Cool_Tools/Retrieval-based-Voice-Conversion-WebUI
1
0
Fork 0
mirror of synced 2024-11-23 23:21:03 +01:00
Code Releases Wiki Activity

Reformat and rewrite _get_name_params (#57)

Browse Source 
* Reformat

* rewrite _get_name_params

* Add workflow for automatic formatting

* Revert "Add workflow for automatic formatting"

This reverts commit 9111c5dbc1.

* revert Retrieval_based_Voice_Conversion_WebUI.ipynb

---------

Co-authored-by: 源文雨 <41315874+fumiama@users.noreply.github.com>
This commit is contained in:
Ftps 2023-04-15 20:44:24 +09:00 committed by GitHub
parent aaa893c4b1
commit c8261b2ccc
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
45 changed files with 4878 additions and 2456 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
config.py
View File

@ -16,12 +16,17 @@ n_cpu = 0
########################命令行参数########################
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--port", type=int, default=7865, help="Listen port")
parser.add_argument("--pycmd", type=str, default="python", help="Python command")
parser.add_argument("--colab", action='store_true', help="Launch in colab")
parser.add_argument("--noparallel", action='store_true', help="Disable parallel processing")
parser.add_argument("--noautoopen", action='store_true', help="Do not open in browser automatically")
parser.add_argument("--colab", action="store_true", help="Launch in colab")
parser.add_argument(
"--noparallel", action="store_true", help="Disable parallel processing"
)
parser.add_argument(
"--noautoopen", action="store_true", help="Do not open in browser automatically"
)
cmd_opts = parser.parse_args()
python_cmd = cmd_opts.pycmd
@ -34,13 +39,15 @@ noautoopen=cmd_opts.noautoopen
import sys
import torch
# has_mps is only available in nightly pytorch (for now) and MasOS 12.3+.
# check `getattr` and try it for compatibility
def has_mps() -> bool:
if sys.platform != "darwin":
return False
else:
if not getattr(torch, 'has_mps', False): return False
if not getattr(torch, "has_mps", False):
return False
try:
torch.zeros(1).to(torch.device("mps"))
return True
@ -48,7 +55,7 @@ def has_mps() -> bool:
return False
if(not torch.cuda.is_available()):
if not torch.cuda.is_available():
if has_mps():
print("没有发现支持的N卡, 使用MPS进行推理")
device = "mps"
@ -57,15 +64,17 @@ if(not torch.cuda.is_available()):
device = "cpu"
is_half = False
if(device not in ["cpu", "mps"]):
if device not in ["cpu", "mps"]:
gpu_name = torch.cuda.get_device_name(int(device.split(":")[-1]))
if("16" in gpu_name or "MX" in gpu_name):
if "16" in gpu_name or "MX" in gpu_name:
print("16系显卡/MX系显卡强制单精度")
is_half = False
from multiprocessing import cpu_count
if(n_cpu==0): n_cpu=cpu_count()
if(is_half):
if n_cpu == 0:
n_cpu = cpu_count()
if is_half:
# 6G显存配置
x_pad = 3
x_query = 10

13
export_onnx.py
View File

@ -5,7 +5,6 @@ person = "Shiroha/shiroha.pth"
exported_path = "model.onnx"
cpt = torch.load(person, map_location="cpu")
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
print(*cpt["config"])
@ -19,16 +18,19 @@ test_pitchf = torch.rand(1, 200)
test_ds = torch.LongTensor([0])
test_rnd = torch.rand(1, 192, 200)
input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
output_names = ["audio", ]
output_names = [
"audio",
]
device = "cpu"
torch.onnx.export(net_g,
torch.onnx.export(
net_g,
(
test_phone.to(device),
test_phone_lengths.to(device),
test_pitch.to(device),
test_pitchf.to(device),
test_ds.to(device),
test_rnd.to(device)
test_rnd.to(device),
),
exported_path,
dynamic_axes={
@ -41,4 +43,5 @@ torch.onnx.export(net_g,
opset_version=16,
verbose=False,
input_names=input_names,
output_names=output_names)
output_names=output_names,
)

70
extract_f0_print.py
View File

@ -3,19 +3,24 @@ import librosa
import pyworld
from scipy.io import wavfile
import numpy as np, logging
logging.getLogger('numba').setLevel(logging.WARNING)
logging.getLogger("numba").setLevel(logging.WARNING)
from multiprocessing import Process
exp_dir = sys.argv[1]
f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
def printt(strr):
print(strr)
f.write("%s\n" % strr)
f.flush()
n_p = int(sys.argv[2])
f0method = sys.argv[3]
class FeatureInput(object):
def __init__(self, samplerate=16000, hop_size=160):
self.fs = samplerate
@ -31,17 +36,26 @@ class FeatureInput(object):
x, sr = librosa.load(path, self.fs)
p_len = x.shape[0] // self.hop
assert sr == self.fs
if(f0_method=="pm"):
if f0_method == "pm":
time_step = 160 / 16000 * 1000
f0_min = 50
f0_max = 1100
f0 = parselmouth.Sound(x, sr).to_pitch_ac(
time_step=time_step / 1000, voicing_threshold=0.6,
pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
f0 = (
parselmouth.Sound(x, sr)
.to_pitch_ac(
time_step=time_step / 1000,
voicing_threshold=0.6,
pitch_floor=f0_min,
pitch_ceiling=f0_max,
)
.selected_array["frequency"]
)
pad_size = (p_len - len(f0) + 1) // 2
if(pad_size>0 or p_len - len(f0) - pad_size>0):
f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
elif(f0_method=="harvest"):
if pad_size > 0 or p_len - len(f0) - pad_size > 0:
f0 = np.pad(
f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
)
elif f0_method == "harvest":
f0, t = pyworld.harvest(
x.astype(np.double),
fs=sr,
@ -50,7 +64,7 @@ class FeatureInput(object):
frame_period=1000 * self.hop / sr,
)
f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
elif(f0_method=="dio"):
elif f0_method == "dio":
f0, t = pyworld.dio(
x.astype(np.double),
fs=sr,
@ -78,22 +92,37 @@ class FeatureInput(object):
return f0_coarse
def go(self, paths, f0_method):
if (len(paths) == 0): printt("no-f0-todo")
if len(paths) == 0:
printt("no-f0-todo")
else:
printt("todo-f0-%s" % len(paths))
n = max(len(paths) // 5, 1) # 每个进程最多打印5条
for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
try:
if(idx%n==0):printt("f0ing,now-%s,all-%s,-%s"%(idx,len(paths),inp_path))
if(os.path.exists(opt_path1+".npy")==True and os.path.exists(opt_path2+".npy")==True):continue
if idx % n == 0:
printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
if (
os.path.exists(opt_path1 + ".npy") == True
and os.path.exists(opt_path2 + ".npy") == True
):
continue
featur_pit = self.compute_f0(inp_path, f0_method)
np.save(opt_path2,featur_pit,allow_pickle=False,)#nsf
np.save(
opt_path2,
featur_pit,
allow_pickle=False,
) # nsf
coarse_pit = self.coarse_f0(featur_pit)
np.save(opt_path1,coarse_pit,allow_pickle=False,)#ori
np.save(
opt_path1,
coarse_pit,
allow_pickle=False,
) # ori
except:
printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
if __name__=='__main__':
if __name__ == "__main__":
# exp_dir=r"E:\codes\py39\dataset\mi-test"
# n_p=16
# f = open("%s/log_extract_f0.log"%exp_dir, "w")
@ -108,14 +137,21 @@ if __name__=='__main__':
os.makedirs(opt_root2, exist_ok=True)
for name in sorted(list(os.listdir(inp_root))):
inp_path = "%s/%s" % (inp_root, name)
if ("spec" in inp_path): continue
if "spec" in inp_path:
continue
opt_path1 = "%s/%s" % (opt_root1, name)
opt_path2 = "%s/%s" % (opt_root2, name)
paths.append([inp_path, opt_path1, opt_path2])
ps = []
for i in range(n_p):
p=Process(target=featureInput.go,args=(paths[i::n_p],f0method,))
p = Process(
target=featureInput.go,
args=(
paths[i::n_p],
f0method,
),
)
p.start()
ps.append(p)
for p in ps:

28
extract_feature_print.py
View File

@ -1,4 +1,5 @@
import os, sys, traceback
# device=sys.argv[1]
n_part = int(sys.argv[2])
i_part = int(sys.argv[3])
@ -14,13 +15,18 @@ import torch.nn.functional as F
import soundfile as sf
import numpy as np
from fairseq import checkpoint_utils
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
def printt(strr):
print(strr)
f.write("%s\n" % strr)
f.flush()
printt(sys.argv)
model_path = "hubert_base.pt"
@ -28,6 +34,8 @@ printt(exp_dir)
wavPath = "%s/1_16k_wavs" % exp_dir
outPath = "%s/3_feature256" % exp_dir
os.makedirs(outPath, exist_ok=True)
# wave must be 16k, hop_size=320
def readwave(wav_path, normalize=False):
wav, sr = sf.read(wav_path)
@ -41,6 +49,8 @@ def readwave(wav_path, normalize=False):
feats = F.layer_norm(feats, feats.shape)
feats = feats.view(1, -1)
return feats
# HuBERT model
printt("load model(s) from {}".format(model_path))
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
@ -50,12 +60,14 @@ models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
model = models[0]
model = model.to(device)
printt("move model to %s" % device)
if device != "cpu": model = model.half()
if device != "cpu":
model = model.half()
model.eval()
todo = sorted(list(os.listdir(wavPath)))[i_part::n_part]
n = max(1, len(todo) // 10) # 最多打印十条
if(len(todo)==0):printt("no-feature-todo")
if len(todo) == 0:
printt("no-feature-todo")
else:
printt("all-feature-%s" % len(todo))
for idx, file in enumerate(todo):
@ -64,12 +76,15 @@ else:
wav_path = "%s/%s" % (wavPath, file)
out_path = "%s/%s" % (outPath, file.replace("wav", "npy"))
if(os.path.exists(out_path)):continue
if os.path.exists(out_path):
continue
feats = readwave(wav_path, normalize=saved_cfg.task.normalize)
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
inputs = {
"source": feats.half().to(device) if device != "cpu" else feats.to(device),
"source": feats.half().to(device)
if device != "cpu"
else feats.to(device),
"padding_mask": padding_mask.to(device),
"output_layer": 9, # layer 9
}
@ -78,11 +93,12 @@ else:
feats = model.final_proj(logits[0])
feats = feats.squeeze(0).float().cpu().numpy()
if(np.isnan(feats).sum()==0):
if np.isnan(feats).sum() == 0:
np.save(out_path, feats, allow_pickle=False)
else:
printt("%s-contains nan" % file)
if (idx % n == 0):printt("now-%s,all-%s,%s,%s"%(len(todo),idx,file,feats.shape))
if idx % n == 0:
printt("now-%s,all-%s,%s,%s" % (len(todo), idx, file, feats.shape))
except:
printt(traceback.format_exc())
printt("all-feature-done")

10
extract_locale.py
View File

@ -7,9 +7,10 @@ pattern = r"""i18n\((["'][^"']+["'])\)"""
# Initialize the dictionary to store key-value pairs
data = {}
def process(fn: str):
global data
with open(fn, 'r', encoding='utf-8') as f:
with open(fn, "r", encoding="utf-8") as f:
contents = f.read()
matches = re.findall(pattern, contents)
for key in matches:
@ -17,12 +18,13 @@ def process(fn: str):
print("extract:", key)
data[key] = key
print("processing infer-web.py")
process('infer-web.py')
process("infer-web.py")
print("processing gui.py")
process('gui.py')
process("gui.py")
# Save as a JSON file
with open('./locale/zh_CN.json', 'w', encoding='utf-8') as f:
with open("./locale/zh_CN.json", "w", encoding="utf-8") as f:
json.dump(data, f, ensure_ascii=False, indent=4)

404
gui.py
View File

@ -10,15 +10,19 @@ import torchaudio.transforms as tat
# import matplotlib.pyplot as plt
from infer_pack.models import SynthesizerTrnMs256NSFsid, SynthesizerTrnMs256NSFsid_nono
from webui_locale import I18nAuto
i18n = I18nAuto()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class RVC:
def __init__(self,key,hubert_path,pth_path,index_path,npy_path,index_rate) -> None:
'''
def __init__(
self, key, hubert_path, pth_path, index_path, npy_path, index_rate
) -> None:
"""
初始化
'''
"""
self.f0_up_key = key
self.time_step = 160 / 16000 * 1000
self.f0_min = 50
@ -27,7 +31,7 @@ class RVC:
self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
self.index = faiss.read_index(index_path)
self.index_rate = index_rate
'''NOT YET USED'''
"""NOT YET USED"""
self.big_npy = np.load(npy_path)
model_path = hubert_path
print("load model(s) from {}".format(model_path))
@ -43,7 +47,7 @@ class RVC:
tgt_sr = cpt["config"][-1]
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
if_f0 = cpt.get("f0", 1)
if(if_f0==1):
if if_f0 == 1:
self.net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=True)
else:
self.net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
@ -52,10 +56,11 @@ class RVC:
self.net_g.eval().to(device)
self.net_g.half()
def get_f0_coarse(self, f0):
f0_mel = 1127 * np.log(1 + f0 / 700)
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (self.f0_mel_max - self.f0_mel_min) + 1
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (
self.f0_mel_max - self.f0_mel_min
) + 1
f0_mel[f0_mel <= 1] = 1
f0_mel[f0_mel > 255] = 255
# f0_mel[f0_mel > 188] = 188
@ -63,24 +68,30 @@ class RVC:
return f0_coarse
def get_f0(self, x, p_len, f0_up_key=0):
f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
time_step=self.time_step / 1000, voicing_threshold=0.6,
pitch_floor=self.f0_min, pitch_ceiling=self.f0_max).selected_array['frequency']
f0 = (
parselmouth.Sound(x, 16000)
.to_pitch_ac(
time_step=self.time_step / 1000,
voicing_threshold=0.6,
pitch_floor=self.f0_min,
pitch_ceiling=self.f0_max,
)
.selected_array["frequency"]
)
pad_size = (p_len - len(f0) + 1) // 2
if(pad_size>0 or p_len - len(f0) - pad_size>0):
f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
if pad_size > 0 or p_len - len(f0) - pad_size > 0:
f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
f0 *= pow(2, f0_up_key / 12)
# f0=suofang(f0)
f0bak = f0.copy()
f0_coarse = self.get_f0_coarse(f0)
return f0_coarse, f0bak
def infer(self, feats: torch.Tensor) -> np.ndarray:
'''
"""
推理函数
'''
"""
audio = feats.clone().cpu().numpy()
assert feats.dim() == 1, feats.dim()
feats = feats.view(1, -1)
@ -96,11 +107,18 @@ class RVC:
feats = self.model.final_proj(logits[0])
####索引优化
if(isinstance(self.index,type(None))==False and isinstance(self.big_npy,type(None))==False and self.index_rate!=0):
if (
isinstance(self.index, type(None)) == False
and isinstance(self.big_npy, type(None)) == False
and self.index_rate != 0
):
npy = feats[0].cpu().numpy().astype("float32")
_, I = self.index.search(npy, 1)
npy = self.big_npy[I.squeeze()].astype("float16")
feats = torch.from_numpy(npy).unsqueeze(0).to(device)*self.index_rate + (1-self.index_rate)*feats
feats = (
torch.from_numpy(npy).unsqueeze(0).to(device) * self.index_rate
+ (1 - self.index_rate) * feats
)
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
torch.cuda.synchronize()
@ -120,17 +138,21 @@ class RVC:
ii = 0 # sid
sid = torch.LongTensor([ii]).to(device)
with torch.no_grad():
infered_audio = self.net_g.infer(feats, p_len,pitch,pitchf,sid)[0][0, 0].data.cpu().float()#nsf
infered_audio = (
self.net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]
.data.cpu()
.float()
) # nsf
torch.cuda.synchronize()
return infered_audio
class Config:
def __init__(self) -> None:
self.hubert_path:str=''
self.pth_path:str=''
self.index_path:str=''
self.npy_path:str=''
self.hubert_path: str = ""
self.pth_path: str = ""
self.index_path: str = ""
self.npy_path: str = ""
self.pitch: int = 12
self.samplerate: int = 44100
self.block_time: float = 1.0 # s
@ -142,6 +164,7 @@ class Config:
self.O_noise_reduce = False
self.index_rate = 0.3
class GUI:
def __init__(self) -> None:
self.config = Config()
@ -150,37 +173,145 @@ class GUI:
self.launcher()
def launcher(self):
sg.theme('LightBlue3')
sg.theme("LightBlue3")
input_devices, output_devices, _, _ = self.get_devices()
layout = [
[
sg.Frame(title=i18n('加载模型'),layout=[
[sg.Input(default_text='TEMP\\hubert_base.pt',key='hubert_path'),sg.FileBrowse(i18n('Hubert模型'))],
[sg.Input(default_text='TEMP\\atri.pth',key='pth_path'),sg.FileBrowse(i18n('选择.pth文件'))],
[sg.Input(default_text='TEMP\\added_IVF512_Flat_atri_baseline_src_feat.index',key='index_path'),sg.FileBrowse(i18n('选择.index文件'))],
[sg.Input(default_text='TEMP\\big_src_feature_atri.npy',key='npy_path'),sg.FileBrowse(i18n('选择.npy文件'))]
])
sg.Frame(
title=i18n("加载模型"),
layout=[
[
sg.Input(
default_text="TEMP\\hubert_base.pt", key="hubert_path"
),
sg.FileBrowse(i18n("Hubert模型")),
],
[
sg.Frame(layout=[
[sg.Text(i18n("输入设备")),sg.Combo(input_devices,key='sg_input_device',default_value=input_devices[sd.default.device[0]])],
[sg.Text(i18n("输出设备")),sg.Combo(output_devices,key='sg_output_device',default_value=output_devices[sd.default.device[1]])]
],title=i18n("音频设备(请使用同种类驱动)"))
sg.Input(default_text="TEMP\\atri.pth", key="pth_path"),
sg.FileBrowse(i18n("选择.pth文件")),
],
[
sg.Frame(layout=[
[sg.Text(i18n("响应阈值")),sg.Slider(range=(-60,0),key='threhold',resolution=1,orientation='h',default_value=-30)],
[sg.Text(i18n("音调设置")),sg.Slider(range=(-24,24),key='pitch',resolution=1,orientation='h',default_value=12)],
[sg.Text(i18n('Index Rate')),sg.Slider(range=(0.0,1.0),key='index_rate',resolution=0.01,orientation='h',default_value=0.5)]
],title=i18n("常规设置")),
sg.Frame(layout=[
[sg.Text(i18n("采样长度")),sg.Slider(range=(0.1,3.0),key='block_time',resolution=0.1,orientation='h',default_value=1.0)],
[sg.Text(i18n("淡入淡出长度")),sg.Slider(range=(0.01,0.15),key='crossfade_length',resolution=0.01,orientation='h',default_value=0.08)],
[sg.Text(i18n("额外推理时长")),sg.Slider(range=(0.05,3.00),key='extra_time',resolution=0.01,orientation='h',default_value=0.05)],
[sg.Checkbox(i18n('输入降噪'),key='I_noise_reduce'),sg.Checkbox(i18n('输出降噪'),key='O_noise_reduce')]
],title=i18n("性能设置"))
sg.Input(
default_text="TEMP\\added_IVF512_Flat_atri_baseline_src_feat.index",
key="index_path",
),
sg.FileBrowse(i18n("选择.index文件")),
],
[
sg.Input(
default_text="TEMP\\big_src_feature_atri.npy",
key="npy_path",
),
sg.FileBrowse(i18n("选择.npy文件")),
],
],
)
],
[
sg.Frame(
layout=[
[
sg.Text(i18n("输入设备")),
sg.Combo(
input_devices,
key="sg_input_device",
default_value=input_devices[sd.default.device[0]],
),
],
[
sg.Text(i18n("输出设备")),
sg.Combo(
output_devices,
key="sg_output_device",
default_value=output_devices[sd.default.device[1]],
),
],
],
title=i18n("音频设备(请使用同种类驱动)"),
)
],
[
sg.Frame(
layout=[
[
sg.Text(i18n("响应阈值")),
sg.Slider(
range=(-60, 0),
key="threhold",
resolution=1,
orientation="h",
default_value=-30,
),
],
[
sg.Text(i18n("音调设置")),
sg.Slider(
range=(-24, 24),
key="pitch",
resolution=1,
orientation="h",
default_value=12,
),
],
[
sg.Text(i18n("Index Rate")),
sg.Slider(
range=(0.0, 1.0),
key="index_rate",
resolution=0.01,
orientation="h",
default_value=0.5,
),
],
],
title=i18n("常规设置"),
),
sg.Frame(
layout=[
[
sg.Text(i18n("采样长度")),
sg.Slider(
range=(0.1, 3.0),
key="block_time",
resolution=0.1,
orientation="h",
default_value=1.0,
),
],
[
sg.Text(i18n("淡入淡出长度")),
sg.Slider(
range=(0.01, 0.15),
key="crossfade_length",
resolution=0.01,
orientation="h",
default_value=0.08,
),
],
[
sg.Text(i18n("额外推理时长")),
sg.Slider(
range=(0.05, 3.00),
key="extra_time",
resolution=0.01,
orientation="h",
default_value=0.05,
),
],
[
sg.Checkbox(i18n("输入降噪"), key="I_noise_reduce"),
sg.Checkbox(i18n("输出降噪"), key="O_noise_reduce"),
],
],
title=i18n("性能设置"),
),
],
[
sg.Button(i18n("开始音频转换"), key="start_vc"),
sg.Button(i18n("停止音频转换"), key="stop_vc"),
sg.Text(i18n("推理时间(ms):")),
sg.Text("0", key="infer_time"),
],
[sg.Button(i18n("开始音频转换"),key='start_vc'),sg.Button(i18n("停止音频转换"),key='stop_vc'),sg.Text(i18n("推理时间(ms):")),sg.Text("0",key='infer_time')]
]
self.window = sg.Window("RVC - GUI", layout=layout)
@ -192,29 +323,28 @@ class GUI:
if event == sg.WINDOW_CLOSED:
self.flag_vc = False
exit()
if event == 'start_vc' and self.flag_vc==False:
if event == "start_vc" and self.flag_vc == False:
self.set_values(values)
print(str(self.config.__dict__))
print('using_cuda:'+str(torch.cuda.is_available()))
print("using_cuda:" + str(torch.cuda.is_available()))
self.start_vc()
if event=='stop_vc'and self.flag_vc==True:
if event == "stop_vc" and self.flag_vc == True:
self.flag_vc = False
def set_values(self, values):
self.set_devices(values["sg_input_device"],values['sg_output_device'])
self.config.hubert_path=values['hubert_path']
self.config.pth_path=values['pth_path']
self.config.index_path=values['index_path']
self.config.npy_path=values['npy_path']
self.config.threhold=values['threhold']
self.config.pitch=values['pitch']
self.config.block_time=values['block_time']
self.config.crossfade_time=values['crossfade_length']
self.config.extra_time=values['extra_time']
self.config.I_noise_reduce=values['I_noise_reduce']
self.config.O_noise_reduce=values['O_noise_reduce']
self.config.index_rate=values['index_rate']
self.set_devices(values["sg_input_device"], values["sg_output_device"])
self.config.hubert_path = values["hubert_path"]
self.config.pth_path = values["pth_path"]
self.config.index_path = values["index_path"]
self.config.npy_path = values["npy_path"]
self.config.threhold = values["threhold"]
self.config.pitch = values["pitch"]
self.config.block_time = values["block_time"]
self.config.crossfade_time = values["crossfade_length"]
self.config.extra_time = values["extra_time"]
self.config.I_noise_reduce = values["I_noise_reduce"]
self.config.O_noise_reduce = values["O_noise_reduce"]
self.config.index_rate = values["index_rate"]
def start_vc(self):
torch.cuda.empty_cache()
@ -223,44 +353,76 @@ class GUI:
self.crossfade_frame = int(self.config.crossfade_time * self.config.samplerate)
self.sola_search_frame = int(0.012 * self.config.samplerate)
self.delay_frame = int(0.02 * self.config.samplerate) # 往前预留0.02s
self.extra_frame=int(self.config.extra_time*self.config.samplerate)#往后预留0.04s
self.extra_frame = int(
self.config.extra_time * self.config.samplerate
) # 往后预留0.04s
self.rvc = None
self.rvc=RVC(self.config.pitch,self.config.hubert_path,self.config.pth_path,self.config.index_path,self.config.npy_path,self.config.index_rate)
self.input_wav:np.ndarray=np.zeros(self.extra_frame+self.crossfade_frame+self.sola_search_frame+self.block_frame,dtype='float32')
self.output_wav:torch.Tensor=torch.zeros(self.block_frame,device=device,dtype=torch.float32)
self.sola_buffer:torch.Tensor=torch.zeros(self.crossfade_frame,device=device,dtype=torch.float32)
self.fade_in_window:torch.Tensor=torch.linspace(0.0,1.0,steps=self.crossfade_frame,device=device,dtype=torch.float32)
self.rvc = RVC(
self.config.pitch,
self.config.hubert_path,
self.config.pth_path,
self.config.index_path,
self.config.npy_path,
self.config.index_rate,
)
self.input_wav: np.ndarray = np.zeros(
self.extra_frame
+ self.crossfade_frame
+ self.sola_search_frame
+ self.block_frame,
dtype="float32",
)
self.output_wav: torch.Tensor = torch.zeros(
self.block_frame, device=device, dtype=torch.float32
)
self.sola_buffer: torch.Tensor = torch.zeros(
self.crossfade_frame, device=device, dtype=torch.float32
)
self.fade_in_window: torch.Tensor = torch.linspace(
0.0, 1.0, steps=self.crossfade_frame, device=device, dtype=torch.float32
)
self.fade_out_window: torch.Tensor = 1 - self.fade_in_window
self.resampler1=tat.Resample(orig_freq=self.config.samplerate,new_freq=16000,dtype=torch.float32)
self.resampler2=tat.Resample(orig_freq=40000,new_freq=self.config.samplerate,dtype=torch.float32)
self.resampler1 = tat.Resample(
orig_freq=self.config.samplerate, new_freq=16000, dtype=torch.float32
)
self.resampler2 = tat.Resample(
orig_freq=40000, new_freq=self.config.samplerate, dtype=torch.float32
)
thread_vc = threading.Thread(target=self.soundinput)
thread_vc.start()
def soundinput(self):
'''
"""
接受音频输入
'''
with sd.Stream(callback=self.audio_callback, blocksize=self.block_frame,samplerate=self.config.samplerate,dtype='float32'):
"""
with sd.Stream(
callback=self.audio_callback,
blocksize=self.block_frame,
samplerate=self.config.samplerate,
dtype="float32",
):
while self.flag_vc:
time.sleep(self.config.block_time)
print('Audio block passed.')
print('ENDing VC')
print("Audio block passed.")
print("ENDing VC")
def audio_callback(self,indata:np.ndarray,outdata:np.ndarray, frames, times, status):
'''
def audio_callback(
self, indata: np.ndarray, outdata: np.ndarray, frames, times, status
):
"""
音频处理
'''
"""
start_time = time.perf_counter()
indata = librosa.to_mono(indata.T)
if self.config.I_noise_reduce:
indata[:] = nr.reduce_noise(y=indata, sr=self.config.samplerate)
'''noise gate'''
"""noise gate"""
frame_length = 2048
hop_length = 1024
rms=librosa.feature.rms(y=indata,frame_length=frame_length,hop_length=hop_length)
rms = librosa.feature.rms(
y=indata, frame_length=frame_length, hop_length=hop_length
)
db_threhold = librosa.amplitude_to_db(rms, ref=1.0)[0] < self.config.threhold
# print(rms.shape,db.shape,db)
for i in range(db_threhold.shape[0]):
@ -269,36 +431,65 @@ class GUI:
self.input_wav[:] = np.append(self.input_wav[self.block_frame :], indata)
# infer
print('input_wav:'+str(self.input_wav.shape))
print("input_wav:" + str(self.input_wav.shape))
# print('infered_wav:'+str(infer_wav.shape))
infer_wav:torch.Tensor=self.resampler2(self.rvc.infer(self.resampler1(torch.from_numpy(self.input_wav))))[-self.crossfade_frame-self.sola_search_frame-self.block_frame:].to(device)
print('infer_wav:'+str(infer_wav.shape))
infer_wav: torch.Tensor = self.resampler2(
self.rvc.infer(self.resampler1(torch.from_numpy(self.input_wav)))
)[-self.crossfade_frame - self.sola_search_frame - self.block_frame :].to(
device
)
print("infer_wav:" + str(infer_wav.shape))
# SOLA algorithm from https://github.com/yxlllc/DDSP-SVC
cor_nom=F.conv1d(infer_wav[None,None,:self.crossfade_frame + self.sola_search_frame],self.sola_buffer[None,None,:])
cor_den=torch.sqrt(F.conv1d(infer_wav[None,None,:self.crossfade_frame + self.sola_search_frame]**2,torch.ones(1, 1,self.crossfade_frame,device=device))+1e-8)
cor_nom = F.conv1d(
infer_wav[None, None, : self.crossfade_frame + self.sola_search_frame],
self.sola_buffer[None, None, :],
)
cor_den = torch.sqrt(
F.conv1d(
infer_wav[None, None, : self.crossfade_frame + self.sola_search_frame]
** 2,
torch.ones(1, 1, self.crossfade_frame, device=device),
)
+ 1e-8
)
sola_offset = torch.argmax(cor_nom[0, 0] / cor_den[0, 0])
print('sola offset: ' + str(int(sola_offset)))
print("sola offset: " + str(int(sola_offset)))
# crossfade
self.output_wav[:] = infer_wav[sola_offset : sola_offset + self.block_frame]
self.output_wav[: self.crossfade_frame] *= self.fade_in_window
self.output_wav[: self.crossfade_frame] += self.sola_buffer[:]
if sola_offset < self.sola_search_frame:
self.sola_buffer[:] = infer_wav[-self.sola_search_frame - self.crossfade_frame + sola_offset: -self.sola_search_frame + sola_offset]* self.fade_out_window
self.sola_buffer[:] = (
infer_wav[
-self.sola_search_frame
- self.crossfade_frame
+ sola_offset : -self.sola_search_frame
+ sola_offset
]
* self.fade_out_window
)
else:
self.sola_buffer[:] = infer_wav[- self.crossfade_frame :]* self.fade_out_window
self.sola_buffer[:] = (
infer_wav[-self.crossfade_frame :] * self.fade_out_window
)
if self.config.O_noise_reduce:
outdata[:]=np.tile(nr.reduce_noise(y=self.output_wav[:].cpu().numpy(),sr=self.config.samplerate),(2,1)).T
outdata[:] = np.tile(
nr.reduce_noise(
y=self.output_wav[:].cpu().numpy(), sr=self.config.samplerate
),
(2, 1),
).T
else:
outdata[:] = self.output_wav[:].repeat(2, 1).t().cpu().numpy()
total_time = time.perf_counter() - start_time
print('infer time:'+str(total_time))
self.window['infer_time'].update(int(total_time*1000))
print("infer time:" + str(total_time))
self.window["infer_time"].update(int(total_time * 1000))
def get_devices(self, update: bool = True):
'''获取设备列表'''
"""获取设备列表"""
if update:
sd._terminate()
sd._initialize()
@ -317,18 +508,33 @@ class GUI:
for d in devices
if d["max_output_channels"] > 0
]
input_devices_indices = [d["index"] for d in devices if d["max_input_channels"] > 0]
input_devices_indices = [
d["index"] for d in devices if d["max_input_channels"] > 0
]
output_devices_indices = [
d["index"] for d in devices if d["max_output_channels"] > 0
]
return input_devices, output_devices, input_devices_indices, output_devices_indices
return (
input_devices,
output_devices,
input_devices_indices,
output_devices_indices,
)
def set_devices(self, input_device, output_device):
'''设置输出设备'''
input_devices,output_devices,input_device_indices, output_device_indices=self.get_devices()
"""设置输出设备"""
(
input_devices,
output_devices,
input_device_indices,
output_device_indices,
) = self.get_devices()
sd.default.device[0] = input_device_indices[input_devices.index(input_device)]
sd.default.device[1]=output_device_indices[output_devices.index(output_device)]
sd.default.device[1] = output_device_indices[
output_devices.index(output_device)
]
print("input device:" + str(sd.default.device[0]) + ":" + str(input_device))
print("output device:" + str(sd.default.device[1]) + ":" + str(output_device))
gui = GUI()

