blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2025-10-07 07:02:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge 0f960d2f7f79b6a979af9e9a38cb0269be083c3d into 7881c01c55f2e2913be17323ad69edd3b5ae8b9c

Browse Source
This commit is contained in:
jax 2024-02-20 09:59:52 +08:00 committed by GitHub
commit faac509040
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 1242 additions and 1224 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

605
GPT_SoVITS/inference_base.py Normal file
View File

@ -0,0 +1,605 @@
'''
按中英混合识别
按日英混合识别
多语种启动切分识别语种
全部按中文识别
全部按英文识别
全部按日文识别
'''
import os, re, logging
import LangSegment
logging.getLogger("markdown_it").setLevel(logging.ERROR)
logging.getLogger("urllib3").setLevel(logging.ERROR)
logging.getLogger("httpcore").setLevel(logging.ERROR)
logging.getLogger("httpx").setLevel(logging.ERROR)
logging.getLogger("asyncio").setLevel(logging.ERROR)
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
import pdb
if os.path.exists("./gweight.txt"):
with open("./gweight.txt", 'r', encoding="utf-8") as file:
gweight_data = file.read()
gpt_path = os.environ.get(
"gpt_path", gweight_data)
else:
gpt_path = os.environ.get(
"gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt")
if os.path.exists("./sweight.txt"):
with open("./sweight.txt", 'r', encoding="utf-8") as file:
sweight_data = file.read()
sovits_path = os.environ.get("sovits_path", sweight_data)
else:
sovits_path = os.environ.get("sovits_path", "GPT_SoVITS/pretrained_models/s2G488k.pth")
# gpt_path = os.environ.get(
# "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
# )
# sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth")
cnhubert_base_path = os.environ.get(
"cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base"
)
bert_path = os.environ.get(
"bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"
)
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
infer_ttswebui = int(infer_ttswebui)
is_share = os.environ.get("is_share", "False")
is_share = eval(is_share)
if "_CUDA_VISIBLE_DEVICES" in os.environ:
os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
is_half = eval(os.environ.get("is_half", "True"))
from transformers import AutoModelForMaskedLM, AutoTokenizer
import numpy as np
import librosa, torch
from feature_extractor import cnhubert
cnhubert.cnhubert_base_path = cnhubert_base_path
from module.models import SynthesizerTrn
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
from text import cleaned_text_to_sequence
from text.cleaner import clean_text
from time import time as ttime
from module.mel_processing import spectrogram_torch
from my_utils import load_audio
from tools.i18n.i18n import I18nAuto
i18n = I18nAuto()
os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。
if torch.cuda.is_available():
device = "cuda"
elif torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
tokenizer = AutoTokenizer.from_pretrained(bert_path)
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
if is_half == True:
bert_model = bert_model.half().to(device)
else:
bert_model = bert_model.to(device)
def get_bert_feature(text, word2ph):
with torch.no_grad():
inputs = tokenizer(text, return_tensors="pt")
for i in inputs:
inputs[i] = inputs[i].to(device)
res = bert_model(**inputs, output_hidden_states=True)
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
assert len(word2ph) == len(text)
phone_level_feature = []
for i in range(len(word2ph)):
repeat_feature = res[i].repeat(word2ph[i], 1)
phone_level_feature.append(repeat_feature)
phone_level_feature = torch.cat(phone_level_feature, dim=0)
return phone_level_feature.T
class DictToAttrRecursive(dict):
def __init__(self, input_dict):
super().__init__(input_dict)
for key, value in input_dict.items():
if isinstance(value, dict):
value = DictToAttrRecursive(value)
self[key] = value
setattr(self, key, value)
def __getattr__(self, item):
try:
return self[item]
except KeyError:
raise AttributeError(f"Attribute {item} not found")
def __setattr__(self, key, value):
if isinstance(value, dict):
value = DictToAttrRecursive(value)
super(DictToAttrRecursive, self).__setitem__(key, value)
super().__setattr__(key, value)
def __delattr__(self, item):
try:
del self[item]
except KeyError:
raise AttributeError(f"Attribute {item} not found")
ssl_model = cnhubert.get_model()
if is_half == True:
ssl_model = ssl_model.half().to(device)
else:
ssl_model = ssl_model.to(device)
def change_sovits_weights(sovits_path):
global vq_model, hps
dict_s2 = torch.load(sovits_path, map_location="cpu")
hps = dict_s2["config"]
hps = DictToAttrRecursive(hps)
hps.model.semantic_frame_rate = "25hz"
vq_model = SynthesizerTrn(
hps.data.filter_length // 2 + 1,
hps.train.segment_size // hps.data.hop_length,
n_speakers=hps.data.n_speakers,
**hps.model
)
if ("pretrained" not in sovits_path):
del vq_model.enc_q
if is_half == True:
vq_model = vq_model.half().to(device)
else:
vq_model = vq_model.to(device)
vq_model.eval()
print(vq_model.load_state_dict(dict_s2["weight"], strict=False))
with open("./sweight.txt", "w", encoding="utf-8") as f:
f.write(sovits_path)
change_sovits_weights(sovits_path)
def change_gpt_weights(gpt_path):
global hz, max_sec, t2s_model, config
hz = 50
dict_s1 = torch.load(gpt_path, map_location="cpu")
config = dict_s1["config"]
max_sec = config["data"]["max_sec"]
t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
t2s_model.load_state_dict(dict_s1["weight"])
if is_half == True:
t2s_model = t2s_model.half()
t2s_model = t2s_model.to(device)
t2s_model.eval()
total = sum([param.nelement() for param in t2s_model.parameters()])
print("Number of parameter: %.2fM" % (total / 1e6))
with open("./gweight.txt", "w", encoding="utf-8") as f: f.write(gpt_path)
change_gpt_weights(gpt_path)
def get_spepc(hps, filename):
audio = load_audio(filename, int(hps.data.sampling_rate))
audio = torch.FloatTensor(audio)
audio_norm = audio
audio_norm = audio_norm.unsqueeze(0)
spec = spectrogram_torch(
audio_norm,
hps.data.filter_length,
hps.data.sampling_rate,
hps.data.hop_length,
hps.data.win_length,
center=False,
)
return spec
dict_language = {
i18n("中文"): "all_zh",#全部按中文识别
i18n("英文"): "en",#全部按英文识别#######不变
i18n("日文"): "all_ja",#全部按日文识别
i18n("中英混合"): "zh",#按中英混合识别####不变
i18n("日英混合"): "ja",#按日英混合识别####不变
i18n("多语种混合"): "auto",#多语种启动切分识别语种
}
def splite_en_inf(sentence, language):
pattern = re.compile(r'[a-zA-Z ]+')
textlist = []
langlist = []
pos = 0
for match in pattern.finditer(sentence):
start, end = match.span()
if start > pos:
textlist.append(sentence[pos:start])
langlist.append(language)
textlist.append(sentence[start:end])
langlist.append("en")
pos = end
if pos < len(sentence):
textlist.append(sentence[pos:])
langlist.append(language)
# Merge punctuation into previous word
for i in range(len(textlist)-1, 0, -1):
if re.match(r'^[\W_]+$', textlist[i]):
textlist[i-1] += textlist[i]
del textlist[i]
del langlist[i]
# Merge consecutive words with the same language tag
i = 0
while i < len(langlist) - 1:
if langlist[i] == langlist[i+1]:
textlist[i] += textlist[i+1]
del textlist[i+1]
del langlist[i+1]
else:
i += 1
return textlist, langlist
def clean_text_inf(text, language):
formattext = ""
language = language.replace("all_","")
for tmp in LangSegment.getTexts(text):
if language == "ja":
if tmp["lang"] == language or tmp["lang"] == "zh":
formattext += tmp["text"] + " "
continue
if tmp["lang"] == language:
formattext += tmp["text"] + " "
while " " in formattext:
formattext = formattext.replace(" ", " ")
phones, word2ph, norm_text = clean_text(formattext, language)
phones = cleaned_text_to_sequence(phones)
return phones, word2ph, norm_text
dtype=torch.float16 if is_half == True else torch.float32
def get_bert_inf(phones, word2ph, norm_text, language):
language=language.replace("all_","")
if language == "zh":
bert = get_bert_feature(norm_text, word2ph).to(device)#.to(dtype)
else:
bert = torch.zeros(
(1024, len(phones)),
dtype=torch.float16 if is_half == True else torch.float32,
).to(device)
return bert
def nonen_clean_text_inf(text, language):
if(language!="auto"):
textlist, langlist = splite_en_inf(text, language)
else:
textlist=[]
langlist=[]
for tmp in LangSegment.getTexts(text):
langlist.append(tmp["lang"])
textlist.append(tmp["text"])
phones_list = []
word2ph_list = []
norm_text_list = []
for i in range(len(textlist)):
lang = langlist[i]
phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
phones_list.append(phones)
if lang == "zh":
word2ph_list.append(word2ph)
norm_text_list.append(norm_text)
print(word2ph_list)
phones = sum(phones_list, [])
word2ph = sum(word2ph_list, [])
norm_text = ' '.join(norm_text_list)
return phones, word2ph, norm_text
def nonen_get_bert_inf(text, language):
if(language!="auto"):
textlist, langlist = splite_en_inf(text, language)
else:
textlist=[]
langlist=[]
for tmp in LangSegment.getTexts(text):
langlist.append(tmp["lang"])
textlist.append(tmp["text"])
print(textlist)
print(langlist)
bert_list = []
for i in range(len(textlist)):
lang = langlist[i]
phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
bert = get_bert_inf(phones, word2ph, norm_text, lang)
bert_list.append(bert)
bert = torch.cat(bert_list, dim=1)
return bert
splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", }
def get_first(text):
pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
text = re.split(pattern, text)[0].strip()
return text
def get_cleaned_text_final(text,language):
if language in {"en","all_zh","all_ja"}:
phones, word2ph, norm_text = clean_text_inf(text, language)
elif language in {"zh", "ja","auto"}:
phones, word2ph, norm_text = nonen_clean_text_inf(text, language)
return phones, word2ph, norm_text
def get_bert_final(phones, word2ph, text,language,device):
if language == "en":
bert = get_bert_inf(phones, word2ph, text, language)
elif language in {"zh", "ja","auto"}:
bert = nonen_get_bert_inf(text, language)
elif language == "all_zh":
bert = get_bert_feature(text, word2ph).to(device)
else:
bert = torch.zeros((1024, len(phones))).to(device)
return bert
def merge_short_text_in_array(texts, threshold):
if (len(texts)) < 2:
return texts
result = []
text = ""
for ele in texts:
text += ele
if len(text) >= threshold:
result.append(text)
text = ""
if (len(text) > 0):
if len(result) == 0:
result.append(text)
else:
result[len(result) - 1] += text
return result
def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False):
if prompt_text is None or len(prompt_text) == 0:
ref_free = True
t0 = ttime()
prompt_language = dict_language[prompt_language]
text_language = dict_language[text_language]
if not ref_free:
prompt_text = prompt_text.strip("\n")
if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "."
