, |
| 14 | + SynthesizerTrnMs256NSFsid_nono, |
| 15 | + SynthesizerTrnMs768NSFsid, |
| 16 | + SynthesizerTrnMs768NSFsid_nono, |
| 17 | +) |
| 18 | +from infer.modules.vc.pipeline import Pipeline |
| 19 | +from infer.modules.vc.utils import * |
| 20 | + |
| 21 | + |
| 22 | +class VC: |
| 23 | + def __init__(self, config): |
| 24 | + self.n_spk = None |
| 25 | + self.tgt_sr = None |
| 26 | + self.net_g = None |
| 27 | + self.pipeline = None |
| 28 | + self.cpt = None |
| 29 | + self.version = None |
| 30 | + self.if_f0 = None |
| 31 | + self.version = None |
| 32 | + self.hubert_model = None |
| 33 | + |
| 34 | + self.config = config |
| 35 | + |
| 36 | + def get_vc(self, sid, *to_return_protect): |
| 37 | + logger.info("Get sid: " + sid) |
| 38 | + |
| 39 | + to_return_protect0 = { |
| 40 | + "visible": self.if_f0 != 0, |
| 41 | + "value": to_return_protect[0] |
| 42 | + if self.if_f0 != 0 and to_return_protect |
| 43 | + else 0.5, |
| 44 | + "__type__": "update", |
| 45 | + } |
| 46 | + to_return_protect1 = { |
| 47 | + "visible": self.if_f0 != 0, |
| 48 | + "value": to_return_protect[1] |
| 49 | + if self.if_f0 != 0 and to_return_protect |
| 50 | + else 0.33, |
| 51 | + "__type__": "update", |
| 52 | + } |
| 53 | + |
| 54 | + if sid == "" or sid == []: |
| 55 | + if self.hubert_model is not None: # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的 |
| 56 | + logger.info("Clean model cache") |
| 57 | + del ( |
| 58 | + self.net_g, |
| 59 | + self.n_spk, |
| 60 | + self.vc, |
| 61 | + self.hubert_model, |
| 62 | + self.tgt_sr, |
| 63 | + ) # ,cpt |
| 64 | + self.hubert_model = ( |
| 65 | + self.net_g |
| 66 | + ) = self.n_spk = self.vc = self.hubert_model = self.tgt_sr = None |
| 67 | + if torch.cuda.is_available(): |
| 68 | + torch.cuda.empty_cache() |
| 69 | + ###楼下不这么折腾清理不干净 |
| 70 | + self.if_f0 = self.cpt.get("f0", 1) |
| 71 | + self.version = self.cpt.get("version", "v1") |
| 72 | + if self.version == "v1": |
| 73 | + if self.if_f0 == 1: |
| 74 | + self.net_g = SynthesizerTrnMs256NSFsid( |
| 75 | + *self.cpt["config"], is_half=self.config.is_half |
| 76 | + ) |
| 77 | + else: |
| 78 | + self.net_g = SynthesizerTrnMs256NSFsid_nono(*self.cpt["config"]) |
| 79 | + elif self.version == "v2": |
| 80 | + if self.if_f0 == 1: |
| 81 | + self.net_g = SynthesizerTrnMs768NSFsid( |
| 82 | + *self.cpt["config"], is_half=self.config.is_half |
| 83 | + ) |
| 84 | + else: |
| 85 | + self.net_g = SynthesizerTrnMs768NSFsid_nono(*self.cpt["config"]) |
| 86 | + del self.net_g, self.cpt |
| 87 | + if torch.cuda.is_available(): |
| 88 | + torch.cuda.empty_cache() |
| 89 | + return ( |
| 90 | + {"visible": False, "__type__": "update"}, |
| 91 | + { |
| 92 | + "visible": True, |
| 93 | + "value": to_return_protect0, |
| 94 | + "__type__": "update", |
| 95 | + }, |
| 96 | + { |
| 97 | + "visible": True, |
| 98 | + "value": to_return_protect1, |
| 99 | + "__type__": "update", |
| 100 | + }, |
| 101 | + "", |
| 102 | + "", |
| 103 | + ) |
| 104 | + person = f'{os.getenv("weight_root")}/{sid}' |
| 105 | + logger.info(f"Loading: {person}") |
| 106 | + |
| 107 | + self.cpt = torch.load(person, map_location="cpu") |
