blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2025-10-07 23:48:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge 04014b959e19ef150f54997f406575d36e14040e into 95354647c87fc86a416361b39e7268afd95e193d

Browse Source
This commit is contained in:
PocketDoc Labs 2024-07-02 02:01:34 +08:00 committed by GitHub
commit ae339ef8d9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 15 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
GPT_SoVITS/AR/models/t2s_model.py
View File

@ -328,6 +328,7 @@ class Text2SemanticDecoder(nn.Module):
prompts, ####参考音频token
bert_feature,
top_k: int = -100,
min_p: float = 0.0,
top_p: int = 100,
early_stop_num: int = -1,
temperature: float = 1.0,
@ -397,7 +398,7 @@ class Text2SemanticDecoder(nn.Module):
if(idx==0):###第一次跑不能EOS否则没有了
logits = logits[:, :-1] ###刨除1024终止符号的概率
samples = sample(
logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
logits[0], y, top_k=top_k, min_p=min_p, top_p=top_p, repetition_penalty=1.35, temperature=temperature
)[0].unsqueeze(0)
# 本次生成的 semantic_ids 和之前的 y 构成新的 y
# print(samples.shape)#[1,1]#第一个1是bs

9
GPT_SoVITS/AR/models/utils.py
View File

@ -112,6 +112,7 @@ def logits_to_probs(
previous_tokens: Optional[torch.Tensor] = None,
temperature: float = 1.0,
top_k: Optional[int] = None,
min_p: Optional[float] = None,
top_p: Optional[int] = None,
repetition_penalty: float = 1.0,
):
@ -127,6 +128,12 @@ def logits_to_probs(
)
logits.scatter_(dim=0, index=previous_tokens, src=score)
# Min-p sampling, we sample tokens with a probability above the min_p * max_prob where max_prob is the maximum probability in the distribution
if min_p is not None and min_p < 1.0 and min_p > 0.0:
max_prob = torch.max(logits)
logits = torch.where(logits < min_p * max_prob, -float("Inf"), logits)
# Top-p sampling, we sample from the smallest set of tokens whose cumulative probability mass exceeds the threshold `p`
if top_p is not None and top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cum_probs = torch.cumsum(
@ -139,8 +146,10 @@ def logits_to_probs(
)
logits = logits.masked_fill(indices_to_remove, -float("Inf"))
# Use temperature to smooth the distribution
logits = logits / max(temperature, 1e-5)
# Top-k sampling, we sample only from the top k most likely tokens
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
pivot = v.select(-1, -1).unsqueeze(-1)

6
GPT_SoVITS/inference_webui.py
View File

@ -312,7 +312,7 @@ def merge_short_text_in_array(texts, threshold):
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):
def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=i18n("不切"), top_k=20, min_p=0.0, 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()
@ -402,6 +402,7 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,
bert,
# prompt_phone_len=ph_offset,
top_k=top_k,
min_p=min_p,
top_p=top_p,
temperature=temperature,
early_stop_num=hz * max_sec,
@ -612,6 +613,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
with gr.Row():
gr.Markdown(value=i18n("gpt采样参数(无参考文本时不要太低)："))
top_k = gr.Slider(minimum=1,maximum=100,step=1,label=i18n("top_k"),value=5,interactive=True)
min_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("min_p"),value=0.0,interactive=True)
top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("top_p"),value=1,interactive=True)
temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("temperature"),value=1,interactive=True)
inference_button = gr.Button(i18n("合成语音"), variant="primary")
@ -619,7 +621,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
inference_button.click(
get_tts_wav,
[inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free],
[inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, min_p, top_p, temperature, ref_text_free],
[output],
)