tuna2134/sbv2-v2
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Refactor: moved the module for extracting BERT features from text in each language to style_bert_vits2/text_processing/(language)/bert_feature.py

Browse Source
This commit is contained in:
tsukumi
2024-03-07 03:32:07 +00:00
parent c3c0dd8b32
commit 62919e904e
16 changed files with 172 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
text/__init__.py
View File

@@ -1,43 +0,0 @@
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing.symbols import *
_symbol_to_id = {s: i for i, s in enumerate(SYMBOLS)}
def cleaned_text_to_sequence(cleaned_text: str, tones: list[int], language: Languages):
"""
Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
Args:
text: string to convert to a sequence
Returns:
List of integers corresponding to the symbols in the text
"""
phones = [_symbol_to_id[symbol] for symbol in cleaned_text]
tone_start = LANGUAGE_TONE_START_MAP[language]
tones = [i + tone_start for i in tones]
lang_id = LANGUAGE_ID_MAP[language]
lang_ids = [lang_id for i in phones]
return phones, tones, lang_ids
def get_bert(
text: str,
word2ph,
language: Languages,
device: str,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
):
if language == Languages.ZH:
from .chinese_bert import get_bert_feature
elif language == Languages.EN:
from .english_bert_mock import get_bert_feature
elif language == Languages.JP:
from .japanese_bert import get_bert_feature
else:
raise ValueError(f"Language {language} not supported")
return get_bert_feature(text, word2ph, device, assist_text, assist_text_weight)

10
text/chinese.py
View File

@@ -176,20 +176,14 @@ def normalize_text(text):
return text
def get_bert_feature(text, word2ph):
from text import chinese_bert
return chinese_bert.get_bert_feature(text, word2ph)
if __name__ == "__main__":
from text.chinese_bert import get_bert_feature
from style_bert_vits2.text_processing.chinese.bert_feature import extract_bert_feature
text = "啊！但是《原神》是由,米哈\游自主， [研发]的一款全.新开放世界.冒险游戏"
text = normalize_text(text)
print(text)
phones, tones, word2ph = g2p(text)
bert = get_bert_feature(text, word2ph)
bert = extract_bert_feature(text, word2ph, 'cuda')
print(phones, tones, word2ph, bert.shape)

120
text/chinese_bert.py
View File

@@ -1,120 +0,0 @@
import sys
import torch
from config import config
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing import bert_models
models = dict()
def get_bert_feature(
text: str,
word2ph,
device = config.bert_gen_config.device,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
):
if (
sys.platform == "darwin"
and torch.backends.mps.is_available()
and device == "cpu"
):
device = "mps"
if not device:
device = "cuda"
if device == "cuda" and not torch.cuda.is_available():
device = "cpu"
if device not in models.keys():
models[device] = bert_models.load_model(Languages.ZH).to(device)
with torch.no_grad():
tokenizer = bert_models.load_tokenizer(Languages.ZH)
inputs = tokenizer(text, return_tensors="pt")
for i in inputs:
inputs[i] = inputs[i].to(device)
res = models[device](**inputs, output_hidden_states=True)
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
if assist_text:
style_inputs = tokenizer(assist_text, return_tensors="pt")
for i in style_inputs:
style_inputs[i] = style_inputs[i].to(device)
style_res = models[device](**style_inputs, output_hidden_states=True)
style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
style_res_mean = style_res.mean(0)
assert len(word2ph) == len(text) + 2
word2phone = word2ph
phone_level_feature = []
for i in range(len(word2phone)):
if assist_text:
repeat_feature = (
res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+ style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
)
else:
repeat_feature = res[i].repeat(word2phone[i], 1)
phone_level_feature.append(repeat_feature)
phone_level_feature = torch.cat(phone_level_feature, dim=0)
return phone_level_feature.T
if __name__ == "__main__":
word_level_feature = torch.rand(38, 1024) # 12个词,每个词1024维特征
word2phone = [
1,
2,
1,
2,
2,
1,
2,
2,
1,
2,
2,
1,
2,
2,
2,
2,
2,
1,
1,
2,
2,
1,
2,
2,
2,
2,
1,
2,
2,
2,
2,
2,
1,
2,
2,
2,
2,
1,
]
# 计算总帧数
total_frames = sum(word2phone)
print(word_level_feature.shape)
print(word2phone)
phone_level_feature = []
for i in range(len(word2phone)):
print(word_level_feature[i].shape)
# 对每个词重复word2phone[i]次
repeat_feature = word_level_feature[i].repeat(word2phone[i], 1)
phone_level_feature.append(repeat_feature)
phone_level_feature = torch.cat(phone_level_feature, dim=0)
print(phone_level_feature.shape) # torch.Size([36, 1024])