print(i18n("实际输入的参考文本:"), prompt_text)
text = text.strip("\n")
if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text
print(i18n("实际输入的目标文本:"), text)
zero_wav = np.zeros(
int(hps.data.sampling_rate * 0.3),
dtype=np.float16 if is_half == True else np.float32,
)
with torch.no_grad():
wav16k, sr = librosa.load(ref_wav_path, sr=16000)
if (wav16k.shape[0] > 160000 or wav16k.shape[0] < 48000):
raise OSError(i18n("参考音频在3~10秒范围外，请更换！"))
wav16k = torch.from_numpy(wav16k)
zero_wav_torch = torch.from_numpy(zero_wav)
if is_half == True:
wav16k = wav16k.half().to(device)
zero_wav_torch = zero_wav_torch.half().to(device)
else:
wav16k = wav16k.to(device)
zero_wav_torch = zero_wav_torch.to(device)
wav16k = torch.cat([wav16k, zero_wav_torch])
ssl_content = ssl_model.model(wav16k.unsqueeze(0))[
"last_hidden_state"
].transpose(
1, 2
) # .float()
codes = vq_model.extract_latent(ssl_content)
prompt_semantic = codes[0, 0]
t1 = ttime()
if (how_to_cut == i18n("凑四句一切")):
text = cut1(text)
elif (how_to_cut == i18n("凑50字一切")):
text = cut2(text)
elif (how_to_cut == i18n("按中文句号。切")):
text = cut3(text)
elif (how_to_cut == i18n("按英文句号.切")):
text = cut4(text)
elif (how_to_cut == i18n("按标点符号切")):
text = cut5(text)
while "\n\n" in text:
text = text.replace("\n\n", "\n")
print(i18n("实际输入的目标文本(切句后):"), text)
texts = text.split("\n")
texts = merge_short_text_in_array(texts, 5)
audio_opt = []
if not ref_free:
phones1, word2ph1, norm_text1=get_cleaned_text_final(prompt_text, prompt_language)
bert1=get_bert_final(phones1, word2ph1, norm_text1,prompt_language,device).to(dtype)
for text in texts:
# 解决输入目标文本的空行导致报错的问题
if (len(text.strip()) == 0):
continue
if (text[-1] not in splits): text += "。" if text_language != "en" else "."
print(i18n("实际输入的目标文本(每句):"), text)
phones2, word2ph2, norm_text2 = get_cleaned_text_final(text, text_language)
bert2 = get_bert_final(phones2, word2ph2, norm_text2, text_language, device).to(dtype)
if not ref_free:
bert = torch.cat([bert1, bert2], 1)
all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0)
else:
bert = bert2
all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0)
bert = bert.to(device).unsqueeze(0)
all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
prompt = prompt_semantic.unsqueeze(0).to(device)
t2 = ttime()
with torch.no_grad():
# pred_semantic = t2s_model.model.infer(
pred_semantic, idx = t2s_model.model.infer_panel(
all_phoneme_ids,
all_phoneme_len,
None if ref_free else prompt,
bert,
# prompt_phone_len=ph_offset,
top_k=top_k,
top_p=top_p,
temperature=temperature,
early_stop_num=hz * max_sec,
)
t3 = ttime()
# print(pred_semantic.shape,idx)
pred_semantic = pred_semantic[:, -idx:].unsqueeze(
0
) # .unsqueeze(0)#mq要多unsqueeze一次
refer = get_spepc(hps, ref_wav_path) # .to(device)
if is_half == True:
refer = refer.half().to(device)
else:
refer = refer.to(device)
# audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
audio = (
vq_model.decode(
pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
)
.detach()
.cpu()
.numpy()[0, 0]
) ###试试重建不带上prompt部分
max_audio=np.abs(audio).max()#简单防止16bit爆音
if max_audio>1:audio/=max_audio
audio_opt.append(audio)
audio_opt.append(zero_wav)
t4 = ttime()
print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(
np.int16
)
def split(todo_text):
todo_text = todo_text.replace("……", "。").replace("——", "，")
if todo_text[-1] not in splits:
todo_text += "。"
i_split_head = i_split_tail = 0
len_text = len(todo_text)
todo_texts = []
while 1:
if i_split_head >= len_text:
break # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入
if todo_text[i_split_head] in splits:
i_split_head += 1
todo_texts.append(todo_text[i_split_tail:i_split_head])
i_split_tail = i_split_head
else:
i_split_head += 1
return todo_texts
def cut1(inp):
inp = inp.strip("\n")
inps = split(inp)
split_idx = list(range(0, len(inps), 4))
split_idx[-1] = None
if len(split_idx) > 1:
opts = []
for idx in range(len(split_idx) - 1):
opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]]))
else:
opts = [inp]
return "\n".join(opts)
def cut2(inp):
inp = inp.strip("\n")
inps = split(inp)
if len(inps) < 2:
return inp
opts = []
summ = 0
tmp_str = ""
for i in range(len(inps)):
summ += len(inps[i])
tmp_str += inps[i]
if summ > 50:
summ = 0
opts.append(tmp_str)
tmp_str = ""
if tmp_str != "":
opts.append(tmp_str)
# print(opts)
if len(opts) > 1 and len(opts[-1]) < 50: ##如果最后一个太短了，和前一个合一起
opts[-2] = opts[-2] + opts[-1]
opts = opts[:-1]
return "\n".join(opts)
def cut3(inp):
inp = inp.strip("\n")
return "\n".join(["%s" % item for item in inp.strip("。").split("。")])
def cut4(inp):
inp = inp.strip("\n")
return "\n".join(["%s" % item for item in inp.strip(".").split(".")])
# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py
def cut5(inp):
# if not re.search(r'[^\w\s]', inp[-1]):
# inp += '。'
inp = inp.strip("\n")
punds = r'[,.;?!、，。？！;：…]'
items = re.split(f'({punds})', inp)
mergeitems = ["".join(group) for group in zip(items[::2], items[1::2])]
# 在句子不存在符号或句尾无符号的时候保证文本完整
if len(items)%2 == 1:
mergeitems.append(items[-1])
opt = "\n".join(mergeitems)
return opt
def custom_sort_key(s):
# 使用正则表达式提取字符串中的数字部分和非数字部分
parts = re.split('(\d+)', s)
# 将数字部分转换为整数，非数字部分保持不变
parts = [int(part) if part.isdigit() else part for part in parts]
return parts
def change_choices():
SoVITS_names, GPT_names = get_weights_names()
return {"choices": sorted(SoVITS_names, key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names, key=custom_sort_key), "__type__": "update"}
pretrained_sovits_name = "GPT_SoVITS/pretrained_models/s2G488k.pth"
pretrained_gpt_name = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
SoVITS_weight_root = "SoVITS_weights"
GPT_weight_root = "GPT_weights"
os.makedirs(SoVITS_weight_root, exist_ok=True)
os.makedirs(GPT_weight_root, exist_ok=True)
def get_weights_names():
SoVITS_names = [pretrained_sovits_name]
for name in os.listdir(SoVITS_weight_root):
if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (SoVITS_weight_root, name))
GPT_names = [pretrained_gpt_name]
for name in os.listdir(GPT_weight_root):
if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (GPT_weight_root, name))
return SoVITS_names, GPT_names
SoVITS_names, GPT_names = get_weights_names()

610
GPT_SoVITS/inference_webui.py
View File

@ -1,609 +1,13 @@
'''
按中英混合识别
按日英混合识别
多语种启动切分识别语种
全部按中文识别
全部按英文识别
全部按日文识别
'''
import os, re, logging
import LangSegment
logging.getLogger("markdown_it").setLevel(logging.ERROR)
logging.getLogger("urllib3").setLevel(logging.ERROR)
logging.getLogger("httpcore").setLevel(logging.ERROR)
logging.getLogger("httpx").setLevel(logging.ERROR)
logging.getLogger("asyncio").setLevel(logging.ERROR)
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
import pdb
if os.path.exists("./gweight.txt"):
with open("./gweight.txt", 'r', encoding="utf-8") as file:
gweight_data = file.read()
gpt_path = os.environ.get(
"gpt_path", gweight_data)
else:
gpt_path = os.environ.get(
"gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt")
if os.path.exists("./sweight.txt"):
with open("./sweight.txt", 'r', encoding="utf-8") as file:
sweight_data = file.read()
sovits_path = os.environ.get("sovits_path", sweight_data)
else:
sovits_path = os.environ.get("sovits_path", "GPT_SoVITS/pretrained_models/s2G488k.pth")
# gpt_path = os.environ.get(
# "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
# )
# sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth")
cnhubert_base_path = os.environ.get(
"cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base"
)
bert_path = os.environ.get(
"bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"
)
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
infer_ttswebui = int(infer_ttswebui)
is_share = os.environ.get("is_share", "False")
is_share = eval(is_share)
if "_CUDA_VISIBLE_DEVICES" in os.environ:
os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
is_half = eval(os.environ.get("is_half", "True"))
import gradio as gr
from transformers import AutoModelForMaskedLM, AutoTokenizer
import numpy as np
import librosa, torch
from feature_extractor import cnhubert
cnhubert.cnhubert_base_path = cnhubert_base_path
from module.models import SynthesizerTrn
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
from text import cleaned_text_to_sequence
from text.cleaner import clean_text
from time import time as ttime
from module.mel_processing import spectrogram_torch
from my_utils import load_audio
import os
from tools.i18n.i18n import I18nAuto
from inference_base import sovits_path, gpt_path, change_choices, GPT_names, custom_sort_key, SoVITS_names, change_sovits_weights, change_gpt_weights, get_tts_wav, cut1, cut2, cut3, cut4, cut5
import gradio as gr
i18n = I18nAuto()
os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 确保直接启动推理UI时也能够设置。
if torch.cuda.is_available():
device = "cuda"
elif torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
tokenizer = AutoTokenizer.from_pretrained(bert_path)
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
if is_half == True:
bert_model = bert_model.half().to(device)
else:
bert_model = bert_model.to(device)
def get_bert_feature(text, word2ph):
with torch.no_grad():
inputs = tokenizer(text, return_tensors="pt")
for i in inputs:
inputs[i] = inputs[i].to(device)
res = bert_model(**inputs, output_hidden_states=True)
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
assert len(word2ph) == len(text)
phone_level_feature = []
for i in range(len(word2ph)):
repeat_feature = res[i].repeat(word2ph[i], 1)
phone_level_feature.append(repeat_feature)
phone_level_feature = torch.cat(phone_level_feature, dim=0)
return phone_level_feature.T
class DictToAttrRecursive(dict):
def __init__(self, input_dict):
super().__init__(input_dict)
for key, value in input_dict.items():
if isinstance(value, dict):
value = DictToAttrRecursive(value)
self[key] = value
setattr(self, key, value)
def __getattr__(self, item):
try:
return self[item]
except KeyError:
raise AttributeError(f"Attribute {item} not found")
def __setattr__(self, key, value):
if isinstance(value, dict):
value = DictToAttrRecursive(value)
super(DictToAttrRecursive, self).__setitem__(key, value)
super().__setattr__(key, value)
def __delattr__(self, item):
try:
del self[item]
except KeyError:
raise AttributeError(f"Attribute {item} not found")
ssl_model = cnhubert.get_model()
if is_half == True:
ssl_model = ssl_model.half().to(device)
else:
ssl_model = ssl_model.to(device)
def change_sovits_weights(sovits_path):
global vq_model, hps
dict_s2 = torch.load(sovits_path, map_location="cpu")
hps = dict_s2["config"]
hps = DictToAttrRecursive(hps)
hps.model.semantic_frame_rate = "25hz"
vq_model = SynthesizerTrn(
hps.data.filter_length // 2 + 1,
hps.train.segment_size // hps.data.hop_length,
n_speakers=hps.data.n_speakers,
**hps.model
)
if ("pretrained" not in sovits_path):
del vq_model.enc_q
if is_half == True:
vq_model = vq_model.half().to(device)
else:
vq_model = vq_model.to(device)
vq_model.eval()
print(vq_model.load_state_dict(dict_s2["weight"], strict=False))
with open("./sweight.txt", "w", encoding="utf-8") as f:
f.write(sovits_path)
change_sovits_weights(sovits_path)
def change_gpt_weights(gpt_path):
global hz, max_sec, t2s_model, config
hz = 50
dict_s1 = torch.load(gpt_path, map_location="cpu")
config = dict_s1["config"]
max_sec = config["data"]["max_sec"]
t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
t2s_model.load_state_dict(dict_s1["weight"])
if is_half == True:
t2s_model = t2s_model.half()
t2s_model = t2s_model.to(device)
t2s_model.eval()
total = sum([param.nelement() for param in t2s_model.parameters()])
print("Number of parameter: %.2fM" % (total / 1e6))
with open("./gweight.txt", "w", encoding="utf-8") as f: f.write(gpt_path)
change_gpt_weights(gpt_path)
def get_spepc(hps, filename):
audio = load_audio(filename, int(hps.data.sampling_rate))
audio = torch.FloatTensor(audio)
audio_norm = audio
audio_norm = audio_norm.unsqueeze(0)
spec = spectrogram_torch(
audio_norm,
hps.data.filter_length,
hps.data.sampling_rate,
hps.data.hop_length,
hps.data.win_length,
center=False,
)
return spec
dict_language = {
i18n("中文"): "all_zh",#全部按中文识别
i18n("英文"): "en",#全部按英文识别#######不变
i18n("日文"): "all_ja",#全部按日文识别
i18n("中英混合"): "zh",#按中英混合识别####不变
i18n("日英混合"): "ja",#按日英混合识别####不变
i18n("多语种混合"): "auto",#多语种启动切分识别语种
}
def splite_en_inf(sentence, language):
pattern = re.compile(r'[a-zA-Z ]+')
textlist = []
langlist = []
pos = 0
for match in pattern.finditer(sentence):
start, end = match.span()
if start > pos:
textlist.append(sentence[pos:start])
langlist.append(language)
textlist.append(sentence[start:end])
langlist.append("en")
pos = end
if pos < len(sentence):
textlist.append(sentence[pos:])
langlist.append(language)
# Merge punctuation into previous word
for i in range(len(textlist)-1, 0, -1):
if re.match(r'^[\W_]+$', textlist[i]):
textlist[i-1] += textlist[i]
del textlist[i]
del langlist[i]
# Merge consecutive words with the same language tag
i = 0
while i < len(langlist) - 1:
if langlist[i] == langlist[i+1]:
textlist[i] += textlist[i+1]
del textlist[i+1]
del langlist[i+1]
else:
i += 1
return textlist, langlist
def clean_text_inf(text, language):
formattext = ""
language = language.replace("all_","")
for tmp in LangSegment.getTexts(text):
if language == "ja":
if tmp["lang"] == language or tmp["lang"] == "zh":
formattext += tmp["text"] + " "
continue
if tmp["lang"] == language:
formattext += tmp["text"] + " "
while " " in formattext:
formattext = formattext.replace(" ", " ")
phones, word2ph, norm_text = clean_text(formattext, language)
phones = cleaned_text_to_sequence(phones)
return phones, word2ph, norm_text
dtype=torch.float16 if is_half == True else torch.float32
def get_bert_inf(phones, word2ph, norm_text, language):
language=language.replace("all_","")
if language == "zh":
bert = get_bert_feature(norm_text, word2ph).to(device)#.to(dtype)
else:
bert = torch.zeros(
(1024, len(phones)),
dtype=torch.float16 if is_half == True else torch.float32,
).to(device)
return bert
def nonen_clean_text_inf(text, language):
if(language!="auto"):
textlist, langlist = splite_en_inf(text, language)
else:
textlist=[]
langlist=[]
for tmp in LangSegment.getTexts(text):
langlist.append(tmp["lang"])
textlist.append(tmp["text"])
phones_list = []
word2ph_list = []
norm_text_list = []
for i in range(len(textlist)):
lang = langlist[i]
phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
phones_list.append(phones)
if lang == "zh":
word2ph_list.append(word2ph)
norm_text_list.append(norm_text)
print(word2ph_list)
phones = sum(phones_list, [])
word2ph = sum(word2ph_list, [])
norm_text = ' '.join(norm_text_list)
return phones, word2ph, norm_text
def nonen_get_bert_inf(text, language):
if(language!="auto"):
textlist, langlist = splite_en_inf(text, language)
else:
textlist=[]
langlist=[]
for tmp in LangSegment.getTexts(text):
langlist.append(tmp["lang"])
textlist.append(tmp["text"])
print(textlist)
print(langlist)
bert_list = []
for i in range(len(textlist)):
lang = langlist[i]
phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
bert = get_bert_inf(phones, word2ph, norm_text, lang)
bert_list.append(bert)
bert = torch.cat(bert_list, dim=1)
return bert
splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", }
def get_first(text):
pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
text = re.split(pattern, text)[0].strip()
return text
def get_cleaned_text_final(text,language):
if language in {"en","all_zh","all_ja"}:
phones, word2ph, norm_text = clean_text_inf(text, language)
elif language in {"zh", "ja","auto"}:
phones, word2ph, norm_text = nonen_clean_text_inf(text, language)
return phones, word2ph, norm_text
def get_bert_final(phones, word2ph, text,language,device):
if language == "en":
bert = get_bert_inf(phones, word2ph, text, language)
elif language in {"zh", "ja","auto"}:
bert = nonen_get_bert_inf(text, language)
elif language == "all_zh":
bert = get_bert_feature(text, word2ph).to(device)
else:
bert = torch.zeros((1024, len(phones))).to(device)
return bert
def merge_short_text_in_array(texts, threshold):
if (len(texts)) < 2:
return texts
result = []
text = ""
for ele in texts:
text += ele
if len(text) >= threshold:
result.append(text)
text = ""
if (len(text) > 0):
if len(result) == 0:
result.append(text)
else:
result[len(result) - 1] += text
return result
def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, top_p=0.6, temperature=0.6, ref_free = False):
if prompt_text is None or len(prompt_text) == 0:
ref_free = True
t0 = ttime()
prompt_language = dict_language[prompt_language]
text_language = dict_language[text_language]
if not ref_free:
prompt_text = prompt_text.strip("\n")
if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "."
