Cool_Tools/Retrieval-based-Voice-Conversion-WebUI
1
0
Fork 0
mirror of synced 2024-11-15 03:07:40 +01:00
Code Releases Wiki Activity

Update rvc_for_realtime.py

Browse Source
This commit is contained in:
RVC-Boss 2023-07-11 22:09:41 +08:00 committed by GitHub
parent 0f9d2e6cac
commit cd1d76aac2
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 269 additions and 269 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

538
rvc_for_realtime.py
View File

@ -1,269 +1,269 @@
import faiss,torch,traceback,parselmouth,numpy as np,torchcrepe,torch.nn as nn,pyworld import faiss,torch,traceback,parselmouth,numpy as np,torchcrepe,torch.nn as nn,pyworld
from fairseq import checkpoint_utils from fairseq import checkpoint_utils
from lib.infer_pack.models import ( from lib.infer_pack.models import (
SynthesizerTrnMs256NSFsid, SynthesizerTrnMs256NSFsid,
SynthesizerTrnMs256NSFsid_nono, SynthesizerTrnMs256NSFsid_nono,
SynthesizerTrnMs768NSFsid, SynthesizerTrnMs768NSFsid,
SynthesizerTrnMs768NSFsid_nono, SynthesizerTrnMs768NSFsid_nono,
) )
import os,sys import os,sys
from time import time as ttime from time import time as ttime
import torch.nn.functional as F import torch.nn.functional as F
import scipy.signal as signal import scipy.signal as signal
now_dir = os.getcwd() now_dir = os.getcwd()
sys.path.append(now_dir) sys.path.append(now_dir)
from config import Config from config import Config
from multiprocessing import Manager as M from multiprocessing import Manager as M
mm = M() mm = M()
config = Config() config = Config()
class RVC: class RVC:
def __init__( def __init__(
self, key, pth_path, index_path, index_rate, n_cpu,inp_q,opt_q,device self, key, pth_path, index_path, index_rate, n_cpu,inp_q,opt_q,device
) -> None: ) -> None:
""" """
初始化 初始化
""" """
try: try:
global config global config
self.inp_q=inp_q self.inp_q=inp_q
self.opt_q=opt_q self.opt_q=opt_q
self.device=device self.device=device
self.f0_up_key = key self.f0_up_key = key
self.time_step = 160 / 16000 * 1000 self.time_step = 160 / 16000 * 1000
self.f0_min = 50 self.f0_min = 50
self.f0_max = 1100 self.f0_max = 1100
self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700) 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.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
self.sr = 16000 self.sr = 16000
self.window = 160 self.window = 160
self.n_cpu = n_cpu self.n_cpu = n_cpu
if index_rate != 0: if index_rate != 0:
self.index = faiss.read_index(index_path) self.index = faiss.read_index(index_path)
self.big_npy = self.index.reconstruct_n(0, self.index.ntotal) self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
print("index search enabled") print("index search enabled")
self.index_rate = index_rate self.index_rate = index_rate
models, _, _ = checkpoint_utils.load_model_ensemble_and_task( models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
["hubert_base.pt"], ["hubert_base.pt"],
suffix="", suffix="",
) )
hubert_model = models[0] hubert_model = models[0]
hubert_model = hubert_model.to(config.device) hubert_model = hubert_model.to(config.device)
if config.is_half: if config.is_half:
hubert_model = hubert_model.half() hubert_model = hubert_model.half()
else: else:
hubert_model = hubert_model.float() hubert_model = hubert_model.float()
hubert_model.eval() hubert_model.eval()
self.model = hubert_model self.model = hubert_model
cpt = torch.load(pth_path, map_location="cpu") cpt = torch.load(pth_path, map_location="cpu")
self.tgt_sr = cpt["config"][-1] self.tgt_sr = cpt["config"][-1]
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
self.if_f0 = cpt.get("f0", 1) self.if_f0 = cpt.get("f0", 1)
self.version = cpt.get("version", "v1") self.version = cpt.get("version", "v1")
if self.version == "v1": if self.version == "v1":
if self.if_f0 == 1: if self.if_f0 == 1:
self.net_g = SynthesizerTrnMs256NSFsid( self.net_g = SynthesizerTrnMs256NSFsid(
*cpt["config"], is_half=config.is_half *cpt["config"], is_half=config.is_half