1105
infer-web.py
View File

File diff suppressed because it is too large Load Diff

96
infer/infer-pm-index256.py
View File

@ -1,14 +1,19 @@
'''
"""
对源特征进行检索
'''
"""
import torch, pdb, os, parselmouth
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import numpy as np
import soundfile as sf
# from models import SynthesizerTrn256#hifigan_nonsf
# from infer_pack.models import SynthesizerTrn256NSF as SynthesizerTrn256#hifigan_nsf
from infer_pack.models import SynthesizerTrnMs256NSFsid as SynthesizerTrn256#hifigan_nsf
from infer_pack.models import (
SynthesizerTrnMs256NSFsid as SynthesizerTrn256,
) # hifigan_nsf
# from infer_pack.models import SynthesizerTrnMs256NSFsid_sim as SynthesizerTrn256#hifigan_nsf
# from models import SynthesizerTrn256NSFsim as SynthesizerTrn256#hifigan_nsf
# from models import SynthesizerTrn256NSFsimFlow as SynthesizerTrn256#hifigan_nsf
@ -16,10 +21,12 @@ from infer_pack.models import SynthesizerTrnMs256NSFsid as SynthesizerTrn256#hif
from scipy.io import wavfile
from fairseq import checkpoint_utils
# import pyworld
import librosa
import torch.nn.functional as F
import scipy.signal as signal
# import torchcrepe
from time import time as ttime
@ -37,7 +44,26 @@ model.eval()
# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],183,256,is_half=True)#hifigan#512#256
# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],109,256,is_half=True)#hifigan#512#256
net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],183,256,is_half=True)#hifigan#512#256#no_dropout
net_g = SynthesizerTrn256(
1025,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 10, 2, 2],
512,
[16, 16, 4, 4],
183,
256,
is_half=True,
) # hifigan#512#256#no_dropout
# net_g = SynthesizerTrn256(1025,32,192,192,768,2,3,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],0)#ts3
# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2],512,[16,16,4],0)#hifigan-ps-sr
#
@ -53,37 +79,53 @@ print(net_g.load_state_dict(weights,strict=True))
net_g.eval().to(device)
net_g.half()
def get_f0(x, p_len,f0_up_key=0):
def get_f0(x, p_len, f0_up_key=0):
time_step = 160 / 16000 * 1000
f0_min = 50
f0_max = 1100
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
time_step=time_step / 1000, voicing_threshold=0.6,
pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
f0 = (
parselmouth.Sound(x, 16000)
.to_pitch_ac(
time_step=time_step / 1000,
voicing_threshold=0.6,
pitch_floor=f0_min,
pitch_ceiling=f0_max,
)
.selected_array["frequency"]
)
pad_size = (p_len - len(f0) + 1) // 2
if(pad_size>0 or p_len - len(f0) - pad_size>0):
f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
if pad_size > 0 or p_len - len(f0) - pad_size > 0:
f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
f0 *= pow(2, f0_up_key / 12)
f0bak = f0.copy()
f0_mel = 1127 * np.log(1 + f0 / 700)
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
f0_mel_max - f0_mel_min
) + 1
f0_mel[f0_mel <= 1] = 1
f0_mel[f0_mel > 255] = 255
# f0_mel[f0_mel > 188] = 188
f0_coarse = np.rint(f0_mel).astype(np.int)
return f0_coarse, f0bak
import faiss
index = faiss.read_index("infer/added_IVF512_Flat_mi_baseline_src_feat.index")
big_npy = np.load("infer/big_src_feature_mi.npy")
ta0 = ta1 = ta2 = 0
for idx,name in enumerate(["冬之花clip1.wav",]):##
for idx, name in enumerate(
[
"冬之花clip1.wav",
]
): ##
wav_path = "todo-songs/%s" % name #
f0_up_key = -2 #
audio, sampling_rate = sf.read(wav_path)
@ -92,7 +134,6 @@ for idx,name in enumerate(["冬之花clip1.wav",]):##
if sampling_rate != 16000:
audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
feats = torch.from_numpy(audio).float()
if feats.dim() == 2: # double channels
feats = feats.mean(-1)
@ -104,7 +145,8 @@ for idx,name in enumerate(["冬之花clip1.wav",]):##
"padding_mask": padding_mask.to(device),
"output_layer": 9, # layer 9
}
if torch.cuda.is_available(): torch.cuda.synchronize()
if torch.cuda.is_available():
torch.cuda.synchronize()
t0 = ttime()
with torch.no_grad():
logits = model.extract_features(**inputs)
@ -113,16 +155,20 @@ for idx,name in enumerate(["冬之花clip1.wav",]):##
####索引优化
npy = feats[0].cpu().numpy().astype("float32")
D, I = index.search(npy, 1)
feats = torch.from_numpy(big_npy[I.squeeze()].astype("float16")).unsqueeze(0).to(device)
feats = (
torch.from_numpy(big_npy[I.squeeze()].astype("float16")).unsqueeze(0).to(device)
)
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
if torch.cuda.is_available(): torch.cuda.synchronize()
if torch.cuda.is_available():
torch.cuda.synchronize()
t1 = ttime()
# p_len = min(feats.shape[1],10000,pitch.shape[0])#太大了爆显存
p_len = min(feats.shape[1], 10000) #
pitch, pitchf = get_f0(audio, p_len, f0_up_key)
p_len = min(feats.shape[1], 10000, pitch.shape[0]) # 太大了爆显存
if torch.cuda.is_available(): torch.cuda.synchronize()
if torch.cuda.is_available():
torch.cuda.synchronize()
t2 = ttime()
feats = feats[:, :p_len, :]
pitch = pitch[:p_len]
@ -132,12 +178,18 @@ for idx,name in enumerate(["冬之花clip1.wav",]):##
sid = torch.LongTensor([0]).to(device)
pitchf = torch.FloatTensor(pitchf).unsqueeze(0).to(device)
with torch.no_grad():
audio = net_g.infer(feats, p_len,pitch,pitchf,sid)[0][0, 0].data.cpu().float().numpy()#nsf
if torch.cuda.is_available(): torch.cuda.synchronize()
audio = (
net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]
.data.cpu()
.float()
.numpy()
) # nsf
if torch.cuda.is_available():
torch.cuda.synchronize()
t3 = ttime()
ta0+=(t1-t0)
ta1+=(t2-t1)
ta2+=(t3-t2)
ta0 += t1 - t0
ta1 += t2 - t1
ta2 += t3 - t2
# wavfile.write("ft-mi_1k-index256-noD-%s.wav"%name, 40000, audio)##
# wavfile.write("ft-mi-freeze-vocoder-flow-enc_q_1k-%s.wav"%name, 40000, audio)##
# wavfile.write("ft-mi-sim1k-%s.wav"%name, 40000, audio)##

10
infer/train-index.py
View File

@ -1,6 +1,6 @@
'''
"""
格式：直接cid为自带的index位；aid放不下了，通过字典来查，反正就5w个
'''
"""
import faiss, numpy as np, os
# ###########如果是原始特征要先写save
@ -21,11 +21,11 @@ print("training")
index_ivf = faiss.extract_index_ivf(index) #
index_ivf.nprobe = 9
index.train(big_npy)
faiss.write_index(index, 'infer/trained_IVF512_Flat_mi_baseline_src_feat.index')
faiss.write_index(index, "infer/trained_IVF512_Flat_mi_baseline_src_feat.index")
print("adding")
index.add(big_npy)
faiss.write_index(index, "infer/added_IVF512_Flat_mi_baseline_src_feat.index")
'''
"""
大小（都是FP32）
big_src_feature 2.95G
(3098036, 256)
@ -33,4 +33,4 @@ big_emb 4.43G
(6196072, 192)
big_emb双倍是因为求特征要repeat后再加pitch
'''
"""

9
infer/trans_weights.py
View File

@ -4,8 +4,13 @@ import torch,pdb
# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-freeze-vocoder-flow-enc_q\G_1000.pth")["model"]#sim_nsf#
# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-freeze-vocoder\G_1000.pth")["model"]#sim_nsf#
# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-test\G_1000.pth")["model"]#sim_nsf#
a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-no_opt-no_dropout\G_1000.pth")["model"]#sim_nsf#
for key in a.keys():a[key]=a[key].half()
a = torch.load(
r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-no_opt-no_dropout\G_1000.pth"
)[
"model"
] # sim_nsf#
for key in a.keys():
a[key] = a[key].half()
# torch.save(a,"ft-mi-freeze-vocoder_true_1k.pt")#
# torch.save(a,"ft-mi-sim1k.pt")#
torch.save(a, "ft-mi-no_opt-no_dropout.pt") #

2
infer_pack/commons.py
View File

@ -48,6 +48,8 @@ def slice_segments(x, ids_str, segment_size=4):
idx_end = idx_str + segment_size
ret[i] = x[i, :, idx_str:idx_end]
return ret
def slice_segments2(x, ids_str, segment_size=4):
ret = torch.zeros_like(x[:, :segment_size])
for i in range(x.size(0)):

180
infer_pack/models.py
View File

@ -12,9 +12,20 @@ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
from infer_pack.commons import init_weights
import numpy as np
from infer_pack import commons
class TextEncoder256(nn.Module):
def __init__(
self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=True ):
self,
out_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout,
f0=True,
):
super().__init__()
self.out_channels = out_channels
self.hidden_channels = hidden_channels
@ -25,7 +36,7 @@ class TextEncoder256(nn.Module):
self.p_dropout = p_dropout
self.emb_phone = nn.Linear(256, hidden_channels)
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
if(f0==True):
if f0 == True:
self.emb_pitch = nn.Embedding(256, hidden_channels) # pitch 256
self.encoder = attentions.Encoder(
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
@ -33,7 +44,7 @@ class TextEncoder256(nn.Module):
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
def forward(self, phone, pitch, lengths):
if(pitch==None):
if pitch == None:
x = self.emb_phone(phone)
else:
x = self.emb_phone(phone) + self.emb_pitch(pitch)
@ -48,8 +59,20 @@ class TextEncoder256(nn.Module):
m, logs = torch.split(stats, self.out_channels, dim=1)
return m, logs, x_mask
class TextEncoder256Sim(nn.Module):
def __init__( self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=True):
def __init__(
self,
out_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout,
f0=True,
):
super().__init__()
self.out_channels = out_channels
self.hidden_channels = hidden_channels
@ -60,7 +83,7 @@ class TextEncoder256Sim(nn.Module):
self.p_dropout = p_dropout
self.emb_phone = nn.Linear(256, hidden_channels)
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
if(f0==True):
if f0 == True:
self.emb_pitch = nn.Embedding(256, hidden_channels) # pitch 256
self.encoder = attentions.Encoder(
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
@ -68,17 +91,21 @@ class TextEncoder256Sim(nn.Module):
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
def forward(self, phone, pitch, lengths):
if(pitch==None):
if pitch == None:
x = self.emb_phone(phone)
else:
x = self.emb_phone(phone) + self.emb_pitch(pitch)
x = x * math.sqrt(self.hidden_channels) # [b, t, h]
x = self.lrelu(x)
x = torch.transpose(x, 1, -1) # [b, h, t]
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
x.dtype
)
x = self.encoder(x * x_mask, x_mask)
x = self.proj(x) * x_mask
return x, x_mask
class ResidualCouplingBlock(nn.Module):
def __init__(
self,
@ -126,6 +153,8 @@ class ResidualCouplingBlock(nn.Module):
def remove_weight_norm(self):
for i in range(self.n_flows):
self.flows[i * 2].remove_weight_norm()
class PosteriorEncoder(nn.Module):
def __init__(
self,
@ -169,6 +198,8 @@ class PosteriorEncoder(nn.Module):
def remove_weight_norm(self):
self.enc.remove_weight_norm()
class Generator(torch.nn.Module):
def __init__(
self,
@ -243,6 +274,8 @@ class Generator(torch.nn.Module):
remove_weight_norm(l)
for l in self.resblocks:
l.remove_weight_norm()
class SineGen(torch.nn.Module):
"""Definition of sine generator
SineGen(samp_rate, harmonic_num = 0,
@ -259,10 +292,15 @@ class SineGen(torch.nn.Module):
segment is always sin(np.pi) or cos(0)
"""
def __init__(self, samp_rate, harmonic_num=0,
sine_amp=0.1, noise_std=0.003,
def __init__(
self,
samp_rate,
harmonic_num=0,
sine_amp=0.1,
noise_std=0.003,
voiced_threshold=0,
flag_for_pulse=False):
flag_for_pulse=False,
):
super(SineGen, self).__init__()
self.sine_amp = sine_amp
self.noise_std = noise_std
@ -289,27 +327,47 @@ class SineGen(torch.nn.Module):
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
# fundamental component
f0_buf[:, :, 0] = f0[:, :, 0]
for idx in np.arange(self.harmonic_num):f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2)# idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
for idx in np.arange(self.harmonic_num):
f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
idx + 2
) # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
rad_values = (f0_buf / self.sampling_rate) % 1 ###%1意味着n_har的乘积无法后处理优化
rand_ini = torch.rand(f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device)
rand_ini = torch.rand(
f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
)
rand_ini[:, 0] = 0
rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
tmp_over_one = torch.cumsum(rad_values, 1) # % 1 #####%1意味着后面的cumsum无法再优化
tmp_over_one *= upp
tmp_over_one=F.interpolate(tmp_over_one.transpose(2, 1), scale_factor=upp, mode='linear', align_corners=True).transpose(2, 1)
rad_values=F.interpolate(rad_values.transpose(2, 1), scale_factor=upp, mode='nearest').transpose(2, 1)#######
tmp_over_one = F.interpolate(
tmp_over_one.transpose(2, 1),
scale_factor=upp,
mode="linear",
align_corners=True,
).transpose(2, 1)
rad_values = F.interpolate(
rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
).transpose(
2, 1
) #######
tmp_over_one %= 1
tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
cumsum_shift = torch.zeros_like(rad_values)
cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
sine_waves = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
sine_waves = torch.sin(
torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
)
sine_waves = sine_waves * self.sine_amp
uv = self._f02uv(f0)
uv = F.interpolate(uv.transpose(2, 1), scale_factor=upp, mode='nearest').transpose(2, 1)
uv = F.interpolate(
uv.transpose(2, 1), scale_factor=upp, mode="nearest"
).transpose(2, 1)
noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
noise = noise_amp * torch.randn_like(sine_waves)
sine_waves = sine_waves * uv + noise
return sine_waves, uv, noise
class SourceModuleHnNSF(torch.nn.Module):
"""SourceModule for hn-nsf
SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
@ -328,16 +386,24 @@ class SourceModuleHnNSF(torch.nn.Module):
uv (batchsize, length, 1)
"""
def __init__(self, sampling_rate, harmonic_num=0, sine_amp=0.1,
add_noise_std=0.003, voiced_threshod=0,is_half=True):
def __init__(
self,
sampling_rate,
harmonic_num=0,
sine_amp=0.1,
add_noise_std=0.003,
voiced_threshod=0,
is_half=True,
):
super(SourceModuleHnNSF, self).__init__()
self.sine_amp = sine_amp
self.noise_std = add_noise_std
self.is_half = is_half
# to produce sine waveforms
self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
sine_amp, add_noise_std, voiced_threshod)
self.l_sin_gen = SineGen(
sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
)
# to merge source harmonics into a single excitation
self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
@ -345,9 +411,12 @@ class SourceModuleHnNSF(torch.nn.Module):
def forward(self, x, upp=None):
sine_wavs, uv, _ = self.l_sin_gen(x, upp)
if(self.is_half):sine_wavs=sine_wavs.half()
if self.is_half:
sine_wavs = sine_wavs.half()
sine_merge = self.l_tanh(self.l_linear(sine_wavs))
return sine_merge, None, None # noise, uv
class GeneratorNSF(torch.nn.Module):
def __init__(
self,
@ -360,7 +429,7 @@ class GeneratorNSF(torch.nn.Module):
upsample_kernel_sizes,
gin_channels,
sr,
is_half=False
is_half=False,
):
super(GeneratorNSF, self).__init__()
self.num_kernels = len(resblock_kernel_sizes)
@ -368,9 +437,7 @@ class GeneratorNSF(torch.nn.Module):
self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
self.m_source = SourceModuleHnNSF(
sampling_rate=sr,
harmonic_num=0,
is_half=is_half
sampling_rate=sr, harmonic_num=0, is_half=is_half
)
self.noise_convs = nn.ModuleList()
self.conv_pre = Conv1d(
@ -394,8 +461,15 @@ class GeneratorNSF(torch.nn.Module):
)
if i + 1 < len(upsample_rates):
stride_f0 = np.prod(upsample_rates[i + 1 :])
self.noise_convs.append(Conv1d(
1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=stride_f0 // 2))
self.noise_convs.append(
Conv1d(
1,
c_cur,
kernel_size=stride_f0 * 2,
stride=stride_f0,
padding=stride_f0 // 2,
)
)
else:
self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
@ -444,11 +518,15 @@ class GeneratorNSF(torch.nn.Module):
remove_weight_norm(l)
for l in self.resblocks:
l.remove_weight_norm()
sr2sr = {
"32k": 32000,
"40k": 40000,
"48k": 48000,
}
class SynthesizerTrnMs256NSFsid(nn.Module):
def __init__(
self,
@ -472,9 +550,8 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
sr,
**kwargs
):
super().__init__()
if(type(sr)==type("strr")):
if type(sr) == type("strr"):
sr = sr2sr[sr]
self.spec_channels = spec_channels
self.inter_channels = inter_channels
@ -511,7 +588,9 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
gin_channels=gin_channels, sr=sr, is_half=kwargs["is_half"]
gin_channels=gin_channels,
sr=sr,
is_half=kwargs["is_half"],
)
self.enc_q = PosteriorEncoder(
spec_channels,
@ -527,12 +606,15 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
)
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
def remove_weight_norm(self):
self.dec.remove_weight_norm()
self.flow.remove_weight_norm()
self.enc_q.remove_weight_norm()
def forward(self, phone, phone_lengths, pitch,pitchf, y, y_lengths,ds):#这里ds是id，[bs,1]
def forward(
self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds
): # 这里ds是id，[bs,1]
# print(1,pitch.shape)#[bs,t]
g = self.emb_g(ds).unsqueeze(-1) # [b, 256, 1]##1是t，广播的
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
@ -542,9 +624,7 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
z, y_lengths, self.segment_size
)
# print(-1,pitchf.shape,ids_slice,self.segment_size,self.hop_length,self.segment_size//self.hop_length)
pitchf = commons.slice_segments2(
pitchf, ids_slice, self.segment_size
)
pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
# print(-2,pitchf.shape,z_slice.shape)
o = self.dec(z_slice, pitchf, g=g)
return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
@ -556,6 +636,8 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
z = self.flow(z_p, x_mask, g=g, reverse=True)
o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
return o, x_mask, (z, z_p, m_p, logs_p)
class SynthesizerTrnMs256NSFsid_nono(nn.Module):
def __init__(
self,
@ -579,7 +661,6 @@ class SynthesizerTrnMs256NSFsid_nono(nn.Module):
sr=None,
**kwargs
):
super().__init__()
self.spec_channels = spec_channels
self.inter_channels = inter_channels
@ -606,7 +687,8 @@ class SynthesizerTrnMs256NSFsid_nono(nn.Module):
n_heads,
n_layers,
kernel_size,
p_dropout,f0=False
p_dropout,
f0=False,
)
self.dec = Generator(
inter_channels,
@ -616,7 +698,7 @@ class SynthesizerTrnMs256NSFsid_nono(nn.Module):
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
gin_channels=gin_channels
gin_channels=gin_channels,
)
self.enc_q = PosteriorEncoder(
spec_channels,
@ -656,6 +738,8 @@ class SynthesizerTrnMs256NSFsid_nono(nn.Module):
z = self.flow(z_p, x_mask, g=g, reverse=True)
o = self.dec((z * x_mask)[:, :, :max_len], g=g)
return o, x_mask, (z, z_p, m_p, logs_p)
class SynthesizerTrnMs256NSFsid_sim(nn.Module):
"""
Synthesizer for Training
@ -684,7 +768,6 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
use_sdp=True,
**kwargs
):
super().__init__()
self.spec_channels = spec_channels
self.inter_channels = inter_channels
@ -721,7 +804,8 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
gin_channels=gin_channels,is_half=kwargs["is_half"]
gin_channels=gin_channels,
is_half=kwargs["is_half"],
)
self.flow = ResidualCouplingBlock(
@ -729,12 +813,15 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
)
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
def remove_weight_norm(self):
self.dec.remove_weight_norm()
self.flow.remove_weight_norm()
self.enc_q.remove_weight_norm()
def forward(self, phone, phone_lengths, pitch, pitchf, y_lengths,ds): # y是spec不需要了现在
def forward(
self, phone, phone_lengths, pitch, pitchf, y_lengths, ds
): # y是spec不需要了现在
g = self.emb_g(ds).unsqueeze(-1) # [b, 256, 1]##1是t，广播的
x, x_mask = self.enc_p(phone, pitch, phone_lengths)
x = self.flow(x, x_mask, g=g, reverse=True)
@ -742,18 +829,20 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
x, y_lengths, self.segment_size
)
pitchf = commons.slice_segments2(
pitchf, ids_slice, self.segment_size
)
pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
o = self.dec(z_slice, pitchf, g=g)
return o, ids_slice
def infer(self, phone, phone_lengths, pitch, pitchf, ds,max_len=None): # y是spec不需要了现在
def infer(
self, phone, phone_lengths, pitch, pitchf, ds, max_len=None
): # y是spec不需要了现在
g = self.emb_g(ds).unsqueeze(-1) # [b, 256, 1]##1是t，广播的
x, x_mask = self.enc_p(phone, pitch, phone_lengths)
x = self.flow(x, x_mask, g=g, reverse=True)
o = self.dec((x * x_mask)[:, :, :max_len], pitchf, g=g)
return o, o
class MultiPeriodDiscriminator(torch.nn.Module):
def __init__(self, use_spectral_norm=False):
super(MultiPeriodDiscriminator, self).__init__()
@ -783,6 +872,7 @@ class MultiPeriodDiscriminator(torch.nn.Module):
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
class DiscriminatorS(torch.nn.Module):
def __init__(self, use_spectral_norm=False):
super(DiscriminatorS, self).__init__()
@ -812,6 +902,7 @@ class DiscriminatorS(torch.nn.Module):
return x, fmap
class DiscriminatorP(torch.nn.Module):
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
super(DiscriminatorP, self).__init__()
@ -889,4 +980,3 @@ class DiscriminatorP(torch.nn.Module):
x = torch.flatten(x, 1, -1)
return x, fmap

