Merge pull request #81 from litagin02/dev

2.3.1

.dockerignore

@@ -1,43 +1,25 @@
-# Dockerfile.deploy用
-
-*.pyc
-*.pyo
-*.pyd
-__pycache__
-*.pyc
-
-venv/
-.vscode/
-
-.ipynb_checkpoints/
-*.ipynb
-
-.git/
-.gitignore
-
-Dockerfile*
-.dockerignore
-*.md
-*.bat
-LICENSE
-
-*.wav
-*.zip
-*.csv
-
-# 中国語と英語が必要な場合はコメントアウト
-/bert/chinese-roberta-wwm-ext-large/
-/bert/deberta-v3-large/
-
-Data/
-dict_data/user_dic.json
-dict_data/user_dic.dic
-docs/
-inputs/
-mos_results/
-pretrained/
-pretrained_jp_extra/
-scripts/
-slm/
-static/
-tools/
+# Dockerfile.deploy用の.dockerignore
+# 日本語のJP-Extraのエディター稼働のみに必要なファイルを指定する
+
+*
+
+!/bert/deberta-v2-large-japanese-char-wwm/
+!/common/
+!/configs/
+!/dict_data/default.csv
+!/model_assets/
+!/monotonic_align/
+!/text/
+
+!/attentions.py
+!/commons.py
+!/config.py
+!/default_config.yml
+!/infer.py
+!/models.py
+!/models_jp_extra.py
+!/modules.py
+!/requirements.txt
+!/server_editor.py
+!/transforms.py
+!/utils.py

app.py

@@ -96,6 +96,8 @@ def tts_fn(
 
     start_time = datetime.datetime.now()
 
+    assert model_holder.current_model is not None
+
     try:
         sr, audio = model_holder.current_model.infer(
             text=text,

colab.ipynb

@@ -4,7 +4,7 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "# Style-Bert-VITS2 (ver 2.3) のGoogle Colabでの学習\n",
+        "# Style-Bert-VITS2 (ver 2.3.1) のGoogle Colabでの学習\n",
         "\n",
         "Google Colab上でStyle-Bert-VITS2の学習を行うことができます。\n",
         "\n",
@@ -118,8 +118,8 @@
         "# こういうふうに書き起こして欲しいという例文（句読点の入れ方・笑い方や固有名詞等）\n",
         "initial_prompt = \"こんにちは。元気、ですかー？ふふっ、私は……ちゃんと元気だよ！\"\n",
         "\n",
-        "!python slice.py -i {input_dir} -o {dataset_root}/{model_name}/raw\n",
-        "!python transcribe.py -i {dataset_root}/{model_name}/raw -o {dataset_root}/{model_name}/esd.list --speaker_name {model_name} --compute_type float16 --initial_prompt {initial_prompt}"
+        "!python slice.py -i {input_dir} --model_name {model_name}\n",
+        "!python transcribe.py --model_name {model_name} --compute_type float16 --initial_prompt {initial_prompt}"
       ]
     },
     {
@@ -229,7 +229,11 @@
         "normalize = False\n",
         "\n",
         "# 音声ファイルの開始・終了にある無音区間を削除するかどうか\n",
-        "trim = False"
+        "trim = False\n",
+        "\n",
+        "# 読みのエラーが出た場合にどうするか。\n",
+        "# \"raise\"ならテキスト前処理が終わったら中断、\"skip\"なら読めない行は学習に使わない、\"use\"なら無理やり使う\n",
+        "yomi_error = \"skip\""
       ]
     },
     {
@@ -269,6 +273,7 @@
         "    use_jp_extra=use_jp_extra,\n",
         "    val_per_lang=0,\n",
         "    log_interval=200,\n",
+        "    yomi_error=yomi_error\n",
         ")"
       ]
     },

common/constants.py

@@ -4,7 +4,7 @@
 # See https://huggingface.co/spaces/gradio/theme-gallery for more themes
 GRADIO_THEME: str = "NoCrypt/miku"
 
