blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2025-10-07 15:19:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: add instance pool for api_v3.py

Browse Source
This commit is contained in:
kevin.zhang 2024-05-27 16:20:06 +08:00
parent c5dc7697a8
commit c24398df8a
6 changed files with 306 additions and 127 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

223
GPT_SoVITS/TTS_infer_pack/TTS.py
View File

@ -70,7 +70,7 @@ def set_seed(seed:int):
except: except:
pass pass
return seed return seed
class TTS_Config: class TTS_Config:
default_configs={ default_configs={
"device": "cpu", "device": "cpu",
@ -82,30 +82,30 @@ class TTS_Config:
} }
configs:dict = None configs:dict = None
def __init__(self, configs: Union[dict, str]=None): def __init__(self, configs: Union[dict, str]=None):
# 设置默认配置文件路径 # 设置默认配置文件路径
configs_base_path:str = "GPT_SoVITS/configs/" configs_base_path:str = "GPT_SoVITS/configs/"
os.makedirs(configs_base_path, exist_ok=True) os.makedirs(configs_base_path, exist_ok=True)
self.configs_path:str = os.path.join(configs_base_path, "tts_infer.yaml") self.configs_path:str = os.path.join(configs_base_path, "tts_infer.yaml")
if configs in ["", None]: if configs in ["", None]:
if not os.path.exists(self.configs_path): if not os.path.exists(self.configs_path):
self.save_configs() self.save_configs()
print(f"Create default config file at {self.configs_path}") print(f"Create default config file at {self.configs_path}")
configs:dict = {"default": deepcopy(self.default_configs)} configs:dict = {"default": deepcopy(self.default_configs)}
if isinstance(configs, str): if isinstance(configs, str):
self.configs_path = configs self.configs_path = configs
configs:dict = self._load_configs(self.configs_path) configs:dict = self._load_configs(self.configs_path)
assert isinstance(configs, dict) assert isinstance(configs, dict)
default_configs:dict = configs.get("default", None) default_configs:dict = configs.get("default", None)
if default_configs is not None: if default_configs is not None:
self.default_configs = default_configs self.default_configs = default_configs
self.configs:dict = configs.get("custom", deepcopy(self.default_configs)) self.configs:dict = configs.get("custom", deepcopy(self.default_configs))
self.device = self.configs.get("device", torch.device("cpu")) self.device = self.configs.get("device", torch.device("cpu"))
self.is_half = self.configs.get("is_half", False) self.is_half = self.configs.get("is_half", False)
self.t2s_weights_path = self.configs.get("t2s_weights_path", None) self.t2s_weights_path = self.configs.get("t2s_weights_path", None)
@ -113,7 +113,7 @@ class TTS_Config:
self.bert_base_path = self.configs.get("bert_base_path", None) self.bert_base_path = self.configs.get("bert_base_path", None)
self.cnhuhbert_base_path = self.configs.get("cnhuhbert_base_path", None) self.cnhuhbert_base_path = self.configs.get("cnhuhbert_base_path", None)
if (self.t2s_weights_path in [None, ""]) or (not os.path.exists(self.t2s_weights_path)): if (self.t2s_weights_path in [None, ""]) or (not os.path.exists(self.t2s_weights_path)):
self.t2s_weights_path = self.default_configs['t2s_weights_path'] self.t2s_weights_path = self.default_configs['t2s_weights_path']
print(f"fall back to default t2s_weights_path: {self.t2s_weights_path}") print(f"fall back to default t2s_weights_path: {self.t2s_weights_path}")
@ -127,8 +127,8 @@ class TTS_Config:
self.cnhuhbert_base_path = self.default_configs['cnhuhbert_base_path'] self.cnhuhbert_base_path = self.default_configs['cnhuhbert_base_path']
print(f"fall back to default cnhuhbert_base_path: {self.cnhuhbert_base_path}") print(f"fall back to default cnhuhbert_base_path: {self.cnhuhbert_base_path}")
self.update_configs() self.update_configs()
self.max_sec = None self.max_sec = None
self.hz:int = 50 self.hz:int = 50
self.semantic_frame_rate:str = "25hz" self.semantic_frame_rate:str = "25hz"
@ -138,14 +138,14 @@ class TTS_Config:
self.hop_length:int = 640 self.hop_length:int = 640
self.win_length:int = 2048 self.win_length:int = 2048
self.n_speakers:int = 300 self.n_speakers:int = 300
self.languages:list = ["auto", "en", "zh", "ja", "all_zh", "all_ja"] self.languages:list = ["auto", "en", "zh", "ja", "all_zh", "all_ja"]
def _load_configs(self, configs_path: str)->dict: def _load_configs(self, configs_path: str)->dict:
with open(configs_path, 'r') as f: with open(configs_path, 'r') as f:
configs = yaml.load(f, Loader=yaml.FullLoader) configs = yaml.load(f, Loader=yaml.FullLoader)
return configs return configs
def save_configs(self, configs_path:str=None)->None: def save_configs(self, configs_path:str=None)->None:
@ -154,7 +154,7 @@ class TTS_Config:
} }
if self.configs is not None: if self.configs is not None:
configs["custom"] = self.update_configs() configs["custom"] = self.update_configs()
if configs_path is None: if configs_path is None:
configs_path = self.configs_path configs_path = self.configs_path
with open(configs_path, 'w') as f: with open(configs_path, 'w') as f:
@ -170,7 +170,7 @@ class TTS_Config:
"cnhuhbert_base_path": self.cnhuhbert_base_path, "cnhuhbert_base_path": self.cnhuhbert_base_path,
} }
return self.config return self.config
def __str__(self): def __str__(self):
self.configs = self.update_configs() self.configs = self.update_configs()