155
infer_pack/models_onnx.py
View File

@ -12,9 +12,20 @@ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
from infer_pack.commons import init_weights
import numpy as np
from infer_pack import commons
class TextEncoder256(nn.Module):
def __init__(
self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=True ):
self,
out_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout,
f0=True,
):
super().__init__()
self.out_channels = out_channels
self.hidden_channels = hidden_channels
@ -25,7 +36,7 @@ class TextEncoder256(nn.Module):
self.p_dropout = p_dropout
self.emb_phone = nn.Linear(256, hidden_channels)
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
if(f0==True):
if f0 == True:
self.emb_pitch = nn.Embedding(256, hidden_channels) # pitch 256
self.encoder = attentions.Encoder(
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
@ -33,7 +44,7 @@ class TextEncoder256(nn.Module):
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
def forward(self, phone, pitch, lengths):
if(pitch==None):
if pitch == None:
x = self.emb_phone(phone)
else:
x = self.emb_phone(phone) + self.emb_pitch(pitch)
@ -48,8 +59,20 @@ class TextEncoder256(nn.Module):
m, logs = torch.split(stats, self.out_channels, dim=1)
return m, logs, x_mask
class TextEncoder256Sim(nn.Module):
def __init__( self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=True):
def __init__(
self,
out_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout,
f0=True,
):
super().__init__()
self.out_channels = out_channels
self.hidden_channels = hidden_channels
@ -60,7 +83,7 @@ class TextEncoder256Sim(nn.Module):
self.p_dropout = p_dropout
self.emb_phone = nn.Linear(256, hidden_channels)
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
if(f0==True):
if f0 == True:
self.emb_pitch = nn.Embedding(256, hidden_channels) # pitch 256
self.encoder = attentions.Encoder(
hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
@ -68,17 +91,21 @@ class TextEncoder256Sim(nn.Module):
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
def forward(self, phone, pitch, lengths):
if(pitch==None):
if pitch == None:
x = self.emb_phone(phone)
else:
x = self.emb_phone(phone) + self.emb_pitch(pitch)
x = x * math.sqrt(self.hidden_channels) # [b, t, h]
x = self.lrelu(x)
x = torch.transpose(x, 1, -1) # [b, h, t]
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
x.dtype
)
x = self.encoder(x * x_mask, x_mask)
x = self.proj(x) * x_mask
return x, x_mask
class ResidualCouplingBlock(nn.Module):
def __init__(
self,
@ -126,6 +153,8 @@ class ResidualCouplingBlock(nn.Module):
def remove_weight_norm(self):
for i in range(self.n_flows):
self.flows[i * 2].remove_weight_norm()
class PosteriorEncoder(nn.Module):
def __init__(
self,
@ -169,6 +198,8 @@ class PosteriorEncoder(nn.Module):
def remove_weight_norm(self):
self.enc.remove_weight_norm()
class Generator(torch.nn.Module):
def __init__(
self,
@ -243,6 +274,8 @@ class Generator(torch.nn.Module):
remove_weight_norm(l)
for l in self.resblocks:
l.remove_weight_norm()
class SineGen(torch.nn.Module):
"""Definition of sine generator
SineGen(samp_rate, harmonic_num = 0,
@ -259,10 +292,15 @@ class SineGen(torch.nn.Module):
segment is always sin(np.pi) or cos(0)
"""
def __init__(self, samp_rate, harmonic_num=0,
sine_amp=0.1, noise_std=0.003,
def __init__(
self,
samp_rate,
harmonic_num=0,
sine_amp=0.1,
noise_std=0.003,
voiced_threshold=0,
flag_for_pulse=False):
flag_for_pulse=False,
):
super(SineGen, self).__init__()
self.sine_amp = sine_amp
self.noise_std = noise_std
@ -289,27 +327,47 @@ class SineGen(torch.nn.Module):
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
# fundamental component
f0_buf[:, :, 0] = f0[:, :, 0]
for idx in np.arange(self.harmonic_num):f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2)# idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
for idx in np.arange(self.harmonic_num):
f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
idx + 2
) # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
rad_values = (f0_buf / self.sampling_rate) % 1 ###%1意味着n_har的乘积无法后处理优化
rand_ini = torch.rand(f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device)
rand_ini = torch.rand(
f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
)
rand_ini[:, 0] = 0
rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
tmp_over_one = torch.cumsum(rad_values, 1) # % 1 #####%1意味着后面的cumsum无法再优化
tmp_over_one *= upp
tmp_over_one=F.interpolate(tmp_over_one.transpose(2, 1), scale_factor=upp, mode='linear', align_corners=True).transpose(2, 1)
rad_values=F.interpolate(rad_values.transpose(2, 1), scale_factor=upp, mode='nearest').transpose(2, 1)#######
tmp_over_one = F.interpolate(
tmp_over_one.transpose(2, 1),
scale_factor=upp,
mode="linear",
align_corners=True,
).transpose(2, 1)
rad_values = F.interpolate(
rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
).transpose(
2, 1
) #######
tmp_over_one %= 1
tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
cumsum_shift = torch.zeros_like(rad_values)
cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
sine_waves = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
sine_waves = torch.sin(
torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
)
sine_waves = sine_waves * self.sine_amp
uv = self._f02uv(f0)
uv = F.interpolate(uv.transpose(2, 1), scale_factor=upp, mode='nearest').transpose(2, 1)
uv = F.interpolate(
uv.transpose(2, 1), scale_factor=upp, mode="nearest"
).transpose(2, 1)
noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
noise = noise_amp * torch.randn_like(sine_waves)
sine_waves = sine_waves * uv + noise
return sine_waves, uv, noise
class SourceModuleHnNSF(torch.nn.Module):
"""SourceModule for hn-nsf
SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
@ -328,16 +386,24 @@ class SourceModuleHnNSF(torch.nn.Module):
uv (batchsize, length, 1)
"""
def __init__(self, sampling_rate, harmonic_num=0, sine_amp=0.1,
add_noise_std=0.003, voiced_threshod=0,is_half=True):
def __init__(
self,
sampling_rate,
harmonic_num=0,
sine_amp=0.1,
add_noise_std=0.003,
voiced_threshod=0,
is_half=True,
):
super(SourceModuleHnNSF, self).__init__()
self.sine_amp = sine_amp
self.noise_std = add_noise_std
self.is_half = is_half
# to produce sine waveforms
self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
sine_amp, add_noise_std, voiced_threshod)
self.l_sin_gen = SineGen(
sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
)
# to merge source harmonics into a single excitation
self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
@ -345,9 +411,12 @@ class SourceModuleHnNSF(torch.nn.Module):
def forward(self, x, upp=None):
sine_wavs, uv, _ = self.l_sin_gen(x, upp)
if(self.is_half):sine_wavs=sine_wavs.half()
if self.is_half:
sine_wavs = sine_wavs.half()
sine_merge = self.l_tanh(self.l_linear(sine_wavs))
return sine_merge, None, None # noise, uv
class GeneratorNSF(torch.nn.Module):
def __init__(
self,
@ -360,7 +429,7 @@ class GeneratorNSF(torch.nn.Module):
upsample_kernel_sizes,
gin_channels,
sr,
is_half=False
is_half=False,
):
super(GeneratorNSF, self).__init__()
self.num_kernels = len(resblock_kernel_sizes)
@ -368,9 +437,7 @@ class GeneratorNSF(torch.nn.Module):
self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
self.m_source = SourceModuleHnNSF(
sampling_rate=sr,
harmonic_num=0,
is_half=is_half
sampling_rate=sr, harmonic_num=0, is_half=is_half
)
self.noise_convs = nn.ModuleList()
self.conv_pre = Conv1d(
@ -394,8 +461,15 @@ class GeneratorNSF(torch.nn.Module):
)
if i + 1 < len(upsample_rates):
stride_f0 = np.prod(upsample_rates[i + 1 :])
self.noise_convs.append(Conv1d(
1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=stride_f0 // 2))
self.noise_convs.append(
Conv1d(
1,
c_cur,
kernel_size=stride_f0 * 2,
stride=stride_f0,
padding=stride_f0 // 2,
)
)
else:
self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
@ -444,11 +518,15 @@ class GeneratorNSF(torch.nn.Module):
remove_weight_norm(l)
for l in self.resblocks:
l.remove_weight_norm()
sr2sr = {
"32k": 32000,
"40k": 40000,
"48k": 48000,
}
class SynthesizerTrnMs256NSFsid(nn.Module):
def __init__(
self,
@ -472,9 +550,8 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
sr,
**kwargs
):
super().__init__()
if(type(sr)==type("strr")):
if type(sr) == type("strr"):
sr = sr2sr[sr]
self.spec_channels = spec_channels
self.inter_channels = inter_channels
@ -511,7 +588,9 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
gin_channels=gin_channels, sr=sr, is_half=kwargs["is_half"]
gin_channels=gin_channels,
sr=sr,
is_half=kwargs["is_half"],
)
self.enc_q = PosteriorEncoder(
spec_channels,
@ -527,13 +606,13 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
)
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
def remove_weight_norm(self):
self.dec.remove_weight_norm()
self.flow.remove_weight_norm()
self.enc_q.remove_weight_norm()
def forward(self, phone, phone_lengths, pitch, nsff0, sid, rnd, max_len=None):
g = self.emb_g(sid).unsqueeze(-1)
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
@ -541,6 +620,7 @@ class SynthesizerTrnMs256NSFsid(nn.Module):
o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
return o
class SynthesizerTrnMs256NSFsid_sim(nn.Module):
"""
Synthesizer for Training
@ -569,7 +649,6 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
use_sdp=True,
**kwargs
):
super().__init__()
self.spec_channels = spec_channels
self.inter_channels = inter_channels
@ -606,7 +685,8 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
gin_channels=gin_channels,is_half=kwargs["is_half"]
gin_channels=gin_channels,
is_half=kwargs["is_half"],
)
self.flow = ResidualCouplingBlock(
@ -614,18 +694,22 @@ class SynthesizerTrnMs256NSFsid_sim(nn.Module):
)
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
print("gin_channels:", gin_channels, "self.spk_embed_dim:", self.spk_embed_dim)
def remove_weight_norm(self):
self.dec.remove_weight_norm()
self.flow.remove_weight_norm()
self.enc_q.remove_weight_norm()
def forward(self, phone, phone_lengths, pitch, pitchf, ds,max_len=None): # y是spec不需要了现在
def forward(
self, phone, phone_lengths, pitch, pitchf, ds, max_len=None
): # y是spec不需要了现在
g = self.emb_g(ds.unsqueeze(0)).unsqueeze(-1) # [b, 256, 1]##1是t，广播的
x, x_mask = self.enc_p(phone, pitch, phone_lengths)
x = self.flow(x, x_mask, g=g, reverse=True)
o = self.dec((x * x_mask)[:, :, :max_len], pitchf, g=g)
return o
class MultiPeriodDiscriminator(torch.nn.Module):
def __init__(self, use_spectral_norm=False):
super(MultiPeriodDiscriminator, self).__init__()
@ -655,6 +739,7 @@ class MultiPeriodDiscriminator(torch.nn.Module):
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
class DiscriminatorS(torch.nn.Module):
def __init__(self, use_spectral_norm=False):
super(DiscriminatorS, self).__init__()
@ -684,6 +769,7 @@ class DiscriminatorS(torch.nn.Module):
return x, fmap
class DiscriminatorP(torch.nn.Module):
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
super(DiscriminatorP, self).__init__()
@ -761,4 +847,3 @@ class DiscriminatorP(torch.nn.Module):
x = torch.flatten(x, 1, -1)
return x, fmap

110
infer_pack/transforms.py
View File

@ -9,26 +9,24 @@ DEFAULT_MIN_BIN_HEIGHT = 1e-3
DEFAULT_MIN_DERIVATIVE = 1e-3
def piecewise_rational_quadratic_transform(inputs,
def piecewise_rational_quadratic_transform(
inputs,
unnormalized_widths,
unnormalized_heights,
unnormalized_derivatives,
inverse=False,
tails=None,
tail_bound=1.,
tail_bound=1.0,
min_bin_width=DEFAULT_MIN_BIN_WIDTH,
min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
min_derivative=DEFAULT_MIN_DERIVATIVE):
min_derivative=DEFAULT_MIN_DERIVATIVE,
):
if tails is None:
spline_fn = rational_quadratic_spline
spline_kwargs = {}
else:
spline_fn = unconstrained_rational_quadratic_spline
spline_kwargs = {
'tails': tails,
'tail_bound': tail_bound
}
spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
outputs, logabsdet = spline_fn(
inputs=inputs,
@ -46,29 +44,28 @@ def piecewise_rational_quadratic_transform(inputs,
def searchsorted(bin_locations, inputs, eps=1e-6):
bin_locations[..., -1] += eps
return torch.sum(
inputs[..., None] >= bin_locations,
dim=-1
) - 1
return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
def unconstrained_rational_quadratic_spline(inputs,
def unconstrained_rational_quadratic_spline(
inputs,
unnormalized_widths,
unnormalized_heights,
unnormalized_derivatives,
inverse=False,
tails='linear',
tail_bound=1.,
tails="linear",
tail_bound=1.0,
min_bin_width=DEFAULT_MIN_BIN_WIDTH,
min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
min_derivative=DEFAULT_MIN_DERIVATIVE):
min_derivative=DEFAULT_MIN_DERIVATIVE,
):
inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
outside_interval_mask = ~inside_interval_mask
outputs = torch.zeros_like(inputs)
logabsdet = torch.zeros_like(inputs)
if tails == 'linear':
if tails == "linear":
unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
constant = np.log(np.exp(1 - min_derivative) - 1)
unnormalized_derivatives[..., 0] = constant
@ -77,45 +74,57 @@ def unconstrained_rational_quadratic_spline(inputs,
outputs[outside_interval_mask] = inputs[outside_interval_mask]
logabsdet[outside_interval_mask] = 0
else:
raise RuntimeError('{} tails are not implemented.'.format(tails))
raise RuntimeError("{} tails are not implemented.".format(tails))
outputs[inside_interval_mask], logabsdet[inside_interval_mask] = rational_quadratic_spline(
(
outputs[inside_interval_mask],
logabsdet[inside_interval_mask],
) = rational_quadratic_spline(
inputs=inputs[inside_interval_mask],
unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
inverse=inverse,
left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound,
left=-tail_bound,
right=tail_bound,
bottom=-tail_bound,
top=tail_bound,
min_bin_width=min_bin_width,
min_bin_height=min_bin_height,
min_derivative=min_derivative
min_derivative=min_derivative,
)
return outputs, logabsdet
def rational_quadratic_spline(inputs,
def rational_quadratic_spline(
inputs,
unnormalized_widths,
unnormalized_heights,
unnormalized_derivatives,
inverse=False,
left=0., right=1., bottom=0., top=1.,
left=0.0,
right=1.0,
bottom=0.0,
top=1.0,
min_bin_width=DEFAULT_MIN_BIN_WIDTH,
min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
min_derivative=DEFAULT_MIN_DERIVATIVE):
min_derivative=DEFAULT_MIN_DERIVATIVE,
):
if torch.min(inputs) < left or torch.max(inputs) > right:
raise ValueError('Input to a transform is not within its domain')
raise ValueError("Input to a transform is not within its domain")
num_bins = unnormalized_widths.shape[-1]
if min_bin_width * num_bins > 1.0:
raise ValueError('Minimal bin width too large for the number of bins')
raise ValueError("Minimal bin width too large for the number of bins")
if min_bin_height * num_bins > 1.0:
raise ValueError('Minimal bin height too large for the number of bins')
raise ValueError("Minimal bin height too large for the number of bins")
widths = F.softmax(unnormalized_widths, dim=-1)
widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
cumwidths = torch.cumsum(widths, dim=-1)
cumwidths = F.pad(cumwidths, pad=(1, 0), mode='constant', value=0.0)
cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
cumwidths = (right - left) * cumwidths + left
cumwidths[..., 0] = left
cumwidths[..., -1] = right
@ -126,7 +135,7 @@ def rational_quadratic_spline(inputs,
heights = F.softmax(unnormalized_heights, dim=-1)
heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
cumheights = torch.cumsum(heights, dim=-1)
cumheights = F.pad(cumheights, pad=(1, 0), mode='constant', value=0.0)
cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
cumheights = (top - bottom) * cumheights + bottom
cumheights[..., 0] = bottom
cumheights[..., -1] = top
@ -150,14 +159,12 @@ def rational_quadratic_spline(inputs,
input_heights = heights.gather(-1, bin_idx)[..., 0]
if inverse:
a = (((inputs - input_cumheights) * (input_derivatives
+ input_derivatives_plus_one
- 2 * input_delta)
+ input_heights * (input_delta - input_derivatives)))
b = (input_heights * input_derivatives
- (inputs - input_cumheights) * (input_derivatives
+ input_derivatives_plus_one
- 2 * input_delta))
a = (inputs - input_cumheights) * (
input_derivatives + input_derivatives_plus_one - 2 * input_delta
) + input_heights * (input_delta - input_derivatives)
b = input_heights * input_derivatives - (inputs - input_cumheights) * (
input_derivatives + input_derivatives_plus_one - 2 * input_delta
)
c = -input_delta * (inputs - input_cumheights)
discriminant = b.pow(2) - 4 * a * c
@ -167,11 +174,15 @@ def rational_quadratic_spline(inputs,
outputs = root * input_bin_widths + input_cumwidths
theta_one_minus_theta = root * (1 - root)
denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
* theta_one_minus_theta)
derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * root.pow(2)
denominator = input_delta + (
(input_derivatives + input_derivatives_plus_one - 2 * input_delta)
* theta_one_minus_theta
)
derivative_numerator = input_delta.pow(2) * (
input_derivatives_plus_one * root.pow(2)
+ 2 * input_delta * theta_one_minus_theta
+ input_derivatives * (1 - root).pow(2))
+ input_derivatives * (1 - root).pow(2)
)
logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
return outputs, -logabsdet
@ -179,15 +190,20 @@ def rational_quadratic_spline(inputs,
theta = (inputs - input_cumwidths) / input_bin_widths
theta_one_minus_theta = theta * (1 - theta)
numerator = input_heights * (input_delta * theta.pow(2)
+ input_derivatives * theta_one_minus_theta)
denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
* theta_one_minus_theta)
numerator = input_heights * (
input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
)
denominator = input_delta + (
(input_derivatives + input_derivatives_plus_one - 2 * input_delta)
* theta_one_minus_theta
)
outputs = input_cumheights + numerator / denominator
derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2)
derivative_numerator = input_delta.pow(2) * (
input_derivatives_plus_one * theta.pow(2)
+ 2 * input_delta * theta_one_minus_theta
+ input_derivatives * (1 - theta).pow(2))
+ input_derivatives * (1 - theta).pow(2)
)
logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
return outputs, logabsdet