| 108 | + self.tgt_sr = self.cpt["config"][-1] |
| 109 | + self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0] # n_spk |
| 110 | + self.if_f0 = self.cpt.get("f0", 1) |
| 111 | + self.version = self.cpt.get("version", "v1") |
| 112 | + |
| 113 | + synthesizer_class = { |
| 114 | + ("v1", 1): SynthesizerTrnMs256NSFsid, |
| 115 | + ("v1", 0): SynthesizerTrnMs256NSFsid_nono, |
| 116 | + ("v2", 1): SynthesizerTrnMs768NSFsid, |
| 117 | + ("v2", 0): SynthesizerTrnMs768NSFsid_nono, |
| 118 | + } |
| 119 | + |
| 120 | + self.net_g = synthesizer_class.get( |
| 121 | + (self.version, self.if_f0), SynthesizerTrnMs256NSFsid |
| 122 | + )(*self.cpt["config"], is_half=self.config.is_half) |
| 123 | + |
| 124 | + del self.net_g.enc_q |
| 125 | + |
| 126 | + self.net_g.load_state_dict(self.cpt["weight"], strict=False) |
| 127 | + self.net_g.eval().to(self.config.device) |
| 128 | + if self.config.is_half: |
| 129 | + self.net_g = self.net_g.half() |
| 130 | + else: |
| 131 | + self.net_g = self.net_g.float() |
| 132 | + |
| 133 | + self.pipeline = Pipeline(self.tgt_sr, self.config) |
| 134 | + n_spk = self.cpt["config"][-3] |
| 135 | + index = {"value": get_index_path_from_model(sid), "__type__": "update"} |
| 136 | + logger.info("Select index: " + index["value"]) |
| 137 | + |
| 138 | + return ( |
| 139 | + ( |
| 140 | + {"visible": True, "maximum": n_spk, "__type__": "update"}, |
| 141 | + to_return_protect0, |
| 142 | + to_return_protect1, |
| 143 | + index, |
| 144 | + index, |
| 145 | + ) |
| 146 | + if to_return_protect |
| 147 | + else {"visible": True, "maximum": n_spk, "__type__": "update"} |
| 148 | + ) |
| 149 | + |
| 150 | + def vc_single( |
| 151 | + self, |
| 152 | + sid, |
| 153 | + input_audio_path, |
| 154 | + f0_up_key, |
| 155 | + f0_file, |
| 156 | + f0_method, |
| 157 | + file_index, |
| 158 | + file_index2, |
| 159 | + index_rate, |
| 160 | + filter_radius, |
| 161 | + resample_sr, |
| 162 | + rms_mix_rate, |
| 163 | + protect, |
| 164 | + ): |
| 165 | + if input_audio_path is None: |
| 166 | + return "You need to upload an audio", None |
| 167 | + f0_up_key = int(f0_up_key) |
| 168 | + try: |
| 169 | + audio = load_audio(input_audio_path, 16000) |
| 170 | + audio_max = np.abs(audio).max() / 0.95 |
| 171 | + if audio_max > 1: |
| 172 | + audio /= audio_max |
| 173 | + times = [0, 0, 0] |
| 174 | + |
| 175 | + if self.hubert_model is None: |
| 176 | + self.hubert_model = load_hubert(self.config) |
| 177 | + |
| 178 | + file_index = ( |
| 179 | + ( |
| 180 | + file_index.strip(" ") |
| 181 | + .strip('"') |
| 182 | + .strip("\n") |
| 183 | + .strip('"') |
| 184 | + .strip(" ") |
| 185 | + .replace("trained", "added") |
| 186 | + ) |
| 187 | + if file_index != "" |
| 188 | + else file_index2 |
| 189 | + ) # 防止小白写错，自动帮他替换掉 |
| 190 | + |
| 191 | + audio_opt = self.pipeline.pipeline( |
| 192 | + self.hubert_model, |
| 193 | + self.net_g, |
| 194 | + sid, |
| 195 | + audio, |
| 196 | + input_audio_path, |
| 197 | + times, |
| 198 | + f0_up_key, |
| 199 | + f0_method, |
| 200 | + file_index, |
| 201 | + index_rate, |
| 202 | + self.if_f0, |