string = "TTS Config".center(100, '-') + '\n' string = "TTS Config".center(100, '-') + '\n'
@ -178,7 +178,7 @@ class TTS_Config:
string += f"{str(k).ljust(20)}: {str(v)}\n" string += f"{str(k).ljust(20)}: {str(v)}\n"
string += "-" * 100 + '\n' string += "-" * 100 + '\n'
return string return string
def __repr__(self): def __repr__(self):
return self.__str__() return self.__str__()
@ -195,21 +195,21 @@ class TTS:
self.configs = configs self.configs = configs
else: else:
self.configs:TTS_Config = TTS_Config(configs) self.configs:TTS_Config = TTS_Config(configs)
self.t2s_model:Text2SemanticLightningModule = None self.t2s_model:Text2SemanticLightningModule = None
self.vits_model:SynthesizerTrn = None self.vits_model:SynthesizerTrn = None
self.bert_tokenizer:AutoTokenizer = None self.bert_tokenizer:AutoTokenizer = None
self.bert_model:AutoModelForMaskedLM = None self.bert_model:AutoModelForMaskedLM = None
self.cnhuhbert_model:CNHubert = None self.cnhuhbert_model:CNHubert = None
self._init_models() self._init_models()
self.text_preprocessor:TextPreprocessor = \ self.text_preprocessor:TextPreprocessor = \
TextPreprocessor(self.bert_model, TextPreprocessor(self.bert_model,
self.bert_tokenizer, self.bert_tokenizer,
self.configs.device) self.configs.device)
self.prompt_cache:dict = { self.prompt_cache:dict = {
"ref_audio_path" : None, "ref_audio_path" : None,
"prompt_semantic": None, "prompt_semantic": None,
@ -220,8 +220,8 @@ class TTS:
"bert_features" : None, "bert_features" : None,
"norm_text" : None, "norm_text" : None,
} }
self.stop_flag:bool = False self.stop_flag:bool = False
self.precision:torch.dtype = torch.float16 if self.configs.is_half else torch.float32 self.precision:torch.dtype = torch.float16 if self.configs.is_half else torch.float32
@ -231,9 +231,9 @@ class TTS:
self.init_bert_weights(self.configs.bert_base_path) self.init_bert_weights(self.configs.bert_base_path)
self.init_cnhuhbert_weights(self.configs.cnhuhbert_base_path) self.init_cnhuhbert_weights(self.configs.cnhuhbert_base_path)
# self.enable_half_precision(self.configs.is_half) # self.enable_half_precision(self.configs.is_half)
def init_cnhuhbert_weights(self, base_path: str): def init_cnhuhbert_weights(self, base_path: str):
print(f"Loading CNHuBERT weights from {base_path}") print(f"Loading CNHuBERT weights from {base_path}")
self.cnhuhbert_model = CNHubert(base_path) self.cnhuhbert_model = CNHubert(base_path)
@ -241,9 +241,9 @@ class TTS:
self.cnhuhbert_model = self.cnhuhbert_model.to(self.configs.device) self.cnhuhbert_model = self.cnhuhbert_model.to(self.configs.device)
if self.configs.is_half and str(self.configs.device)!="cpu": if self.configs.is_half and str(self.configs.device)!="cpu":
self.cnhuhbert_model = self.cnhuhbert_model.half() self.cnhuhbert_model = self.cnhuhbert_model.half()
def init_bert_weights(self, base_path: str): def init_bert_weights(self, base_path: str):
print(f"Loading BERT weights from {base_path}") print(f"Loading BERT weights from {base_path}")
self.bert_tokenizer = AutoTokenizer.from_pretrained(base_path) self.bert_tokenizer = AutoTokenizer.from_pretrained(base_path)
@ -252,16 +252,17 @@ class TTS:
self.bert_model = self.bert_model.to(self.configs.device) self.bert_model = self.bert_model.to(self.configs.device)
if self.configs.is_half and str(self.configs.device)!="cpu": if self.configs.is_half and str(self.configs.device)!="cpu":
self.bert_model = self.bert_model.half() self.bert_model = self.bert_model.half()
def init_vits_weights(self, weights_path: str): def init_vits_weights(self, weights_path: str, save_configs: bool = True):
print(f"Loading VITS weights from {weights_path}") print(f"Loading VITS weights from {weights_path}")
self.configs.vits_weights_path = weights_path self.configs.vits_weights_path = weights_path
self.configs.save_configs() if save_configs:
self.configs.save_configs()
dict_s2 = torch.load(weights_path, map_location=self.configs.device) dict_s2 = torch.load(weights_path, map_location=self.configs.device)
hps = dict_s2["config"] hps = dict_s2["config"]
self.configs.filter_length = hps["data"]["filter_length"] self.configs.filter_length = hps["data"]["filter_length"]
self.configs.segment_size = hps["train"]["segment_size"] self.configs.segment_size = hps["train"]["segment_size"]
self.configs.sampling_rate = hps["data"]["sampling_rate"] self.configs.sampling_rate = hps["data"]["sampling_rate"]
self.configs.hop_length = hps["data"]["hop_length"] self.configs.hop_length = hps["data"]["hop_length"]
self.configs.win_length = hps["data"]["win_length"] self.configs.win_length = hps["data"]["win_length"]
self.configs.n_speakers = hps["data"]["n_speakers"] self.configs.n_speakers = hps["data"]["n_speakers"]
@ -276,7 +277,7 @@ class TTS:
# if ("pretrained" not in weights_path): # if ("pretrained" not in weights_path):
if hasattr(vits_model, "enc_q"): if hasattr(vits_model, "enc_q"):
del vits_model.enc_q del vits_model.enc_q
vits_model = vits_model.to(self.configs.device) vits_model = vits_model.to(self.configs.device)
vits_model = vits_model.eval() vits_model = vits_model.eval()
vits_model.load_state_dict(dict_s2["weight"], strict=False) vits_model.load_state_dict(dict_s2["weight"], strict=False)
@ -284,11 +285,12 @@ class TTS:
if self.configs.is_half and str(self.configs.device)!="cpu": if self.configs.is_half and str(self.configs.device)!="cpu":
self.vits_model = self.vits_model.half() self.vits_model = self.vits_model.half()
def init_t2s_weights(self, weights_path: str): def init_t2s_weights(self, weights_path: str, save_configs: bool = True):
print(f"Loading Text2Semantic weights from {weights_path}") print(f"Loading Text2Semantic weights from {weights_path}")
self.configs.t2s_weights_path = weights_path self.configs.t2s_weights_path = weights_path
self.configs.save_configs() if save_configs:
self.configs.save_configs()
self.configs.hz = 50 self.configs.hz = 50
dict_s1 = torch.load(weights_path, map_location=self.configs.device) dict_s1 = torch.load(weights_path, map_location=self.configs.device)
config = dict_s1["config"] config = dict_s1["config"]
@ -300,7 +302,7 @@ class TTS:
self.t2s_model = t2s_model self.t2s_model = t2s_model
if self.configs.is_half and str(self.configs.device)!="cpu": if self.configs.is_half and str(self.configs.device)!="cpu":
self.t2s_model = self.t2s_model.half() self.t2s_model = self.t2s_model.half()
def enable_half_precision(self, enable: bool = True): def enable_half_precision(self, enable: bool = True):
''' '''
To enable half precision for the TTS model. To enable half precision for the TTS model.
@ -311,7 +313,7 @@ class TTS:
if str(self.configs.device) == "cpu" and enable: if str(self.configs.device) == "cpu" and enable:
print("Half precision is not supported on CPU.") print("Half precision is not supported on CPU.")
return return
self.configs.is_half = enable self.configs.is_half = enable
self.precision = torch.float16 if enable else torch.float32 self.precision = torch.float16 if enable else torch.float32
self.configs.save_configs() self.configs.save_configs()
@ -333,15 +335,16 @@ class TTS:
self.bert_model = self.bert_model.float() self.bert_model = self.bert_model.float()
if self.cnhuhbert_model is not None: if self.cnhuhbert_model is not None:
self.cnhuhbert_model = self.cnhuhbert_model.float() self.cnhuhbert_model = self.cnhuhbert_model.float()
def set_device(self, device: torch.device): def set_device(self, device: torch.device, save_configs: bool = True):
''' '''
To set the device for all models. To set the device for all models.
Args: Args:
device: torch.device, the device to use for all models. device: torch.device, the device to use for all models.
''' '''
self.configs.device = device self.configs.device = device
self.configs.save_configs() if save_configs:
self.configs.save_configs()
if self.t2s_model is not None: if self.t2s_model is not None:
self.t2s_model = self.t2s_model.to(device) self.t2s_model = self.t2s_model.to(device)
if self.vits_model is not None: if self.vits_model is not None:
@ -350,7 +353,7 @@ class TTS:
self.bert_model = self.bert_model.to(device) self.bert_model = self.bert_model.to(device)
if self.cnhuhbert_model is not None: if self.cnhuhbert_model is not None:
self.cnhuhbert_model = self.cnhuhbert_model.to(device) self.cnhuhbert_model = self.cnhuhbert_model.to(device)
def set_ref_audio(self, ref_audio_path:str): def set_ref_audio(self, ref_audio_path:str):
''' '''
To set the reference audio for the TTS model, To set the reference audio for the TTS model,
@ -360,7 +363,7 @@ class TTS:
''' '''
self._set_prompt_semantic(ref_audio_path) self._set_prompt_semantic(ref_audio_path)
self._set_ref_spec(ref_audio_path) self._set_ref_spec(ref_audio_path)
def _set_ref_spec(self, ref_audio_path): def _set_ref_spec(self, ref_audio_path):
audio = load_audio(ref_audio_path, int(self.configs.sampling_rate)) audio = load_audio(ref_audio_path, int(self.configs.sampling_rate))
audio = torch.FloatTensor(audio) audio = torch.FloatTensor(audio)
@ -379,8 +382,8 @@ class TTS:
spec = spec.half() spec = spec.half()
# self.refer_spec = spec # self.refer_spec = spec
self.prompt_cache["refer_spec"] = spec self.prompt_cache["refer_spec"] = spec
def _set_prompt_semantic(self, ref_wav_path:str): def _set_prompt_semantic(self, ref_wav_path:str):
zero_wav = np.zeros( zero_wav = np.zeros(
int(self.configs.sampling_rate * 0.3), int(self.configs.sampling_rate * 0.3),
@ -405,10 +408,10 @@ class TTS:
1, 2 1, 2
) # .float() ) # .float()
codes = self.vits_model.extract_latent(hubert_feature) codes = self.vits_model.extract_latent(hubert_feature)
prompt_semantic = codes[0, 0].to(self.configs.device) prompt_semantic = codes[0, 0].to(self.configs.device)
self.prompt_cache["prompt_semantic"] = prompt_semantic self.prompt_cache["prompt_semantic"] = prompt_semantic
def batch_sequences(self, sequences: List[torch.Tensor], axis: int = 0, pad_value: int = 0, max_length:int=None): def batch_sequences(self, sequences: List[torch.Tensor], axis: int = 0, pad_value: int = 0, max_length:int=None):
seq = sequences[0] seq = sequences[0]
ndim = seq.dim() ndim = seq.dim()
@ -429,12 +432,12 @@ class TTS:
padded_sequences.append(padded_seq) padded_sequences.append(padded_seq)
batch = torch.stack(padded_sequences) batch = torch.stack(padded_sequences)
return batch return batch
def to_batch(self, data:list, def to_batch(self, data:list,
prompt_data:dict=None, prompt_data:dict=None,
batch_size:int=5, batch_size:int=5,
threshold:float=0.75, threshold:float=0.75,
split_bucket:bool=True, split_bucket:bool=True,