print(i18n("实际输入的参考文本:"), prompt_text)
text = text.strip("\n")
if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text
print(i18n("实际输入的目标文本:"), text)
zero_wav = np.zeros(
int(hps.data.sampling_rate * 0.3),
dtype=np.float16 if is_half == True else np.float32,
)
with torch.no_grad():
wav16k, sr = librosa.load(ref_wav_path, sr=16000)
if (wav16k.shape[0] > 160000 or wav16k.shape[0] < 48000):
raise OSError(i18n("参考音频在3~10秒范围外，请更换！"))
wav16k = torch.from_numpy(wav16k)
zero_wav_torch = torch.from_numpy(zero_wav)
if is_half == True:
wav16k = wav16k.half().to(device)
zero_wav_torch = zero_wav_torch.half().to(device)
else:
wav16k = wav16k.to(device)
zero_wav_torch = zero_wav_torch.to(device)
wav16k = torch.cat([wav16k, zero_wav_torch])
ssl_content = ssl_model.model(wav16k.unsqueeze(0))[
"last_hidden_state"
].transpose(
1, 2
) # .float()
codes = vq_model.extract_latent(ssl_content)
prompt_semantic = codes[0, 0]
t1 = ttime()
if (how_to_cut == i18n("凑四句一切")):
text = cut1(text)
elif (how_to_cut == i18n("凑50字一切")):
text = cut2(text)
elif (how_to_cut == i18n("按中文句号。切")):
text = cut3(text)
elif (how_to_cut == i18n("按英文句号.切")):
text = cut4(text)
elif (how_to_cut == i18n("按标点符号切")):
text = cut5(text)
while "\n\n" in text:
text = text.replace("\n\n", "\n")
print(i18n("实际输入的目标文本(切句后):"), text)
texts = text.split("\n")
texts = merge_short_text_in_array(texts, 5)
audio_opt = []
if not ref_free:
phones1, word2ph1, norm_text1=get_cleaned_text_final(prompt_text, prompt_language)
bert1=get_bert_final(phones1, word2ph1, norm_text1,prompt_language,device).to(dtype)
for text in texts:
# 解决输入目标文本的空行导致报错的问题
if (len(text.strip()) == 0):
continue
if (text[-1] not in splits): text += "。" if text_language != "en" else "."
print(i18n("实际输入的目标文本(每句):"), text)
phones2, word2ph2, norm_text2 = get_cleaned_text_final(text, text_language)
bert2 = get_bert_final(phones2, word2ph2, norm_text2, text_language, device).to(dtype)
if not ref_free:
bert = torch.cat([bert1, bert2], 1)
all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0)
else:
bert = bert2
all_phoneme_ids = torch.LongTensor(phones2).to(device).unsqueeze(0)
bert = bert.to(device).unsqueeze(0)
all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
prompt = prompt_semantic.unsqueeze(0).to(device)
t2 = ttime()
with torch.no_grad():
# pred_semantic = t2s_model.model.infer(
pred_semantic, idx = t2s_model.model.infer_panel(
all_phoneme_ids,
all_phoneme_len,
None if ref_free else prompt,
bert,
# prompt_phone_len=ph_offset,
top_k=top_k,
top_p=top_p,
temperature=temperature,
early_stop_num=hz * max_sec,
)
t3 = ttime()
# print(pred_semantic.shape,idx)
pred_semantic = pred_semantic[:, -idx:].unsqueeze(
0
) # .unsqueeze(0)#mq要多unsqueeze一次
refer = get_spepc(hps, ref_wav_path) # .to(device)
if is_half == True:
refer = refer.half().to(device)
else:
refer = refer.to(device)
# audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
audio = (
vq_model.decode(
pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
)
.detach()
.cpu()
.numpy()[0, 0]
) ###试试重建不带上prompt部分
max_audio=np.abs(audio).max()#简单防止16bit爆音
if max_audio>1:audio/=max_audio
audio_opt.append(audio)
audio_opt.append(zero_wav)
t4 = ttime()
print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(
np.int16
)
def split(todo_text):
todo_text = todo_text.replace("……", "。").replace("——", "，")
if todo_text[-1] not in splits:
todo_text += "。"
i_split_head = i_split_tail = 0
len_text = len(todo_text)
todo_texts = []
while 1:
if i_split_head >= len_text:
break # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入
if todo_text[i_split_head] in splits:
i_split_head += 1
todo_texts.append(todo_text[i_split_tail:i_split_head])
i_split_tail = i_split_head
else:
i_split_head += 1
return todo_texts
def cut1(inp):
inp = inp.strip("\n")
inps = split(inp)
split_idx = list(range(0, len(inps), 4))
split_idx[-1] = None
if len(split_idx) > 1:
opts = []
for idx in range(len(split_idx) - 1):
opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]]))
else:
opts = [inp]
return "\n".join(opts)
def cut2(inp):
inp = inp.strip("\n")
inps = split(inp)
if len(inps) < 2:
return inp
opts = []
summ = 0
tmp_str = ""
for i in range(len(inps)):
summ += len(inps[i])
tmp_str += inps[i]
if summ > 50:
summ = 0
opts.append(tmp_str)
tmp_str = ""
if tmp_str != "":
opts.append(tmp_str)
# print(opts)
if len(opts) > 1 and len(opts[-1]) < 50: ##如果最后一个太短了，和前一个合一起
opts[-2] = opts[-2] + opts[-1]
opts = opts[:-1]
return "\n".join(opts)
def cut3(inp):
inp = inp.strip("\n")
return "\n".join(["%s" % item for item in inp.strip("。").split("。")])
def cut4(inp):
inp = inp.strip("\n")
return "\n".join(["%s" % item for item in inp.strip(".").split(".")])
# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py
def cut5(inp):
# if not re.search(r'[^\w\s]', inp[-1]):
# inp += '。'
inp = inp.strip("\n")
punds = r'[,.;?!、，。？！;：…]'
items = re.split(f'({punds})', inp)
mergeitems = ["".join(group) for group in zip(items[::2], items[1::2])]
# 在句子不存在符号或句尾无符号的时候保证文本完整
if len(items)%2 == 1:
mergeitems.append(items[-1])
opt = "\n".join(mergeitems)
return opt
def custom_sort_key(s):
# 使用正则表达式提取字符串中的数字部分和非数字部分
parts = re.split('(\d+)', s)
# 将数字部分转换为整数，非数字部分保持不变
parts = [int(part) if part.isdigit() else part for part in parts]
return parts
def change_choices():
SoVITS_names, GPT_names = get_weights_names()
return {"choices": sorted(SoVITS_names, key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names, key=custom_sort_key), "__type__": "update"}
pretrained_sovits_name = "GPT_SoVITS/pretrained_models/s2G488k.pth"
pretrained_gpt_name = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
SoVITS_weight_root = "SoVITS_weights"
GPT_weight_root = "GPT_weights"
os.makedirs(SoVITS_weight_root, exist_ok=True)
os.makedirs(GPT_weight_root, exist_ok=True)
def get_weights_names():
SoVITS_names = [pretrained_sovits_name]
for name in os.listdir(SoVITS_weight_root):
if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (SoVITS_weight_root, name))
GPT_names = [pretrained_gpt_name]
for name in os.listdir(GPT_weight_root):
if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (GPT_weight_root, name))
return SoVITS_names, GPT_names
SoVITS_names, GPT_names = get_weights_names()
is_share = os.environ.get("is_share", "False")
is_share = eval(is_share)
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
infer_ttswebui = int(infer_ttswebui)
with gr.Blocks(title="GPT-SoVITS WebUI") as app:
gr.Markdown(

625
train_base.py Normal file
View File

@ -0,0 +1,625 @@
import os,shutil,sys,pdb,re
now_dir = os.getcwd()
sys.path.append(now_dir)
import json,yaml,warnings,torch
import platform
import psutil
import signal
warnings.filterwarnings("ignore")
torch.manual_seed(233333)
tmp = os.path.join(now_dir, "TEMP")
os.makedirs(tmp, exist_ok=True)
os.environ["TEMP"] = tmp
if(os.path.exists(tmp)):
for name in os.listdir(tmp):
if(name=="jieba.cache"):continue
path="%s/%s"%(tmp,name)
delete=os.remove if os.path.isfile(path) else shutil.rmtree
try:
delete(path)
except Exception as e:
print(str(e))
pass
import site
site_packages_roots = []
for path in site.getsitepackages():
if "packages" in path:
site_packages_roots.append(path)
if(site_packages_roots==[]):site_packages_roots=["%s/runtime/Lib/site-packages" % now_dir]
#os.environ["OPENBLAS_NUM_THREADS"] = "4"
os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
os.environ["all_proxy"] = ""
for site_packages_root in site_packages_roots:
if os.path.exists(site_packages_root):
try:
with open("%s/users.pth" % (site_packages_root), "w") as f:
f.write(
"%s\n%s/tools\n%s/tools/damo_asr\n%s/GPT_SoVITS\n%s/tools/uvr5"
% (now_dir, now_dir, now_dir, now_dir, now_dir)
)
break
except PermissionError:
pass
from tools import my_utils
import traceback
import shutil
import pdb
from subprocess import Popen
import signal
from config import python_exec,infer_device,is_half,exp_root,webui_port_main,webui_port_infer_tts,webui_port_uvr5,webui_port_subfix,is_share
from tools.i18n.i18n import I18nAuto
i18n = I18nAuto()
from scipy.io import wavfile
from tools.my_utils import load_audio
from multiprocessing import cpu_count
os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 当遇到mps不支持的步骤时使用cpu
n_cpu=cpu_count()
ngpu = torch.cuda.device_count()
gpu_infos = []
mem = []
if_gpu_ok = False
# 判断是否有能用来训练和加速推理的N卡
if torch.cuda.is_available() or ngpu != 0:
for i in range(ngpu):
gpu_name = torch.cuda.get_device_name(i)
if any(value in gpu_name.upper()for value in ["10","16","20","30","40","A2","A3","A4","P4","A50","500","A60","70","80","90","M4","T4","TITAN","L4","4060"]):
# A10#A100#V100#A40#P40#M40#K80#A4500
if_gpu_ok = True # 至少有一张能用的N卡
