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

4
style_bert_vits2/models/infer.py
View File

@@ -1,12 +1,12 @@
import torch
import utils
from text import cleaned_text_to_sequence, get_bert
from style_bert_vits2.constants import Languages
from style_bert_vits2.logging import logger
from style_bert_vits2.models import commons
from style_bert_vits2.models.models import SynthesizerTrn
from style_bert_vits2.models.models_jp_extra import SynthesizerTrn as SynthesizerTrnJPExtra
from style_bert_vits2.text_processing import cleaned_text_to_sequence, extract_bert_feature
from style_bert_vits2.text_processing.cleaner import clean_text
from style_bert_vits2.text_processing.symbols import SYMBOLS
@@ -77,7 +77,7 @@ def get_text(
for i in range(len(word2ph)):
word2ph[i] = word2ph[i] * 2
word2ph[0] += 1
bert_ori = get_bert(
bert_ori = extract_bert_feature(
norm_text,
word2ph,
language_str,

68
style_bert_vits2/text_processing/__init__.py Normal file
View File

@@ -0,0 +1,68 @@
import torch
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing.symbols import (
LANGUAGE_ID_MAP,
LANGUAGE_TONE_START_MAP,
SYMBOLS,
)
_symbol_to_id = {s: i for i, s in enumerate(SYMBOLS)}
def cleaned_text_to_sequence(cleaned_text: str, tones: list[int], language: Languages) -> tuple[list[int], list[int], list[int]]:
"""
Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
Args:
cleaned_text (str): string to convert to a sequence
tones (list[int]): List of tones
language (Languages): Language of the text
Returns:
tuple[list[int], list[int], list[int]]: 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 extract_bert_feature(
text: str,
word2ph: list[int],
language: Languages,
device: torch.device | str,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
) -> torch.Tensor:
"""
テキストから BERT の特徴量を抽出する
Args:
text (str): テキスト
word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
language (Languages): テキストの言語
device (torch.device | str): 推論に利用するデバイス
assist_text (str | None, optional): 補助テキスト (デフォルト: None)
assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
Returns:
torch.Tensor: BERT の特徴量
"""
if language == Languages.JP:
from style_bert_vits2.text_processing.japanese.bert_feature import extract_bert_feature
elif language == Languages.EN:
from style_bert_vits2.text_processing.english.bert_feature import extract_bert_feature
elif language == Languages.ZH:
from style_bert_vits2.text_processing.chinese.bert_feature import extract_bert_feature
else:
raise ValueError(f"Language {language} not supported")
return extract_bert_feature(text, word2ph, device, assist_text, assist_text_weight)

138
style_bert_vits2/text_processing/chinese/bert_feature.py Normal file
View File

@@ -0,0 +1,138 @@
import sys
import torch
from transformers import PreTrainedModel
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing import bert_models
models: dict[str, PreTrainedModel] = {}
def extract_bert_feature(
text: str,
word2ph: list[int],
device: torch.device | str,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
) -> torch.Tensor:
"""
中国語のテキストから BERT の特徴量を抽出する
Args:
text (str): 中国語のテキスト
word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
device (torch.device | str): 推論に利用するデバイス
assist_text (str | None, optional): 補助テキスト (デフォルト: None)
assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
Returns:
torch.Tensor: BERT の特徴量
"""
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) # type: ignore
style_res_mean = None
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) # type: ignore
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) # type: ignore
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:
assert style_res_mean is not None
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])

79
style_bert_vits2/text_processing/english/bert_feature.py Normal file
View File

@@ -0,0 +1,79 @@
import sys
import torch
from transformers import PreTrainedModel
from style_bert_vits2.constants import Languages
from style_bert_vits2.text_processing import bert_models
models: dict[str, PreTrainedModel] = {}
def extract_bert_feature(
text: str,
word2ph: list[int],
device: torch.device | str,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
) -> torch.Tensor:
"""
英語のテキストから BERT の特徴量を抽出する
Args:
text (str): 英語のテキスト
word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
device (torch.device | str): 推論に利用するデバイス
assist_text (str | None, optional): 補助テキスト (デフォルト: None)
assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
Returns:
torch.Tensor: BERT の特徴量
"""
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) # type: ignore
style_res_mean = None
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) # type: ignore
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) # type: ignore
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:
assert style_res_mean is not None
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

86
style_bert_vits2/text_processing/japanese/bert_feature.py Normal file
View File

@@ -0,0 +1,86 @@
import sys
import torch
from transformers import PreTrainedModel
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[str, PreTrainedModel] = {}
def extract_bert_feature(
text: str,
word2ph: list[int],
device: torch.device | str,
assist_text: str | None = None,
assist_text_weight: float = 0.7,
) -> torch.Tensor:
"""
日本語のテキストから BERT の特徴量を抽出する
Args:
text (str): 日本語のテキスト
word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
device (torch.device | str): 推論に利用するデバイス
assist_text (str | None, optional): 補助テキスト (デフォルト: None)
assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
Returns:
torch.Tensor: BERT の特徴量
"""
# 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) # type: ignore
style_res_mean = None
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) # type: ignore
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) # type: ignore
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:
assert style_res_mean is not None
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