159
infer_uvr5.py
View File

@ -1,4 +1,5 @@
import os, sys, torch, warnings, pdb
warnings.filterwarnings("ignore")
import librosa
import importlib
@ -10,99 +11,161 @@ from uvr5_pack.utils import _get_name_params,inference
from uvr5_pack.lib_v5.model_param_init import ModelParameters
from scipy.io import wavfile
class _audio_pre_():
class _audio_pre_:
def __init__(self, model_path, device, is_half):
self.model_path = model_path
self.device = device
self.data = {
# Processing Options
'postprocess': False,
'tta': False,
"postprocess": False,
"tta": False,
# Constants
'window_size': 512,
'agg': 10,
'high_end_process': 'mirroring',
"window_size": 512,
"agg": 10,
"high_end_process": "mirroring",
}
nn_arch_sizes = [
31191, # default
33966,61968, 123821, 123812, 537238 # custom
33966,
61968,
123821,
123812,
537238, # custom
]
self.nn_architecture = list('{}KB'.format(s) for s in nn_arch_sizes)
self.nn_architecture = list("{}KB".format(s) for s in nn_arch_sizes)
model_size = math.ceil(os.stat(model_path).st_size / 1024)
nn_architecture = '{}KB'.format(min(nn_arch_sizes, key=lambda x:abs(x-model_size)))
nets = importlib.import_module('uvr5_pack.lib_v5.nets' + f'_{nn_architecture}'.replace('_{}KB'.format(nn_arch_sizes[0]), ''), package=None)
model_hash = hashlib.md5(open(model_path,'rb').read()).hexdigest()
nn_architecture = "{}KB".format(
min(nn_arch_sizes, key=lambda x: abs(x - model_size))
)
nets = importlib.import_module(
"uvr5_pack.lib_v5.nets"
+ f"_{nn_architecture}".replace("_{}KB".format(nn_arch_sizes[0]), ""),
package=None,
)
model_hash = hashlib.md5(open(model_path, "rb").read()).hexdigest()
param_name, model_params_d = _get_name_params(model_path, model_hash)
mp = ModelParameters(model_params_d)
model = nets.CascadedASPPNet(mp.param['bins'] * 2)
cpk = torch.load( model_path , map_location='cpu')
model = nets.CascadedASPPNet(mp.param["bins"] * 2)
cpk = torch.load(model_path, map_location="cpu")
model.load_state_dict(cpk)
model.eval()
if(is_half):model = model.half().to(device)
else:model = model.to(device)
if is_half:
model = model.half().to(device)
else:
model = model.to(device)
self.mp = mp
self.model = model
def _path_audio_(self, music_file, ins_root=None, vocal_root=None):
if(ins_root is None and vocal_root is None):return "No save root."
if ins_root is None and vocal_root is None:
return "No save root."
name = os.path.basename(music_file)
if(ins_root is not None):os.makedirs(ins_root, exist_ok=True)
if(vocal_root is not None):os.makedirs(vocal_root , exist_ok=True)
if ins_root is not None:
os.makedirs(ins_root, exist_ok=True)
if vocal_root is not None:
os.makedirs(vocal_root, exist_ok=True)
X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
bands_n = len(self.mp.param['band'])
bands_n = len(self.mp.param["band"])
# print(bands_n)
for d in range(bands_n, 0, -1):
bp = self.mp.param['band'][d]
bp = self.mp.param["band"][d]
if d == bands_n: # high-end band
X_wave[d], _ = librosa.core.load(#理论上librosa读取可能对某些音频有bug，应该上ffmpeg读取，但是太麻烦了弃坑
music_file, bp['sr'], False, dtype=np.float32, res_type=bp['res_type'])
(
X_wave[d],
_,
) = librosa.core.load( # 理论上librosa读取可能对某些音频有bug，应该上ffmpeg读取，但是太麻烦了弃坑
music_file,
bp["sr"],
False,
dtype=np.float32,
res_type=bp["res_type"],
)
if X_wave[d].ndim == 1:
X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
else: # lower bands
X_wave[d] = librosa.core.resample(X_wave[d+1], self.mp.param['band'][d+1]['sr'], bp['sr'], res_type=bp['res_type'])
X_wave[d] = librosa.core.resample(
X_wave[d + 1],
self.mp.param["band"][d + 1]["sr"],
bp["sr"],
res_type=bp["res_type"],
)
# Stft of wave source
X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(X_wave[d], bp['hl'], bp['n_fft'], self.mp.param['mid_side'], self.mp.param['mid_side_b2'], self.mp.param['reverse'])
X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
X_wave[d],
bp["hl"],
bp["n_fft"],
self.mp.param["mid_side"],
self.mp.param["mid_side_b2"],
self.mp.param["reverse"],
)
# pdb.set_trace()
if d == bands_n and self.data['high_end_process'] != 'none':
input_high_end_h = (bp['n_fft']//2 - bp['crop_stop']) + ( self.mp.param['pre_filter_stop'] - self.mp.param['pre_filter_start'])
input_high_end = X_spec_s[d][:, bp['n_fft']//2-input_high_end_h:bp['n_fft']//2, :]
if d == bands_n and self.data["high_end_process"] != "none":
input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
)
input_high_end = X_spec_s[d][
:, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
]
X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
aggresive_set = float(self.data['agg']/100)
aggressiveness = {'value': aggresive_set, 'split_bin': self.mp.param['band'][1]['crop_stop']}
aggresive_set = float(self.data["agg"] / 100)
aggressiveness = {
"value": aggresive_set,
"split_bin": self.mp.param["band"][1]["crop_stop"],
}
with torch.no_grad():
pred, X_mag, X_phase = inference(X_spec_m,self.device,self.model, aggressiveness,self.data)
pred, X_mag, X_phase = inference(
X_spec_m, self.device, self.model, aggressiveness, self.data
)
# Postprocess
if self.data['postprocess']:
if self.data["postprocess"]:
pred_inv = np.clip(X_mag - pred, 0, np.inf)
pred = spec_utils.mask_silence(pred, pred_inv)
y_spec_m = pred * X_phase
v_spec_m = X_spec_m - y_spec_m
if (ins_root is not None):
if self.data['high_end_process'].startswith('mirroring'):
input_high_end_ = spec_utils.mirroring(self.data['high_end_process'], y_spec_m, input_high_end, self.mp)
wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp,input_high_end_h, input_high_end_)
if ins_root is not None:
if self.data["high_end_process"].startswith("mirroring"):
input_high_end_ = spec_utils.mirroring(
self.data["high_end_process"], y_spec_m, input_high_end, self.mp
)
wav_instrument = spec_utils.cmb_spectrogram_to_wave(
y_spec_m, self.mp, input_high_end_h, input_high_end_
)
else:
wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
print ('%s instruments done'%name)
wavfile.write(os.path.join(ins_root, 'instrument_{}.wav'.format(name) ), self.mp.param['sr'], (np.array(wav_instrument)*32768).astype("int16")) #
if (vocal_root is not None):
if self.data['high_end_process'].startswith('mirroring'):
input_high_end_ = spec_utils.mirroring(self.data['high_end_process'], v_spec_m, input_high_end, self.mp)
wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp, input_high_end_h, input_high_end_)
print("%s instruments done" % name)
wavfile.write(
os.path.join(ins_root, "instrument_{}.wav".format(name)),
self.mp.param["sr"],
(np.array(wav_instrument) * 32768).astype("int16"),
) #
if vocal_root is not None:
if self.data["high_end_process"].startswith("mirroring"):
input_high_end_ = spec_utils.mirroring(
self.data["high_end_process"], v_spec_m, input_high_end, self.mp
)
wav_vocals = spec_utils.cmb_spectrogram_to_wave(
v_spec_m, self.mp, input_high_end_h, input_high_end_
)
else:
wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
print ('%s vocals done'%name)
wavfile.write(os.path.join(vocal_root , 'vocal_{}.wav'.format(name) ), self.mp.param['sr'], (np.array(wav_vocals)*32768).astype("int16"))
print("%s vocals done" % name)
wavfile.write(
os.path.join(vocal_root, "vocal_{}.wav".format(name)),
self.mp.param["sr"],
(np.array(wav_vocals) * 32768).astype("int16"),
)
if __name__ == '__main__':
device = 'cuda'
if __name__ == "__main__":
device = "cuda"
is_half = True
model_path='uvr5_weights/2_HP-UVR.pth'
model_path = "uvr5_weights/2_HP-UVR.pth"
pre_fun = _audio_pre_(model_path=model_path, device=device, is_half=True)
audio_path = '神女劈观.aac'
save_path = 'opt'
audio_path = "神女劈观.aac"
save_path = "opt"
pre_fun._path_audio_(audio_path, save_path, save_path)

4
locale/locale_diff.py
View File

@ -31,7 +31,9 @@ for lang_file in languages:
del lang_data[key]
# Sort the keys of the language file to match the order of the standard file
lang_data = OrderedDict(sorted(lang_data.items(), key=lambda x: list(standard_data.keys()).index(x[0])))
lang_data = OrderedDict(
sorted(lang_data.items(), key=lambda x: list(standard_data.keys()).index(x[0]))
)
# Save the updated language file
with open(lang_file, "w", encoding="utf-8") as f:

6
my_utils.py
View File

@ -1,11 +1,15 @@
import ffmpeg
import numpy as np
def load_audio(file, sr):
try:
# https://github.com/openai/whisper/blob/main/whisper/audio.py#L26
# This launches a subprocess to decode audio while down-mixing and resampling as necessary.
# Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
file=file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")#防止小白拷路径头尾带了空格和"和回车
file = (
file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
) # 防止小白拷路径头尾带了空格和"和回车
out, _ = (
ffmpeg.input(file, threads=0)
.output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=sr)

147
slicer2.py
View File

@ -18,9 +18,7 @@ def get_rms(
x_shape_trimmed = list(y.shape)
x_shape_trimmed[axis] -= frame_length - 1
out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
xw = np.lib.stride_tricks.as_strided(
y, shape=out_shape, strides=out_strides
)
xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
if axis < 0:
target_axis = axis - 1
else:
@ -38,19 +36,25 @@ def get_rms(
class Slicer:
def __init__(self,
def __init__(
self,
sr: int,
threshold: float = -40.,
threshold: float = -40.0,
min_length: int = 5000,
min_interval: int = 300,
hop_size: int = 20,
max_sil_kept: int = 5000):
max_sil_kept: int = 5000,
):
if not min_length >= min_interval >= hop_size:
raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size')
raise ValueError(
"The following condition must be satisfied: min_length >= min_interval >= hop_size"
)
if not max_sil_kept >= hop_size:
raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size')
raise ValueError(
"The following condition must be satisfied: max_sil_kept >= hop_size"
)
min_interval = sr * min_interval / 1000
self.threshold = 10 ** (threshold / 20.)
self.threshold = 10 ** (threshold / 20.0)
self.hop_size = round(sr * hop_size / 1000)
self.win_size = min(round(min_interval), 4 * self.hop_size)
self.min_length = round(sr * min_length / 1000 / self.hop_size)
@ -59,9 +63,13 @@ class Slicer:
def _apply_slice(self, waveform, begin, end):
if len(waveform.shape) > 1:
return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)]
return waveform[
:, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)
]
else:
return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)]
return waveform[
begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)
]
# @timeit
def slice(self, waveform):
@ -71,7 +79,9 @@ class Slicer:
samples = waveform
if samples.shape[0] <= self.min_length:
return [waveform]
rms_list = get_rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
rms_list = get_rms(
y=samples, frame_length=self.win_size, hop_length=self.hop_size
).squeeze(0)
sil_tags = []
silence_start = None
clip_start = 0
@ -87,7 +97,10 @@ class Slicer:
continue
# Clear recorded silence start if interval is not enough or clip is too short
is_leading_silence = silence_start == 0 and i > self.max_sil_kept
need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length
need_slice_middle = (
i - silence_start >= self.min_interval
and i - clip_start >= self.min_length
)
if not is_leading_silence and not need_slice_middle:
silence_start = None
continue
@ -100,10 +113,21 @@ class Slicer:
sil_tags.append((pos, pos))
clip_start = pos
elif i - silence_start <= self.max_sil_kept * 2:
pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin()
pos = rms_list[
i - self.max_sil_kept : silence_start + self.max_sil_kept + 1
].argmin()
pos += i - self.max_sil_kept
pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
pos_l = (
rms_list[
silence_start : silence_start + self.max_sil_kept + 1
].argmin()
+ silence_start
)
pos_r = (
rms_list[i - self.max_sil_kept : i + 1].argmin()
+ i
- self.max_sil_kept
)
if silence_start == 0:
sil_tags.append((0, pos_r))
clip_start = pos_r
@ -111,8 +135,17 @@ class Slicer:
sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
clip_start = max(pos_r, pos)
else:
pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
pos_l = (
rms_list[
silence_start : silence_start + self.max_sil_kept + 1
].argmin()
+ silence_start
)
pos_r = (
rms_list[i - self.max_sil_kept : i + 1].argmin()
+ i
- self.max_sil_kept
)
if silence_start == 0:
sil_tags.append((0, pos_r))
else:
@ -121,7 +154,10 @@ class Slicer:
silence_start = None
# Deal with trailing silence.
total_frames = rms_list.shape[0]
if silence_start is not None and total_frames - silence_start >= self.min_interval:
if (
silence_start is not None
and total_frames - silence_start >= self.min_interval
):
silence_end = min(total_frames, silence_start + self.max_sil_kept)
pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
sil_tags.append((pos, total_frames + 1))
@ -133,9 +169,13 @@ class Slicer:
if sil_tags[0][0] > 0:
chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
for i in range(len(sil_tags) - 1):
chunks.append(self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0]))
chunks.append(
self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0])
)
if sil_tags[-1][1] < total_frames:
chunks.append(self._apply_slice(waveform, sil_tags[-1][1], total_frames))
chunks.append(
self._apply_slice(waveform, sil_tags[-1][1], total_frames)
)
return chunks
@ -147,18 +187,45 @@ def main():
import soundfile
parser = ArgumentParser()
parser.add_argument('audio', type=str, help='The audio to be sliced')
parser.add_argument('--out', type=str, help='Output directory of the sliced audio clips')
parser.add_argument('--db_thresh', type=float, required=False, default=-40,
help='The dB threshold for silence detection')
parser.add_argument('--min_length', type=int, required=False, default=5000,
help='The minimum milliseconds required for each sliced audio clip')
parser.add_argument('--min_interval', type=int, required=False, default=300,
help='The minimum milliseconds for a silence part to be sliced')
parser.add_argument('--hop_size', type=int, required=False, default=10,
help='Frame length in milliseconds')
parser.add_argument('--max_sil_kept', type=int, required=False, default=500,
help='The maximum silence length kept around the sliced clip, presented in milliseconds')
parser.add_argument("audio", type=str, help="The audio to be sliced")
parser.add_argument(
"--out", type=str, help="Output directory of the sliced audio clips"
)
parser.add_argument(
"--db_thresh",
type=float,
required=False,
default=-40,
help="The dB threshold for silence detection",
)
parser.add_argument(
"--min_length",
type=int,
required=False,
default=5000,
help="The minimum milliseconds required for each sliced audio clip",
)
parser.add_argument(
"--min_interval",
type=int,
required=False,
default=300,
help="The minimum milliseconds for a silence part to be sliced",
)
parser.add_argument(
"--hop_size",
type=int,
required=False,
default=10,
help="Frame length in milliseconds",
)
parser.add_argument(
"--max_sil_kept",
type=int,
required=False,
default=500,
help="The maximum silence length kept around the sliced clip, presented in milliseconds",
)
args = parser.parse_args()
out = args.out
if out is None:
@ -170,7 +237,7 @@ def main():
min_length=args.min_length,
min_interval=args.min_interval,
hop_size=args.hop_size,
max_sil_kept=args.max_sil_kept
max_sil_kept=args.max_sil_kept,
)
chunks = slicer.slice(audio)
if not os.path.exists(out):
@ -178,8 +245,16 @@ def main():
for i, chunk in enumerate(chunks):
if len(chunk.shape) > 1:
chunk = chunk.T
soundfile.write(os.path.join(out, f'%s_%d.wav' % (os.path.basename(args.audio).rsplit('.', maxsplit=1)[0], i)), chunk, sr)
soundfile.write(
os.path.join(
out,
f"%s_%d.wav"
% (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
),
chunk,
sr,
)
if __name__ == '__main__':
if __name__ == "__main__":
main()

45
train/data_utils.py
View File

@ -6,6 +6,7 @@ import torch.utils.data
from mel_processing import spectrogram_torch
from utils import load_wav_to_torch, load_filepaths_and_text
class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
"""
1) loads audio, text pairs
@ -40,6 +41,7 @@ class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
self.audiopaths_and_text = audiopaths_and_text_new
self.lengths = lengths
def get_sid(self, sid):
sid = torch.LongTensor([int(sid)])
return sid
@ -104,9 +106,14 @@ class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
spec = torch.load(spec_filename)
except:
print(spec_filename, traceback.format_exc())
spec = spectrogram_torch(audio_norm, self.filter_length,
self.sampling_rate, self.hop_length, self.win_length,
center=False)
spec = spectrogram_torch(
audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False,
)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
else:
@ -127,6 +134,8 @@ class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
def __len__(self):
return len(self.audiopaths_and_text)
class TextAudioCollateMultiNSFsid:
"""Zero-pads model inputs and targets"""
@ -155,7 +164,9 @@ class TextAudioCollateMultiNSFsid:
max_phone_len = max([x[2].size(0) for x in batch])
phone_lengths = torch.LongTensor(len(batch))
phone_padded = torch.FloatTensor(len(batch), max_phone_len, batch[0][2].shape[1])#(spec, wav, phone, pitch)
phone_padded = torch.FloatTensor(
len(batch), max_phone_len, batch[0][2].shape[1]
) # (spec, wav, phone, pitch)
pitch_padded = torch.LongTensor(len(batch), max_phone_len)
pitchf_padded = torch.FloatTensor(len(batch), max_phone_len)
phone_padded.zero_()
@ -187,7 +198,6 @@ class TextAudioCollateMultiNSFsid:
# dv[i] = row[5]
sid[i] = row[5]
return (
phone_padded,
phone_lengths,
@ -198,9 +208,10 @@ class TextAudioCollateMultiNSFsid:
wave_padded,
wave_lengths,
# dv
sid
sid,
)
class TextAudioLoader(torch.utils.data.Dataset):
"""
1) loads audio, text pairs
@ -235,6 +246,7 @@ class TextAudioLoader(torch.utils.data.Dataset):
lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
self.audiopaths_and_text = audiopaths_and_text_new
self.lengths = lengths
def get_sid(self, sid):
sid = torch.LongTensor([int(sid)])
return sid
@ -283,9 +295,14 @@ class TextAudioLoader(torch.utils.data.Dataset):
spec = torch.load(spec_filename)
except:
print(spec_filename, traceback.format_exc())
spec = spectrogram_torch(audio_norm, self.filter_length,
self.sampling_rate, self.hop_length, self.win_length,
center=False)
spec = spectrogram_torch(
audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False,
)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
else:
@ -306,6 +323,8 @@ class TextAudioLoader(torch.utils.data.Dataset):
def __len__(self):
return len(self.audiopaths_and_text)
class TextAudioCollate:
"""Zero-pads model inputs and targets"""
@ -334,7 +353,9 @@ class TextAudioCollate:
max_phone_len = max([x[2].size(0) for x in batch])
phone_lengths = torch.LongTensor(len(batch))
phone_padded = torch.FloatTensor(len(batch), max_phone_len, batch[0][2].shape[1])
phone_padded = torch.FloatTensor(
len(batch), max_phone_len, batch[0][2].shape[1]
)
phone_padded.zero_()
sid = torch.LongTensor(len(batch))
@ -355,7 +376,6 @@ class TextAudioCollate:
sid[i] = row[3]
return (
phone_padded,
phone_lengths,
@ -363,9 +383,10 @@ class TextAudioCollate:
spec_lengths,
wave_padded,
wave_lengths,
sid
sid,
)
class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
"""
Maintain similar input lengths in a batch.

1
train/losses.py
View File

@ -1,6 +1,7 @@
import torch
from torch.nn import functional as F
def feature_loss(fmap_r, fmap_g):
loss = 0
for dr, dg in zip(fmap_r, fmap_g):

17
train/mel_processing.py
View File

@ -78,7 +78,8 @@ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False)
center=center,
pad_mode="reflect",
normalized=False,
onesided=True,return_complex=False
onesided=True,
return_complex=False,
)
spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
@ -139,8 +140,18 @@ def mel_spectrogram_torch(
# normalized=False,
# onesided=True,
# )
spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
spec = torch.stft(
y,
n_fft,
hop_length=hop_size,
win_length=win_size,
window=hann_window[wnsize_dtype_device],
center=center,
pad_mode="reflect",
normalized=False,
onesided=True,
return_complex=False,
)
spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
spec = torch.matmul(mel_basis[fmax_dtype_device], spec)

193
train/process_ckpt.py
View File

@ -1,16 +1,78 @@
import torch, traceback, os, pdb
from collections import OrderedDict
def savee(ckpt, sr, if_f0, name, epoch):
try:
opt = OrderedDict()
opt["weight"] = {}
for key in ckpt.keys():
if ("enc_q" in key): continue
if "enc_q" in key:
continue
opt["weight"][key] = ckpt[key].half()
if(sr=="40k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 10, 2, 2], 512, [16, 16, 4, 4], 109, 256, 40000]
elif(sr=="48k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,6,2,2,2], 512, [16, 16, 4, 4,4], 109, 256, 48000]
elif(sr=="32k"):opt["config"] = [513, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 4, 2, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 32000]
if sr == "40k":
opt["config"] = [
1025,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 10, 2, 2],
512,
[16, 16, 4, 4],
109,
256,
40000,
]
elif sr == "48k":
opt["config"] = [
1025,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 6, 2, 2, 2],
512,
[16, 16, 4, 4, 4],
109,
256,
48000,
]
elif sr == "32k":
opt["config"] = [
513,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 4, 2, 2, 2],
512,
[16, 16, 4, 4, 4],
109,
256,
32000,
]
opt["info"] = "%sepoch" % epoch
opt["sr"] = sr
opt["f0"] = if_f0
@ -19,26 +81,95 @@ def savee(ckpt,sr,if_f0,name,epoch):
except:
return traceback.format_exc()
def show_info(path):
try:
a = torch.load(path, map_location="cpu")
return "模型信息:%s\n采样率:%s\n模型是否输入音高引导:%s"%(a.get("info","None"),a.get("sr","None"),a.get("f0","None"),)
return "模型信息:%s\n采样率:%s\n模型是否输入音高引导:%s" % (
a.get("info", "None"),
a.get("sr", "None"),
a.get("f0", "None"),
)
except:
return traceback.format_exc()
def extract_small_model(path, name, sr, if_f0, info):
try:
ckpt = torch.load(path, map_location="cpu")
if("model"in ckpt):ckpt=ckpt["model"]
if "model" in ckpt:
ckpt = ckpt["model"]
opt = OrderedDict()
opt["weight"] = {}
for key in ckpt.keys():
if ("enc_q" in key): continue
if "enc_q" in key:
continue
opt["weight"][key] = ckpt[key].half()
if(sr=="40k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 10, 2, 2], 512, [16, 16, 4, 4], 109, 256, 40000]
elif(sr=="48k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,6,2,2,2], 512, [16, 16, 4, 4,4], 109, 256, 48000]
elif(sr=="32k"):opt["config"] = [513, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 4, 2, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 32000]
if(info==""):info="Extracted model."
if sr == "40k":
opt["config"] = [
1025,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 10, 2, 2],
512,
[16, 16, 4, 4],
109,
256,
40000,
]
elif sr == "48k":
opt["config"] = [
1025,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 6, 2, 2, 2],
512,
[16, 16, 4, 4, 4],
109,
256,
48000,
]
elif sr == "32k":
opt["config"] = [
513,
32,
192,
192,
768,
2,
6,
3,
0,
"1",
[3, 7, 11],
[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
[10, 4, 2, 2, 2],
512,
[16, 16, 4, 4, 4],
109,
256,
32000,
]
if info == "":
info = "Extracted model."
opt["info"] = info
opt["sr"] = sr
opt["f0"] = int(if_f0)
@ -47,51 +178,67 @@ def extract_small_model(path,name,sr,if_f0,info):
except:
return traceback.format_exc()
def change_info(path, info, name):
try:
ckpt = torch.load(path, map_location="cpu")
ckpt["info"] = info
if(name==""):name=os.path.basename(path)
if name == "":
name = os.path.basename(path)
torch.save(ckpt, "weights/%s" % name)
return "Success."
except:
return traceback.format_exc()
def merge(path1, path2, alpha1, sr, f0, info, name):
try:
def extract(ckpt):
a = ckpt["model"]
opt = OrderedDict()
opt["weight"] = {}
for key in a.keys():
if ("enc_q" in key): continue
if "enc_q" in key:
continue
opt["weight"][key] = a[key]
return opt
ckpt1 = torch.load(path1, map_location="cpu")
ckpt2 = torch.load(path2, map_location="cpu")
cfg = ckpt1["config"]
if("model"in ckpt1): ckpt1=extract(ckpt1)
else: ckpt1=ckpt1["weight"]
if("model"in ckpt2): ckpt2=extract(ckpt2)
else: ckpt2=ckpt2["weight"]
if(sorted(list(ckpt1.keys()))!=sorted(list(ckpt2.keys()))):return "Fail to merge the models. The model architectures are not the same."
if "model" in ckpt1:
ckpt1 = extract(ckpt1)
else:
ckpt1 = ckpt1["weight"]
if "model" in ckpt2:
ckpt2 = extract(ckpt2)
else:
ckpt2 = ckpt2["weight"]
if sorted(list(ckpt1.keys())) != sorted(list(ckpt2.keys())):
return "Fail to merge the models. The model architectures are not the same."
opt = OrderedDict()
opt["weight"] = {}
for key in ckpt1.keys():
# try:
if(key=="emb_g.weight"and ckpt1[key].shape!=ckpt2[key].shape):
if key == "emb_g.weight" and ckpt1[key].shape != ckpt2[key].shape:
min_shape0 = min(ckpt1[key].shape[0], ckpt2[key].shape[0])
opt["weight"][key] = (alpha1 * (ckpt1[key][:min_shape0].float()) + (1 - alpha1) * (ckpt2[key][:min_shape0].float())).half()
opt["weight"][key] = (
alpha1 * (ckpt1[key][:min_shape0].float())
+ (1 - alpha1) * (ckpt2[key][:min_shape0].float())
).half()
else:
opt["weight"][key] = (alpha1*(ckpt1[key].float())+(1-alpha1)*(ckpt2[key].float())).half()
opt["weight"][key] = (
alpha1 * (ckpt1[key].float()) + (1 - alpha1) * (ckpt2[key].float())
).half()
# except:
# pdb.set_trace()
opt["config"] = cfg
'''
"""
if(sr=="40k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 10, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 40000]
elif(sr=="48k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,6,2,2,2], 512, [16, 16, 4, 4], 109, 256, 48000]
elif(sr=="32k"):opt["config"] = [513, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 4, 2, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 32000]
'''
"""
opt["sr"] = sr
opt["f0"] = 1 if f0 == "是" else 0
opt["info"] = info