6
text/english.py
View File

@@ -477,12 +477,6 @@ def g2p(text):
return phones, tones, word2ph
def get_bert_feature(text, word2ph):
from text import english_bert_mock
return english_bert_mock.get_bert_feature(text, word2ph)
if __name__ == "__main__":
# print(get_dict())
# print(eng_word_to_phoneme("hello"))

61
text/english_bert_mock.py
View File

@@ -1,61 +0,0 @@
import sys
import torch
from config import config
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing import bert_models
models = dict()
def get_bert_feature(
text: str,
word2ph,
device = config.bert_gen_config.device,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
):
if (
sys.platform == "darwin"
and torch.backends.mps.is_available()
and device == "cpu"
):
device = "mps"
if not device:
device = "cuda"
if device == "cuda" and not torch.cuda.is_available():
device = "cpu"
if device not in models.keys():
models[device] = bert_models.load_model(Languages.EN).to(device)
with torch.no_grad():
tokenizer = bert_models.load_tokenizer(Languages.EN)
inputs = tokenizer(text, return_tensors="pt")
for i in inputs:
inputs[i] = inputs[i].to(device)
res = models[device](**inputs, output_hidden_states=True)
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
if assist_text:
style_inputs = tokenizer(assist_text, return_tensors="pt")
for i in style_inputs:
style_inputs[i] = style_inputs[i].to(device)
style_res = models[device](**style_inputs, output_hidden_states=True)
style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
style_res_mean = style_res.mean(0)
assert len(word2ph) == res.shape[0], (text, res.shape[0], len(word2ph))
word2phone = word2ph
phone_level_feature = []
for i in range(len(word2phone)):
if assist_text:
repeat_feature = (
res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+ style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
)
else:
repeat_feature = res[i].repeat(word2phone[i], 1)
phone_level_feature.append(repeat_feature)
phone_level_feature = torch.cat(phone_level_feature, dim=0)
return phone_level_feature.T

69
text/japanese_bert.py
View File

@@ -1,69 +0,0 @@
import sys
import torch
from config import config
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing import bert_models
from style_bert_vits2.text_processing.japanese.g2p import text_to_sep_kata
models = dict()
def get_bert_feature(
text: str,
word2ph,
device = config.bert_gen_config.device,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
):
# 各単語が何文字かを作る`word2ph`を使う必要があるので、読めない文字は必ず無視する
# でないと`word2ph`の結果とテキストの文字数結果が整合性が取れない
text = "".join(text_to_sep_kata(text, raise_yomi_error=False)[0])
if assist_text:
assist_text = "".join(text_to_sep_kata(assist_text, raise_yomi_error=False)[0])
if (
sys.platform == "darwin"
and torch.backends.mps.is_available()
and device == "cpu"
):
device = "mps"
if not device:
device = "cuda"
if device == "cuda" and not torch.cuda.is_available():
device = "cpu"
if device not in models.keys():
models[device] = bert_models.load_model(Languages.JP).to(device)
with torch.no_grad():
tokenizer = bert_models.load_tokenizer(Languages.JP)
inputs = tokenizer(text, return_tensors="pt")
for i in inputs:
inputs[i] = inputs[i].to(device)
res = models[device](**inputs, output_hidden_states=True)
res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
if assist_text:
style_inputs = tokenizer(assist_text, return_tensors="pt")
for i in style_inputs:
style_inputs[i] = style_inputs[i].to(device)
style_res = models[device](**style_inputs, output_hidden_states=True)
style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
style_res_mean = style_res.mean(0)
assert len(word2ph) == len(text) + 2, text
word2phone = word2ph
phone_level_feature = []
for i in range(len(word2phone)):
if assist_text:
repeat_feature = (
res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+ style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
)
else:
repeat_feature = res[i].repeat(word2phone[i], 1)
phone_level_feature.append(repeat_feature)
phone_level_feature = torch.cat(phone_level_feature, dim=0)
return phone_level_feature.T