) )
else: else:
self.net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"]) self.net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
elif self.version == "v2": elif self.version == "v2":
if self.if_f0 == 1: if self.if_f0 == 1:
self.net_g = SynthesizerTrnMs768NSFsid( self.net_g = SynthesizerTrnMs768NSFsid(
*cpt["config"], is_half=config.is_half *cpt["config"], is_half=config.is_half
) )
else: else:
self.net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"]) self.net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
del self.net_g.enc_q del self.net_g.enc_q
print(self.net_g.load_state_dict(cpt["weight"], strict=False)) print(self.net_g.load_state_dict(cpt["weight"], strict=False))
self.net_g.eval().to(device) self.net_g.eval().to(device)
if config.is_half: if config.is_half:
self.net_g = self.net_g.half() self.net_g = self.net_g.half()
else: else:
self.net_g = self.net_g.float() self.net_g = self.net_g.float()
self.is_half=config.is_half self.is_half=config.is_half
except: except:
print(traceback.format_exc()) print(traceback.format_exc())
def get_f0_post(self, f0): def get_f0_post(self, f0):
f0_min = self.f0_min f0_min = self.f0_min
f0_max = self.f0_max f0_max = self.f0_max
f0_mel_min = 1127 * np.log(1 + f0_min / 700) f0_mel_min = 1127 * np.log(1 + f0_min / 700)
f0_mel_max = 1127 * np.log(1 + f0_max / 700) f0_mel_max = 1127 * np.log(1 + f0_max / 700)
f0bak = f0.copy() f0bak = f0.copy()
f0_mel = 1127 * np.log(1 + f0 / 700) f0_mel = 1127 * np.log(1 + f0 / 700)
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / ( f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
f0_mel_max - f0_mel_min f0_mel_max - f0_mel_min
) + 1 ) + 1
f0_mel[f0_mel <= 1] = 1 f0_mel[f0_mel <= 1] = 1
f0_mel[f0_mel > 255] = 255 f0_mel[f0_mel > 255] = 255
f0_coarse = np.rint(f0_mel).astype(np.int) f0_coarse = np.rint(f0_mel).astype(np.int)
return f0_coarse, f0bak return f0_coarse, f0bak
def get_f0(self, x, f0_up_key, n_cpu, method="harvest"): def get_f0(self, x, f0_up_key, n_cpu, method="harvest"):
n_cpu = int(n_cpu) n_cpu = int(n_cpu)
if (method == "crepe"): return self.get_f0_crepe(x, f0_up_key) if (method == "crepe"): return self.get_f0_crepe(x, f0_up_key)
if (method == "rmvpe"): return self.get_f0_rmvpe(x, f0_up_key) if (method == "rmvpe"): return self.get_f0_rmvpe(x, f0_up_key)
if (method == "pm"): if (method == "pm"):
p_len = x.shape[0] // 160 p_len = x.shape[0] // 160
f0 = ( f0 = (
parselmouth.Sound(x, 16000) parselmouth.Sound(x, 16000)
.to_pitch_ac( .to_pitch_ac(
time_step=0.01, time_step=0.01,
voicing_threshold=0.6, voicing_threshold=0.6,
pitch_floor=50, pitch_floor=50,
pitch_ceiling=1100, pitch_ceiling=1100,
) )
.selected_array["frequency"] .selected_array["frequency"]
) )
pad_size = (p_len - len(f0) + 1) // 2 pad_size = (p_len - len(f0) + 1) // 2
if pad_size > 0 or p_len - len(f0) - pad_size > 0: if pad_size > 0 or p_len - len(f0) - pad_size > 0:
print(pad_size, p_len - len(f0) - pad_size) print(pad_size, p_len - len(f0) - pad_size)
f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant") f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
f0 *= pow(2, f0_up_key / 12) f0 *= pow(2, f0_up_key / 12)
return self.get_f0_post(f0) return self.get_f0_post(f0)
if (n_cpu == 1): if (n_cpu == 1):
f0, t = pyworld.harvest( f0, t = pyworld.harvest(
x.astype(np.double), x.astype(np.double),
fs=16000, fs=16000,
f0_ceil=1100, f0_ceil=1100,
f0_floor=50, f0_floor=50,
frame_period=10, frame_period=10,
) )
f0 = signal.medfilt(f0, 3) f0 = signal.medfilt(f0, 3)
f0 *= pow(2, f0_up_key / 12) f0 *= pow(2, f0_up_key / 12)
return self.get_f0_post(f0) return self.get_f0_post(f0)
f0bak = np.zeros(x.shape[0] // 160, dtype=np.float64) f0bak = np.zeros(x.shape[0] // 160, dtype=np.float64)
length = len(x) length = len(x)
part_length = int(length / n_cpu / 160) * 160 part_length = int(length / n_cpu / 160) * 160
ts = ttime() ts = ttime()
