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

app.py

@@ -10,6 +10,7 @@ import gradio as gr
 import torch
 import yaml
 
+from common.tts_model import ModelHolder
 from style_bert_vits2.constants import (
     DEFAULT_ASSIST_TEXT_WEIGHT,
     DEFAULT_LENGTH,
@@ -25,11 +26,11 @@ from style_bert_vits2.constants import (
     Languages,
 )
 from style_bert_vits2.logging import logger
-from common.tts_model import ModelHolder
 from style_bert_vits2.models.infer import InvalidToneError
 from style_bert_vits2.text_processing.japanese.g2p_utils import g2kata_tone, kata_tone2phone_tone
 from style_bert_vits2.text_processing.japanese.normalizer import normalize_text
 
+
 # Get path settings
 with open(os.path.join("configs", "paths.yml"), "r", encoding="utf-8") as f:
     path_config: dict[str, str] = yaml.safe_load(f.read())

bert_gen.py

@@ -5,12 +5,12 @@ import torch
 import torch.multiprocessing as mp
 from tqdm import tqdm
 
-from style_bert_vits2.models import commons
 import utils
-from style_bert_vits2.logging import logger
-from style_bert_vits2.utils.stdout_wrapper import SAFE_STDOUT
 from config import config
-from text import cleaned_text_to_sequence, get_bert
+from style_bert_vits2.logging import logger
+from style_bert_vits2.models import commons
+from style_bert_vits2.text_processing import cleaned_text_to_sequence, extract_bert_feature
+from style_bert_vits2.utils.stdout_wrapper import SAFE_STDOUT
 
 
 def process_line(x):
@@ -45,7 +45,7 @@ def process_line(x):
         bert = torch.load(bert_path)
         assert bert.shape[-1] == len(phone)
     except Exception:
-        bert = get_bert(text, word2ph, language_str, device)
+        bert = extract_bert_feature(text, word2ph, language_str, device)
         assert bert.shape[-1] == len(phone)
         torch.save(bert, bert_path)
 

data_utils.py

@@ -7,12 +7,12 @@ import torch
 import torch.utils.data
 from tqdm import tqdm
 
-from style_bert_vits2.models import commons
 from config import config
 from mel_processing import mel_spectrogram_torch, spectrogram_torch
-from text import cleaned_text_to_sequence
-from style_bert_vits2.logging import logger
 from utils import load_filepaths_and_text, load_wav_to_torch
+from style_bert_vits2.logging import logger
+from style_bert_vits2.models import commons
+from style_bert_vits2.text_processing import cleaned_text_to_sequence
 
 """Multi speaker version"""
 

server_fastapi.py

@@ -20,6 +20,8 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, Response
 from scipy.io import wavfile
 
+from common.tts_model import Model, ModelHolder
+from config import config
 from style_bert_vits2.constants import (
     DEFAULT_ASSIST_TEXT_WEIGHT,
     DEFAULT_LENGTH,
@@ -33,8 +35,6 @@ from style_bert_vits2.constants import (
     Languages,
 )
 from style_bert_vits2.logging import logger
-from common.tts_model import Model, ModelHolder
-from config import config
 
 ln = config.server_config.language
 

style_bert_vits2/models/infer.py

@@ -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,

style_bert_vits2/text_processing/__init__.py

@@ -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)

style_bert_vits2/text_processing/chinese/bert_feature.py

@@ -1,22 +1,36 @@
 import sys
 
 import torch
+from transformers import PreTrainedModel
 
-from config import config
 from style_bert_vits2.constants import Languages
 from style_bert_vits2.text_processing import bert_models
 
 
-models = dict()
+models: dict[str, PreTrainedModel] = {}
 
 
-def get_bert_feature(
+def extract_bert_feature(
     text: str,
-    word2ph,
+    word2ph: list[int],
-    device = config.bert_gen_config.device,
+    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()
@@ -28,26 +42,30 @@ def get_bert_feature(
     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)
+        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)
+            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)
+                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

style_bert_vits2/text_processing/english/bert_feature.py

@@ -1,22 +1,36 @@
 import sys
 
 import torch
+from transformers import PreTrainedModel
 
-from config import config
 from style_bert_vits2.constants import Languages
 from style_bert_vits2.text_processing import bert_models
 
 
-models = dict()
+models: dict[str, PreTrainedModel] = {}
 
 
-def get_bert_feature(
+def extract_bert_feature(
     text: str,
-    word2ph,
+    word2ph: list[int],
-    device = config.bert_gen_config.device,
+    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()
@@ -28,26 +42,30 @@ def get_bert_feature(
     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)
+        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)
+            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)
+                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

style_bert_vits2/text_processing/japanese/bert_feature.py

@@ -1,25 +1,39 @@
 import sys
 
 import torch
+from transformers import PreTrainedModel
 
-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()
+models: dict[str, PreTrainedModel] = {}
 
 
-def get_bert_feature(
+def extract_bert_feature(
     text: str,
-    word2ph,
+    word2ph: list[int],
-    device = config.bert_gen_config.device,
+    device: torch.device | str,
     assist_text: str | None = None,
     assist_text_weight: float = 0.7,
-):
+) -> torch.Tensor:
-    # 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
+    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:
@@ -35,18 +49,20 @@ def get_bert_feature(
     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)
+        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)
+            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)
+                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)
@@ -56,6 +72,7 @@ def get_bert_feature(
     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

text/__init__.py

@@ -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)

text/chinese.py

@@ -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)
 

text/english.py

@@ -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"))

webui_dataset.py

@@ -8,6 +8,7 @@ from style_bert_vits2.constants import GRADIO_THEME
 from style_bert_vits2.logging import logger
 from style_bert_vits2.utils.subprocess import run_script_with_log
 
+
 # Get path settings
 with open(os.path.join("configs", "paths.yml"), "r", encoding="utf-8") as f:
     path_config: dict[str, str] = yaml.safe_load(f.read())

webui_merge.py

@@ -11,9 +11,10 @@ import yaml
 from safetensors import safe_open
 from safetensors.torch import save_file
 
+from common.tts_model import Model, ModelHolder
 from style_bert_vits2.constants import DEFAULT_STYLE, GRADIO_THEME
 from style_bert_vits2.logging import logger
-from common.tts_model import Model, ModelHolder
+
 
 voice_keys = ["dec"]
 voice_pitch_keys = ["flow"]

webui_style_vectors.py

@@ -12,9 +12,10 @@ from sklearn.cluster import DBSCAN, AgglomerativeClustering, KMeans
 from sklearn.manifold import TSNE
 from umap import UMAP
 
+from config import config
 from style_bert_vits2.constants import DEFAULT_STYLE, GRADIO_THEME
 from style_bert_vits2.logging import logger
-from config import config
+
 
 # Get path settings
 with open(os.path.join("configs", "paths.yml"), "r", encoding="utf-8") as f:

webui_train.py

@@ -19,6 +19,7 @@ from style_bert_vits2.logging import logger
 from style_bert_vits2.utils.stdout_wrapper import SAFE_STDOUT
 from style_bert_vits2.utils.subprocess import run_script_with_log, second_elem_of
 
+
 logger_handler = None
 tensorboard_executed = False