gpu_infos.append("%s\t%s" % (i, gpu_name))
mem.append(int(torch.cuda.get_device_properties(i).total_memory/ 1024/ 1024/ 1024+ 0.4))
# 判断是否支持mps加速
if torch.backends.mps.is_available():
if_gpu_ok = True
gpu_infos.append("%s\t%s" % ("0", "Apple GPU"))
mem.append(psutil.virtual_memory().total/ 1024 / 1024 / 1024) # 实测使用系统内存作为显存不会爆显存
if if_gpu_ok and len(gpu_infos) > 0:
gpu_info = "\n".join(gpu_infos)
default_batch_size = min(mem) // 2
else:
gpu_info = i18n("很遗憾您这没有能用的显卡来支持您训练")
default_batch_size = 1
gpus = "-".join([i[0] for i in gpu_infos])
pretrained_sovits_name="GPT_SoVITS/pretrained_models/s2G488k.pth"
pretrained_gpt_name="GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
def get_weights_names():
SoVITS_names = [pretrained_sovits_name]
for name in os.listdir(SoVITS_weight_root):
if name.endswith(".pth"):SoVITS_names.append(name)
GPT_names = [pretrained_gpt_name]
for name in os.listdir(GPT_weight_root):
if name.endswith(".ckpt"): GPT_names.append(name)
return SoVITS_names,GPT_names
SoVITS_weight_root="SoVITS_weights"
GPT_weight_root="GPT_weights"
os.makedirs(SoVITS_weight_root,exist_ok=True)
os.makedirs(GPT_weight_root,exist_ok=True)
SoVITS_names,GPT_names = get_weights_names()
def custom_sort_key(s):
# 使用正则表达式提取字符串中的数字部分和非数字部分
parts = re.split('(\d+)', s)
# 将数字部分转换为整数，非数字部分保持不变
parts = [int(part) if part.isdigit() else part for part in parts]
return parts
def change_choices():
SoVITS_names, GPT_names = get_weights_names()
return {"choices": sorted(SoVITS_names,key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names,key=custom_sort_key), "__type__": "update"}
p_label=None
p_uvr5=None
p_asr=None
def kill_proc_tree(pid, including_parent=True):
try:
parent = psutil.Process(pid)
except psutil.NoSuchProcess:
# Process already terminated
return
children = parent.children(recursive=True)
for child in children:
try:
os.kill(child.pid, signal.SIGTERM) # or signal.SIGKILL
except OSError:
pass
if including_parent:
try:
os.kill(parent.pid, signal.SIGTERM) # or signal.SIGKILL
except OSError:
pass
system=platform.system()
def kill_process(pid):
if(system=="Windows"):
cmd = "taskkill /t /f /pid %s" % pid
os.system(cmd)
else:
kill_proc_tree(pid)
def change_label(if_label,path_list):
global p_label
if(if_label==True and p_label==None):
path_list=my_utils.clean_path(path_list)
cmd = '"%s" tools/subfix_webui.py --load_list "%s" --webui_port %s --is_share %s'%(python_exec,path_list,webui_port_subfix,is_share)
yield i18n("打标工具WebUI已开启")
print(cmd)
p_label = Popen(cmd, shell=True)
elif(if_label==False and p_label!=None):
kill_process(p_label.pid)
p_label=None
yield i18n("打标工具WebUI已关闭")
def change_uvr5(if_uvr5):
global p_uvr5
if(if_uvr5==True and p_uvr5==None):
cmd = '"%s" tools/uvr5/webui.py "%s" %s %s %s'%(python_exec,infer_device,is_half,webui_port_uvr5,is_share)
yield i18n("UVR5已开启")
print(cmd)
p_uvr5 = Popen(cmd, shell=True)
elif(if_uvr5==False and p_uvr5!=None):
kill_process(p_uvr5.pid)
p_uvr5=None
yield i18n("UVR5已关闭")
from tools.asr.config import asr_dict
def open_asr(asr_inp_dir, asr_opt_dir, asr_model, asr_model_size, asr_lang):
global p_asr
if(p_asr==None):
asr_inp_dir=my_utils.clean_path(asr_inp_dir)
cmd = f'"{python_exec}" tools/asr/{asr_dict[asr_model]["path"]}'
cmd += f' -i "{asr_inp_dir}"'
cmd += f' -o "{asr_opt_dir}"'
cmd += f' -s {asr_model_size}'
cmd += f' -l {asr_lang}'
cmd += " -p %s"%("float16"if is_half==True else "float32")
yield "ASR任务开启：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True}
print(cmd)
p_asr = Popen(cmd, shell=True)
p_asr.wait()
p_asr=None
yield f"ASR任务完成, 查看终端进行下一步",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的ASR任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True}
# return None
def close_asr():
global p_asr
if(p_asr!=None):
kill_process(p_asr.pid)
p_asr=None
return "已终止ASR进程",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
p_train_SoVITS=None
def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D):
global p_train_SoVITS
if(p_train_SoVITS==None):
with open("GPT_SoVITS/configs/s2.json")as f:
data=f.read()
data=json.loads(data)
s2_dir="%s/%s"%(exp_root,exp_name)
os.makedirs("%s/logs_s2"%(s2_dir),exist_ok=True)
if(is_half==False):
data["train"]["fp16_run"]=False
batch_size=max(1,batch_size//2)
data["train"]["batch_size"]=batch_size
data["train"]["epochs"]=total_epoch
data["train"]["text_low_lr_rate"]=text_low_lr_rate
data["train"]["pretrained_s2G"]=pretrained_s2G
data["train"]["pretrained_s2D"]=pretrained_s2D
data["train"]["if_save_latest"]=if_save_latest
data["train"]["if_save_every_weights"]=if_save_every_weights
data["train"]["save_every_epoch"]=save_every_epoch
data["train"]["gpu_numbers"]=gpu_numbers1Ba
data["data"]["exp_dir"]=data["s2_ckpt_dir"]=s2_dir
data["save_weight_dir"]=SoVITS_weight_root
data["name"]=exp_name
tmp_config_path="%s/tmp_s2.json"%tmp
with open(tmp_config_path,"w")as f:f.write(json.dumps(data))
cmd = '"%s" GPT_SoVITS/s2_train.py --config "%s"'%(python_exec,tmp_config_path)
yield "SoVITS训练开始：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True}
print(cmd)
p_train_SoVITS = Popen(cmd, shell=True)
p_train_SoVITS.wait()
p_train_SoVITS=None
yield "SoVITS训练完成",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的SoVITS训练任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True}
def close1Ba():
global p_train_SoVITS
if(p_train_SoVITS!=None):
kill_process(p_train_SoVITS.pid)
p_train_SoVITS=None
return "已终止SoVITS训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
p_train_GPT=None
def open1Bb(batch_size,total_epoch,exp_name,if_dpo,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers,pretrained_s1):
global p_train_GPT
if(p_train_GPT==None):
with open("GPT_SoVITS/configs/s1longer.yaml")as f:
data=f.read()
data=yaml.load(data, Loader=yaml.FullLoader)
s1_dir="%s/%s"%(exp_root,exp_name)
os.makedirs("%s/logs_s1"%(s1_dir),exist_ok=True)
if(is_half==False):
data["train"]["precision"]="32"
batch_size = max(1, batch_size // 2)
data["train"]["batch_size"]=batch_size
data["train"]["epochs"]=total_epoch
data["pretrained_s1"]=pretrained_s1
data["train"]["save_every_n_epoch"]=save_every_epoch
data["train"]["if_save_every_weights"]=if_save_every_weights
data["train"]["if_save_latest"]=if_save_latest
data["train"]["if_dpo"]=if_dpo
data["train"]["half_weights_save_dir"]=GPT_weight_root
data["train"]["exp_name"]=exp_name
data["train_semantic_path"]="%s/6-name2semantic.tsv"%s1_dir
data["train_phoneme_path"]="%s/2-name2text.txt"%s1_dir
data["output_dir"]="%s/logs_s1"%s1_dir
os.environ["_CUDA_VISIBLE_DEVICES"]=gpu_numbers.replace("-",",")
os.environ["hz"]="25hz"
tmp_config_path="%s/tmp_s1.yaml"%tmp
with open(tmp_config_path, "w") as f:f.write(yaml.dump(data, default_flow_style=False))
# cmd = '"%s" GPT_SoVITS/s1_train.py --config_file "%s" --train_semantic_path "%s/6-name2semantic.tsv" --train_phoneme_path "%s/2-name2text.txt" --output_dir "%s/logs_s1"'%(python_exec,tmp_config_path,s1_dir,s1_dir,s1_dir)
cmd = '"%s" GPT_SoVITS/s1_train.py --config_file "%s" '%(python_exec,tmp_config_path)
yield "GPT训练开始：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True}
print(cmd)
p_train_GPT = Popen(cmd, shell=True)
p_train_GPT.wait()
p_train_GPT=None
yield "GPT训练完成",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的GPT训练任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True}
def close1Bb():
global p_train_GPT
if(p_train_GPT!=None):
kill_process(p_train_GPT.pid)
p_train_GPT=None
return "已终止GPT训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
ps_slice=[]
def open_slice(inp,opt_root,threshold,min_length,min_interval,hop_size,max_sil_kept,_max,alpha,n_parts):
global ps_slice
inp = my_utils.clean_path(inp)
opt_root = my_utils.clean_path(opt_root)
if(os.path.exists(inp)==False):
yield "输入路径不存在",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
return
if os.path.isfile(inp):n_parts=1
elif os.path.isdir(inp):pass
else:
yield "输入路径存在但既不是文件也不是文件夹",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
return
if (ps_slice == []):
for i_part in range(n_parts):
cmd = '"%s" tools/slice_audio.py "%s" "%s" %s %s %s %s %s %s %s %s %s''' % (python_exec,inp, opt_root, threshold, min_length, min_interval, hop_size, max_sil_kept, _max, alpha, i_part, n_parts)
print(cmd)
p = Popen(cmd, shell=True)
ps_slice.append(p)
yield "切割执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps_slice:
p.wait()
ps_slice=[]
yield "切割结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的切割任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close_slice():
global ps_slice
if (ps_slice != []):
for p_slice in ps_slice:
try:
kill_process(p_slice.pid)
except:
traceback.print_exc()
ps_slice=[]