device:torch.device=torch.device("cpu"), device:torch.device=torch.device("cpu"),
precision:torch.dtype=torch.float32, precision:torch.dtype=torch.float32,
): ):
@ -447,8 +450,8 @@ class TTS:
batch_index_list = [] batch_index_list = []
if split_bucket: if split_bucket:
index_and_len_list.sort(key=lambda x: x[1]) index_and_len_list.sort(key=lambda x: x[1])
index_and_len_list = np.array(index_and_len_list, dtype=np.int64) index_and_len_list = np.array(index_and_len_list, dtype=np.int64)
batch_index_list_len = 0 batch_index_list_len = 0
pos = 0 pos = 0
while pos <index_and_len_list.shape[0]: while pos <index_and_len_list.shape[0]:
@ -464,16 +467,16 @@ class TTS:
pos = pos_end pos = pos_end
break break
pos_end=pos_end-1 pos_end=pos_end-1
assert batch_index_list_len == len(data) assert batch_index_list_len == len(data)
else: else:
for i in range(len(data)): for i in range(len(data)):
if i%batch_size == 0: if i%batch_size == 0:
batch_index_list.append([]) batch_index_list.append([])
batch_index_list[-1].append(i) batch_index_list[-1].append(i)
for batch_idx, index_list in enumerate(batch_index_list): for batch_idx, index_list in enumerate(batch_index_list):
item_list = [data[idx] for idx in index_list] item_list = [data[idx] for idx in index_list]
phones_list = [] phones_list = []
@ -501,19 +504,19 @@ class TTS:
bert_max_len = max(bert_max_len, all_bert_features.shape[-1]) bert_max_len = max(bert_max_len, all_bert_features.shape[-1])
phones_max_len = max(phones_max_len, phones.shape[-1]) phones_max_len = max(phones_max_len, phones.shape[-1])
phones_list.append(phones) phones_list.append(phones)
phones_len_list.append(phones.shape[-1]) phones_len_list.append(phones.shape[-1])
all_phones_list.append(all_phones) all_phones_list.append(all_phones)
all_phones_len_list.append(all_phones.shape[-1]) all_phones_len_list.append(all_phones.shape[-1])
all_bert_features_list.append(all_bert_features) all_bert_features_list.append(all_bert_features)
norm_text_batch.append(item["norm_text"]) norm_text_batch.append(item["norm_text"])
phones_batch = phones_list phones_batch = phones_list
all_phones_batch = all_phones_list all_phones_batch = all_phones_list
all_bert_features_batch = all_bert_features_list all_bert_features_batch = all_bert_features_list
max_len = max(bert_max_len, phones_max_len) max_len = max(bert_max_len, phones_max_len)
# phones_batch = self.batch_sequences(phones_list, axis=0, pad_value=0, max_length=max_len) # phones_batch = self.batch_sequences(phones_list, axis=0, pad_value=0, max_length=max_len)
#### 直接对phones和bert_features进行pad。（padding策略会影响T2S模型生成的结果，但不直接影响复读概率。影响复读概率的主要因素是mask的策略） #### 直接对phones和bert_features进行pad。（padding策略会影响T2S模型生成的结果，但不直接影响复读概率。影响复读概率的主要因素是mask的策略）
@ -522,16 +525,16 @@ class TTS:
# all_bert_features_batch = torch.zeros((len(all_bert_features_list), 1024, max_len), dtype=precision, device=device) # all_bert_features_batch = torch.zeros((len(all_bert_features_list), 1024, max_len), dtype=precision, device=device)
# for idx, item in enumerate(all_bert_features_list): # for idx, item in enumerate(all_bert_features_list):
# all_bert_features_batch[idx, :, : item.shape[-1]] = item # all_bert_features_batch[idx, :, : item.shape[-1]] = item
# #### 先对phones进行embedding、对bert_features进行project，再pad到相同长度，（padding策略会影响T2S模型生成的结果，但不直接影响复读概率。影响复读概率的主要因素是mask的策略） # #### 先对phones进行embedding、对bert_features进行project，再pad到相同长度，（padding策略会影响T2S模型生成的结果，但不直接影响复读概率。影响复读概率的主要因素是mask的策略）
# all_phones_list = [self.t2s_model.model.ar_text_embedding(item.to(self.t2s_model.device)) for item in all_phones_list] # all_phones_list = [self.t2s_model.model.ar_text_embedding(item.to(self.t2s_model.device)) for item in all_phones_list]
# all_phones_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) for item in all_phones_list] # all_phones_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) for item in all_phones_list]
# all_phones_batch = torch.stack(all_phones_list, dim=0) # all_phones_batch = torch.stack(all_phones_list, dim=0)
# all_bert_features_list = [self.t2s_model.model.bert_proj(item.to(self.t2s_model.device).transpose(0, 1)) for item in all_bert_features_list] # all_bert_features_list = [self.t2s_model.model.bert_proj(item.to(self.t2s_model.device).transpose(0, 1)) for item in all_bert_features_list]
# all_bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) for item in all_bert_features_list] # all_bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) for item in all_bert_features_list]
# all_bert_features_batch = torch.stack(all_bert_features_list, dim=0) # all_bert_features_batch = torch.stack(all_bert_features_list, dim=0)
batch = { batch = {
"phones": phones_batch, "phones": phones_batch,
"phones_len": torch.LongTensor(phones_len_list).to(device), "phones_len": torch.LongTensor(phones_len_list).to(device),
@ -542,9 +545,9 @@ class TTS:
"max_len": max_len, "max_len": max_len,
} }
_data.append(batch) _data.append(batch)
return _data, batch_index_list return _data, batch_index_list
def recovery_order(self, data:list, batch_index_list:list)->list: def recovery_order(self, data:list, batch_index_list:list)->list:
''' '''
Recovery the order of the audio according to the batch_index_list. Recovery the order of the audio according to the batch_index_list.
