blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2025-04-06 03:57:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

【fast_inference_】通过简单添加torch.no_grad()修饰器，可能会使速度加快 (#930)

Browse Source 
* 添加with torch.no_grad()，速度快一大截

* 恢复先前缩进

* 恢复make batch的位置

* 改用修饰器

* 去除没必要的增加的空行

---------

Co-authored-by: XTer <xxoy1234@outlook.com>
This commit is contained in:
箱庭XTer 2024-04-07 18:36:09 +08:00 committed by GitHub
parent ec7647e08d
commit 3706ad1b8b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 43 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
.gitignore vendored
View File

@ -10,6 +10,8 @@ reference
GPT_weights GPT_weights
SoVITS_weights SoVITS_weights
TEMP TEMP
PortableGit
ffmpeg.exe ffmpeg.exe
ffprobe.exe ffprobe.exe
tmp_audio
trained

43
GPT_SoVITS/TTS_infer_pack/TTS.py
View File

@ -249,8 +249,6 @@ 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.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):
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
@ -437,7 +435,8 @@ class TTS:
device:torch.device=torch.device("cpu"), device:torch.device=torch.device("cpu"),
precision:torch.dtype=torch.float32, precision:torch.dtype=torch.float32,
): ):
# 但是这里不能套，反而会负优化
# with torch.no_grad():
_data:list = [] _data:list = []
index_and_len_list = [] index_and_len_list = []
for idx, item in enumerate(data): for idx, item in enumerate(data):
@ -485,6 +484,8 @@ class TTS:
norm_text_batch = [] norm_text_batch = []
bert_max_len = 0 bert_max_len = 0
phones_max_len = 0 phones_max_len = 0
# 但是这里也不能套，反而会负优化
# with torch.no_grad():
for item in item_list: for item in item_list:
if prompt_data is not None: if prompt_data is not None:
all_bert_features = torch.cat([prompt_data["bert_features"], item["bert_features"]], 1)\ all_bert_features = torch.cat([prompt_data["bert_features"], item["bert_features"]], 1)\
@ -568,7 +569,8 @@ class TTS:
''' '''
self.stop_flag = True self.stop_flag = True
# 使用装饰器
@torch.no_grad()
def run(self, inputs:dict): def run(self, inputs:dict):
""" """
Text to speech inference. Text to speech inference.
@ -643,11 +645,10 @@ class TTS:
((self.prompt_cache["prompt_semantic"] is None) or (self.prompt_cache["refer_spec"] is None)): ((self.prompt_cache["prompt_semantic"] is None) or (self.prompt_cache["refer_spec"] is None)):
raise ValueError("ref_audio_path cannot be empty, when the reference audio is not set using set_ref_audio()") raise ValueError("ref_audio_path cannot be empty, when the reference audio is not set using set_ref_audio()")
###### setting reference audio and prompt text preprocessing ######## ###### setting reference audio and prompt text preprocessing ########
t0 = ttime() t0 = ttime()
if (ref_audio_path is not None) and (ref_audio_path != self.prompt_cache["ref_audio_path"]): if (ref_audio_path is not None) and (ref_audio_path != self.prompt_cache["ref_audio_path"]):
self.set_ref_audio(ref_audio_path) self.set_ref_audio(ref_audio_path)
if not no_prompt_text: if not no_prompt_text:
prompt_text = prompt_text.strip("\n") prompt_text = prompt_text.strip("\n")
@ -664,7 +665,6 @@ class TTS:
self.prompt_cache["bert_features"] = bert_features self.prompt_cache["bert_features"] = bert_features
self.prompt_cache["norm_text"] = norm_text self.prompt_cache["norm_text"] = norm_text
###### text preprocessing ######## ###### text preprocessing ########
t1 = ttime() t1 = ttime()
data:list = None data:list = None
@ -718,6 +718,7 @@ class TTS:
) )
return batch[0] return batch[0]
t2 = ttime() t2 = ttime()
try: try:
print("############ 推理 ############") print("############ 推理 ############")
@ -745,18 +746,18 @@ class TTS:
else: else:
prompt = self.prompt_cache["prompt_semantic"].expand(len(all_phoneme_ids), -1).to(self.configs.device) prompt = self.prompt_cache["prompt_semantic"].expand(len(all_phoneme_ids), -1).to(self.configs.device)
with torch.no_grad():
pred_semantic_list, idx_list = self.t2s_model.model.infer_panel( pred_semantic_list, idx_list = self.t2s_model.model.infer_panel(
all_phoneme_ids, all_phoneme_ids,
all_phoneme_lens, all_phoneme_lens,
prompt, prompt,
all_bert_features, all_bert_features,
# prompt_phone_len=ph_offset, # prompt_phone_len=ph_offset,
top_k=top_k, top_k=top_k,
top_p=top_p, top_p=top_p,
temperature=temperature, temperature=temperature,
early_stop_num=self.configs.hz * self.configs.max_sec, early_stop_num=self.configs.hz * self.configs.max_sec,
) )
t4 = ttime() t4 = ttime()
t_34 += t4 - t3 t_34 += t4 - t3
@ -765,6 +766,9 @@ class TTS:
batch_audio_fragment = [] batch_audio_fragment = []
# 这里要记得加 torch.no_grad() 不然速度慢一大截
# 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)
@ -791,7 +795,6 @@ class TTS:
audio_frag_end_idx.insert(0, 0) audio_frag_end_idx.insert(0, 0)
batch_audio_fragment= [_batch_audio_fragment[audio_frag_end_idx[i-1]:audio_frag_end_idx[i]] for i in range(1, len(audio_frag_end_idx))] batch_audio_fragment= [_batch_audio_fragment[audio_frag_end_idx[i-1]:audio_frag_end_idx[i]] for i in range(1, len(audio_frag_end_idx))]
# ## vits串行推理 # ## vits串行推理
# for i, idx in enumerate(idx_list): # for i, idx in enumerate(idx_list):
# phones = batch_phones[i].unsqueeze(0).to(self.configs.device) # phones = batch_phones[i].unsqueeze(0).to(self.configs.device)