228
train/utils.py
View File

@ -14,40 +14,49 @@ MATPLOTLIB_FLAG = False
logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
logger = logging
def load_checkpoint_d(checkpoint_path, combd, sbd, optimizer=None, load_opt=1):
assert os.path.isfile(checkpoint_path)
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
##################
def go(model, bkey):
saved_state_dict = checkpoint_dict[bkey]
if hasattr(model, 'module'):state_dict = model.module.state_dict()
else:state_dict = model.state_dict()
if hasattr(model, "module"):
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
new_state_dict = {}
for k, v in state_dict.items(): # 模型需要的shape
try:
new_state_dict[k] = saved_state_dict[k]
if(saved_state_dict[k].shape!=state_dict[k].shape):
print("shape-%s-mismatch|need-%s|get-%s"%(k,state_dict[k].shape,saved_state_dict[k].shape))#
if saved_state_dict[k].shape != state_dict[k].shape:
print(
"shape-%s-mismatch|need-%s|get-%s"
% (k, state_dict[k].shape, saved_state_dict[k].shape)
) #
raise KeyError
except:
# logger.info(traceback.format_exc())
logger.info("%s is not in the checkpoint" % k) # pretrain缺失的
new_state_dict[k] = v # 模型自带的随机值
if hasattr(model, 'module'):
if hasattr(model, "module"):
model.module.load_state_dict(new_state_dict, strict=False)
else:
model.load_state_dict(new_state_dict, strict=False)
go(combd, "combd")
go(sbd, "sbd")
#############
logger.info("Loaded model weights")
iteration = checkpoint_dict['iteration']
learning_rate = checkpoint_dict['learning_rate']
if optimizer is not None and load_opt==1:###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
iteration = checkpoint_dict["iteration"]
learning_rate = checkpoint_dict["learning_rate"]
if (
optimizer is not None and load_opt == 1
): ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
# try:
optimizer.load_state_dict(checkpoint_dict['optimizer'])
optimizer.load_state_dict(checkpoint_dict["optimizer"])
# except:
# traceback.print_exc()
logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
@ -85,10 +94,10 @@ def load_checkpoint_d(checkpoint_path, combd,sbd, optimizer=None,load_opt=1):
# return model, optimizer, learning_rate, iteration
def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
assert os.path.isfile(checkpoint_path)
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
saved_state_dict = checkpoint_dict['model']
if hasattr(model, 'module'):
saved_state_dict = checkpoint_dict["model"]
if hasattr(model, "module"):
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
@ -96,24 +105,29 @@ def load_checkpoint(checkpoint_path, model, optimizer=None,load_opt=1):
for k, v in state_dict.items(): # 模型需要的shape
try:
new_state_dict[k] = saved_state_dict[k]
if(saved_state_dict[k].shape!=state_dict[k].shape):
print("shape-%s-mismatch|need-%s|get-%s"%(k,state_dict[k].shape,saved_state_dict[k].shape))#
if saved_state_dict[k].shape != state_dict[k].shape:
print(
"shape-%s-mismatch|need-%s|get-%s"
% (k, state_dict[k].shape, saved_state_dict[k].shape)
) #
raise KeyError
except:
# logger.info(traceback.format_exc())
logger.info("%s is not in the checkpoint" % k) # pretrain缺失的
new_state_dict[k] = v # 模型自带的随机值
if hasattr(model, 'module'):
if hasattr(model, "module"):
model.module.load_state_dict(new_state_dict, strict=False)
else:
model.load_state_dict(new_state_dict, strict=False)
logger.info("Loaded model weights")
iteration = checkpoint_dict['iteration']
learning_rate = checkpoint_dict['learning_rate']
if optimizer is not None and load_opt==1:###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
iteration = checkpoint_dict["iteration"]
learning_rate = checkpoint_dict["learning_rate"]
if (
optimizer is not None and load_opt == 1
): ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
# try:
optimizer.load_state_dict(checkpoint_dict['optimizer'])
optimizer.load_state_dict(checkpoint_dict["optimizer"])
# except:
# traceback.print_exc()
logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
@ -121,38 +135,67 @@ def load_checkpoint(checkpoint_path, model, optimizer=None,load_opt=1):
def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
logger.info("Saving model and optimizer state at epoch {} to {}".format(
iteration, checkpoint_path))
if hasattr(model, 'module'):
logger.info(
"Saving model and optimizer state at epoch {} to {}".format(
iteration, checkpoint_path
)
)
if hasattr(model, "module"):
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
torch.save({'model': state_dict,
'iteration': iteration,
'optimizer': optimizer.state_dict(),
'learning_rate': learning_rate}, checkpoint_path)
torch.save(
{
"model": state_dict,
"iteration": iteration,
"optimizer": optimizer.state_dict(),
"learning_rate": learning_rate,
},
checkpoint_path,
)
def save_checkpoint_d(combd, sbd, optimizer, learning_rate, iteration, checkpoint_path):
logger.info("Saving model and optimizer state at epoch {} to {}".format(
iteration, checkpoint_path))
if hasattr(combd, 'module'): state_dict_combd = combd.module.state_dict()
else:state_dict_combd = combd.state_dict()
if hasattr(sbd, 'module'): state_dict_sbd = sbd.module.state_dict()
else:state_dict_sbd = sbd.state_dict()
torch.save({
'combd': state_dict_combd,
'sbd': state_dict_sbd,
'iteration': iteration,
'optimizer': optimizer.state_dict(),
'learning_rate': learning_rate}, checkpoint_path)
logger.info(
"Saving model and optimizer state at epoch {} to {}".format(
iteration, checkpoint_path
)
)
if hasattr(combd, "module"):
state_dict_combd = combd.module.state_dict()
else:
state_dict_combd = combd.state_dict()
if hasattr(sbd, "module"):
state_dict_sbd = sbd.module.state_dict()
else:
state_dict_sbd = sbd.state_dict()
torch.save(
{
"combd": state_dict_combd,
"sbd": state_dict_sbd,
"iteration": iteration,
"optimizer": optimizer.state_dict(),
"learning_rate": learning_rate,
},
checkpoint_path,
)
def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050):
def summarize(
writer,
global_step,
scalars={},
histograms={},
images={},
audios={},
audio_sampling_rate=22050,
):
for k, v in scalars.items():
writer.add_scalar(k, v, global_step)
for k, v in histograms.items():
writer.add_histogram(k, v, global_step)
for k, v in images.items():
writer.add_image(k, v, global_step, dataformats='HWC')
writer.add_image(k, v, global_step, dataformats="HWC")
for k, v in audios.items():
writer.add_audio(k, v, global_step, audio_sampling_rate)
@ -169,23 +212,23 @@ def plot_spectrogram_to_numpy(spectrogram):
global MATPLOTLIB_FLAG
if not MATPLOTLIB_FLAG:
import matplotlib
matplotlib.use("Agg")
MATPLOTLIB_FLAG = True
mpl_logger = logging.getLogger('matplotlib')
mpl_logger = logging.getLogger("matplotlib")
mpl_logger.setLevel(logging.WARNING)
import matplotlib.pylab as plt
import numpy as np
fig, ax = plt.subplots(figsize=(10, 2))
im = ax.imshow(spectrogram, aspect="auto", origin="lower",
interpolation='none')
im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
plt.colorbar(im, ax=ax)
plt.xlabel("Frames")
plt.ylabel("Channels")
plt.tight_layout()
fig.canvas.draw()
data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
plt.close()
return data
@ -195,26 +238,28 @@ def plot_alignment_to_numpy(alignment, info=None):
global MATPLOTLIB_FLAG
if not MATPLOTLIB_FLAG:
import matplotlib
matplotlib.use("Agg")
MATPLOTLIB_FLAG = True
mpl_logger = logging.getLogger('matplotlib')
mpl_logger = logging.getLogger("matplotlib")
mpl_logger.setLevel(logging.WARNING)
import matplotlib.pylab as plt
import numpy as np
fig, ax = plt.subplots(figsize=(6, 4))
im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
interpolation='none')
im = ax.imshow(
alignment.transpose(), aspect="auto", origin="lower", interpolation="none"
)
fig.colorbar(im, ax=ax)
xlabel = 'Decoder timestep'
xlabel = "Decoder timestep"
if info is not None:
xlabel += '\n\n' + info
xlabel += "\n\n" + info
plt.xlabel(xlabel)
plt.ylabel('Encoder timestep')
plt.ylabel("Encoder timestep")
plt.tight_layout()
fig.canvas.draw()
data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
plt.close()
return data
@ -226,13 +271,13 @@ def load_wav_to_torch(full_path):
def load_filepaths_and_text(filename, split="|"):
with open(filename, encoding='utf-8') as f:
with open(filename, encoding="utf-8") as f:
filepaths_and_text = [line.strip().split(split) for line in f]
return filepaths_and_text
def get_hparams(init=True):
'''
"""
todo:
结尾七人组：
保存频率、总epoch done
@ -247,20 +292,56 @@ todo:
-m:
自动决定training_files路径,改掉train_nsf_load_pretrain.py里的hps.data.training_files done
-c不要了
'''
"""
parser = argparse.ArgumentParser()
# parser.add_argument('-c', '--config', type=str, default="configs/40k.json",help='JSON file for configuration')
parser.add_argument('-se', '--save_every_epoch', type=int, required=True,help='checkpoint save frequency (epoch)')
parser.add_argument('-te', '--total_epoch', type=int, required=True,help='total_epoch')
parser.add_argument('-pg', '--pretrainG', type=str, default="",help='Pretrained Discriminator path')
parser.add_argument('-pd', '--pretrainD', type=str, default="",help='Pretrained Generator path')
parser.add_argument('-g', '--gpus', type=str, default="0",help='split by -')
parser.add_argument('-bs', '--batch_size', type=int, required=True,help='batch size')
parser.add_argument('-e', '--experiment_dir', type=str, required=True,help='experiment dir')#-m
parser.add_argument('-sr', '--sample_rate', type=str, required=True,help='sample rate, 32k/40k/48k')
parser.add_argument('-f0', '--if_f0', type=int, required=True,help='use f0 as one of the inputs of the model, 1 or 0')
parser.add_argument('-l', '--if_latest', type=int, required=True,help='if only save the latest G/D pth file, 1 or 0')
parser.add_argument('-c', '--if_cache_data_in_gpu', type=int, required=True,help='if caching the dataset in GPU memory, 1 or 0')
parser.add_argument(
"-se",
"--save_every_epoch",
type=int,
required=True,
help="checkpoint save frequency (epoch)",
)
parser.add_argument(
"-te", "--total_epoch", type=int, required=True, help="total_epoch"
)
parser.add_argument(
"-pg", "--pretrainG", type=str, default="", help="Pretrained Discriminator path"
)
parser.add_argument(
"-pd", "--pretrainD", type=str, default="", help="Pretrained Generator path"
)
parser.add_argument("-g", "--gpus", type=str, default="0", help="split by -")
parser.add_argument(
"-bs", "--batch_size", type=int, required=True, help="batch size"
)
parser.add_argument(
"-e", "--experiment_dir", type=str, required=True, help="experiment dir"
) # -m
parser.add_argument(
"-sr", "--sample_rate", type=str, required=True, help="sample rate, 32k/40k/48k"
)
parser.add_argument(
"-f0",
"--if_f0",
type=int,
required=True,
help="use f0 as one of the inputs of the model, 1 or 0",
)
parser.add_argument(
"-l",
"--if_latest",
type=int,
required=True,
help="if only save the latest G/D pth file, 1 or 0",
)
parser.add_argument(
"-c",
"--if_cache_data_in_gpu",
type=int,
required=True,
help="if caching the dataset in GPU memory, 1 or 0",
)
args = parser.parse_args()
name = args.experiment_dir
@ -321,9 +402,11 @@ def get_hparams_from_file(config_path):
def check_git_hash(model_dir):
source_dir = os.path.dirname(os.path.realpath(__file__))
if not os.path.exists(os.path.join(source_dir, ".git")):
logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
logger.warn(
"{} is not a git repository, therefore hash value comparison will be ignored.".format(
source_dir
))
)
)
return
cur_hash = subprocess.getoutput("git rev-parse HEAD")
@ -332,8 +415,11 @@ def check_git_hash(model_dir):
if os.path.exists(path):
saved_hash = open(path).read()
if saved_hash != cur_hash:
logger.warn("git hash values are different. {}(saved) != {}(current)".format(
saved_hash[:8], cur_hash[:8]))
logger.warn(
"git hash values are different. {}(saved) != {}(current)".format(
saved_hash[:8], cur_hash[:8]
)
)
else:
open(path, "w").write(cur_hash)
@ -353,7 +439,7 @@ def get_logger(model_dir, filename="train.log"):
return logger
class HParams():
class HParams:
def __init__(self, **kwargs):
for k, v in kwargs.items():
if type(v) == dict:

275
train_nsf_sim_cache_sid_load_pretrain.py
View File

@ -1,12 +1,15 @@
import sys, os
now_dir = os.getcwd()
sys.path.append(os.path.join(now_dir, "train"))
import utils
hps = utils.get_hparams()
os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
n_gpus = len(hps.gpus.split("-"))
from random import shuffle
import traceback, json, argparse, itertools, math, torch, pdb
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = False
from torch import nn, optim
@ -20,9 +23,16 @@ from torch.cuda.amp import autocast, GradScaler
from infer_pack import commons
from time import time as ttime
from data_utils import TextAudioLoaderMultiNSFsid,TextAudioLoader, TextAudioCollateMultiNSFsid,TextAudioCollate, DistributedBucketSampler
from data_utils import (
TextAudioLoaderMultiNSFsid,
TextAudioLoader,
TextAudioCollateMultiNSFsid,
TextAudioCollate,
DistributedBucketSampler,
)
from infer_pack.models import (
SynthesizerTrnMs256NSFsid,SynthesizerTrnMs256NSFsid_nono,
SynthesizerTrnMs256NSFsid,
SynthesizerTrnMs256NSFsid_nono,
MultiPeriodDiscriminator,
)
from losses import generator_loss, discriminator_loss, feature_loss, kl_loss
@ -32,13 +42,11 @@ from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
global_step = 0
def main():
# n_gpus = torch.cuda.device_count()
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "5555"
mp.spawn(
run,
nprocs=n_gpus,
@ -62,10 +70,13 @@ def run(rank, n_gpus, hps):
backend="gloo", init_method="env://", world_size=n_gpus, rank=rank
)
torch.manual_seed(hps.train.seed)
if torch.cuda.is_available(): torch.cuda.set_device(rank)
if torch.cuda.is_available():
torch.cuda.set_device(rank)
if (hps.if_f0 == 1):train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
else:train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
if hps.if_f0 == 1:
train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
else:
train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
train_sampler = DistributedBucketSampler(
train_dataset,
hps.train.batch_size * n_gpus,
@ -77,8 +88,10 @@ def run(rank, n_gpus, hps):
)
# It is possible that dataloader's workers are out of shared memory. Please try to raise your shared memory limit.
# num_workers=8 -> num_workers=4
if (hps.if_f0 == 1):collate_fn = TextAudioCollateMultiNSFsid()
else:collate_fn = TextAudioCollate()
if hps.if_f0 == 1:
collate_fn = TextAudioCollateMultiNSFsid()
else:
collate_fn = TextAudioCollate()
train_loader = DataLoader(
train_dataset,
num_workers=4,
@ -89,13 +102,26 @@ def run(rank, n_gpus, hps):
persistent_workers=True,
prefetch_factor=8,
)
if(hps.if_f0==1):
net_g = SynthesizerTrnMs256NSFsid(hps.data.filter_length // 2 + 1,hps.train.segment_size // hps.data.hop_length,**hps.model,is_half=hps.train.fp16_run,sr=hps.sample_rate)
if hps.if_f0 == 1:
net_g = SynthesizerTrnMs256NSFsid(
hps.data.filter_length // 2 + 1,
hps.train.segment_size // hps.data.hop_length,
**hps.model,
is_half=hps.train.fp16_run,
sr=hps.sample_rate,
)
else:
net_g = SynthesizerTrnMs256NSFsid_nono(hps.data.filter_length // 2 + 1,hps.train.segment_size // hps.data.hop_length,**hps.model,is_half=hps.train.fp16_run)
if torch.cuda.is_available(): net_g = net_g.cuda(rank)
net_g = SynthesizerTrnMs256NSFsid_nono(
hps.data.filter_length // 2 + 1,
hps.train.segment_size // hps.data.hop_length,
**hps.model,
is_half=hps.train.fp16_run,
)
if torch.cuda.is_available():
net_g = net_g.cuda(rank)
net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm)
if torch.cuda.is_available(): net_d = net_d.cuda(rank)
if torch.cuda.is_available():
net_d = net_d.cuda(rank)
optim_g = torch.optim.AdamW(
net_g.parameters(),
hps.train.learning_rate,
@ -118,11 +144,15 @@ def run(rank, n_gpus, hps):
net_d = DDP(net_d)
try: # 如果能加载自动resume
_, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d) # D多半加载没事
_, _, _, epoch_str = utils.load_checkpoint(
utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d
) # D多半加载没事
if rank == 0:
logger.info("loaded D")
# _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g,load_opt=0)
_, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g)
_, _, _, epoch_str = utils.load_checkpoint(
utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g
)
global_step = (epoch_str - 1) * len(train_loader)
# epoch_str = 1
# global_step = 0
@ -132,8 +162,16 @@ def run(rank, n_gpus, hps):
global_step = 0
if rank == 0:
logger.info("loaded pretrained %s %s" % (hps.pretrainG, hps.pretrainD))
print(net_g.module.load_state_dict(torch.load(hps.pretrainG,map_location="cpu")["model"]))##测试不加载优化器
print(net_d.module.load_state_dict(torch.load(hps.pretrainD,map_location="cpu")["model"]))
print(
net_g.module.load_state_dict(
torch.load(hps.pretrainG, map_location="cpu")["model"]
)
) ##测试不加载优化器
print(
net_d.module.load_state_dict(
torch.load(hps.pretrainD, map_location="cpu")["model"]
)
)
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
@ -157,7 +195,8 @@ def run(rank, n_gpus, hps):
scaler,
[train_loader, None],
logger,
[writer, writer_eval],cache
[writer, writer_eval],
cache,
)
else:
train_and_evaluate(
@ -170,7 +209,8 @@ def run(rank, n_gpus, hps):
scaler,
[train_loader, None],
None,
None,cache
None,
cache,
)
scheduler_g.step()
scheduler_d.step()
@ -190,25 +230,101 @@ def train_and_evaluate(
net_g.train()
net_d.train()
if(cache==[]or hps.if_cache_data_in_gpu==False):#第一个epoch把cache全部填满训练集
if cache == [] or hps.if_cache_data_in_gpu == False: # 第一个epoch把cache全部填满训练集
# print("caching")
for batch_idx, info in enumerate(train_loader):
if (hps.if_f0 == 1):phone,phone_lengths,pitch,pitchf,spec,spec_lengths,wave,wave_lengths,sid=info
else:phone,phone_lengths,spec,spec_lengths,wave,wave_lengths,sid=info
if hps.if_f0 == 1:
(
phone,
phone_lengths,
pitch,
pitchf,
spec,
spec_lengths,
wave,
wave_lengths,
sid,
) = info
else:
phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
if torch.cuda.is_available():
phone, phone_lengths = phone.cuda(rank, non_blocking=True), phone_lengths.cuda(rank, non_blocking=True )
if (hps.if_f0 == 1):pitch,pitchf = pitch.cuda(rank, non_blocking=True),pitchf.cuda(rank, non_blocking=True)
phone, phone_lengths = phone.cuda(
rank, non_blocking=True
), phone_lengths.cuda(rank, non_blocking=True)
if hps.if_f0 == 1:
pitch, pitchf = pitch.cuda(rank, non_blocking=True), pitchf.cuda(
rank, non_blocking=True
)
sid = sid.cuda(rank, non_blocking=True)
spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
wave, wave_lengths = wave.cuda(rank, non_blocking=True), wave_lengths.cuda(rank, non_blocking=True)
if(hps.if_cache_data_in_gpu==True):
if (hps.if_f0 == 1):cache.append((batch_idx, (phone,phone_lengths,pitch,pitchf,spec,spec_lengths,wave,wave_lengths ,sid)))
else:cache.append((batch_idx, (phone,phone_lengths,spec,spec_lengths,wave,wave_lengths ,sid)))
spec, spec_lengths = spec.cuda(
rank, non_blocking=True
), spec_lengths.cuda(rank, non_blocking=True)
wave, wave_lengths = wave.cuda(
rank, non_blocking=True
), wave_lengths.cuda(rank, non_blocking=True)
if hps.if_cache_data_in_gpu == True:
if hps.if_f0 == 1:
cache.append(
(
batch_idx,
(
phone,
phone_lengths,
pitch,
pitchf,
spec,
spec_lengths,
wave,
wave_lengths,
sid,
),
)
)
else:
cache.append(
(
batch_idx,
(
phone,
phone_lengths,
spec,
spec_lengths,
wave,
wave_lengths,
sid,
),
)
)
with autocast(enabled=hps.train.fp16_run):
if (hps.if_f0 == 1):y_hat,ids_slice,x_mask,z_mask,(z, z_p, m_p, logs_p, m_q, logs_q) = net_g(phone, phone_lengths, pitch,pitchf, spec, spec_lengths,sid)
else:y_hat,ids_slice,x_mask,z_mask,(z, z_p, m_p, logs_p, m_q, logs_q) = net_g(phone, phone_lengths, spec, spec_lengths,sid)
mel = spec_to_mel_torch(spec,hps.data.filter_length,hps.data.n_mel_channels,hps.data.sampling_rate,hps.data.mel_fmin,hps.data.mel_fmax,)
y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
if hps.if_f0 == 1:
(
y_hat,
ids_slice,
x_mask,
z_mask,
(z, z_p, m_p, logs_p, m_q, logs_q),
) = net_g(
phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid
)
else:
(
y_hat,
ids_slice,
x_mask,
z_mask,
(z, z_p, m_p, logs_p, m_q, logs_q),
) = net_g(phone, phone_lengths, spec, spec_lengths, sid)
mel = spec_to_mel_torch(
spec,
hps.data.filter_length,
hps.data.n_mel_channels,
hps.data.sampling_rate,
hps.data.mel_fmin,
hps.data.mel_fmax,
)
y_mel = commons.slice_segments(
mel, ids_slice, hps.train.segment_size // hps.data.hop_length
)
with autocast(enabled=False):
y_hat_mel = mel_spectrogram_torch(
y_hat.float().squeeze(1),
@ -220,7 +336,7 @@ def train_and_evaluate(
hps.data.mel_fmin,
hps.data.mel_fmax,
)
if(hps.train.fp16_run==True):
if hps.train.fp16_run == True:
y_hat_mel = y_hat_mel.half()
wave = commons.slice_segments(
wave, ids_slice * hps.data.hop_length, hps.train.segment_size
@ -280,17 +396,27 @@ def train_and_evaluate(
"grad_norm_g": grad_norm_g,
}
scalar_dict.update(
{"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/kl": loss_kl}
{
"loss/g/fm": loss_fm,
"loss/g/mel": loss_mel,
"loss/g/kl": loss_kl,
}
)
scalar_dict.update(
{"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
)
scalar_dict.update(
{"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
{
"loss/d_r/{}".format(i): v
for i, v in enumerate(losses_disc_r)
}
)
scalar_dict.update(
{"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
{
"loss/d_g/{}".format(i): v
for i, v in enumerate(losses_disc_g)
}
)
image_dict = {
"slice/mel_org": utils.plot_spectrogram_to_numpy(
@ -312,7 +438,7 @@ def train_and_evaluate(
global_step += 1
# if global_step % hps.train.eval_interval == 0:
if epoch % hps.save_every_epoch == 0 and rank == 0:
if(hps.if_latest==0):
if hps.if_latest == 0:
utils.save_checkpoint(
net_g,
optim_g,
@ -347,11 +473,39 @@ def train_and_evaluate(
shuffle(cache)
# print("using cache")
for batch_idx, info in cache:
if (hps.if_f0 == 1):phone,phone_lengths,pitch,pitchf,spec,spec_lengths,wave,wave_lengths,sid=info
else:phone,phone_lengths,spec,spec_lengths,wave,wave_lengths,sid=info
if hps.if_f0 == 1:
(
phone,
phone_lengths,
pitch,
pitchf,
spec,
spec_lengths,
wave,
wave_lengths,
sid,
) = info
else:
phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
with autocast(enabled=hps.train.fp16_run):
if (hps.if_f0 == 1):y_hat,ids_slice,x_mask,z_mask,(z, z_p, m_p, logs_p, m_q, logs_q) = net_g(phone, phone_lengths, pitch,pitchf, spec, spec_lengths,sid)
else:y_hat,ids_slice,x_mask,z_mask,(z, z_p, m_p, logs_p, m_q, logs_q) = net_g(phone, phone_lengths, spec, spec_lengths,sid)
if hps.if_f0 == 1:
(
y_hat,
ids_slice,
x_mask,
z_mask,
(z, z_p, m_p, logs_p, m_q, logs_q),
) = net_g(
phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid
)
else:
(
y_hat,
ids_slice,
x_mask,
z_mask,
(z, z_p, m_p, logs_p, m_q, logs_q),
) = net_g(phone, phone_lengths, spec, spec_lengths, sid)
mel = spec_to_mel_torch(
spec,
hps.data.filter_length,
@ -374,7 +528,7 @@ def train_and_evaluate(
hps.data.mel_fmin,
hps.data.mel_fmax,
)
if(hps.train.fp16_run==True):
if hps.train.fp16_run == True:
y_hat_mel = y_hat_mel.half()
wave = commons.slice_segments(
wave, ids_slice * hps.data.hop_length, hps.train.segment_size
@ -435,17 +589,27 @@ def train_and_evaluate(
"grad_norm_g": grad_norm_g,
}
scalar_dict.update(
{"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/kl": loss_kl}
{
"loss/g/fm": loss_fm,
"loss/g/mel": loss_mel,
"loss/g/kl": loss_kl,
}
)
scalar_dict.update(
{"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
)
scalar_dict.update(
{"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
{
"loss/d_r/{}".format(i): v
for i, v in enumerate(losses_disc_r)
}
)
scalar_dict.update(
{"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
{
"loss/d_g/{}".format(i): v
for i, v in enumerate(losses_disc_g)
}
)
image_dict = {
"slice/mel_org": utils.plot_spectrogram_to_numpy(
@ -467,7 +631,7 @@ def train_and_evaluate(
global_step += 1
# if global_step % hps.train.eval_interval == 0:
if epoch % hps.save_every_epoch == 0 and rank == 0:
if(hps.if_latest==0):
if hps.if_latest == 0:
utils.save_checkpoint(
net_g,
optim_g,
@ -498,15 +662,20 @@ def train_and_evaluate(
os.path.join(hps.model_dir, "D_{}.pth".format(2333333)),
)
if rank == 0:
logger.info("====> Epoch: {}".format(epoch))
if(epoch>=hps.total_epoch and rank == 0):
if epoch >= hps.total_epoch and rank == 0:
logger.info("Training is done. The program is closed.")
from process_ckpt import savee # def savee(ckpt,sr,if_f0,name,epoch):
if hasattr(net_g, 'module'):ckpt = net_g.module.state_dict()
else:ckpt = net_g.state_dict()
logger.info("saving final ckpt:%s"%(savee(ckpt,hps.sample_rate,hps.if_f0,hps.name,epoch)))
if hasattr(net_g, "module"):
ckpt = net_g.module.state_dict()
else:
ckpt = net_g.state_dict()
logger.info(
"saving final ckpt:%s"
% (savee(ckpt, hps.sample_rate, hps.if_f0, hps.name, epoch))
)
os._exit(2333333)