res_f0 = mm.dict() res_f0 = mm.dict()
for idx in range(n_cpu): for idx in range(n_cpu):
tail = part_length * (idx + 1) + 320 tail = part_length * (idx + 1) + 320
if (idx == 0): if (idx == 0):
self.inp_q.put((idx, x[:tail], res_f0, n_cpu, ts)) self.inp_q.put((idx, x[:tail], res_f0, n_cpu, ts))
else: else:
self.inp_q.put((idx, x[part_length * idx - 320:tail], res_f0, n_cpu, ts)) self.inp_q.put((idx, x[part_length * idx - 320:tail], res_f0, n_cpu, ts))
while (1): while (1):
res_ts = self.opt_q.get() res_ts = self.opt_q.get()
if (res_ts == ts): if (res_ts == ts):
break break
f0s = [i[1] for i in sorted(res_f0.items(), key=lambda x: x[0])] f0s = [i[1] for i in sorted(res_f0.items(), key=lambda x: x[0])]
for idx, f0 in enumerate(f0s): for idx, f0 in enumerate(f0s):
if (idx == 0): if (idx == 0):
f0 = f0[:-3] f0 = f0[:-3]
elif (idx != n_cpu - 1): elif (idx != n_cpu - 1):
f0 = f0[2:-3] f0 = f0[2:-3]
else: else:
f0 = f0[2:-1] f0 = f0[2:-1]
f0bak[part_length * idx // 160:part_length * idx // 160 + f0.shape[0]] = f0 f0bak[part_length * idx // 160:part_length * idx // 160 + f0.shape[0]] = f0
f0bak = signal.medfilt(f0bak, 3) f0bak = signal.medfilt(f0bak, 3)
f0bak *= pow(2, f0_up_key / 12) f0bak *= pow(2, f0_up_key / 12)
return self.get_f0_post(f0bak) return self.get_f0_post(f0bak)
def get_f0_crepe(self, x, f0_up_key): def get_f0_crepe(self, x, f0_up_key):
audio = torch.tensor(np.copy(x))[None].float() audio = torch.tensor(np.copy(x))[None].float()
f0, pd = torchcrepe.predict( f0, pd = torchcrepe.predict(
audio, audio,
self.sr, self.sr,
160, 160,
self.f0_min, self.f0_min,
self.f0_max, self.f0_max,
"full", "full",
batch_size=512, batch_size=512,
device=self.device, device=self.device,
return_periodicity=True, return_periodicity=True,
) )
pd = torchcrepe.filter.median(pd, 3) pd = torchcrepe.filter.median(pd, 3)
f0 = torchcrepe.filter.mean(f0, 3) f0 = torchcrepe.filter.mean(f0, 3)
f0[pd < 0.1] = 0 f0[pd < 0.1] = 0
f0 = f0[0].cpu().numpy() f0 = f0[0].cpu().numpy()
f0 *= pow(2, f0_up_key / 12) f0 *= pow(2, f0_up_key / 12)
return self.get_f0_post(f0) return self.get_f0_post(f0)
def get_f0_rmvpe(self, x, f0_up_key): def get_f0_rmvpe(self, x, f0_up_key):
if (hasattr(self, "model_rmvpe") == False): if (hasattr(self, "model_rmvpe") == False):
from rmvpe import RMVPE from rmvpe import RMVPE
print("loading rmvpe model") print("loading rmvpe model")
# self.model_rmvpe = RMVPE("rmvpe.pt", is_half=self.is_half, device=self.device) self.model_rmvpe = RMVPE("rmvpe.pt", is_half=self.is_half, device=self.device)
self.model_rmvpe = RMVPE("aug2_58000_half.pt", is_half=self.is_half, device=self.device) # self.model_rmvpe = RMVPE("aug2_58000_half.pt", is_half=self.is_half, device=self.device)
f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03) f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
f0 *= pow(2, f0_up_key / 12) f0 *= pow(2, f0_up_key / 12)
return self.get_f0_post(f0) return self.get_f0_post(f0)
def infer(self, feats: torch.Tensor, indata: np.ndarray, rate1, rate2, cache_pitch, cache_pitchf, f0method) -> np.ndarray: def infer(self, feats: torch.Tensor, indata: np.ndarray, rate1, rate2, cache_pitch, cache_pitchf, f0method) -> np.ndarray:
feats = feats.view(1, -1) feats = feats.view(1, -1)
if config.is_half: if config.is_half:
feats = feats.half() feats = feats.half()
else: else:
feats = feats.float() feats = feats.float()
feats = feats.to(self.device) feats = feats.to(self.device)
t1 = ttime() t1 = ttime()
with torch.no_grad(): with torch.no_grad():
padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False) padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
inputs = { inputs = {
"source": feats, "source": feats,
"padding_mask": padding_mask, "padding_mask": padding_mask,
"output_layer": 9 if self.version == "v1" else 12, "output_layer": 9 if self.version == "v1" else 12,
} }
logits = self.model.extract_features(**inputs) logits = self.model.extract_features(**inputs)
