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

update_infer

Browse Source
This commit is contained in:
Watchtower-Liu
2024-02-16 16:53:57 +08:00
parent 41041715a4
commit 1803729360
6 changed files with 88 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

107
GPT_SoVITS/AR/models/t2s_model.py
View File

@@ -240,7 +240,7 @@ class Text2SemanticDecoder(nn.Module):
# AR Decoder
y = prompts
prefix_len = y.shape[1]
x_len = x.shape[1]
x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
stop = False
@@ -256,47 +256,41 @@ class Text2SemanticDecoder(nn.Module):
"first_infer": 1,
"stage": 0,
}
for idx in tqdm(range(1500)):
if cache["first_infer"] == 1:
y_emb = self.ar_audio_embedding(y)
else:
y_emb = torch.cat(
[cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], 1
)
cache["y_emb"] = y_emb
################### first step ##########################
if y is not None:
y_emb = self.ar_audio_embedding(y)
y_len = y_emb.shape[1]
prefix_len = y.shape[1]
y_pos = self.ar_audio_position(y_emb)
# x 和逐渐增长的 y 一起输入给模型
if cache["first_infer"] == 1:
xy_pos = torch.concat([x, y_pos], dim=1)
else:
xy_pos = y_pos[:, -1:]
y_len = y_pos.shape[1]
###以下3个不做缓存
if cache["first_infer"] == 1:
x_attn_mask_pad = F.pad(
xy_pos = torch.concat([x, y_pos], dim=1)
cache["y_emb"] = y_emb
ref_free = False
else:
y_emb = None
y_len = 0
prefix_len = 0
y_pos = None
xy_pos = x
y = torch.zeros(x.shape[0], 0, dtype=torch.int, device=x.device)
ref_free = True
x_attn_mask_pad = F.pad(
x_attn_mask,
(0, y_len), ###xx的纯0扩展到xx纯0+xy纯1，(x,x+y)
value=True,
)
y_attn_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),
(x_len, 0),
value=False,
)
xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
y.device
)
else:
###最右边一列（是错的）
# xy_attn_mask=torch.ones((1, x_len+y_len), dtype=torch.bool,device=xy_pos.device)
# xy_attn_mask[:,-1]=False
###最下面一行（是对的）
xy_attn_mask = torch.zeros(
(1, x_len + y_len), dtype=torch.bool, device=xy_pos.device
)
# pdb.set_trace()
###缓存重头戏
# print(1111,xy_pos.shape,xy_attn_mask.shape,x_len,y_len)
y_attn_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),
(x_len, 0),
value=False,
)
xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
x.device
)
for idx in tqdm(range(1500)):
xy_dec, _ = self.h((xy_pos, None), mask=xy_attn_mask, cache=cache)
logits = self.ar_predict_layer(
xy_dec[:, -1]
@@ -307,6 +301,10 @@ class Text2SemanticDecoder(nn.Module):
samples = sample(
logits[0], y, top_k=top_k, top_p=1.0, repetition_penalty=1.35
)[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
y = torch.concat([y, samples], dim=1)
if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
print("use early stop num:", early_stop_num)
stop = True
@@ -315,13 +313,38 @@ class Text2SemanticDecoder(nn.Module):
# print(torch.argmax(logits, dim=-1)[0] == self.EOS, samples[0, 0] == self.EOS)
stop = True
if stop:
if prompts.shape[1] == y.shape[1]:
# if prompts.shape[1] == y.shape[1]:
# y = torch.concat([y, torch.zeros_like(samples)], dim=1)
# print("bad zero prediction")
if y.shape[1]==0:
y = torch.concat([y, torch.zeros_like(samples)], dim=1)
print("bad zero prediction")
print(f"T2S Decoding EOS [{prefix_len} -> {y.shape[1]}]")
break
# 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
y = torch.concat([y, samples], dim=1)
####################### update next step ###################################
cache["first_infer"] = 0
return y, idx
if cache["y_emb"] is not None:
y_emb = torch.cat(
[cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], dim = 1
)
cache["y_emb"] = y_emb
y_pos = self.ar_audio_position(y_emb)
xy_pos = y_pos[:, -1:]
else:
y_emb = self.ar_audio_embedding(y[:, -1:])
cache["y_emb"] = y_emb
y_pos = self.ar_audio_position(y_emb)
xy_pos = y_pos
y_len = y_pos.shape[1]
###最右边一列（是错的）
# xy_attn_mask=torch.ones((1, x_len+y_len), dtype=torch.bool,device=xy_pos.device)
# xy_attn_mask[:,-1]=False
###最下面一行（是对的）
xy_attn_mask = torch.zeros(
(1, x_len + y_len), dtype=torch.bool, device=xy_pos.device
)
if ref_free:
return y[:, :-1], 0
return y[:, :-1], idx-1