47
trainset_preprocess_pipeline_print.py
View File

@ -1,4 +1,5 @@
import sys, os, multiprocessing
now_dir = os.getcwd()
sys.path.append(now_dir)
@ -16,6 +17,8 @@ from my_utils import load_audio
mutex = multiprocessing.Lock()
f = open("%s/preprocess.log" % exp_dir, "a+")
def println(strr):
mutex.acquire()
print(strr)
@ -23,7 +26,8 @@ def println(strr):
f.flush()
mutex.release()
class PreProcess():
class PreProcess:
def __init__(self, sr, exp_dir):
self.slicer = Slicer(
sr=sr,
@ -31,7 +35,7 @@ class PreProcess():
min_length=800,
min_interval=400,
hop_size=15,
max_sil_kept=150
max_sil_kept=150,
)
self.sr = sr
self.per = 3.7
@ -47,10 +51,20 @@ class PreProcess():
os.makedirs(self.wavs16k_dir, exist_ok=True)
def norm_write(self, tmp_audio, idx0, idx1):
tmp_audio = (tmp_audio / np.abs(tmp_audio).max() * (self.max * self.alpha)) + (1 - self.alpha) * tmp_audio
wavfile.write("%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1), self.sr, (tmp_audio*32768).astype(np.int16))
tmp_audio = (tmp_audio / np.abs(tmp_audio).max() * (self.max * self.alpha)) + (
1 - self.alpha
) * tmp_audio
wavfile.write(
"%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1),
self.sr,
(tmp_audio * 32768).astype(np.int16),
)
tmp_audio = librosa.resample(tmp_audio, orig_sr=self.sr, target_sr=16000)
wavfile.write("%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1), 16000, (tmp_audio*32768).astype(np.int16))
wavfile.write(
"%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
16000,
(tmp_audio * 32768).astype(np.int16),
)
def pipeline(self, path, idx0):
try:
@ -58,10 +72,10 @@ class PreProcess():
idx1 = 0
for audio in self.slicer.slice(audio):
i = 0
while (1):
while 1:
start = int(self.sr * (self.per - self.overlap) * i)
i += 1
if (len(audio[start:]) > self.tail * self.sr):
if len(audio[start:]) > self.tail * self.sr:
tmp_audio = audio[start : start + int(self.per * self.sr)]
self.norm_write(tmp_audio, idx0, idx1)
idx1 += 1
@ -79,19 +93,27 @@ class PreProcess():
def pipeline_mp_inp_dir(self, inp_root, n_p):
try:
infos = [("%s/%s" % (inp_root, name), idx) for idx, name in enumerate(sorted(list(os.listdir(inp_root))))]
infos = [
("%s/%s" % (inp_root, name), idx)
for idx, name in enumerate(sorted(list(os.listdir(inp_root))))
]
if noparallel:
for i in range(n_p): self.pipeline_mp(infos[i::n_p])
for i in range(n_p):
self.pipeline_mp(infos[i::n_p])
else:
ps = []
for i in range(n_p):
p=multiprocessing.Process(target=self.pipeline_mp,args=(infos[i::n_p],))
p = multiprocessing.Process(
target=self.pipeline_mp, args=(infos[i::n_p],)
)
p.start()
ps.append(p)
for p in ps:p.join()
for p in ps:
p.join()
except:
println("Fail. %s" % traceback.format_exc())
def preprocess_trainset(inp_root, sr, n_p, exp_dir):
pp = PreProcess(sr, exp_dir)
println("start preprocess")
@ -99,5 +121,6 @@ def preprocess_trainset(inp_root, sr, n_p, exp_dir):
pp.pipeline_mp_inp_dir(inp_root, n_p)
println("end preprocess")
if __name__=='__main__':
if __name__ == "__main__":
preprocess_trainset(inp_root, sr, n_p, exp_dir)

73
uvr5_pack/lib_v5/dataset.py
View File

@ -10,7 +10,6 @@ from uvr5_pack.lib_v5 import spec_utils
class VocalRemoverValidationSet(torch.utils.data.Dataset):
def __init__(self, patch_list):
self.patch_list = patch_list
@ -21,7 +20,7 @@ class VocalRemoverValidationSet(torch.utils.data.Dataset):
path = self.patch_list[idx]
data = np.load(path)
X, y = data['X'], data['y']
X, y = data["X"], data["y"]
X_mag = np.abs(X)
y_mag = np.abs(y)
@ -30,16 +29,22 @@ class VocalRemoverValidationSet(torch.utils.data.Dataset):
def make_pair(mix_dir, inst_dir):
input_exts = ['.wav', '.m4a', '.mp3', '.mp4', '.flac']
input_exts = [".wav", ".m4a", ".mp3", ".mp4", ".flac"]
X_list = sorted([
X_list = sorted(
[
os.path.join(mix_dir, fname)
for fname in os.listdir(mix_dir)
if os.path.splitext(fname)[1] in input_exts])
y_list = sorted([
if os.path.splitext(fname)[1] in input_exts
]
)
y_list = sorted(
[
os.path.join(inst_dir, fname)
for fname in os.listdir(inst_dir)
if os.path.splitext(fname)[1] in input_exts])
if os.path.splitext(fname)[1] in input_exts
]
)
filelist = list(zip(X_list, y_list))
@ -47,10 +52,11 @@ def make_pair(mix_dir, inst_dir):
def train_val_split(dataset_dir, split_mode, val_rate, val_filelist):
if split_mode == 'random':
if split_mode == "random":
filelist = make_pair(
os.path.join(dataset_dir, 'mixtures'),
os.path.join(dataset_dir, 'instruments'))
os.path.join(dataset_dir, "mixtures"),
os.path.join(dataset_dir, "instruments"),
)
random.shuffle(filelist)
@ -60,19 +66,23 @@ def train_val_split(dataset_dir, split_mode, val_rate, val_filelist):
val_filelist = filelist[-val_size:]
else:
train_filelist = [
pair for pair in filelist
if list(pair) not in val_filelist]
elif split_mode == 'subdirs':
pair for pair in filelist if list(pair) not in val_filelist
]
elif split_mode == "subdirs":
if len(val_filelist) != 0:
raise ValueError('The `val_filelist` option is not available in `subdirs` mode')
raise ValueError(
"The `val_filelist` option is not available in `subdirs` mode"
)
train_filelist = make_pair(
os.path.join(dataset_dir, 'training/mixtures'),
os.path.join(dataset_dir, 'training/instruments'))
os.path.join(dataset_dir, "training/mixtures"),
os.path.join(dataset_dir, "training/instruments"),
)
val_filelist = make_pair(
os.path.join(dataset_dir, 'validation/mixtures'),
os.path.join(dataset_dir, 'validation/instruments'))
os.path.join(dataset_dir, "validation/mixtures"),
os.path.join(dataset_dir, "validation/instruments"),
)
return train_filelist, val_filelist
@ -81,7 +91,9 @@ def augment(X, y, reduction_rate, reduction_mask, mixup_rate, mixup_alpha):
perm = np.random.permutation(len(X))
for i, idx in enumerate(tqdm(perm)):
if np.random.uniform() < reduction_rate:
y[idx] = spec_utils.reduce_vocal_aggressively(X[idx], y[idx], reduction_mask)
y[idx] = spec_utils.reduce_vocal_aggressively(
X[idx], y[idx], reduction_mask
)
if np.random.uniform() < 0.5:
# swap channel
@ -116,10 +128,8 @@ def make_padding(width, cropsize, offset):
def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset):
len_dataset = patches * len(filelist)
X_dataset = np.zeros(
(len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
y_dataset = np.zeros(
(len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
X_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
y_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
@ -127,8 +137,8 @@ def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset
X, y = X / coef, y / coef
l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode='constant')
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
starts = np.random.randint(0, X_pad.shape[2] - cropsize, patches)
ends = starts + cropsize
@ -142,7 +152,9 @@ def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset
def make_validation_set(filelist, cropsize, sr, hop_length, n_fft, offset):
patch_list = []
patch_dir = 'cs{}_sr{}_hl{}_nf{}_of{}'.format(cropsize, sr, hop_length, n_fft, offset)
patch_dir = "cs{}_sr{}_hl{}_nf{}_of{}".format(
cropsize, sr, hop_length, n_fft, offset
)
os.makedirs(patch_dir, exist_ok=True)
for i, (X_path, y_path) in enumerate(tqdm(filelist)):
@ -153,18 +165,19 @@ def make_validation_set(filelist, cropsize, sr, hop_length, n_fft, offset):
X, y = X / coef, y / coef
l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode='constant')
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode='constant')
X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
len_dataset = int(np.ceil(X.shape[2] / roi_size))
for j in range(len_dataset):
outpath = os.path.join(patch_dir, '{}_p{}.npz'.format(basename, j))
outpath = os.path.join(patch_dir, "{}_p{}.npz".format(basename, j))
start = j * roi_size
if not os.path.exists(outpath):
np.savez(
outpath,
X=X_pad[:, :, start : start + cropsize],
y=y_pad[:, :, start:start + cropsize])
y=y_pad[:, :, start : start + cropsize],
)
patch_list.append(outpath)
return VocalRemoverValidationSet(patch_list)

50
uvr5_pack/lib_v5/layers.py
View File

@ -6,19 +6,20 @@ from uvr5_pack.lib_v5 import spec_utils
class Conv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(Conv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nout,
nin,
nout,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
bias=False),
bias=False,
),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -26,24 +27,22 @@ class Conv2DBNActiv(nn.Module):
class SeperableConv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(SeperableConv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nin,
nin,
nin,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
groups=nin,
bias=False),
nn.Conv2d(
nin, nout,
kernel_size=1,
bias=False),
bias=False,
),
nn.Conv2d(nin, nout, kernel_size=1, bias=False),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -51,7 +50,6 @@ class SeperableConv2DBNActiv(nn.Module):
class Encoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
super(Encoder, self).__init__()
self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
@ -65,14 +63,15 @@ class Encoder(nn.Module):
class Decoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
def __init__(
self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
):
super(Decoder, self).__init__()
self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
self.dropout = nn.Dropout2d(0.1) if dropout else None
def __call__(self, x, skip=None):
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
if skip is not None:
skip = spec_utils.crop_center(skip, x)
x = torch.cat([x, skip], dim=1)
@ -85,28 +84,31 @@ class Decoder(nn.Module):
class ASPPModule(nn.Module):
def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
super(ASPPModule, self).__init__()
self.conv1 = nn.Sequential(
nn.AdaptiveAvgPool2d((1, None)),
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
)
self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
self.conv3 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
)
self.conv4 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
)
self.conv5 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.bottleneck = nn.Sequential(
Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ),
nn.Dropout2d(0.1)
Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
)
def forward(self, x):
_, _, h, w = x.size()
feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
feat1 = F.interpolate(
self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
)
feat2 = self.conv2(x)
feat3 = self.conv3(x)
feat4 = self.conv4(x)

50
uvr5_pack/lib_v5/layers_123812KB .py
View File

@ -6,19 +6,20 @@ from uvr5_pack.lib_v5 import spec_utils
class Conv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(Conv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nout,
nin,
nout,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
bias=False),
bias=False,
),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -26,24 +27,22 @@ class Conv2DBNActiv(nn.Module):
class SeperableConv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(SeperableConv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nin,
nin,
nin,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
groups=nin,
bias=False),
nn.Conv2d(
nin, nout,
kernel_size=1,
bias=False),
bias=False,
),
nn.Conv2d(nin, nout, kernel_size=1, bias=False),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -51,7 +50,6 @@ class SeperableConv2DBNActiv(nn.Module):
class Encoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
super(Encoder, self).__init__()
self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
@ -65,14 +63,15 @@ class Encoder(nn.Module):
class Decoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
def __init__(
self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
):
super(Decoder, self).__init__()
self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
self.dropout = nn.Dropout2d(0.1) if dropout else None
def __call__(self, x, skip=None):
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
if skip is not None:
skip = spec_utils.crop_center(skip, x)
x = torch.cat([x, skip], dim=1)
@ -85,28 +84,31 @@ class Decoder(nn.Module):
class ASPPModule(nn.Module):
def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
super(ASPPModule, self).__init__()
self.conv1 = nn.Sequential(
nn.AdaptiveAvgPool2d((1, None)),
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
)
self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
self.conv3 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
)
self.conv4 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
)
self.conv5 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.bottleneck = nn.Sequential(
Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ),
nn.Dropout2d(0.1)
Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
)
def forward(self, x):
_, _, h, w = x.size()
feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
feat1 = F.interpolate(
self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
)
feat2 = self.conv2(x)
feat3 = self.conv3(x)
feat4 = self.conv4(x)

50
uvr5_pack/lib_v5/layers_123821KB.py
View File

@ -6,19 +6,20 @@ from uvr5_pack.lib_v5 import spec_utils
class Conv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(Conv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nout,
nin,
nout,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
bias=False),
bias=False,
),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -26,24 +27,22 @@ class Conv2DBNActiv(nn.Module):
class SeperableConv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(SeperableConv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nin,
nin,
nin,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
groups=nin,
bias=False),
nn.Conv2d(
nin, nout,
kernel_size=1,
bias=False),
bias=False,
),
nn.Conv2d(nin, nout, kernel_size=1, bias=False),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -51,7 +50,6 @@ class SeperableConv2DBNActiv(nn.Module):
class Encoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
super(Encoder, self).__init__()
self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
@ -65,14 +63,15 @@ class Encoder(nn.Module):
class Decoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
def __init__(
self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
):
super(Decoder, self).__init__()
self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
self.dropout = nn.Dropout2d(0.1) if dropout else None
def __call__(self, x, skip=None):
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
if skip is not None:
skip = spec_utils.crop_center(skip, x)
x = torch.cat([x, skip], dim=1)
@ -85,28 +84,31 @@ class Decoder(nn.Module):
class ASPPModule(nn.Module):
def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
super(ASPPModule, self).__init__()
self.conv1 = nn.Sequential(
nn.AdaptiveAvgPool2d((1, None)),
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
)
self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
self.conv3 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
)
self.conv4 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
)
self.conv5 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.bottleneck = nn.Sequential(
Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ),
nn.Dropout2d(0.1)
Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
)
def forward(self, x):
_, _, h, w = x.size()
feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
feat1 = F.interpolate(
self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
)
feat2 = self.conv2(x)
feat3 = self.conv3(x)
feat4 = self.conv4(x)

56
uvr5_pack/lib_v5/layers_33966KB.py
View File

@ -6,19 +6,20 @@ from uvr5_pack.lib_v5 import spec_utils
class Conv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(Conv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nout,
nin,
nout,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
bias=False),
bias=False,
),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -26,24 +27,22 @@ class Conv2DBNActiv(nn.Module):
class SeperableConv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(SeperableConv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nin,
nin,
nin,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
groups=nin,
bias=False),
nn.Conv2d(
nin, nout,
kernel_size=1,
bias=False),
bias=False,
),
nn.Conv2d(nin, nout, kernel_size=1, bias=False),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -51,7 +50,6 @@ class SeperableConv2DBNActiv(nn.Module):
class Encoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
super(Encoder, self).__init__()
self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
@ -65,14 +63,15 @@ class Encoder(nn.Module):
class Decoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
def __init__(
self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
):
super(Decoder, self).__init__()
self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
self.dropout = nn.Dropout2d(0.1) if dropout else None
def __call__(self, x, skip=None):
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
if skip is not None:
skip = spec_utils.crop_center(skip, x)
x = torch.cat([x, skip], dim=1)
@ -85,32 +84,37 @@ class Decoder(nn.Module):
class ASPPModule(nn.Module):
def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
super(ASPPModule, self).__init__()
self.conv1 = nn.Sequential(
nn.AdaptiveAvgPool2d((1, None)),
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
)
self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
self.conv3 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
)
self.conv4 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
)
self.conv5 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.conv6 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.conv7 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.bottleneck = nn.Sequential(
Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ),
nn.Dropout2d(0.1)
Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
)
def forward(self, x):
_, _, h, w = x.size()
feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
feat1 = F.interpolate(
self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
)
feat2 = self.conv2(x)
feat3 = self.conv3(x)
feat4 = self.conv4(x)

56
uvr5_pack/lib_v5/layers_537227KB.py
View File

@ -6,19 +6,20 @@ from uvr5_pack.lib_v5 import spec_utils
class Conv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(Conv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nout,
nin,
nout,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
bias=False),
bias=False,
),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -26,24 +27,22 @@ class Conv2DBNActiv(nn.Module):
class SeperableConv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(SeperableConv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nin,
nin,
nin,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
groups=nin,
bias=False),
nn.Conv2d(
nin, nout,
kernel_size=1,
bias=False),
bias=False,
),
nn.Conv2d(nin, nout, kernel_size=1, bias=False),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -51,7 +50,6 @@ class SeperableConv2DBNActiv(nn.Module):
class Encoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
super(Encoder, self).__init__()
self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
@ -65,14 +63,15 @@ class Encoder(nn.Module):
class Decoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
def __init__(
self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
):
super(Decoder, self).__init__()
self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
self.dropout = nn.Dropout2d(0.1) if dropout else None
def __call__(self, x, skip=None):
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
if skip is not None:
skip = spec_utils.crop_center(skip, x)
x = torch.cat([x, skip], dim=1)
@ -85,32 +84,37 @@ class Decoder(nn.Module):
class ASPPModule(nn.Module):
def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
super(ASPPModule, self).__init__()
self.conv1 = nn.Sequential(
nn.AdaptiveAvgPool2d((1, None)),
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
)
self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
self.conv3 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
)
self.conv4 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
)
self.conv5 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.conv6 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.conv7 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.bottleneck = nn.Sequential(
Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ),
nn.Dropout2d(0.1)
Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
)
def forward(self, x):
_, _, h, w = x.size()
feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
feat1 = F.interpolate(
self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
)
feat2 = self.conv2(x)
feat3 = self.conv3(x)
feat4 = self.conv4(x)

56
uvr5_pack/lib_v5/layers_537238KB.py
View File

@ -6,19 +6,20 @@ from uvr5_pack.lib_v5 import spec_utils
class Conv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(Conv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nout,
nin,
nout,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
bias=False),
bias=False,
),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -26,24 +27,22 @@ class Conv2DBNActiv(nn.Module):
class SeperableConv2DBNActiv(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
super(SeperableConv2DBNActiv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(
nin, nin,
nin,
nin,
kernel_size=ksize,
stride=stride,
padding=pad,
dilation=dilation,
groups=nin,
bias=False),
nn.Conv2d(
nin, nout,
kernel_size=1,
bias=False),
bias=False,
),
nn.Conv2d(nin, nout, kernel_size=1, bias=False),
nn.BatchNorm2d(nout),
activ()
activ(),
)
def __call__(self, x):
@ -51,7 +50,6 @@ class SeperableConv2DBNActiv(nn.Module):
class Encoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
super(Encoder, self).__init__()
self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
@ -65,14 +63,15 @@ class Encoder(nn.Module):
class Decoder(nn.Module):
def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False):
def __init__(
self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
):
super(Decoder, self).__init__()
self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
self.dropout = nn.Dropout2d(0.1) if dropout else None
def __call__(self, x, skip=None):
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=True)
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
if skip is not None:
skip = spec_utils.crop_center(skip, x)
x = torch.cat([x, skip], dim=1)
@ -85,32 +84,37 @@ class Decoder(nn.Module):
class ASPPModule(nn.Module):
def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
super(ASPPModule, self).__init__()
self.conv1 = nn.Sequential(
nn.AdaptiveAvgPool2d((1, None)),
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
)
self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
self.conv3 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ)
nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
)
self.conv4 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ)
nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
)
self.conv5 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.conv6 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.conv7 = SeperableConv2DBNActiv(
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ)
nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
)
self.bottleneck = nn.Sequential(
Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ),
nn.Dropout2d(0.1)
Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
)
def forward(self, x):
_, _, h, w = x.size()
feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
feat1 = F.interpolate(
self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
)
feat2 = self.conv2(x)
feat3 = self.conv3(x)
feat4 = self.conv4(x)

69
uvr5_pack/lib_v5/model_param_init.py
View File

@ -3,33 +3,33 @@ import os
import pathlib
default_param = {}
default_param['bins'] = 768
default_param['unstable_bins'] = 9 # training only
default_param['reduction_bins'] = 762 # training only
default_param['sr'] = 44100
default_param['pre_filter_start'] = 757
default_param['pre_filter_stop'] = 768
default_param['band'] = {}
default_param["bins"] = 768
default_param["unstable_bins"] = 9 # training only
default_param["reduction_bins"] = 762 # training only
default_param["sr"] = 44100
default_param["pre_filter_start"] = 757
default_param["pre_filter_stop"] = 768
default_param["band"] = {}
default_param['band'][1] = {
'sr': 11025,
'hl': 128,
'n_fft': 960,
'crop_start': 0,
'crop_stop': 245,
'lpf_start': 61, # inference only
'res_type': 'polyphase'
default_param["band"][1] = {
"sr": 11025,
"hl": 128,
"n_fft": 960,
"crop_start": 0,
"crop_stop": 245,
"lpf_start": 61, # inference only
"res_type": "polyphase",
}
default_param['band'][2] = {
'sr': 44100,
'hl': 512,
'n_fft': 1536,
'crop_start': 24,
'crop_stop': 547,
'hpf_start': 81, # inference only
'res_type': 'sinc_best'
default_param["band"][2] = {
"sr": 44100,
"hl": 512,
"n_fft": 1536,
"crop_start": 24,
"crop_stop": 547,
"hpf_start": 81, # inference only
"res_type": "sinc_best",
}
@ -43,18 +43,27 @@ def int_keys(d):
class ModelParameters(object):
def __init__(self, config_path=''):
if '.pth' == pathlib.Path(config_path).suffix:
def __init__(self, config_path=""):
if ".pth" == pathlib.Path(config_path).suffix:
import zipfile
with zipfile.ZipFile(config_path, 'r') as zip:
self.param = json.loads(zip.read('param.json'), object_pairs_hook=int_keys)
elif '.json' == pathlib.Path(config_path).suffix:
with open(config_path, 'r') as f:
with zipfile.ZipFile(config_path, "r") as zip:
self.param = json.loads(
zip.read("param.json"), object_pairs_hook=int_keys
)
elif ".json" == pathlib.Path(config_path).suffix:
with open(config_path, "r") as f:
self.param = json.loads(f.read(), object_pairs_hook=int_keys)
else:
self.param = default_param
for k in ['mid_side', 'mid_side_b', 'mid_side_b2', 'stereo_w', 'stereo_n', 'reverse']:
for k in [
"mid_side",
"mid_side_b",
"mid_side_b2",
"stereo_w",
"stereo_n",
"reverse",
]:
if not k in self.param:
self.param[k] = False

30
uvr5_pack/lib_v5/nets.py
View File

@ -7,7 +7,6 @@ from uvr5_pack.lib_v5 import spec_utils
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -39,7 +38,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 16)
@ -67,10 +65,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -82,24 +83,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

30
uvr5_pack/lib_v5/nets_123812KB.py
View File

@ -6,7 +6,6 @@ from uvr5_pack.lib_v5 import layers_123821KB as layers
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -38,7 +37,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 32)
@ -66,10 +64,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -81,24 +82,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

30
uvr5_pack/lib_v5/nets_123821KB.py
View File

@ -6,7 +6,6 @@ from uvr5_pack.lib_v5 import layers_123821KB as layers
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -38,7 +37,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 32)
@ -66,10 +64,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -81,24 +82,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

30
uvr5_pack/lib_v5/nets_33966KB.py
View File

@ -6,7 +6,6 @@ from uvr5_pack.lib_v5 import layers_33966KB as layers
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16, 32)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -38,7 +37,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 16)
@ -66,10 +64,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -81,24 +82,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

30
uvr5_pack/lib_v5/nets_537227KB.py
View File

@ -7,7 +7,6 @@ from uvr5_pack.lib_v5 import layers_537238KB as layers
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -39,7 +38,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 64)
@ -67,10 +65,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -82,24 +83,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

30
uvr5_pack/lib_v5/nets_537238KB.py
View File

@ -7,7 +7,6 @@ from uvr5_pack.lib_v5 import layers_537238KB as layers
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -39,7 +38,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 64)
@ -67,10 +65,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -82,24 +83,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

30
uvr5_pack/lib_v5/nets_61968KB.py
View File

@ -6,7 +6,6 @@ from uvr5_pack.lib_v5 import layers_123821KB as layers
class BaseASPPNet(nn.Module):
def __init__(self, nin, ch, dilations=(4, 8, 16)):
super(BaseASPPNet, self).__init__()
self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
@ -38,7 +37,6 @@ class BaseASPPNet(nn.Module):
class CascadedASPPNet(nn.Module):
def __init__(self, n_fft):
super(CascadedASPPNet, self).__init__()
self.stg1_low_band_net = BaseASPPNet(2, 32)
@ -66,10 +64,13 @@ class CascadedASPPNet(nn.Module):
x = x[:, :, : self.max_bin]
bandw = x.size()[2] // 2
aux1 = torch.cat([
aux1 = torch.cat(
[
self.stg1_low_band_net(x[:, :, :bandw]),
self.stg1_high_band_net(x[:, :, bandw:])
], dim=2)
self.stg1_high_band_net(x[:, :, bandw:]),
],
dim=2,
)
h = torch.cat([x, aux1], dim=1)
aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
@ -81,24 +82,33 @@ class CascadedASPPNet(nn.Module):
mask = F.pad(
input=mask,
pad=(0, 0, 0, self.output_bin - mask.size()[2]),
mode='replicate')
mode="replicate",
)
if self.training:
aux1 = torch.sigmoid(self.aux1_out(aux1))
aux1 = F.pad(
input=aux1,
pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
mode='replicate')
mode="replicate",
)
aux2 = torch.sigmoid(self.aux2_out(aux2))
aux2 = F.pad(
input=aux2,
pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
mode='replicate')
mode="replicate",
)
return mask * mix, aux1 * mix, aux2 * mix
else:
if aggressiveness:
mask[:, :, :aggressiveness['split_bin']] = torch.pow(mask[:, :, :aggressiveness['split_bin']], 1 + aggressiveness['value'] / 3)
mask[:, :, aggressiveness['split_bin']:] = torch.pow(mask[:, :, aggressiveness['split_bin']:], 1 + aggressiveness['value'])
mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
mask[:, :, : aggressiveness["split_bin"]],
1 + aggressiveness["value"] / 3,
)
mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
mask[:, :, aggressiveness["split_bin"] :],
1 + aggressiveness["value"],
)
return mask * mix