@ -568,7 +571,7 @@ class TTS:
Stop the inference process. Stop the inference process.
''' '''
self.stop_flag = True self.stop_flag = True
@torch.no_grad() @torch.no_grad()
def run(self, inputs:dict): def run(self, inputs:dict):
""" """
@ -668,7 +671,7 @@ class TTS:
self.prompt_cache["prompt_lang"] = prompt_lang self.prompt_cache["prompt_lang"] = prompt_lang
phones, bert_features, norm_text = \ phones, bert_features, norm_text = \
self.text_preprocessor.segment_and_extract_feature_for_text( self.text_preprocessor.segment_and_extract_feature_for_text(
prompt_text, prompt_text,
prompt_lang) prompt_lang)
self.prompt_cache["phones"] = phones self.prompt_cache["phones"] = phones
self.prompt_cache["bert_features"] = bert_features self.prompt_cache["bert_features"] = bert_features
@ -685,9 +688,9 @@ class TTS:
return return
batch_index_list:list = None batch_index_list:list = None
data, batch_index_list = self.to_batch(data, data, batch_index_list = self.to_batch(data,
prompt_data=self.prompt_cache if not no_prompt_text else None, prompt_data=self.prompt_cache if not no_prompt_text else None,
batch_size=batch_size, batch_size=batch_size,
threshold=batch_threshold, threshold=batch_threshold,
split_bucket=split_bucket, split_bucket=split_bucket,
device=self.configs.device, device=self.configs.device,
@ -701,7 +704,7 @@ class TTS:
if i%batch_size == 0: if i%batch_size == 0:
data.append([]) data.append([])
data[-1].append(texts[i]) data[-1].append(texts[i])
def make_batch(batch_texts): def make_batch(batch_texts):
batch_data = [] batch_data = []
print(i18n("############ 提取文本Bert特征 ############")) print(i18n("############ 提取文本Bert特征 ############"))
@ -717,9 +720,9 @@ class TTS:
batch_data.append(res) batch_data.append(res)
if len(batch_data) == 0: if len(batch_data) == 0:
return None return None
batch, _ = self.to_batch(batch_data, batch, _ = self.to_batch(batch_data,
prompt_data=self.prompt_cache if not no_prompt_text else None, prompt_data=self.prompt_cache if not no_prompt_text else None,
batch_size=batch_size, batch_size=batch_size,
threshold=batch_threshold, threshold=batch_threshold,
split_bucket=False, split_bucket=False,
device=self.configs.device, device=self.configs.device,
@ -781,7 +784,7 @@ class TTS:
# 这里要记得加 torch.no_grad() 不然速度慢一大截 # 这里要记得加 torch.no_grad() 不然速度慢一大截
# with torch.no_grad(): # with torch.no_grad():
# ## vits并行推理 method 1 # ## vits并行推理 method 1
# pred_semantic_list = [item[-idx:] for item, idx in zip(pred_semantic_list, idx_list)] # pred_semantic_list = [item[-idx:] for item, idx in zip(pred_semantic_list, idx_list)]
# pred_semantic_len = torch.LongTensor([item.shape[0] for item in pred_semantic_list]).to(self.configs.device) # pred_semantic_len = torch.LongTensor([item.shape[0] for item in pred_semantic_list]).to(self.configs.device)
@ -823,10 +826,10 @@ class TTS:
t_45 += t5 - t4 t_45 += t5 - t4
if return_fragment: if return_fragment:
print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t4 - t3, t5 - t4)) print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t4 - t3, t5 - t4))
yield self.audio_postprocess([batch_audio_fragment], yield self.audio_postprocess([batch_audio_fragment],
self.configs.sampling_rate, self.configs.sampling_rate,
None, None,
speed_factor, speed_factor,
False, False,
fragment_interval fragment_interval
) )
@ -840,10 +843,10 @@ class TTS:
if not return_fragment: if not return_fragment:
print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t_34, t_45)) print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t_34, t_45))
yield self.audio_postprocess(audio, yield self.audio_postprocess(audio,
self.configs.sampling_rate, self.configs.sampling_rate,
batch_index_list, batch_index_list,
speed_factor, speed_factor,
split_bucket, split_bucket,
fragment_interval fragment_interval
) )
@ -863,21 +866,21 @@ class TTS:
raise e raise e
finally: finally:
self.empty_cache() self.empty_cache()
def empty_cache(self): def empty_cache(self):
try: try:
if "cuda" in str(self.configs.device): if "cuda" in str(self.configs.device):
torch.cuda.empty_cache() torch.cuda.empty_cache()
elif str(self.configs.device) == "mps": elif str(self.configs.device) == "mps":
torch.mps.empty_cache() torch.mps.empty_cache()
except: except:
pass pass
def audio_postprocess(self, def audio_postprocess(self,
audio:List[torch.Tensor], audio:List[torch.Tensor],
sr:int, sr:int,
batch_index_list:list=None, batch_index_list:list=None,
speed_factor:float=1.0, speed_factor:float=1.0,
split_bucket:bool=True, split_bucket:bool=True,
fragment_interval:float=0.3 fragment_interval:float=0.3
)->Tuple[int, np.ndarray]: )->Tuple[int, np.ndarray]:
@ -886,36 +889,36 @@ class TTS:
dtype=self.precision, dtype=self.precision,