return "已终止所有切割进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
ps1a=[]
def open1a(inp_text,inp_wav_dir,exp_name,gpu_numbers,bert_pretrained_dir):
global ps1a
inp_text = my_utils.clean_path(inp_text)
inp_wav_dir = my_utils.clean_path(inp_wav_dir)
if (ps1a == []):
opt_dir="%s/%s"%(exp_root,exp_name)
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":opt_dir,
"bert_pretrained_dir":bert_pretrained_dir,
}
gpu_names=gpu_numbers.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
"is_half": str(is_half)
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/1-get-text.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1a.append(p)
yield "文本进程执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1a:
p.wait()
opt = []
for i_part in range(all_parts):
txt_path = "%s/2-name2text-%s.txt" % (opt_dir, i_part)
with open(txt_path, "r", encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(txt_path)
path_text = "%s/2-name2text.txt" % opt_dir
with open(path_text, "w", encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
ps1a=[]
yield "文本进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的文本任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1a():
global ps1a
if (ps1a != []):
for p1a in ps1a:
try:
kill_process(p1a.pid)
except:
traceback.print_exc()
ps1a=[]
return "已终止所有1a进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
ps1b=[]
def open1b(inp_text,inp_wav_dir,exp_name,gpu_numbers,ssl_pretrained_dir):
global ps1b
inp_text = my_utils.clean_path(inp_text)
inp_wav_dir = my_utils.clean_path(inp_wav_dir)
if (ps1b == []):
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":"%s/%s"%(exp_root,exp_name),
"cnhubert_base_dir":ssl_pretrained_dir,
"is_half": str(is_half)
}
gpu_names=gpu_numbers.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1b.append(p)
yield "SSL提取进程执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1b:
p.wait()
ps1b=[]
yield "SSL提取进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的SSL提取任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1b():
global ps1b
if (ps1b != []):
for p1b in ps1b:
try:
kill_process(p1b.pid)
except:
traceback.print_exc()
ps1b=[]
return "已终止所有1b进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
ps1c=[]
def open1c(inp_text,exp_name,gpu_numbers,pretrained_s2G_path):
global ps1c
inp_text = my_utils.clean_path(inp_text)
if (ps1c == []):
opt_dir="%s/%s"%(exp_root,exp_name)
config={
"inp_text":inp_text,
"exp_name":exp_name,
"opt_dir":opt_dir,
"pretrained_s2G":pretrained_s2G_path,
"s2config_path":"GPT_SoVITS/configs/s2.json",
"is_half": str(is_half)
}
gpu_names=gpu_numbers.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/3-get-semantic.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1c.append(p)
yield "语义token提取进程执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1c:
p.wait()
opt = ["item_name\tsemantic_audio"]
path_semantic = "%s/6-name2semantic.tsv" % opt_dir
for i_part in range(all_parts):
semantic_path = "%s/6-name2semantic-%s.tsv" % (opt_dir, i_part)
with open(semantic_path, "r", encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(semantic_path)
with open(path_semantic, "w", encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
ps1c=[]
yield "语义token提取进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的语义token提取任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1c():
global ps1c
if (ps1c != []):
for p1c in ps1c:
try:
kill_process(p1c.pid)
except:
traceback.print_exc()
ps1c=[]
return "已终止所有语义token进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
#####inp_text,inp_wav_dir,exp_name,gpu_numbers1a,gpu_numbers1Ba,gpu_numbers1c,bert_pretrained_dir,cnhubert_base_dir,pretrained_s2G
ps1abc=[]
def open1abc(inp_text,inp_wav_dir,exp_name,gpu_numbers1a,gpu_numbers1Ba,gpu_numbers1c,bert_pretrained_dir,ssl_pretrained_dir,pretrained_s2G_path):
global ps1abc
inp_text = my_utils.clean_path(inp_text)
inp_wav_dir = my_utils.clean_path(inp_wav_dir)
if (ps1abc == []):
opt_dir="%s/%s"%(exp_root,exp_name)
try:
#############################1a
path_text="%s/2-name2text.txt" % opt_dir
if(os.path.exists(path_text)==False or (os.path.exists(path_text)==True and len(open(path_text,"r",encoding="utf8").read().strip("\n").split("\n"))<2)):
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":opt_dir,
"bert_pretrained_dir":bert_pretrained_dir,
"is_half": str(is_half)
}
gpu_names=gpu_numbers1a.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/1-get-text.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1abc.append(p)
yield "进度：1a-ing", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1abc:p.wait()
opt = []
for i_part in range(all_parts):#txt_path="%s/2-name2text-%s.txt"%(opt_dir,i_part)
txt_path = "%s/2-name2text-%s.txt" % (opt_dir, i_part)
with open(txt_path, "r",encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(txt_path)
with open(path_text, "w",encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
yield "进度：1a-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
ps1abc=[]
#############################1b
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":opt_dir,
"cnhubert_base_dir":ssl_pretrained_dir,
}
gpu_names=gpu_numbers1Ba.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1abc.append(p)
yield "进度：1a-done, 1b-ing", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1abc:p.wait()
yield "进度：1a1b-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
ps1abc=[]
#############################1c
path_semantic = "%s/6-name2semantic.tsv" % opt_dir
if(os.path.exists(path_semantic)==False or (os.path.exists(path_semantic)==True and os.path.getsize(path_semantic)<31)):
config={
"inp_text":inp_text,
"exp_name":exp_name,
"opt_dir":opt_dir,
"pretrained_s2G":pretrained_s2G_path,
"s2config_path":"GPT_SoVITS/configs/s2.json",
}
gpu_names=gpu_numbers1c.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/3-get-semantic.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1abc.append(p)
yield "进度：1a1b-done, 1cing", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1abc:p.wait()
opt = ["item_name\tsemantic_audio"]
for i_part in range(all_parts):
semantic_path = "%s/6-name2semantic-%s.tsv" % (opt_dir, i_part)
with open(semantic_path, "r",encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(semantic_path)
with open(path_semantic, "w",encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
yield "进度：all-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
ps1abc = []
yield "一键三连进程结束", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
except:
traceback.print_exc()
close1abc()
yield "一键三连中途报错", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
else:
yield "已有正在进行的一键三连任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1abc():
global ps1abc
if (ps1abc != []):
for p1abc in ps1abc:
try:
kill_process(p1abc.pid)
except:
traceback.print_exc()
ps1abc=[]
return "已终止所有一键三连进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}

626
webui.py
View File

@ -1,177 +1,14 @@
import os,shutil,sys,pdb,re
now_dir = os.getcwd()
sys.path.append(now_dir)
import json,yaml,warnings,torch
import platform
import psutil
import signal
warnings.filterwarnings("ignore")
torch.manual_seed(233333)
tmp = os.path.join(now_dir, "TEMP")
os.makedirs(tmp, exist_ok=True)
os.environ["TEMP"] = tmp
if(os.path.exists(tmp)):
for name in os.listdir(tmp):
if(name=="jieba.cache"):continue
path="%s/%s"%(tmp,name)
delete=os.remove if os.path.isfile(path) else shutil.rmtree
try:
delete(path)
except Exception as e:
print(str(e))
pass
import site
site_packages_roots = []
for path in site.getsitepackages():
if "packages" in path:
site_packages_roots.append(path)
if(site_packages_roots==[]):site_packages_roots=["%s/runtime/Lib/site-packages" % now_dir]
#os.environ["OPENBLAS_NUM_THREADS"] = "4"
os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
os.environ["all_proxy"] = ""
for site_packages_root in site_packages_roots:
if os.path.exists(site_packages_root):
try:
with open("%s/users.pth" % (site_packages_root), "w") as f:
f.write(
"%s\n%s/tools\n%s/tools/damo_asr\n%s/GPT_SoVITS\n%s/tools/uvr5"
% (now_dir, now_dir, now_dir, now_dir, now_dir)
)
break
except PermissionError:
pass
from tools import my_utils
import traceback
import shutil
import pdb
import os
import gradio as gr
from subprocess import Popen
import signal
from config import python_exec,infer_device,is_half,exp_root,webui_port_main,webui_port_infer_tts,webui_port_uvr5,webui_port_subfix,is_share
from tools.i18n.i18n import I18nAuto
i18n = I18nAuto()
from scipy.io import wavfile
from tools.my_utils import load_audio
from multiprocessing import cpu_count