392
uvr5_pack/lib_v5/spec_utils.py
View File

@ -4,6 +4,7 @@ import soundfile as sf
from tqdm import tqdm
import json, math, hashlib
def crop_center(h1, h2):
h1_shape = h1.size()
h2_shape = h2.size()
@ -11,7 +12,7 @@ def crop_center(h1, h2):
if h1_shape[3] == h2_shape[3]:
return h1
elif h1_shape[3] < h2_shape[3]:
raise ValueError('h1_shape[3] must be greater than h2_shape[3]')
raise ValueError("h1_shape[3] must be greater than h2_shape[3]")
# s_freq = (h2_shape[2] - h1_shape[2]) // 2
# e_freq = s_freq + h1_shape[2]
@ -22,7 +23,9 @@ def crop_center(h1, h2):
return h1
def wave_to_spectrogram(wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False):
def wave_to_spectrogram(
wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False
):
if reverse:
wave_left = np.flip(np.asfortranarray(wave[0]))
wave_right = np.flip(np.asfortranarray(wave[1]))
@ -30,8 +33,8 @@ def wave_to_spectrogram(wave, hop_length, n_fft, mid_side=False, mid_side_b2=Fal
wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
elif mid_side_b2:
wave_left = np.asfortranarray(np.add(wave[1], wave[0] * .5))
wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * .5))
wave_left = np.asfortranarray(np.add(wave[1], wave[0] * 0.5))
wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * 0.5))
else:
wave_left = np.asfortranarray(wave[0])
wave_right = np.asfortranarray(wave[1])
@ -44,7 +47,9 @@ def wave_to_spectrogram(wave, hop_length, n_fft, mid_side=False, mid_side_b2=Fal
return spec
def wave_to_spectrogram_mt(wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False):
def wave_to_spectrogram_mt(
wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False
):
import threading
if reverse:
@ -54,8 +59,8 @@ def wave_to_spectrogram_mt(wave, hop_length, n_fft, mid_side=False, mid_side_b2=
wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
elif mid_side_b2:
wave_left = np.asfortranarray(np.add(wave[1], wave[0] * .5))
wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * .5))
wave_left = np.asfortranarray(np.add(wave[1], wave[0] * 0.5))
wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * 0.5))
else:
wave_left = np.asfortranarray(wave[0])
wave_right = np.asfortranarray(wave[1])
@ -64,7 +69,10 @@ def wave_to_spectrogram_mt(wave, hop_length, n_fft, mid_side=False, mid_side_b2=
global spec_left
spec_left = librosa.stft(**kwargs)
thread = threading.Thread(target=run_thread, kwargs={'y': wave_left, 'n_fft': n_fft, 'hop_length': hop_length})
thread = threading.Thread(
target=run_thread,
kwargs={"y": wave_left, "n_fft": n_fft, "hop_length": hop_length},
)
thread.start()
spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
thread.join()
@ -76,40 +84,50 @@ def wave_to_spectrogram_mt(wave, hop_length, n_fft, mid_side=False, mid_side_b2=
def combine_spectrograms(specs, mp):
l = min([specs[i].shape[2] for i in specs])
spec_c = np.zeros(shape=(2, mp.param['bins'] + 1, l), dtype=np.complex64)
spec_c = np.zeros(shape=(2, mp.param["bins"] + 1, l), dtype=np.complex64)
offset = 0
bands_n = len(mp.param['band'])
bands_n = len(mp.param["band"])
for d in range(1, bands_n + 1):
h = mp.param['band'][d]['crop_stop'] - mp.param['band'][d]['crop_start']
spec_c[:, offset:offset+h, :l] = specs[d][:, mp.param['band'][d]['crop_start']:mp.param['band'][d]['crop_stop'], :l]
h = mp.param["band"][d]["crop_stop"] - mp.param["band"][d]["crop_start"]
spec_c[:, offset : offset + h, :l] = specs[d][
:, mp.param["band"][d]["crop_start"] : mp.param["band"][d]["crop_stop"], :l
]
offset += h
if offset > mp.param['bins']:
raise ValueError('Too much bins')
if offset > mp.param["bins"]:
raise ValueError("Too much bins")
# lowpass fiter
if mp.param['pre_filter_start'] > 0: # and mp.param['band'][bands_n]['res_type'] in ['scipy', 'polyphase']:
if (
mp.param["pre_filter_start"] > 0
): # and mp.param['band'][bands_n]['res_type'] in ['scipy', 'polyphase']:
if bands_n == 1:
spec_c = fft_lp_filter(spec_c, mp.param['pre_filter_start'], mp.param['pre_filter_stop'])
spec_c = fft_lp_filter(
spec_c, mp.param["pre_filter_start"], mp.param["pre_filter_stop"]
)
else:
gp = 1
for b in range(mp.param['pre_filter_start'] + 1, mp.param['pre_filter_stop']):
g = math.pow(10, -(b - mp.param['pre_filter_start']) * (3.5 - gp) / 20.0)
for b in range(
mp.param["pre_filter_start"] + 1, mp.param["pre_filter_stop"]
):
g = math.pow(
10, -(b - mp.param["pre_filter_start"]) * (3.5 - gp) / 20.0
)
gp = g
spec_c[:, b, :] *= g
return np.asfortranarray(spec_c)
def spectrogram_to_image(spec, mode='magnitude'):
if mode == 'magnitude':
def spectrogram_to_image(spec, mode="magnitude"):
if mode == "magnitude":
if np.iscomplexobj(spec):
y = np.abs(spec)
else:
y = spec
y = np.log10(y**2 + 1e-8)
elif mode == 'phase':
elif mode == "phase":
if np.iscomplexobj(spec):
y = np.angle(spec)
else:
@ -121,9 +139,7 @@ def spectrogram_to_image(spec, mode='magnitude'):
if y.ndim == 3:
img = img.transpose(1, 2, 0)
img = np.concatenate([
np.max(img, axis=2, keepdims=True), img
], axis=2)
img = np.concatenate([np.max(img, axis=2, keepdims=True), img], axis=2)
return img
@ -136,12 +152,12 @@ def reduce_vocal_aggressively(X, y, softmask):
v_mask = v_mag_tmp > y_mag_tmp
y_mag = np.clip(y_mag_tmp - v_mag_tmp * v_mask * softmask, 0, np.inf)
return y_mag * np.exp(1.j * np.angle(y))
return y_mag * np.exp(1.0j * np.angle(y))
def mask_silence(mag, ref, thres=0.2, min_range=64, fade_size=32):
if min_range < fade_size * 2:
raise ValueError('min_range must be >= fade_area * 2')
raise ValueError("min_range must be >= fade_area * 2")
mag = mag.copy()
@ -169,7 +185,9 @@ def mask_silence(mag, ref, thres=0.2, min_range=64, fade_size=32):
else:
e += fade_size
mag[:, :, s + fade_size:e - fade_size] += ref[:, :, s + fade_size:e - fade_size]
mag[:, :, s + fade_size : e - fade_size] += ref[
:, :, s + fade_size : e - fade_size
]
old_e = e
return mag
@ -185,15 +203,17 @@ def cache_or_load(mix_path, inst_path, mp):
mix_basename = os.path.splitext(os.path.basename(mix_path))[0]
inst_basename = os.path.splitext(os.path.basename(inst_path))[0]
cache_dir = 'mph{}'.format(hashlib.sha1(json.dumps(mp.param, sort_keys=True).encode('utf-8')).hexdigest())
mix_cache_dir = os.path.join('cache', cache_dir)
inst_cache_dir = os.path.join('cache', cache_dir)
cache_dir = "mph{}".format(
hashlib.sha1(json.dumps(mp.param, sort_keys=True).encode("utf-8")).hexdigest()
)
mix_cache_dir = os.path.join("cache", cache_dir)
inst_cache_dir = os.path.join("cache", cache_dir)
os.makedirs(mix_cache_dir, exist_ok=True)
os.makedirs(inst_cache_dir, exist_ok=True)
mix_cache_path = os.path.join(mix_cache_dir, mix_basename + '.npy')
inst_cache_path = os.path.join(inst_cache_dir, inst_basename + '.npy')
mix_cache_path = os.path.join(mix_cache_dir, mix_basename + ".npy")
inst_cache_path = os.path.join(inst_cache_dir, inst_basename + ".npy")
if os.path.exists(mix_cache_path) and os.path.exists(inst_cache_path):
X_spec_m = np.load(mix_cache_path)
@ -201,22 +221,52 @@ def cache_or_load(mix_path, inst_path, mp):
else:
X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
for d in range(len(mp.param['band']), 0, -1):
bp = mp.param['band'][d]
for d in range(len(mp.param["band"]), 0, -1):
bp = mp.param["band"][d]
if d == len(mp.param['band']): # high-end band
if d == len(mp.param["band"]): # high-end band
X_wave[d], _ = librosa.load(
mix_path, bp['sr'], False, dtype=np.float32, res_type=bp['res_type'])
mix_path, bp["sr"], False, dtype=np.float32, res_type=bp["res_type"]
)
y_wave[d], _ = librosa.load(
inst_path, bp['sr'], False, dtype=np.float32, res_type=bp['res_type'])
inst_path,
bp["sr"],
False,
dtype=np.float32,
res_type=bp["res_type"],
)
else: # lower bands
X_wave[d] = librosa.resample(X_wave[d+1], mp.param['band'][d+1]['sr'], bp['sr'], res_type=bp['res_type'])
y_wave[d] = librosa.resample(y_wave[d+1], mp.param['band'][d+1]['sr'], bp['sr'], res_type=bp['res_type'])
X_wave[d] = librosa.resample(
X_wave[d + 1],
mp.param["band"][d + 1]["sr"],
bp["sr"],
res_type=bp["res_type"],
)
y_wave[d] = librosa.resample(
y_wave[d + 1],
mp.param["band"][d + 1]["sr"],
bp["sr"],
res_type=bp["res_type"],
)
X_wave[d], y_wave[d] = align_wave_head_and_tail(X_wave[d], y_wave[d])
X_spec_s[d] = wave_to_spectrogram(X_wave[d], bp['hl'], bp['n_fft'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse'])
y_spec_s[d] = wave_to_spectrogram(y_wave[d], bp['hl'], bp['n_fft'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse'])
X_spec_s[d] = wave_to_spectrogram(
X_wave[d],
bp["hl"],
bp["n_fft"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
)
y_spec_s[d] = wave_to_spectrogram(
y_wave[d],
bp["hl"],
bp["n_fft"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
)
del X_wave, y_wave
@ -224,7 +274,7 @@ def cache_or_load(mix_path, inst_path, mp):
y_spec_m = combine_spectrograms(y_spec_s, mp)
if X_spec_m.shape != y_spec_m.shape:
raise ValueError('The combined spectrograms are different: ' + mix_path)
raise ValueError("The combined spectrograms are different: " + mix_path)
_, ext = os.path.splitext(mix_path)
@ -244,9 +294,16 @@ def spectrogram_to_wave(spec, hop_length, mid_side, mid_side_b2, reverse):
if reverse:
return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
elif mid_side:
return np.asfortranarray([np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)])
return np.asfortranarray(
[np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)]
)
elif mid_side_b2:
return np.asfortranarray([np.add(wave_right / 1.25, .4 * wave_left), np.subtract(wave_left / 1.25, .4 * wave_right)])
return np.asfortranarray(
[
np.add(wave_right / 1.25, 0.4 * wave_left),
np.subtract(wave_left / 1.25, 0.4 * wave_right),
]
)
else:
return np.asfortranarray([wave_left, wave_right])
@ -261,7 +318,9 @@ def spectrogram_to_wave_mt(spec, hop_length, mid_side, reverse, mid_side_b2):
global wave_left
wave_left = librosa.istft(**kwargs)
thread = threading.Thread(target=run_thread, kwargs={'stft_matrix': spec_left, 'hop_length': hop_length})
thread = threading.Thread(
target=run_thread, kwargs={"stft_matrix": spec_left, "hop_length": hop_length}
)
thread.start()
wave_right = librosa.istft(spec_right, hop_length=hop_length)
thread.join()
@ -269,46 +328,94 @@ def spectrogram_to_wave_mt(spec, hop_length, mid_side, reverse, mid_side_b2):
if reverse:
return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
elif mid_side:
return np.asfortranarray([np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)])
return np.asfortranarray(
[np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)]
)
elif mid_side_b2:
return np.asfortranarray([np.add(wave_right / 1.25, .4 * wave_left), np.subtract(wave_left / 1.25, .4 * wave_right)])
return np.asfortranarray(
[
np.add(wave_right / 1.25, 0.4 * wave_left),
np.subtract(wave_left / 1.25, 0.4 * wave_right),
]
)
else:
return np.asfortranarray([wave_left, wave_right])
def cmb_spectrogram_to_wave(spec_m, mp, extra_bins_h=None, extra_bins=None):
wave_band = {}
bands_n = len(mp.param['band'])
bands_n = len(mp.param["band"])
offset = 0
for d in range(1, bands_n + 1):
bp = mp.param['band'][d]
spec_s = np.ndarray(shape=(2, bp['n_fft'] // 2 + 1, spec_m.shape[2]), dtype=complex)
h = bp['crop_stop'] - bp['crop_start']
spec_s[:, bp['crop_start']:bp['crop_stop'], :] = spec_m[:, offset:offset+h, :]
bp = mp.param["band"][d]
spec_s = np.ndarray(
shape=(2, bp["n_fft"] // 2 + 1, spec_m.shape[2]), dtype=complex
)
h = bp["crop_stop"] - bp["crop_start"]
spec_s[:, bp["crop_start"] : bp["crop_stop"], :] = spec_m[
:, offset : offset + h, :
]
offset += h
if d == bands_n: # higher
if extra_bins_h: # if --high_end_process bypass
max_bin = bp['n_fft'] // 2
spec_s[:, max_bin-extra_bins_h:max_bin, :] = extra_bins[:, :extra_bins_h, :]
if bp['hpf_start'] > 0:
spec_s = fft_hp_filter(spec_s, bp['hpf_start'], bp['hpf_stop'] - 1)
max_bin = bp["n_fft"] // 2
spec_s[:, max_bin - extra_bins_h : max_bin, :] = extra_bins[
:, :extra_bins_h, :
]
if bp["hpf_start"] > 0:
spec_s = fft_hp_filter(spec_s, bp["hpf_start"], bp["hpf_stop"] - 1)
if bands_n == 1:
wave = spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse'])
wave = spectrogram_to_wave(
spec_s,
bp["hl"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
)
else:
wave = np.add(wave, spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse']))
wave = np.add(
wave,
spectrogram_to_wave(
spec_s,
bp["hl"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
),
)
else:
sr = mp.param['band'][d+1]['sr']
sr = mp.param["band"][d + 1]["sr"]
if d == 1: # lower
spec_s = fft_lp_filter(spec_s, bp['lpf_start'], bp['lpf_stop'])
wave = librosa.resample(spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse']), bp['sr'], sr, res_type="sinc_fastest")
spec_s = fft_lp_filter(spec_s, bp["lpf_start"], bp["lpf_stop"])
wave = librosa.resample(
spectrogram_to_wave(
spec_s,
bp["hl"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
),
bp["sr"],
sr,
res_type="sinc_fastest",
)
else: # mid
spec_s = fft_hp_filter(spec_s, bp['hpf_start'], bp['hpf_stop'] - 1)
spec_s = fft_lp_filter(spec_s, bp['lpf_start'], bp['lpf_stop'])
wave2 = np.add(wave, spectrogram_to_wave(spec_s, bp['hl'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse']))
spec_s = fft_hp_filter(spec_s, bp["hpf_start"], bp["hpf_stop"] - 1)
spec_s = fft_lp_filter(spec_s, bp["lpf_start"], bp["lpf_stop"])
wave2 = np.add(
wave,
spectrogram_to_wave(
spec_s,
bp["hl"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
),
)
# wave = librosa.core.resample(wave2, bp['sr'], sr, res_type="sinc_fastest")
wave = librosa.core.resample(wave2, bp['sr'], sr,res_type='scipy')
wave = librosa.core.resample(wave2, bp["sr"], sr, res_type="scipy")
return wave.T
@ -336,14 +443,40 @@ def fft_hp_filter(spec, bin_start, bin_stop):
def mirroring(a, spec_m, input_high_end, mp):
if 'mirroring' == a:
mirror = np.flip(np.abs(spec_m[:, mp.param['pre_filter_start']-10-input_high_end.shape[1]:mp.param['pre_filter_start']-10, :]), 1)
mirror = mirror * np.exp(1.j * np.angle(input_high_end))
if "mirroring" == a:
mirror = np.flip(
np.abs(
spec_m[
:,
mp.param["pre_filter_start"]
- 10
- input_high_end.shape[1] : mp.param["pre_filter_start"]
- 10,
:,
]
),
1,
)
mirror = mirror * np.exp(1.0j * np.angle(input_high_end))
return np.where(np.abs(input_high_end) <= np.abs(mirror), input_high_end, mirror)
return np.where(
np.abs(input_high_end) <= np.abs(mirror), input_high_end, mirror
)
if 'mirroring2' == a:
mirror = np.flip(np.abs(spec_m[:, mp.param['pre_filter_start']-10-input_high_end.shape[1]:mp.param['pre_filter_start']-10, :]), 1)
if "mirroring2" == a:
mirror = np.flip(
np.abs(
spec_m[
:,
mp.param["pre_filter_start"]
- 10
- input_high_end.shape[1] : mp.param["pre_filter_start"]
- 10,
:,
]
),
1,
)
mi = np.multiply(mirror, input_high_end * 1.7)
return np.where(np.abs(input_high_end) <= np.abs(mi), input_high_end, mi)
@ -358,13 +491,14 @@ def ensembling(a, specs):
spec = spec[:, :, :ln]
specs[i] = specs[i][:, :, :ln]
if 'min_mag' == a:
if "min_mag" == a:
spec = np.where(np.abs(specs[i]) <= np.abs(spec), specs[i], spec)
if 'max_mag' == a:
if "max_mag" == a:
spec = np.where(np.abs(specs[i]) >= np.abs(spec), specs[i], spec)
return spec
def stft(wave, nfft, hl):
wave_left = np.asfortranarray(wave[0])
wave_right = np.asfortranarray(wave[1])
@ -374,6 +508,7 @@ def stft(wave, nfft, hl):
return spec
def istft(spec, hl):
spec_left = np.asfortranarray(spec[0])
spec_right = np.asfortranarray(spec[1])
@ -391,20 +526,31 @@ if __name__ == "__main__":
from model_param_init import ModelParameters
p = argparse.ArgumentParser()
p.add_argument('--algorithm', '-a', type=str, choices=['invert', 'invert_p', 'min_mag', 'max_mag', 'deep', 'align'], default='min_mag')
p.add_argument('--model_params', '-m', type=str, default=os.path.join('modelparams', '1band_sr44100_hl512.json'))
p.add_argument('--output_name', '-o', type=str, default='output')
p.add_argument('--vocals_only', '-v', action='store_true')
p.add_argument('input', nargs='+')
p.add_argument(
"--algorithm",
"-a",
type=str,
choices=["invert", "invert_p", "min_mag", "max_mag", "deep", "align"],
default="min_mag",
)
p.add_argument(
"--model_params",
"-m",
type=str,
default=os.path.join("modelparams", "1band_sr44100_hl512.json"),
)
p.add_argument("--output_name", "-o", type=str, default="output")
p.add_argument("--vocals_only", "-v", action="store_true")
p.add_argument("input", nargs="+")
args = p.parse_args()
start_time = time.time()
if args.algorithm.startswith('invert') and len(args.input) != 2:
raise ValueError('There should be two input files.')
if args.algorithm.startswith("invert") and len(args.input) != 2:
raise ValueError("There should be two input files.")
if not args.algorithm.startswith('invert') and len(args.input) < 2:
raise ValueError('There must be at least two input files.')
if not args.algorithm.startswith("invert") and len(args.input) < 2:
raise ValueError("There must be at least two input files.")
wave, specs = {}, {}
mp = ModelParameters(args.model_params)
@ -412,39 +558,60 @@ if __name__ == "__main__":
for i in range(len(args.input)):
spec = {}
for d in range(len(mp.param['band']), 0, -1):
bp = mp.param['band'][d]
for d in range(len(mp.param["band"]), 0, -1):
bp = mp.param["band"][d]
if d == len(mp.param['band']): # high-end band
if d == len(mp.param["band"]): # high-end band
wave[d], _ = librosa.load(
args.input[i], bp['sr'], False, dtype=np.float32, res_type=bp['res_type'])
args.input[i],
bp["sr"],
False,
dtype=np.float32,
res_type=bp["res_type"],
)
if len(wave[d].shape) == 1: # mono to stereo
wave[d] = np.array([wave[d], wave[d]])
else: # lower bands
wave[d] = librosa.resample(wave[d+1], mp.param['band'][d+1]['sr'], bp['sr'], res_type=bp['res_type'])
wave[d] = librosa.resample(
wave[d + 1],
mp.param["band"][d + 1]["sr"],
bp["sr"],
res_type=bp["res_type"],
)
spec[d] = wave_to_spectrogram(wave[d], bp['hl'], bp['n_fft'], mp.param['mid_side'], mp.param['mid_side_b2'], mp.param['reverse'])
spec[d] = wave_to_spectrogram(
wave[d],
bp["hl"],
bp["n_fft"],
mp.param["mid_side"],
mp.param["mid_side_b2"],
mp.param["reverse"],
)
specs[i] = combine_spectrograms(spec, mp)
del wave
if args.algorithm == 'deep':
if args.algorithm == "deep":
d_spec = np.where(np.abs(specs[0]) <= np.abs(spec[1]), specs[0], spec[1])
v_spec = d_spec - specs[1]
sf.write(os.path.join('{}.wav'.format(args.output_name)), cmb_spectrogram_to_wave(v_spec, mp), mp.param['sr'])
sf.write(
os.path.join("{}.wav".format(args.output_name)),
cmb_spectrogram_to_wave(v_spec, mp),
mp.param["sr"],
)
if args.algorithm.startswith('invert'):
if args.algorithm.startswith("invert"):
ln = min([specs[0].shape[2], specs[1].shape[2]])
specs[0] = specs[0][:, :, :ln]
specs[1] = specs[1][:, :, :ln]
if 'invert_p' == args.algorithm:
if "invert_p" == args.algorithm:
X_mag = np.abs(specs[0])
y_mag = np.abs(specs[1])
max_mag = np.where(X_mag >= y_mag, X_mag, y_mag)
v_spec = specs[1] - max_mag * np.exp(1.j * np.angle(specs[0]))
v_spec = specs[1] - max_mag * np.exp(1.0j * np.angle(specs[0]))
else:
specs[1] = reduce_vocal_aggressively(specs[0], specs[1], 0.2)
v_spec = specs[0] - specs[1]
@ -458,24 +625,39 @@ if __name__ == "__main__":
y_image = spectrogram_to_image(y_mag)
v_image = spectrogram_to_image(v_mag)
cv2.imwrite('{}_X.png'.format(args.output_name), X_image)
cv2.imwrite('{}_y.png'.format(args.output_name), y_image)
cv2.imwrite('{}_v.png'.format(args.output_name), v_image)
cv2.imwrite("{}_X.png".format(args.output_name), X_image)
cv2.imwrite("{}_y.png".format(args.output_name), y_image)
cv2.imwrite("{}_v.png".format(args.output_name), v_image)
sf.write('{}_X.wav'.format(args.output_name), cmb_spectrogram_to_wave(specs[0], mp), mp.param['sr'])
sf.write('{}_y.wav'.format(args.output_name), cmb_spectrogram_to_wave(specs[1], mp), mp.param['sr'])
sf.write(
"{}_X.wav".format(args.output_name),
cmb_spectrogram_to_wave(specs[0], mp),
mp.param["sr"],
)
sf.write(
"{}_y.wav".format(args.output_name),
cmb_spectrogram_to_wave(specs[1], mp),
mp.param["sr"],
)
sf.write('{}_v.wav'.format(args.output_name), cmb_spectrogram_to_wave(v_spec, mp), mp.param['sr'])
sf.write(
"{}_v.wav".format(args.output_name),
cmb_spectrogram_to_wave(v_spec, mp),
mp.param["sr"],
)
else:
if not args.algorithm == 'deep':
sf.write(os.path.join('ensembled','{}.wav'.format(args.output_name)), cmb_spectrogram_to_wave(ensembling(args.algorithm, specs), mp), mp.param['sr'])
if args.algorithm == 'align':
if not args.algorithm == "deep":
sf.write(
os.path.join("ensembled", "{}.wav".format(args.output_name)),
cmb_spectrogram_to_wave(ensembling(args.algorithm, specs), mp),
mp.param["sr"],
)
if args.algorithm == "align":
trackalignment = [
{
'file1':'"{}"'.format(args.input[0]),
'file2':'"{}"'.format(args.input[1])
"file1": '"{}"'.format(args.input[0]),
"file2": '"{}"'.format(args.input[1]),
}
]