device=self.configs.device device=self.configs.device
) )
for i, batch in enumerate(audio): for i, batch in enumerate(audio):
for j, audio_fragment in enumerate(batch): for j, audio_fragment in enumerate(batch):
max_audio=torch.abs(audio_fragment).max()#简单防止16bit爆音 max_audio=torch.abs(audio_fragment).max()#简单防止16bit爆音
if max_audio>1: audio_fragment/=max_audio if max_audio>1: audio_fragment/=max_audio
audio_fragment:torch.Tensor = torch.cat([audio_fragment, zero_wav], dim=0) audio_fragment:torch.Tensor = torch.cat([audio_fragment, zero_wav], dim=0)
audio[i][j] = audio_fragment.cpu().numpy() audio[i][j] = audio_fragment.cpu().numpy()
if split_bucket: if split_bucket:
audio = self.recovery_order(audio, batch_index_list) audio = self.recovery_order(audio, batch_index_list)
else: else:
# audio = [item for batch in audio for item in batch] # audio = [item for batch in audio for item in batch]
audio = sum(audio, []) audio = sum(audio, [])
audio = np.concatenate(audio, 0) audio = np.concatenate(audio, 0)
audio = (audio * 32768).astype(np.int16) audio = (audio * 32768).astype(np.int16)
try: try:
if speed_factor != 1.0: if speed_factor != 1.0:
audio = speed_change(audio, speed=speed_factor, sr=int(sr)) audio = speed_change(audio, speed=speed_factor, sr=int(sr))
except Exception as e: except Exception as e:
print(f"Failed to change speed of audio: \n{e}") print(f"Failed to change speed of audio: \n{e}")
return sr, audio return sr, audio
def speed_change(input_audio:np.ndarray, speed:float, sr:int): def speed_change(input_audio:np.ndarray, speed:float, sr:int):
# 将 NumPy 数组转换为原始 PCM 流 # 将 NumPy 数组转换为原始 PCM 流
raw_audio = input_audio.astype(np.int16).tobytes() raw_audio = input_audio.astype(np.int16).tobytes()

133
GPT_SoVITS/TTS_infer_pack/tts_instance_pool.py Normal file
View File

@ -0,0 +1,133 @@
import threading
from time import perf_counter
import traceback
from typing import Dict, Union
from GPT_SoVITS.TTS_infer_pack.TTS import TTS, TTS_Config
class TTSWrapper(TTS):
heat: float = 0
usage_count: int = 0
usage_counter: int = 0
usage_time: float = 0.0
first_used_time: float = 0.0
def __init__(self, configs: Union[dict, str, TTS_Config]):
super(TTSWrapper, self).__init__(configs)
self.first_used_time = perf_counter()
def __hash__(self) -> int:
return hash(self.first_used_time)
def run(self, *args, **kwargs):
self.usage_counter += 1
t0 = perf_counter()
for result in super(TTSWrapper, self).run(*args, **kwargs):
yield result
t1 = perf_counter()
self.usage_time += t1 - t0
idle_time = self.usage_time - self.first_used_time
self.heat = self.usage_counter / idle_time
def reset_heat(self):
self.heat: int = 0
self.usage_count: int = 0
self.usage_time: float = 0.0
self.first_used_time: float = perf_counter()
class TTSInstancePool:
def __init__(self, max_size):
self.max_size: int = max_size
self.semaphore: threading.Semaphore = threading.Semaphore(max_size)
self.pool_lock: threading.Lock = threading.Lock()
self.pool: Dict[int, TTSWrapper] = dict()
self.current_index: int = 0
self.size: int = 0
def acquire(self, configs: TTS_Config):
self.semaphore.acquire()
try:
with self.pool_lock:
# 查询最匹配的实例
indexed_key = None
rank = []
for key, tts_instance in self.pool.items():
if tts_instance.configs.vits_weights_path == configs.vits_weights_path \
and tts_instance.configs.t2s_weights_path == configs.t2s_weights_path:
indexed_key = key
rank.append((tts_instance.heat, key))
rank.sort(key=lambda x: x[0])
matched_key = None if len(rank) == 0 else rank[0][1]
# 如果已有实例匹配，则直接复用
if indexed_key is not None:
tts_instance = self._reuse_instance(indexed_key, configs)
print(f"如果已有实例匹配，则直接复用: {configs.vits_weights_path} {configs.t2s_weights_path}")
return tts_instance
# 如果pool未满，则创建一个新实例
if self.size < self.max_size:
tts_instance = TTSWrapper(configs)
self.size += 1
print(f"如果pool未满，则创建一个新实例: {configs.vits_weights_path} {configs.t2s_weights_path}")
return tts_instance
else:
# 否则用最合适的实例进行复用
tts_instance = self._reuse_instance(matched_key, configs)
print(f"否则用最合适的实例进行复用: {configs.vits_weights_path} {configs.t2s_weights_path}")
return tts_instance
except Exception as e:
self.semaphore.release()
traceback.print_exc()
raise e
def release(self, tts_instance: TTSWrapper):
assert tts_instance is not None
with self.pool_lock:
key = hash(tts_instance)
if key in self.pool.keys():
return
self.pool[key] = tts_instance
self.semaphore.release()
def clear_pool(self):
for i in range(self.max_size):
self.semaphore.acquire()
with self.pool_lock:
self.pool.clear()
# for i in range(self.max_size):
self.semaphore.release(self.max_size)
def _reuse_instance(self, instance_key: int, configs: TTS_Config) -> TTSWrapper:
"""
复用已有实例
args:
instance_key: int, 已有实例的Key