from train_base import gpu_info, n_cpu, SoVITS_names, pretrained_sovits_name, pretrained_gpt_name, custom_sort_key, GPT_names, default_batch_size, kill_process, SoVITS_weight_root, GPT_weight_root, change_choices, change_label, change_uvr5, open_asr, open1Ba, open1Bb, close1Bb, open_slice, close_asr, open1a, close1a, open1b, close1Ba, close_slice, close1b, open1c, close1c, open1abc, close1abc, gpus
from tools.asr.config import asr_dict
from config import python_exec,infer_device,is_half,exp_root,webui_port_main,webui_port_infer_tts,webui_port_uvr5,webui_port_subfix,is_share
from subprocess import Popen
os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1' # 当遇到mps不支持的步骤时使用cpu
n_cpu=cpu_count()
ngpu = torch.cuda.device_count()
gpu_infos = []
mem = []
if_gpu_ok = False
# 判断是否有能用来训练和加速推理的N卡
if torch.cuda.is_available() or ngpu != 0:
for i in range(ngpu):
gpu_name = torch.cuda.get_device_name(i)
if any(value in gpu_name.upper()for value in ["10","16","20","30","40","A2","A3","A4","P4","A50","500","A60","70","80","90","M4","T4","TITAN","L4","4060"]):
# A10#A100#V100#A40#P40#M40#K80#A4500
if_gpu_ok = True # 至少有一张能用的N卡
gpu_infos.append("%s\t%s" % (i, gpu_name))
mem.append(int(torch.cuda.get_device_properties(i).total_memory/ 1024/ 1024/ 1024+ 0.4))
# 判断是否支持mps加速
if torch.backends.mps.is_available():
if_gpu_ok = True
gpu_infos.append("%s\t%s" % ("0", "Apple GPU"))
mem.append(psutil.virtual_memory().total/ 1024 / 1024 / 1024) # 实测使用系统内存作为显存不会爆显存
if if_gpu_ok and len(gpu_infos) > 0:
gpu_info = "\n".join(gpu_infos)
default_batch_size = min(mem) // 2
else:
gpu_info = i18n("很遗憾您这没有能用的显卡来支持您训练")
default_batch_size = 1
gpus = "-".join([i[0] for i in gpu_infos])
pretrained_sovits_name="GPT_SoVITS/pretrained_models/s2G488k.pth"
pretrained_gpt_name="GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
def get_weights_names():
SoVITS_names = [pretrained_sovits_name]
for name in os.listdir(SoVITS_weight_root):
if name.endswith(".pth"):SoVITS_names.append(name)
GPT_names = [pretrained_gpt_name]
for name in os.listdir(GPT_weight_root):
if name.endswith(".ckpt"): GPT_names.append(name)
return SoVITS_names,GPT_names
SoVITS_weight_root="SoVITS_weights"
GPT_weight_root="GPT_weights"
os.makedirs(SoVITS_weight_root,exist_ok=True)
os.makedirs(GPT_weight_root,exist_ok=True)
SoVITS_names,GPT_names = get_weights_names()
def custom_sort_key(s):
# 使用正则表达式提取字符串中的数字部分和非数字部分
parts = re.split('(\d+)', s)
# 将数字部分转换为整数，非数字部分保持不变
parts = [int(part) if part.isdigit() else part for part in parts]
return parts
def change_choices():
SoVITS_names, GPT_names = get_weights_names()
return {"choices": sorted(SoVITS_names,key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names,key=custom_sort_key), "__type__": "update"}
p_label=None
p_uvr5=None
p_asr=None
p_tts_inference=None
def kill_proc_tree(pid, including_parent=True):
try:
parent = psutil.Process(pid)
except psutil.NoSuchProcess:
# Process already terminated
return
children = parent.children(recursive=True)
for child in children:
try:
os.kill(child.pid, signal.SIGTERM) # or signal.SIGKILL
except OSError:
pass
if including_parent:
try:
os.kill(parent.pid, signal.SIGTERM) # or signal.SIGKILL
except OSError:
pass
system=platform.system()
def kill_process(pid):
if(system=="Windows"):
cmd = "taskkill /t /f /pid %s" % pid
os.system(cmd)
else:
kill_proc_tree(pid)
def change_label(if_label,path_list):
global p_label
if(if_label==True and p_label==None):
path_list=my_utils.clean_path(path_list)
cmd = '"%s" tools/subfix_webui.py --load_list "%s" --webui_port %s --is_share %s'%(python_exec,path_list,webui_port_subfix,is_share)
yield i18n("打标工具WebUI已开启")
print(cmd)
p_label = Popen(cmd, shell=True)
elif(if_label==False and p_label!=None):
kill_process(p_label.pid)
p_label=None
yield i18n("打标工具WebUI已关闭")
def change_uvr5(if_uvr5):
global p_uvr5
if(if_uvr5==True and p_uvr5==None):
cmd = '"%s" tools/uvr5/webui.py "%s" %s %s %s'%(python_exec,infer_device,is_half,webui_port_uvr5,is_share)
yield i18n("UVR5已开启")
print(cmd)
p_uvr5 = Popen(cmd, shell=True)
elif(if_uvr5==False and p_uvr5!=None):
kill_process(p_uvr5.pid)
p_uvr5=None
yield i18n("UVR5已关闭")
def change_tts_inference(if_tts,bert_path,cnhubert_base_path,gpu_number,gpt_path,sovits_path):
global p_tts_inference
if(if_tts==True and p_tts_inference==None):
@ -192,459 +29,6 @@ def change_tts_inference(if_tts,bert_path,cnhubert_base_path,gpu_number,gpt_path
p_tts_inference=None
yield i18n("TTS推理进程已关闭")
from tools.asr.config import asr_dict
def open_asr(asr_inp_dir, asr_opt_dir, asr_model, asr_model_size, asr_lang):
global p_asr
if(p_asr==None):
asr_inp_dir=my_utils.clean_path(asr_inp_dir)
cmd = f'"{python_exec}" tools/asr/{asr_dict[asr_model]["path"]}'
cmd += f' -i "{asr_inp_dir}"'
cmd += f' -o "{asr_opt_dir}"'
cmd += f' -s {asr_model_size}'
cmd += f' -l {asr_lang}'
cmd += " -p %s"%("float16"if is_half==True else "float32")
yield "ASR任务开启：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True}
print(cmd)
p_asr = Popen(cmd, shell=True)
p_asr.wait()
p_asr=None
yield f"ASR任务完成, 查看终端进行下一步",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的ASR任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True}
# return None
def close_asr():
global p_asr
if(p_asr!=None):
kill_process(p_asr.pid)
p_asr=None
return "已终止ASR进程",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
p_train_SoVITS=None
def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D):
global p_train_SoVITS
if(p_train_SoVITS==None):
with open("GPT_SoVITS/configs/s2.json")as f:
data=f.read()
data=json.loads(data)
s2_dir="%s/%s"%(exp_root,exp_name)
os.makedirs("%s/logs_s2"%(s2_dir),exist_ok=True)
if(is_half==False):
data["train"]["fp16_run"]=False
batch_size=max(1,batch_size//2)
data["train"]["batch_size"]=batch_size
data["train"]["epochs"]=total_epoch
data["train"]["text_low_lr_rate"]=text_low_lr_rate
data["train"]["pretrained_s2G"]=pretrained_s2G
data["train"]["pretrained_s2D"]=pretrained_s2D
data["train"]["if_save_latest"]=if_save_latest
data["train"]["if_save_every_weights"]=if_save_every_weights
data["train"]["save_every_epoch"]=save_every_epoch
data["train"]["gpu_numbers"]=gpu_numbers1Ba
data["data"]["exp_dir"]=data["s2_ckpt_dir"]=s2_dir
data["save_weight_dir"]=SoVITS_weight_root
data["name"]=exp_name
tmp_config_path="%s/tmp_s2.json"%tmp
with open(tmp_config_path,"w")as f:f.write(json.dumps(data))
cmd = '"%s" GPT_SoVITS/s2_train.py --config "%s"'%(python_exec,tmp_config_path)
yield "SoVITS训练开始：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True}
print(cmd)
p_train_SoVITS = Popen(cmd, shell=True)
p_train_SoVITS.wait()
p_train_SoVITS=None
yield "SoVITS训练完成",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的SoVITS训练任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True}
def close1Ba():
global p_train_SoVITS
if(p_train_SoVITS!=None):
kill_process(p_train_SoVITS.pid)
p_train_SoVITS=None
return "已终止SoVITS训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
p_train_GPT=None
def open1Bb(batch_size,total_epoch,exp_name,if_dpo,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers,pretrained_s1):
global p_train_GPT
if(p_train_GPT==None):
with open("GPT_SoVITS/configs/s1longer.yaml")as f:
data=f.read()
data=yaml.load(data, Loader=yaml.FullLoader)
s1_dir="%s/%s"%(exp_root,exp_name)
os.makedirs("%s/logs_s1"%(s1_dir),exist_ok=True)
if(is_half==False):
data["train"]["precision"]="32"
batch_size = max(1, batch_size // 2)
data["train"]["batch_size"]=batch_size
data["train"]["epochs"]=total_epoch
data["pretrained_s1"]=pretrained_s1
data["train"]["save_every_n_epoch"]=save_every_epoch
data["train"]["if_save_every_weights"]=if_save_every_weights
data["train"]["if_save_latest"]=if_save_latest
data["train"]["if_dpo"]=if_dpo
data["train"]["half_weights_save_dir"]=GPT_weight_root
data["train"]["exp_name"]=exp_name
data["train_semantic_path"]="%s/6-name2semantic.tsv"%s1_dir
data["train_phoneme_path"]="%s/2-name2text.txt"%s1_dir
data["output_dir"]="%s/logs_s1"%s1_dir
os.environ["_CUDA_VISIBLE_DEVICES"]=gpu_numbers.replace("-",",")
os.environ["hz"]="25hz"
tmp_config_path="%s/tmp_s1.yaml"%tmp
with open(tmp_config_path, "w") as f:f.write(yaml.dump(data, default_flow_style=False))
# cmd = '"%s" GPT_SoVITS/s1_train.py --config_file "%s" --train_semantic_path "%s/6-name2semantic.tsv" --train_phoneme_path "%s/2-name2text.txt" --output_dir "%s/logs_s1"'%(python_exec,tmp_config_path,s1_dir,s1_dir,s1_dir)
cmd = '"%s" GPT_SoVITS/s1_train.py --config_file "%s" '%(python_exec,tmp_config_path)
yield "GPT训练开始：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True}
print(cmd)
p_train_GPT = Popen(cmd, shell=True)
p_train_GPT.wait()
p_train_GPT=None
yield "GPT训练完成",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的GPT训练任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True}