-LATEST_VERSION: str = "2.3"
+LATEST_VERSION: str = "2.3.1"
 
 USER_DICT_DIR = "dict_data"
 

common/tts_model.py

@@ -136,7 +136,6 @@ def infer(
         given_tone: Optional[list[int]] = None,
         pitch_scale: float = 1.0,
         intonation_scale: float = 1.0,
-        ignore_unknown: bool = False,
     ) -> tuple[int, np.ndarray]:
         logger.info(f"Start generating audio data from text:\n{text}")
         if language != "JP" and self.hps.version.endswith("JP-Extra"):
@@ -174,7 +173,6 @@ def infer(
                     assist_text_weight=assist_text_weight,
                     style_vec=style_vector,
                     given_tone=given_tone,
-                    ignore_unknown=ignore_unknown,
                 )
         else:
             texts = text.split("\n")
@@ -197,7 +195,6 @@ def infer(
                             assist_text=assist_text,
                             assist_text_weight=assist_text_weight,
                             style_vec=style_vector,
-                            ignore_unknown=ignore_unknown,
                         )
                     )
                     if i != len(texts) - 1:

configs/config.json

@@ -68,5 +68,5 @@
     "use_spectral_norm": false,
     "gin_channels": 256
   },
-  "version": "2.3"
+  "version": "2.3.1"
 }

configs/configs_jp_extra.json

@@ -75,5 +75,5 @@
       "initial_channel": 64
     }
   },
-  "version": "2.3-JP-Extra"
+  "version": "2.3.1-JP-Extra"
 }

docs/CHANGELOG.md

@@ -1,5 +1,20 @@
 # Changelog
 
+## v2.3.1 (2024-02-27)
+
+### バグ修正
+- colabの学習用ノートブックが動かなかったのを修正
+- `App.bat`や`server_fastapi.py`では読めない文字でまだエラーが発生するようになっていたので、推論時は必ず読めない文字を無視して強引に読むように挙動を変更
+
+### 改善
+- 読みが取得できない場合に、テキスト前処理完了時にエラーで中断する今までの挙動に加えて、「読み取得失敗ファイルを学習に使わずに進める」もしくは「読めない文字を無視して読んでファイルを学習に使い進める」というオプションを追加。
+- マージ方法に線形補間の他に球面線形補完を追加 （[@frodo821](https://github.com/frodo821) さんによるPRです、ありがとうございます！）
+- デプロイ用`.dockerignore`を更新
+
+### アップデート手順
+- 2.3未満からのアップデートの場合は、[Update-to-Dict-Editor.bat](https://github.com/litagin02/Style-Bert-VITS2/releases/download/2.3/Update-to-Dict-Editor.bat)をダウンロードし、`Style-Bert-VITS2`フォルダがある場所（インストールbatファイルとかがあったところ）においてダブルクリックしてください。
+- 2.3からのアップデートの場合は、単純に今までの`Update-Style-Bert-VITS2.bat`でアップデートできます。
+
 ## v2.3 (2024-02-26)
 
 ### 大きな変更

docs/CLI.md

@@ -55,7 +55,7 @@ Optional
 ## 2. Preprocess
 
 ```bash
-python preprocess_all.py -m <model_name> [--use_jp_extra] [-b <batch_size>] [-e <epochs>] [-s <save_every_steps>] [--num_processes <num_processes>] [--normalize] [--trim] [--val_per_lang <val_per_lang>] [--log_interval <log_interval>] [--freeze_EN_bert] [--freeze_JP_bert] [--freeze_ZH_bert] [--freeze_style] [--freeze_decoder]
+python preprocess_all.py -m <model_name> [--use_jp_extra] [-b <batch_size>] [-e <epochs>] [-s <save_every_steps>] [--num_processes <num_processes>] [--normalize] [--trim] [--val_per_lang <val_per_lang>] [--log_interval <log_interval>] [--freeze_EN_bert] [--freeze_JP_bert] [--freeze_ZH_bert] [--freeze_style] [--freeze_decoder] [--yomi_error <yomi_error>]
 ```
 
 Required:
@@ -76,6 +76,7 @@ Optional:
 - `--use_jp_extra`: Use JP-Extra model.
 - `--val_per_lang`: Validation data per language (default: 0).
 - `--log_interval`: Log interval (default: 200).
+- `--yomi_error`: How to handle yomi errors (default: `raise`: raise an error after preprocessing all texts, `skip`: skip the texts with errors, `use`: use the texts with errors by ignoring unknown characters).
 