config: TTS_Config
return:
TTS_Wrapper: 返回复用的TTS实例
"""
# 复用已有实例
tts_instance = self.pool.pop(instance_key, None)
if tts_instance is None:
raise ValueError("Instance not found")
tts_instance.configs.device = configs.device
if tts_instance.configs.vits_weights_path != configs.vits_weights_path \
or tts_instance.configs.t2s_weights_path != configs.t2s_weights_path:
tts_instance.reset_heat()
if tts_instance.configs.vits_weights_path != configs.vits_weights_path:
tts_instance.init_vits_weights(configs.vits_weights_path, False)
tts_instance.configs.vits_weights_path = configs.vits_weights_path
if tts_instance.configs.t2s_weights_path != configs.t2s_weights_path:
tts_instance.init_t2s_weights(configs.t2s_weights_path, False)
tts_instance.configs.t2s_weights_path = configs.t2s_weights_path
tts_instance.set_device(configs.device, False)
return tts_instance

14
GPT_SoVITS/configs/voices/voice1.yaml Normal file
View File

@ -0,0 +1,14 @@
custom:
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
device: cpu
is_half: false
t2s_weights_path: GPT_weights/liyunlong-e15.ckpt
vits_weights_path: SoVITS_weights/liyunlong_e8_s176.pth
default:
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
device: cpu
is_half: false
t2s_weights_path: GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt
vits_weights_path: GPT_SoVITS/pretrained_models/s2G488k.pth

14
GPT_SoVITS/configs/voices/voice2.yaml Normal file
View File

@ -0,0 +1,14 @@
custom:
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
device: cpu
is_half: false
t2s_weights_path: GPT_weights/jackma-e10.ckpt
vits_weights_path: SoVITS_weights/jackma_e8_s192.pth
default:
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
device: cpu
is_half: false
t2s_weights_path: GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt
vits_weights_path: GPT_SoVITS/pretrained_models/s2G488k.pth

14
GPT_SoVITS/configs/voices/voice3.yaml Normal file
View File

@ -0,0 +1,14 @@
custom:
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
device: cpu
is_half: false
t2s_weights_path: GPT_weights/stephenchow-e15.ckpt
vits_weights_path: SoVITS_weights/stephenchow_e8_s112.pth
default:
bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
device: cpu
is_half: false
t2s_weights_path: GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt
vits_weights_path: GPT_SoVITS/pretrained_models/s2G488k.pth

35
api_v3.py
View File

@ -104,6 +104,8 @@ from typing import Generator
import torch import torch
from TTS_infer_pack.tts_instance_pool import TTSInstancePool, TTSWrapper
now_dir = os.getcwd() now_dir = os.getcwd()
sys.path.append(now_dir) sys.path.append(now_dir)
sys.path.append("%s/GPT_SoVITS" % (now_dir)) sys.path.append("%s/GPT_SoVITS" % (now_dir))
@ -119,11 +121,10 @@ from fastapi.responses import JSONResponse
from fastapi import FastAPI from fastapi import FastAPI
import uvicorn import uvicorn
from io import BytesIO from io import BytesIO
from GPT_SoVITS.TTS_infer_pack.TTS import TTS, TTS_Config from GPT_SoVITS.TTS_infer_pack.TTS import TTS_Config
from GPT_SoVITS.TTS_infer_pack.text_segmentation_method import get_method_names as get_cut_method_names from GPT_SoVITS.TTS_infer_pack.text_segmentation_method import get_method_names as get_cut_method_names
from fastapi.responses import StreamingResponse from fastapi.responses import StreamingResponse
from pydantic import BaseModel from pydantic import BaseModel
from functools import lru_cache
cut_method_names = get_cut_method_names() cut_method_names = get_cut_method_names()
@ -137,6 +138,9 @@ argv = sys.argv
APP = FastAPI() APP = FastAPI()
max_size = 10
tts_instance_pool = TTSInstancePool(max_size)
class TTS_Request(BaseModel): class TTS_Request(BaseModel):
text: str = None text: str = None
@ -162,12 +166,6 @@ class TTS_Request(BaseModel):
"""推理时需要加载的声音模型的yaml配置文件路径，如：GPT_SoVITS/configs/tts_infer.yaml""" """推理时需要加载的声音模型的yaml配置文件路径，如：GPT_SoVITS/configs/tts_infer.yaml"""
@lru_cache(maxsize=10)
def get_tts_instance(tts_config: TTS_Config) -> TTS:
print(f"load tts config from {tts_config.configs_path}")
return TTS(tts_config)
def pack_ogg(io_buffer: BytesIO, data: np.ndarray, rate: int): def pack_ogg(io_buffer: BytesIO, data: np.ndarray, rate: int):
"""modify from https://github.com/RVC-Boss/GPT-SoVITS/pull/894/files""" """modify from https://github.com/RVC-Boss/GPT-SoVITS/pull/894/files"""
with sf.SoundFile(io_buffer, mode='w', samplerate=rate, channels=1, format='ogg') as audio_file: with sf.SoundFile(io_buffer, mode='w', samplerate=rate, channels=1, format='ogg') as audio_file:
@ -318,7 +316,7 @@ async def tts_handle(req: dict):
req["return_fragment"] = True req["return_fragment"] = True
try: try:
tts_instance = get_tts_instance(tts_config) tts_instance = tts_instance_pool.acquire(tts_config)