263
uvr5_pack/name_params.json Normal file
View File

@ -0,0 +1,263 @@
{
"equivalent" : [
{
"model_hash_name" : [
{
"hash_name": "47939caf0cfe52a0e81442b85b971dfd",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100"
},
{
"hash_name": "4e4ecb9764c50a8c414fee6e10395bbe",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_v2.json",
"param_name": "4band_v2"
},
{
"hash_name": "ca106edd563e034bde0bdec4bb7a4b36",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_v2.json",
"param_name": "4band_v2"
},
{
"hash_name": "e60a1e84803ce4efc0a6551206cc4b71",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100"
},
{
"hash_name": "a82f14e75892e55e994376edbf0c8435",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100"
},
{
"hash_name": "6dd9eaa6f0420af9f1d403aaafa4cc06",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
"param_name": "4band_v2_sn"
},
{
"hash_name": "08611fb99bd59eaa79ad27c58d137727",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
"param_name": "4band_v2_sn"
},
{
"hash_name": "5c7bbca45a187e81abbbd351606164e5",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
"param_name": "3band_44100_msb2"
},
{
"hash_name": "d6b2cb685a058a091e5e7098192d3233",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
"param_name": "3band_44100_msb2"
},
{
"hash_name": "c1b9f38170a7c90e96f027992eb7c62b",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100"
},
{
"hash_name": "c3448ec923fa0edf3d03a19e633faa53",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100"
},
{
"hash_name": "68aa2c8093d0080704b200d140f59e54",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100.json",
"param_name": "3band_44100"
},
{
"hash_name": "fdc83be5b798e4bd29fe00fe6600e147",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
"param_name": "3band_44100_mid.json"
},
{
"hash_name": "2ce34bc92fd57f55db16b7a4def3d745",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
"param_name": "3band_44100_mid.json"
},
{
"hash_name": "52fdca89576f06cf4340b74a4730ee5f",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100.json"
},
{
"hash_name": "41191165b05d38fc77f072fa9e8e8a30",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100.json"
},
{
"hash_name": "89e83b511ad474592689e562d5b1f80e",
"model_params": "uvr5_pack/lib_v5/modelparams/2band_32000.json",
"param_name": "2band_32000.json"
},
{
"hash_name": "0b954da81d453b716b114d6d7c95177f",
"model_params": "uvr5_pack/lib_v5/modelparams/2band_32000.json",
"param_name": "2band_32000.json"
}
],
"v4 Models": [
{
"hash_name": "6a00461c51c2920fd68937d4609ed6c8",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json",
"param_name": "1band_sr16000_hl512"
},
{
"hash_name": "0ab504864d20f1bd378fe9c81ef37140",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
"param_name": "1band_sr32000_hl512"
},
{
"hash_name": "7dd21065bf91c10f7fccb57d7d83b07f",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
"param_name": "1band_sr32000_hl512"
},
{
"hash_name": "80ab74d65e515caa3622728d2de07d23",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
"param_name": "1band_sr32000_hl512"
},
{
"hash_name": "edc115e7fc523245062200c00caa847f",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
"param_name": "1band_sr33075_hl384"
},
{
"hash_name": "28063e9f6ab5b341c5f6d3c67f2045b7",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
"param_name": "1band_sr33075_hl384"
},
{
"hash_name": "b58090534c52cbc3e9b5104bad666ef2",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
"param_name": "1band_sr44100_hl512"
},
{
"hash_name": "0cdab9947f1b0928705f518f3c78ea8f",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
"param_name": "1band_sr44100_hl512"
},
{
"hash_name": "ae702fed0238afb5346db8356fe25f13",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json",
"param_name": "1band_sr44100_hl1024"
}
]
}
],
"User Models" : [
{
"1 Band": [
{
"hash_name": "1band_sr16000_hl512",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json",
"param_name": "1band_sr16000_hl512"
},
{
"hash_name": "1band_sr32000_hl512",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
"param_name": "1band_sr16000_hl512"
},
{
"hash_name": "1band_sr33075_hl384",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
"param_name": "1band_sr33075_hl384"
},
{
"hash_name": "1band_sr44100_hl256",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json",
"param_name": "1band_sr44100_hl256"
},
{
"hash_name": "1band_sr44100_hl512",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
"param_name": "1band_sr44100_hl512"
},
{
"hash_name": "1band_sr44100_hl1024",
"model_params": "uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json",
"param_name": "1band_sr44100_hl1024"
}
],
"2 Band": [
{
"hash_name": "2band_44100_lofi",
"model_params": "uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json",
"param_name": "2band_44100_lofi"
},
{
"hash_name": "2band_32000",
"model_params": "uvr5_pack/lib_v5/modelparams/2band_32000.json",
"param_name": "2band_32000"
},
{
"hash_name": "2band_48000",
"model_params": "uvr5_pack/lib_v5/modelparams/2band_48000.json",
"param_name": "2band_48000"
}
],
"3 Band": [
{
"hash_name": "3band_44100",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100.json",
"param_name": "3band_44100"
},
{
"hash_name": "3band_44100_mid",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
"param_name": "3band_44100_mid"
},
{
"hash_name": "3band_44100_msb2",
"model_params": "uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
"param_name": "3band_44100_msb2"
}
],
"4 Band": [
{
"hash_name": "4band_44100",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100.json",
"param_name": "4band_44100"
},
{
"hash_name": "4band_44100_mid",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100_mid.json",
"param_name": "4band_44100_mid"
},
{
"hash_name": "4band_44100_msb",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100_msb.json",
"param_name": "4band_44100_msb"
},
{
"hash_name": "4band_44100_msb2",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json",
"param_name": "4band_44100_msb2"
},
{
"hash_name": "4band_44100_reverse",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json",
"param_name": "4band_44100_reverse"
},
{
"hash_name": "4band_44100_sw",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_44100_sw.json",
"param_name": "4band_44100_sw"
},
{
"hash_name": "4band_v2",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_v2.json",
"param_name": "4band_v2"
},
{
"hash_name": "4band_v2_sn",
"model_params": "uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
"param_name": "4band_v2_sn"
},
{
"hash_name": "tmodelparam",
"model_params": "uvr5_pack/lib_v5/modelparams/tmodelparam.json",
"param_name": "User Model Param Set"
}
]
}
]
}

232
uvr5_pack/utils.py
View File

@ -1,6 +1,15 @@
import torch
import numpy as np
from tqdm import tqdm
import json
def load_data(file_name: str = "./uvr5_pack/data.json") -> dict:
with open(file_name, "r") as f:
data = json.load(f)
return data
def make_padding(width, cropsize, offset):
left = offset
@ -10,12 +19,16 @@ def make_padding(width, cropsize, offset):
right = roi_size - (width % roi_size) + left
return left, right, roi_size
def inference(X_spec, device, model, aggressiveness,data):
'''
data ： dic configs
'''
def _execute(X_mag_pad, roi_size, n_window, device, model, aggressiveness,is_half=True):
def inference(X_spec, device, model, aggressiveness, data):
"""
data ： dic configs
"""
def _execute(
X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half=True
):
model.eval()
with torch.no_grad():
preds = []
@ -25,9 +38,12 @@ def inference(X_spec, device, model, aggressiveness,data):
total_iterations = sum(iterations)
for i in tqdm(range(n_window)):
start = i * roi_size
X_mag_window = X_mag_pad[None, :, :, start:start + data['window_size']]
X_mag_window = X_mag_pad[
None, :, :, start : start + data["window_size"]
]
X_mag_window = torch.from_numpy(X_mag_window)
if(is_half):X_mag_window=X_mag_window.half()
if is_half:
X_mag_window = X_mag_window.half()
X_mag_window = X_mag_window.to(device)
pred = model.predict(X_mag_window, aggressiveness)
@ -50,193 +66,55 @@ def inference(X_spec, device, model, aggressiveness,data):
X_mag_pre = X_mag / coef
n_frame = X_mag_pre.shape[2]
pad_l, pad_r, roi_size = make_padding(n_frame,
data['window_size'], model.offset)
pad_l, pad_r, roi_size = make_padding(n_frame, data["window_size"], model.offset)
n_window = int(np.ceil(n_frame / roi_size))
X_mag_pad = np.pad(
X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode="constant")
if(list(model.state_dict().values())[0].dtype==torch.float16):is_half=True
else:is_half=False
pred = _execute(X_mag_pad, roi_size, n_window,
device, model, aggressiveness,is_half)
if list(model.state_dict().values())[0].dtype == torch.float16:
is_half = True
else:
is_half = False
pred = _execute(
X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half
)
pred = pred[:, :, :n_frame]
if data['tta']:
if data["tta"]:
pad_l += roi_size // 2
pad_r += roi_size // 2
n_window += 1
X_mag_pad = np.pad(
X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode="constant")
pred_tta = _execute(X_mag_pad, roi_size, n_window,
device, model, aggressiveness,is_half)
pred_tta = _execute(
X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half
)
pred_tta = pred_tta[:, :, roi_size // 2 :]
pred_tta = pred_tta[:, :, :n_frame]
return (pred + pred_tta) * 0.5 * coef, X_mag, np.exp(1.j * X_phase)
return (pred + pred_tta) * 0.5 * coef, X_mag, np.exp(1.0j * X_phase)
else:
return pred * coef, X_mag, np.exp(1.j * X_phase)
return pred * coef, X_mag, np.exp(1.0j * X_phase)
def _get_name_params(model_path, model_hash):
data = load_data()
flag = False
ModelName = model_path
if model_hash == '47939caf0cfe52a0e81442b85b971dfd':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100')
if model_hash == '4e4ecb9764c50a8c414fee6e10395bbe':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_v2.json')
param_name_auto=str('4band_v2')
if model_hash == 'ca106edd563e034bde0bdec4bb7a4b36':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_v2.json')
param_name_auto=str('4band_v2')
if model_hash == 'e60a1e84803ce4efc0a6551206cc4b71':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100')
if model_hash == 'a82f14e75892e55e994376edbf0c8435':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100')
if model_hash == '6dd9eaa6f0420af9f1d403aaafa4cc06':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_v2_sn.json')
param_name_auto=str('4band_v2_sn')
if model_hash == '08611fb99bd59eaa79ad27c58d137727':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_v2_sn.json')
param_name_auto=str('4band_v2_sn')
if model_hash == '5c7bbca45a187e81abbbd351606164e5':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json')
param_name_auto=str('3band_44100_msb2')
if model_hash == 'd6b2cb685a058a091e5e7098192d3233':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json')
param_name_auto=str('3band_44100_msb2')
if model_hash == 'c1b9f38170a7c90e96f027992eb7c62b':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100')
if model_hash == 'c3448ec923fa0edf3d03a19e633faa53':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100')
if model_hash == '68aa2c8093d0080704b200d140f59e54':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100.json')
param_name_auto=str('3band_44100.json')
if model_hash == 'fdc83be5b798e4bd29fe00fe6600e147':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100_mid.json')
param_name_auto=str('3band_44100_mid.json')
if model_hash == '2ce34bc92fd57f55db16b7a4def3d745':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100_mid.json')
param_name_auto=str('3band_44100_mid.json')
if model_hash == '52fdca89576f06cf4340b74a4730ee5f':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100.json')
if model_hash == '41191165b05d38fc77f072fa9e8e8a30':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100.json')
if model_hash == '89e83b511ad474592689e562d5b1f80e':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/2band_32000.json')
param_name_auto=str('2band_32000.json')
if model_hash == '0b954da81d453b716b114d6d7c95177f':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/2band_32000.json')
param_name_auto=str('2band_32000.json')
for type in list(data):
for model in list(data[type][0]):
for i in range(len(data[type][0][model])):
if str(data[type][0][model][i]["hash_name"]) == model_hash:
flag = True
elif str(data[type][0][model][i]["hash_name"]) in ModelName:
flag = True
#v4 Models
if model_hash == '6a00461c51c2920fd68937d4609ed6c8':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json')
param_name_auto=str('1band_sr16000_hl512')
if model_hash == '0ab504864d20f1bd378fe9c81ef37140':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json')
param_name_auto=str('1band_sr32000_hl512')
if model_hash == '7dd21065bf91c10f7fccb57d7d83b07f':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json')
param_name_auto=str('1band_sr32000_hl512')
if model_hash == '80ab74d65e515caa3622728d2de07d23':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json')
param_name_auto=str('1band_sr32000_hl512')
if model_hash == 'edc115e7fc523245062200c00caa847f':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json')
param_name_auto=str('1band_sr33075_hl384')
if model_hash == '28063e9f6ab5b341c5f6d3c67f2045b7':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json')
param_name_auto=str('1band_sr33075_hl384')
if model_hash == 'b58090534c52cbc3e9b5104bad666ef2':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json')
param_name_auto=str('1band_sr44100_hl512')
if model_hash == '0cdab9947f1b0928705f518f3c78ea8f':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json')
param_name_auto=str('1band_sr44100_hl512')
if model_hash == 'ae702fed0238afb5346db8356fe25f13':
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json')
param_name_auto=str('1band_sr44100_hl1024')
#User Models
#1 Band
if '1band_sr16000_hl512' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json')
param_name_auto=str('1band_sr16000_hl512')
if '1band_sr32000_hl512' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json')
param_name_auto=str('1band_sr32000_hl512')
if '1band_sr33075_hl384' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json')
param_name_auto=str('1band_sr33075_hl384')
if '1band_sr44100_hl256' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json')
param_name_auto=str('1band_sr44100_hl256')
if '1band_sr44100_hl512' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json')
param_name_auto=str('1band_sr44100_hl512')
if '1band_sr44100_hl1024' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json')
param_name_auto=str('1band_sr44100_hl1024')
#2 Band
if '2band_44100_lofi' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json')
param_name_auto=str('2band_44100_lofi')
if '2band_32000' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/2band_32000.json')
param_name_auto=str('2band_32000')
if '2band_48000' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/2band_48000.json')
param_name_auto=str('2band_48000')
#3 Band
if '3band_44100' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100.json')
param_name_auto=str('3band_44100')
if '3band_44100_mid' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100_mid.json')
param_name_auto=str('3band_44100_mid')
if '3band_44100_msb2' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json')
param_name_auto=str('3band_44100_msb2')
#4 Band
if '4band_44100' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100.json')
param_name_auto=str('4band_44100')
if '4band_44100_mid' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100_mid.json')
param_name_auto=str('4band_44100_mid')
if '4band_44100_msb' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100_msb.json')
param_name_auto=str('4band_44100_msb')
if '4band_44100_msb2' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json')
param_name_auto=str('4band_44100_msb2')
if '4band_44100_reverse' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json')
param_name_auto=str('4band_44100_reverse')
if '4band_44100_sw' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_44100_sw.json')
param_name_auto=str('4band_44100_sw')
if '4band_v2' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_v2.json')
param_name_auto=str('4band_v2')
if '4band_v2_sn' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/4band_v2_sn.json')
param_name_auto=str('4band_v2_sn')
if 'tmodelparam' in ModelName:
model_params_auto=str('uvr5_pack/lib_v5/modelparams/tmodelparam.json')
param_name_auto=str('User Model Param Set')
if flag:
model_params_auto = data[type][0][model][i]["model_params"]
param_name_auto = data[type][0][model][i]["param_name"]
if type == "equivalent":
return param_name_auto, model_params_auto
else:
flag = False
return param_name_auto, model_params_auto

224
vc_infer_pipeline.py
View File

@ -4,6 +4,8 @@ import torch.nn.functional as F
from config import x_pad, x_query, x_center, x_max
import scipy.signal as signal
import pyworld, os, traceback, faiss
class VC(object):
def __init__(self, tgt_sr, device, is_half):
self.sr = 16000 # hubert输入采样率
@ -23,14 +25,23 @@ class VC(object):
f0_max = 1100
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
if(f0_method=="pm"):
f0 = parselmouth.Sound(x, self.sr).to_pitch_ac(
time_step=time_step / 1000, voicing_threshold=0.6,
pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
if f0_method == "pm":
f0 = (
parselmouth.Sound(x, self.sr)
.to_pitch_ac(
time_step=time_step / 1000,
voicing_threshold=0.6,
pitch_floor=f0_min,
pitch_ceiling=f0_max,
)
.selected_array["frequency"]
)
pad_size = (p_len - len(f0) + 1) // 2
if(pad_size>0 or p_len - len(f0) - pad_size>0):
f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
elif(f0_method=="harvest"):
if pad_size > 0 or p_len - len(f0) - pad_size > 0:
f0 = np.pad(
f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
)
elif f0_method == "harvest":
f0, t = pyworld.harvest(
x.astype(np.double),
fs=self.sr,
@ -43,24 +54,44 @@ class VC(object):
f0 *= pow(2, f0_up_key / 12)
# with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
tf0 = self.sr // self.window # 每秒f0点数
if (inp_f0 is not None):
delta_t=np.round((inp_f0[:,0].max()-inp_f0[:,0].min())*tf0+1).astype("int16")
replace_f0=np.interp(list(range(delta_t)), inp_f0[:, 0]*100, inp_f0[:, 1])
if inp_f0 is not None:
delta_t = np.round(
(inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
).astype("int16")
replace_f0 = np.interp(
list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
)
shape = f0[x_pad * tf0 : x_pad * tf0 + len(replace_f0)].shape[0]
f0[x_pad * tf0 : x_pad * tf0 + len(replace_f0)] = replace_f0[:shape]
# with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
f0bak = f0.copy()
f0_mel = 1127 * np.log(1 + f0 / 700)
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
f0_mel_max - f0_mel_min
) + 1
f0_mel[f0_mel <= 1] = 1
f0_mel[f0_mel > 255] = 255
f0_coarse = np.rint(f0_mel).astype(np.int)
return f0_coarse, f0bak # 1-0
def vc(self,model,net_g,sid,audio0,pitch,pitchf,times,index,big_npy,index_rate):#,file_index,file_big_npy
def vc(
self,
model,
net_g,
sid,
audio0,
pitch,
pitchf,
times,
index,
big_npy,
index_rate,
): # ,file_index,file_big_npy
feats = torch.from_numpy(audio0)
if(self.is_half):feats=feats.half()
else:feats=feats.float()
if self.is_half:
feats = feats.half()
else:
feats = feats.float()
if feats.dim() == 2: # double channels
feats = feats.mean(-1)
assert feats.dim() == 1, feats.dim()
@ -77,37 +108,79 @@ class VC(object):
logits = model.extract_features(**inputs)
feats = model.final_proj(logits[0])
if(isinstance(index,type(None))==False and isinstance(big_npy,type(None))==False and index_rate!=0):
if (
isinstance(index, type(None)) == False
and isinstance(big_npy, type(None)) == False
and index_rate != 0
):
npy = feats[0].cpu().numpy()
if(self.is_half):npy=npy.astype("float32")
if self.is_half:
npy = npy.astype("float32")
_, I = index.search(npy, 1)
npy = big_npy[I.squeeze()]
if(self.is_half):npy=npy.astype("float16")
feats = torch.from_numpy(npy).unsqueeze(0).to(self.device)*index_rate + (1-index_rate)*feats
if self.is_half:
npy = npy.astype("float16")
feats = (
torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
+ (1 - index_rate) * feats
)
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
t1 = ttime()
p_len = audio0.shape[0] // self.window
if(feats.shape[1]<p_len):
if feats.shape[1] < p_len:
p_len = feats.shape[1]
if(pitch!=None and pitchf!=None):
if pitch != None and pitchf != None:
pitch = pitch[:, :p_len]
pitchf = pitchf[:, :p_len]
p_len = torch.tensor([p_len], device=self.device).long()
with torch.no_grad():
if(pitch!=None and pitchf!=None):
audio1 = (net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0] * 32768).data.cpu().float().numpy().astype(np.int16)
if pitch != None and pitchf != None:
audio1 = (
(net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0] * 32768)
.data.cpu()
.float()
.numpy()
.astype(np.int16)
)
else:
audio1 = (net_g.infer(feats, p_len, sid)[0][0, 0] * 32768).data.cpu().float().numpy().astype(np.int16)
audio1 = (
(net_g.infer(feats, p_len, sid)[0][0, 0] * 32768)
.data.cpu()
.float()
.numpy()
.astype(np.int16)
)
del feats, p_len, padding_mask
if torch.cuda.is_available(): torch.cuda.empty_cache()
if torch.cuda.is_available():
torch.cuda.empty_cache()
t2 = ttime()
times[0] += (t1 - t0)
times[2] += (t2 - t1)
times[0] += t1 - t0
times[2] += t2 - t1
return audio1
def pipeline(self,model,net_g,sid,audio,times,f0_up_key,f0_method,file_index,file_big_npy,index_rate,if_f0,f0_file=None):
if(file_big_npy!=""and file_index!=""and os.path.exists(file_big_npy)==True and os.path.exists(file_index)==True and index_rate!=0):
def pipeline(
self,
model,
net_g,
sid,
audio,
times,
f0_up_key,
f0_method,
file_index,
file_big_npy,
index_rate,
if_f0,
f0_file=None,
):
if (
file_big_npy != ""
and file_index != ""
and os.path.exists(file_big_npy) == True
and os.path.exists(file_index) == True
and index_rate != 0
):
try:
index = faiss.read_index(file_index)
big_npy = np.load(file_big_npy)
@ -116,50 +189,113 @@ class VC(object):
index = big_npy = None
else:
index = big_npy = None
audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode='reflect')
audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
opt_ts = []
if(audio_pad.shape[0]>self.t_max):
if audio_pad.shape[0] > self.t_max:
audio_sum = np.zeros_like(audio)
for i in range(self.window): audio_sum += audio_pad[i:i - self.window]
for t in range(self.t_center, audio.shape[0],self.t_center):opt_ts.append(t - self.t_query + np.where(np.abs(audio_sum[t - self.t_query:t + self.t_query]) == np.abs(audio_sum[t - self.t_query:t + self.t_query]).min())[0][0])
for i in range(self.window):
audio_sum += audio_pad[i : i - self.window]
for t in range(self.t_center, audio.shape[0], self.t_center):
opt_ts.append(
t
- self.t_query
+ np.where(
np.abs(audio_sum[t - self.t_query : t + self.t_query])
== np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
)[0][0]
)
s = 0
audio_opt = []
t = None
t1 = ttime()
audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode='reflect')
audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
p_len = audio_pad.shape[0] // self.window
inp_f0 = None
if(hasattr(f0_file,'name') ==True):
if hasattr(f0_file, "name") == True:
try:
with open(f0_file.name, "r") as f:
lines = f.read().strip("\n").split("\n")
inp_f0 = []
for line in lines:inp_f0.append([float(i)for i in line.split(",")])
for line in lines:
inp_f0.append([float(i) for i in line.split(",")])
inp_f0 = np.array(inp_f0, dtype="float32")
except:
traceback.print_exc()
sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
pitch, pitchf = None, None
if(if_f0==1):
if if_f0 == 1:
pitch, pitchf = self.get_f0(audio_pad, p_len, f0_up_key, f0_method, inp_f0)
pitch = pitch[:p_len]
pitchf = pitchf[:p_len]
pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
t2 = ttime()
times[1] += (t2 - t1)
times[1] += t2 - t1
for t in opt_ts:
t = t // self.window * self.window
if (if_f0 == 1):
audio_opt.append(self.vc(model,net_g,sid,audio_pad[s:t+self.t_pad2+self.window],pitch[:,s//self.window:(t+self.t_pad2)//self.window],pitchf[:,s//self.window:(t+self.t_pad2)//self.window],times,index,big_npy,index_rate)[self.t_pad_tgt:-self.t_pad_tgt])
if if_f0 == 1:
audio_opt.append(
self.vc(
model,
net_g,
sid,
audio_pad[s : t + self.t_pad2 + self.window],
pitch[:, s // self.window : (t + self.t_pad2) // self.window],
pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
times,
index,
big_npy,
index_rate,
)[self.t_pad_tgt : -self.t_pad_tgt]
)
else:
audio_opt.append(self.vc(model,net_g,sid,audio_pad[s:t+self.t_pad2+self.window],None,None,times,index,big_npy,index_rate)[self.t_pad_tgt:-self.t_pad_tgt])
audio_opt.append(
self.vc(
model,
net_g,
sid,
audio_pad[s : t + self.t_pad2 + self.window],
None,
None,
times,
index,
big_npy,
index_rate,
)[self.t_pad_tgt : -self.t_pad_tgt]
)
s = t
if (if_f0 == 1):
audio_opt.append(self.vc(model,net_g,sid,audio_pad[t:],pitch[:,t//self.window:]if t is not None else pitch,pitchf[:,t//self.window:]if t is not None else pitchf,times,index,big_npy,index_rate)[self.t_pad_tgt:-self.t_pad_tgt])
if if_f0 == 1:
audio_opt.append(
self.vc(
model,
net_g,
sid,
audio_pad[t:],
pitch[:, t // self.window :] if t is not None else pitch,
pitchf[:, t // self.window :] if t is not None else pitchf,
times,
index,
big_npy,
index_rate,
)[self.t_pad_tgt : -self.t_pad_tgt]
)
else:
audio_opt.append(self.vc(model,net_g,sid,audio_pad[t:],None,None,times,index,big_npy,index_rate)[self.t_pad_tgt:-self.t_pad_tgt])
audio_opt.append(
self.vc(
model,
net_g,
sid,
audio_pad[t:],
None,
None,
times,
index,
big_npy,
index_rate,
)[self.t_pad_tgt : -self.t_pad_tgt]
)
audio_opt = np.concatenate(audio_opt)
del pitch, pitchf, sid
if torch.cuda.is_available(): torch.cuda.empty_cache()
if torch.cuda.is_available():
torch.cuda.empty_cache()
return audio_opt

6
webui_locale.py
View File

@ -1,16 +1,18 @@
import locale
import json
def load_language_list(language):
with open(f"./locale/{language}.json", "r", encoding="utf-8") as f:
language_list = json.load(f)
return language_list
class I18nAuto:
def __init__(self, language=None):
if language is None:
language = 'auto'
if language == 'auto':
language = "auto"
if language == "auto":
language = locale.getdefaultlocale()[0]
self.language = language
print("Use Language:", language)