def close1Bb():
global p_train_GPT
if(p_train_GPT!=None):
kill_process(p_train_GPT.pid)
p_train_GPT=None
return "已终止GPT训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
ps_slice=[]
def open_slice(inp,opt_root,threshold,min_length,min_interval,hop_size,max_sil_kept,_max,alpha,n_parts):
global ps_slice
inp = my_utils.clean_path(inp)
opt_root = my_utils.clean_path(opt_root)
if(os.path.exists(inp)==False):
yield "输入路径不存在",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
return
if os.path.isfile(inp):n_parts=1
elif os.path.isdir(inp):pass
else:
yield "输入路径存在但既不是文件也不是文件夹",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
return
if (ps_slice == []):
for i_part in range(n_parts):
cmd = '"%s" tools/slice_audio.py "%s" "%s" %s %s %s %s %s %s %s %s %s''' % (python_exec,inp, opt_root, threshold, min_length, min_interval, hop_size, max_sil_kept, _max, alpha, i_part, n_parts)
print(cmd)
p = Popen(cmd, shell=True)
ps_slice.append(p)
yield "切割执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps_slice:
p.wait()
ps_slice=[]
yield "切割结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的切割任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close_slice():
global ps_slice
if (ps_slice != []):
for p_slice in ps_slice:
try:
kill_process(p_slice.pid)
except:
traceback.print_exc()
ps_slice=[]
return "已终止所有切割进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
ps1a=[]
def open1a(inp_text,inp_wav_dir,exp_name,gpu_numbers,bert_pretrained_dir):
global ps1a
inp_text = my_utils.clean_path(inp_text)
inp_wav_dir = my_utils.clean_path(inp_wav_dir)
if (ps1a == []):
opt_dir="%s/%s"%(exp_root,exp_name)
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":opt_dir,
"bert_pretrained_dir":bert_pretrained_dir,
}
gpu_names=gpu_numbers.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
"is_half": str(is_half)
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/1-get-text.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1a.append(p)
yield "文本进程执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1a:
p.wait()
opt = []
for i_part in range(all_parts):
txt_path = "%s/2-name2text-%s.txt" % (opt_dir, i_part)
with open(txt_path, "r", encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(txt_path)
path_text = "%s/2-name2text.txt" % opt_dir
with open(path_text, "w", encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
ps1a=[]
yield "文本进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的文本任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1a():
global ps1a
if (ps1a != []):
for p1a in ps1a:
try:
kill_process(p1a.pid)
except:
traceback.print_exc()
ps1a=[]
return "已终止所有1a进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
ps1b=[]
def open1b(inp_text,inp_wav_dir,exp_name,gpu_numbers,ssl_pretrained_dir):
global ps1b
inp_text = my_utils.clean_path(inp_text)
inp_wav_dir = my_utils.clean_path(inp_wav_dir)
if (ps1b == []):
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":"%s/%s"%(exp_root,exp_name),
"cnhubert_base_dir":ssl_pretrained_dir,
"is_half": str(is_half)
}
gpu_names=gpu_numbers.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1b.append(p)
yield "SSL提取进程执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1b:
p.wait()
ps1b=[]
yield "SSL提取进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的SSL提取任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1b():
global ps1b
if (ps1b != []):
for p1b in ps1b:
try:
kill_process(p1b.pid)
except:
traceback.print_exc()
ps1b=[]
return "已终止所有1b进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
ps1c=[]
def open1c(inp_text,exp_name,gpu_numbers,pretrained_s2G_path):
global ps1c
inp_text = my_utils.clean_path(inp_text)
if (ps1c == []):
opt_dir="%s/%s"%(exp_root,exp_name)
config={
"inp_text":inp_text,
"exp_name":exp_name,
"opt_dir":opt_dir,
"pretrained_s2G":pretrained_s2G_path,
"s2config_path":"GPT_SoVITS/configs/s2.json",
"is_half": str(is_half)
}
gpu_names=gpu_numbers.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/3-get-semantic.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1c.append(p)
yield "语义token提取进程执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1c:
p.wait()
opt = ["item_name\tsemantic_audio"]
path_semantic = "%s/6-name2semantic.tsv" % opt_dir
for i_part in range(all_parts):
semantic_path = "%s/6-name2semantic-%s.tsv" % (opt_dir, i_part)
with open(semantic_path, "r", encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(semantic_path)
with open(path_semantic, "w", encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
ps1c=[]
yield "语义token提取进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False}
else:
yield "已有正在进行的语义token提取任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1c():
global ps1c
if (ps1c != []):
for p1c in ps1c:
try:
kill_process(p1c.pid)
except:
traceback.print_exc()
ps1c=[]
return "已终止所有语义token进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
#####inp_text,inp_wav_dir,exp_name,gpu_numbers1a,gpu_numbers1Ba,gpu_numbers1c,bert_pretrained_dir,cnhubert_base_dir,pretrained_s2G
ps1abc=[]
def open1abc(inp_text,inp_wav_dir,exp_name,gpu_numbers1a,gpu_numbers1Ba,gpu_numbers1c,bert_pretrained_dir,ssl_pretrained_dir,pretrained_s2G_path):
global ps1abc
inp_text = my_utils.clean_path(inp_text)
inp_wav_dir = my_utils.clean_path(inp_wav_dir)
if (ps1abc == []):
opt_dir="%s/%s"%(exp_root,exp_name)
try:
#############################1a
path_text="%s/2-name2text.txt" % opt_dir
if(os.path.exists(path_text)==False or (os.path.exists(path_text)==True and len(open(path_text,"r",encoding="utf8").read().strip("\n").split("\n"))<2)):
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":opt_dir,
"bert_pretrained_dir":bert_pretrained_dir,
"is_half": str(is_half)
}
gpu_names=gpu_numbers1a.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/1-get-text.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1abc.append(p)
yield "进度：1a-ing", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1abc:p.wait()
opt = []
for i_part in range(all_parts):#txt_path="%s/2-name2text-%s.txt"%(opt_dir,i_part)
txt_path = "%s/2-name2text-%s.txt" % (opt_dir, i_part)
with open(txt_path, "r",encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(txt_path)
with open(path_text, "w",encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
yield "进度：1a-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
ps1abc=[]
#############################1b
config={
"inp_text":inp_text,
"inp_wav_dir":inp_wav_dir,
"exp_name":exp_name,
"opt_dir":opt_dir,
"cnhubert_base_dir":ssl_pretrained_dir,
}
gpu_names=gpu_numbers1Ba.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1abc.append(p)
yield "进度：1a-done, 1b-ing", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1abc:p.wait()
yield "进度：1a1b-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
ps1abc=[]
#############################1c
path_semantic = "%s/6-name2semantic.tsv" % opt_dir
if(os.path.exists(path_semantic)==False or (os.path.exists(path_semantic)==True and os.path.getsize(path_semantic)<31)):
config={
"inp_text":inp_text,
"exp_name":exp_name,
"opt_dir":opt_dir,
"pretrained_s2G":pretrained_s2G_path,
"s2config_path":"GPT_SoVITS/configs/s2.json",
}
gpu_names=gpu_numbers1c.split("-")
all_parts=len(gpu_names)
for i_part in range(all_parts):
config.update(
{
"i_part": str(i_part),
"all_parts": str(all_parts),
"_CUDA_VISIBLE_DEVICES": gpu_names[i_part],
}
)
os.environ.update(config)
cmd = '"%s" GPT_SoVITS/prepare_datasets/3-get-semantic.py'%python_exec
print(cmd)
p = Popen(cmd, shell=True)
ps1abc.append(p)
yield "进度：1a1b-done, 1cing", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
for p in ps1abc:p.wait()
opt = ["item_name\tsemantic_audio"]
for i_part in range(all_parts):
semantic_path = "%s/6-name2semantic-%s.tsv" % (opt_dir, i_part)
with open(semantic_path, "r",encoding="utf8") as f:
opt += f.read().strip("\n").split("\n")
os.remove(semantic_path)
with open(path_semantic, "w",encoding="utf8") as f:
f.write("\n".join(opt) + "\n")
yield "进度：all-done", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
ps1abc = []
yield "一键三连进程结束", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
except:
traceback.print_exc()
close1abc()
yield "一键三连中途报错", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
else:
yield "已有正在进行的一键三连任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}
def close1abc():
global ps1abc
if (ps1abc != []):
for p1abc in ps1abc:
try:
kill_process(p1abc.pid)
except:
traceback.print_exc()
ps1abc=[]
return "已终止所有一键三连进程", {"__type__": "update", "visible": True}, {"__type__": "update", "visible": False}
with gr.Blocks(title="GPT-SoVITS WebUI") as app:
gr.Markdown(
value=