move_to_gpu(tts_instance, tts_config) move_to_gpu(tts_instance, tts_config)
@ -332,27 +330,30 @@ async def tts_handle(req: dict):
for sr, chunk in tts_generator: for sr, chunk in tts_generator:
yield pack_audio(BytesIO(), chunk, sr, media_type).getvalue() yield pack_audio(BytesIO(), chunk, sr, media_type).getvalue()
move_to_cpu(tts_instance) move_to_cpu(tts_instance)
tts_instance_pool.release(tts_instance)
# _media_type = f"audio/{media_type}" if not (streaming_mode and media_type in ["wav", "raw"]) else f"audio/x-{media_type}" # _media_type = f"audio/{media_type}" if not (streaming_mode and media_type in ["wav", "raw"]) else
# f"audio/x-{media_type}"
return StreamingResponse(streaming_generator(tts_generator, media_type, ), media_type=f"audio/{media_type}") return StreamingResponse(streaming_generator(tts_generator, media_type, ), media_type=f"audio/{media_type}")
else: else:
sr, audio_data = next(tts_generator) sr, audio_data = next(tts_generator)
audio_data = pack_audio(BytesIO(), audio_data, sr, media_type).getvalue() audio_data = pack_audio(BytesIO(), audio_data, sr, media_type).getvalue()
move_to_cpu(tts_instance) move_to_cpu(tts_instance)
tts_instance_pool.release(tts_instance)
return Response(audio_data, media_type=f"audio/{media_type}") return Response(audio_data, media_type=f"audio/{media_type}")
except Exception as e: except Exception as e:
return JSONResponse(status_code=400, content={"message": f"tts failed", "Exception": str(e)}) return JSONResponse(status_code=400, content={"message": f"tts failed", "Exception": str(e)})
def move_to_cpu(tts): def move_to_cpu(tts: TTSWrapper):
cpu_device = torch.device('cpu') cpu_device = torch.device('cpu')
tts.set_device(cpu_device) tts.set_device(cpu_device, False)
print("Moved TTS models to CPU to save GPU memory.") print("Moved TTS models to CPU to save GPU memory.")
def move_to_gpu(tts: TTS, tts_config: TTS_Config): def move_to_gpu(tts: TTSWrapper, tts_config: TTS_Config):
tts.set_device(tts_config.device) tts.set_device(tts_config.device, False)
print("Moved TTS models back to GPU for performance.") print("Moved TTS models back to GPU for performance.")
@ -422,7 +423,7 @@ async def tts_post_endpoint(request: TTS_Request):
async def set_refer_audio(refer_audio_path: str = None, tts_infer_yaml_path: str = "GPT_SoVITS/configs/tts_infer.yaml"): async def set_refer_audio(refer_audio_path: str = None, tts_infer_yaml_path: str = "GPT_SoVITS/configs/tts_infer.yaml"):
try: try:
tts_config = TTS_Config(tts_infer_yaml_path) tts_config = TTS_Config(tts_infer_yaml_path)
tts_instance = get_tts_instance(tts_config) tts_instance = tts_instance_pool.acquire(tts_config)
tts_instance.set_ref_audio(refer_audio_path) tts_instance.set_ref_audio(refer_audio_path)
except Exception as e: except Exception as e:
return JSONResponse(status_code=400, content={"message": f"set refer audio failed", "Exception": str(e)}) return JSONResponse(status_code=400, content={"message": f"set refer audio failed", "Exception": str(e)})
@ -436,7 +437,7 @@ async def set_gpt_weights(weights_path: str = None, tts_infer_yaml_path: str = "
return JSONResponse(status_code=400, content={"message": "gpt weight path is required"}) return JSONResponse(status_code=400, content={"message": "gpt weight path is required"})
tts_config = TTS_Config(tts_infer_yaml_path) tts_config = TTS_Config(tts_infer_yaml_path)
tts_instance = get_tts_instance(tts_config) tts_instance = tts_instance_pool.acquire(tts_config)
tts_instance.init_t2s_weights(weights_path) tts_instance.init_t2s_weights(weights_path)
except Exception as e: except Exception as e:
return JSONResponse(status_code=400, content={"message": f"change gpt weight failed", "Exception": str(e)}) return JSONResponse(status_code=400, content={"message": f"change gpt weight failed", "Exception": str(e)})
@ -451,7 +452,7 @@ async def set_sovits_weights(weights_path: str = None, tts_infer_yaml_path: str
return JSONResponse(status_code=400, content={"message": "sovits weight path is required"}) return JSONResponse(status_code=400, content={"message": "sovits weight path is required"})
tts_config = TTS_Config(tts_infer_yaml_path) tts_config = TTS_Config(tts_infer_yaml_path)
tts_instance = get_tts_instance(tts_config) tts_instance = tts_instance_pool.acquire(tts_config)
tts_instance.init_vits_weights(weights_path) tts_instance.init_vits_weights(weights_path)
except Exception as e: except Exception as e:
return JSONResponse(status_code=400, content={"message": f"change sovits weight failed", "Exception": str(e)}) return JSONResponse(status_code=400, content={"message": f"change sovits weight failed", "Exception": str(e)})