feats = self.model.final_proj(logits[0]) if self.version == "v1" else logits[0] feats = self.model.final_proj(logits[0]) if self.version == "v1" else logits[0]
t2 = ttime() t2 = ttime()
try: try:
if ( if (
hasattr(self, "index") hasattr(self, "index")
and self.index_rate != 0 and self.index_rate != 0
): ):
leng_replace_head = int(rate1 * feats[0].shape[0]) leng_replace_head = int(rate1 * feats[0].shape[0])
npy = feats[0][-leng_replace_head:].cpu().numpy().astype("float32") npy = feats[0][-leng_replace_head:].cpu().numpy().astype("float32")
score, ix = self.index.search(npy, k=8) score, ix = self.index.search(npy, k=8)
weight = np.square(1 / score) weight = np.square(1 / score)
weight /= weight.sum(axis=1, keepdims=True) weight /= weight.sum(axis=1, keepdims=True)
npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1) npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
if config.is_half: if config.is_half:
npy = npy.astype("float16") npy = npy.astype("float16")
feats[0][-leng_replace_head:] = ( feats[0][-leng_replace_head:] = (
torch.from_numpy(npy).unsqueeze(0).to(self.device) * self.index_rate torch.from_numpy(npy).unsqueeze(0).to(self.device) * self.index_rate
+ (1 - self.index_rate) * feats[0][-leng_replace_head:] + (1 - self.index_rate) * feats[0][-leng_replace_head:]
) )
else: else:
print("index search FAIL or disabled") print("index search FAIL or disabled")
except: except:
traceback.print_exc() traceback.print_exc()
print("index search FAIL") print("index search FAIL")
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1) feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
t3 = ttime() t3 = ttime()
if self.if_f0 == 1: if self.if_f0 == 1:
pitch, pitchf = self.get_f0(indata, self.f0_up_key, self.n_cpu, f0method) pitch, pitchf = self.get_f0(indata, self.f0_up_key, self.n_cpu, f0method)
cache_pitch[:] = np.append(cache_pitch[pitch[:-1].shape[0]:], pitch[:-1]) cache_pitch[:] = np.append(cache_pitch[pitch[:-1].shape[0]:], pitch[:-1])
cache_pitchf[:] = np.append(cache_pitchf[pitchf[:-1].shape[0]:], pitchf[:-1]) cache_pitchf[:] = np.append(cache_pitchf[pitchf[:-1].shape[0]:], pitchf[:-1])
p_len = min(feats.shape[1], 13000, cache_pitch.shape[0]) p_len = min(feats.shape[1], 13000, cache_pitch.shape[0])
else: else:
cache_pitch, cache_pitchf = None, None cache_pitch, cache_pitchf = None, None
p_len = min(feats.shape[1], 13000) p_len = min(feats.shape[1], 13000)
t4 = ttime() t4 = ttime()
feats = feats[:, :p_len, :] feats = feats[:, :p_len, :]
if self.if_f0 == 1: if self.if_f0 == 1:
cache_pitch = cache_pitch[:p_len] cache_pitch = cache_pitch[:p_len]
cache_pitchf = cache_pitchf[:p_len] cache_pitchf = cache_pitchf[:p_len]
cache_pitch = torch.LongTensor(cache_pitch).unsqueeze(0).to(self.device) cache_pitch = torch.LongTensor(cache_pitch).unsqueeze(0).to(self.device)
cache_pitchf = torch.FloatTensor(cache_pitchf).unsqueeze(0).to(self.device) cache_pitchf = torch.FloatTensor(cache_pitchf).unsqueeze(0).to(self.device)
p_len = torch.LongTensor([p_len]).to(self.device) p_len = torch.LongTensor([p_len]).to(self.device)
ii = 0 # sid ii = 0 # sid
sid = torch.LongTensor([ii]).to(self.device) sid = torch.LongTensor([ii]).to(self.device)
with torch.no_grad(): with torch.no_grad():
if self.if_f0 == 1: if self.if_f0 == 1:
infered_audio = ( infered_audio = (
self.net_g.infer(feats, p_len, cache_pitch, cache_pitchf, sid, rate2)[0][0, 0] self.net_g.infer(feats, p_len, cache_pitch, cache_pitchf, sid, rate2)[0][0, 0]
.data.cpu() .data.cpu()
.float() .float()
) )
else: else:
infered_audio = ( infered_audio = (
self.net_g.infer(feats, p_len, sid, rate2)[0][0, 0].data.cpu().float() self.net_g.infer(feats, p_len, sid, rate2)[0][0, 0].data.cpu().float()
) )
t5 = ttime() t5 = ttime()
print("time->fea-index-f0-model:", t2 - t1, t3 - t2, t4 - t3, t5 - t4) print("time->fea-index-f0-model:", t2 - t1, t3 - t2, t4 - t3, t5 - t4)
return infered_audio return infered_audio