| 203 | + filter_radius, |
| 204 | + self.tgt_sr, |
| 205 | + resample_sr, |
| 206 | + rms_mix_rate, |
| 207 | + self.version, |
| 208 | + protect, |
| 209 | + f0_file, |
| 210 | + ) |
| 211 | + if self.tgt_sr != resample_sr >= 16000: |
| 212 | + tgt_sr = resample_sr |
| 213 | + else: |
| 214 | + tgt_sr = self.tgt_sr |
| 215 | + index_info = ( |
| 216 | + "Index:\n%s." % file_index |
| 217 | + if os.path.exists(file_index) |
| 218 | + else "Index not used." |
| 219 | + ) |
| 220 | + return ( |
| 221 | + "Success.\n%s\nTime:\nnpy: %.2fs, f0: %.2fs, infer: %.2fs." |
| 222 | + % (index_info, *times), |
| 223 | + (tgt_sr, audio_opt), |
| 224 | + ) |
| 225 | + except: |
| 226 | + info = traceback.format_exc() |
| 227 | + logger.warning(info) |
| 228 | + return info, (None, None) |
| 229 | + |
| 230 | + def vc_multi( |
| 231 | + self, |
| 232 | + sid, |
| 233 | + dir_path, |
| 234 | + opt_root, |
| 235 | + paths, |
| 236 | + f0_up_key, |
| 237 | + f0_method, |
| 238 | + file_index, |
| 239 | + file_index2, |
| 240 | + index_rate, |
| 241 | + filter_radius, |
| 242 | + resample_sr, |
| 243 | + rms_mix_rate, |
| 244 | + protect, |
| 245 | + format1, |
| 246 | + ): |
| 247 | + try: |
| 248 | + dir_path = ( |
| 249 | + dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ") |
| 250 | + ) # 防止小白拷路径头尾带了空格和"和回车 |
| 251 | + opt_root = opt_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ") |
| 252 | + os.makedirs(opt_root, exist_ok=True) |
| 253 | + try: |
| 254 | + if dir_path != "": |
| 255 | + paths = [ |
| 256 | + os.path.join(dir_path, name) for name in os.listdir(dir_path) |
| 257 | + ] |
| 258 | + else: |
| 259 | + paths = [path.name for path in paths] |
| 260 | + except: |
| 261 | + traceback.print_exc() |
| 262 | + paths = [path.name for path in paths] |
| 263 | + infos = [] |
| 264 | + for path in paths: |
| 265 | + info, opt = self.vc_single( |
| 266 | + sid, |
| 267 | + path, |
| 268 | + f0_up_key, |
| 269 | + None, |
| 270 | + f0_method, |
| 271 | + file_index, |
| 272 | + file_index2, |
| 273 | + # file_big_npy, |
| 274 | + index_rate, |
| 275 | + filter_radius, |
| 276 | + resample_sr, |
| 277 | + rms_mix_rate, |
| 278 | + protect, |
| 279 | + ) |
| 280 | + if "Success" in info: |
| 281 | + try: |
| 282 | + tgt_sr, audio_opt = opt |
| 283 | + if format1 in ["wav", "flac"]: |
| 284 | + sf.write( |
| 285 | + "%s/%s.%s" |
| 286 | + % (opt_root, os.path.basename(path), format1), |
| 287 | + audio_opt, |
| 288 | + tgt_sr, |
| 289 | + ) |
| 290 | + else: |
| 291 | + path = "%s/%s.%s" % ( |
| 292 | + opt_root, |
| 293 | + os.path.basename(path), |
| 294 | + format1, |
| 295 | + ) |
| 296 | + with BytesIO() as wavf: |
| 297 | + sf.write(wavf, audio_opt, tgt_sr, format="wav") |
| 298 | + wavf.seek(0, 0) |
| 299 | + with open(path, "wb") as outf: |
| 300 | + wav2(wavf, outf, format1) |
| 301 | + except: |
| 302 | + info += traceback.format_exc() |
| 303 | + infos.append("%s->%s" % (os.path.basename(path), info)) |
| 304 | + yield "\n".join(infos) |
| 305 | + yield "\n".join(infos) |
| 306 | + except: |
| 307 | + yield traceback.format_exc() |
0 commit comments