bacow/GPT-SoVITS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

改变训练和推理时的mask策略，以修复当batch_size>1时，产生的复读现象 (#966)

Browse Source
This commit is contained in:
ChasonJiang
2024-04-12 18:00:50 +08:00
committed by GitHub
parent 3706ad1b8b
commit 959269b5ae
2 changed files with 48 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
GPT_SoVITS/AR/models/t2s_model.py
View File

@@ -297,7 +297,8 @@ class Text2SemanticDecoder(nn.Module):
(0, y_len),
value=True,
)
# 取消对y[0]的mask,以防止复读，详见https://github.com/RVC-Boss/GPT-SoVITS/issues/965
x_attn_mask[:, x_len]=False
y_attn_mask = F.pad(
torch.triu(
torch.ones(y_len, y_len, dtype=torch.bool, device=x.device),
@@ -393,6 +394,8 @@ class Text2SemanticDecoder(nn.Module):
(0, y_len),
value=True,
)
# 取消对y[0]的mask,以防止复读，详见https://github.com/RVC-Boss/GPT-SoVITS/issues/965
x_attn_mask[:, x_len]=False
y_attn_mask = F.pad(
torch.triu(
torch.ones(y_len, y_len, dtype=torch.bool, device=x.device),
@@ -458,7 +461,7 @@ class Text2SemanticDecoder(nn.Module):
value=True,
)
y_attn_mask = F.pad(
torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=0),# diagonal必须为0，否则会导致batch_size>1时的复读情况
(x_len, 0),
value=False,
)
@@ -504,29 +507,29 @@ class Text2SemanticDecoder(nn.Module):
def infer_panel_batch_infer_with_flash_attn(
self,
x:List[torch.LongTensor], #####全部文本token
x:torch.LongTensor, #####全部文本token
x_lens:torch.LongTensor,
prompts:torch.LongTensor, ####参考音频token
bert_feature:List[torch.LongTensor],
bert_feature:torch.LongTensor,
top_k: int = -100,
top_p: int = 100,
early_stop_num: int = -1,
temperature: float = 1.0,
):
# 先对phones进行embedding、对bert_features进行project，再pad到相同长度，以缓解复读问题。（可能还有其他因素导致复读）
max_len = 0
for x_item, bert_item in zip(x, bert_feature):
max_len = max(max_len, x_item.shape[0], bert_item.shape[1])
x_list = [self.ar_text_embedding(item) for item in x]
x_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) if item.shape[0]<max_len else item for item in x_list]
x = torch.stack(x_list, dim=0)
## 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）
# max_len = 0
# for x_item, bert_item in zip(x, bert_feature):
# max_len = max(max_len, x_item.shape[0], bert_item.shape[1])
# x_list = [self.ar_text_embedding(item) for item in x]
# x_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) if item.shape[0]<max_len else item for item in x_list]
# x = torch.stack(x_list, dim=0)
bert_features_list = [self.bert_proj(item.transpose(0, 1)) for item in bert_feature]
bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) if item.shape[0]<max_len else item for item in bert_features_list]
bert_feature = torch.stack(bert_features_list, dim=0)
# bert_features_list = [self.bert_proj(item.transpose(0, 1)) for item in bert_feature]
# bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) if item.shape[0]<max_len else item for item in bert_features_list]
# bert_feature = torch.stack(bert_features_list, dim=0)
# bert_feature = self.bert_proj(bert_feature.transpose(1, 2))
# x = self.ar_text_embedding(x)
bert_feature = self.bert_proj(bert_feature.transpose(1, 2))
x = self.ar_text_embedding(x)
x = x + bert_feature
x = self.ar_text_position(x)
@@ -573,8 +576,8 @@ class Text2SemanticDecoder(nn.Module):
(0, y_len), ###xx的纯0扩展到xx纯0+xy纯1，(x,x+y)
value=True,
)
y_mask = F.pad( ###yy的右上1扩展到左边xy的0,(y,x+y)
torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
y_mask = F.pad( ###yy的右上0扩展到左边xy的0,(y,x+y)
torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=0), # diagonal必须为0，否则会导致batch_size>1时的复读情况
(x_len, 0),
value=False,
)
@@ -669,29 +672,29 @@ class Text2SemanticDecoder(nn.Module):
def infer_panel_batch_only(
self,
x:List[torch.LongTensor], #####全部文本token
x:torch.LongTensor, #####全部文本token
x_lens:torch.LongTensor,
prompts:torch.LongTensor, ####参考音频token
bert_feature:List[torch.LongTensor],
bert_feature:torch.LongTensor,
top_k: int = -100,
top_p: int = 100,
early_stop_num: int = -1,
temperature: float = 1.0,
):
# 先对phones进行embedding、对bert_features进行project，再pad到相同长度，以缓解复读问题。（可能还有其他因素导致复读）
max_len = 0
for x_item, bert_item in zip(x, bert_feature):
max_len = max(max_len, x_item.shape[0], bert_item.shape[1])
x_list = [self.ar_text_embedding(item) for item in x]
x_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) if item.shape[0]<max_len else item for item in x_list]
x = torch.stack(x_list, dim=0)
## 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）
# max_len = 0
# for x_item, bert_item in zip(x, bert_feature):
# max_len = max(max_len, x_item.shape[0], bert_item.shape[1])
# x_list = [self.ar_text_embedding(item) for item in x]
# x_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) if item.shape[0]<max_len else item for item in x_list]
# x = torch.stack(x_list, dim=0)
bert_features_list = [self.bert_proj(item.transpose(0, 1)) for item in bert_feature]
bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) if item.shape[0]<max_len else item for item in bert_features_list]
bert_feature = torch.stack(bert_features_list, dim=0)
# bert_features_list = [self.bert_proj(item.transpose(0, 1)) for item in bert_feature]
# bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) if item.shape[0]<max_len else item for item in bert_features_list]
# bert_feature = torch.stack(bert_features_list, dim=0)
# bert_feature = self.bert_proj(bert_feature.transpose(1, 2))
# x = self.ar_text_embedding(x)
bert_feature = self.bert_proj(bert_feature.transpose(1, 2))
x = self.ar_text_embedding(x)
x = x + bert_feature
x = self.ar_text_position(x)
@@ -747,7 +750,7 @@ class Text2SemanticDecoder(nn.Module):
value=True,
)
y_mask = F.pad( ###yy的右上1扩展到左边xy的0,(y,x+y)
torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=0), # diagonal必须为0，否则会导致batch_size>1时的复读情况
(x_len, 0),
value=False,
)