 ## 3. Train
 

infer.py

@@ -52,11 +52,11 @@ def get_text(
     assist_text=None,
     assist_text_weight=0.7,
     given_tone=None,
-    ignore_unknown=False,
 ):
     use_jp_extra = hps.version.endswith("JP-Extra")
+    # 推論のときにのみ呼び出されるので、raise_yomi_errorはFalseに設定
     norm_text, phone, tone, word2ph = clean_text(
-        text, language_str, use_jp_extra, ignore_unknown=ignore_unknown
+        text, language_str, use_jp_extra, raise_yomi_error=False
     )
     if given_tone is not None:
         if len(given_tone) != len(phone):
@@ -80,7 +80,6 @@ def get_text(
         device,
         assist_text,
         assist_text_weight,
-        ignore_unknown,
     )
     del word2ph
     assert bert_ori.shape[-1] == len(phone), phone
@@ -127,7 +126,6 @@ def infer(
     assist_text=None,
     assist_text_weight=0.7,
     given_tone=None,
-    ignore_unknown=False,
 ):
     is_jp_extra = hps.version.endswith("JP-Extra")
     bert, ja_bert, en_bert, phones, tones, lang_ids = get_text(
@@ -138,7 +136,6 @@ def infer(
         assist_text=assist_text,
         assist_text_weight=assist_text_weight,
         given_tone=given_tone,
-        ignore_unknown=ignore_unknown,
     )
     if skip_start:
         phones = phones[3:]

preprocess_all.py

@@ -74,6 +74,9 @@
         help="Log interval",
         default=200,
     )
+    parser.add_argument(
+        "--yomi_error", type=str, help="Yomi error. raise, skip, use", default="raise"
+    )
 
     args = parser.parse_args()
 
@@ -93,4 +96,5 @@
         use_jp_extra=args.use_jp_extra,
         val_per_lang=args.val_per_lang,
         log_interval=args.log_interval,
+        yomi_error=args.yomi_error,
     )

preprocess_text.py

@@ -40,6 +40,7 @@ def count_lines(file_path: str):
 @click.option("--clean/--no-clean", default=preprocess_text_config.clean)
 @click.option("-y", "--yml_config")
 @click.option("--use_jp_extra", is_flag=True)
+@click.option("--yomi_error", default="raise")
 def preprocess(
     transcription_path: str,
     cleaned_path: Optional[str],
@@ -51,11 +52,15 @@ def preprocess(
     clean: bool,
     yml_config: str,  # 这个不要删
     use_jp_extra: bool,
+    yomi_error: str,
 ):
+    assert yomi_error in ["raise", "skip", "use"]
     if cleaned_path == "" or cleaned_path is None:
         cleaned_path = transcription_path + ".cleaned"
 
     error_log_path = os.path.join(os.path.dirname(cleaned_path), "text_error.log")
+    if os.path.exists(error_log_path):
+        os.remove(error_log_path)
     error_count = 0
 
     if clean:
@@ -66,8 +71,12 @@ def preprocess(
                     try:
                         utt, spk, language, text = line.strip().split("|")
                         norm_text, phones, tones, word2ph = clean_text(
-                            text, language, use_jp_extra
+                            text=text,
+                            language=language,
+                            use_jp_extra=use_jp_extra,
+                            raise_yomi_error=(yomi_error != "use"),
                         )
+
                         out_file.write(
                             "{}|{}|{}|{}|{}|{}|{}\n".format(
                                 utt,
@@ -151,12 +160,20 @@ def preprocess(
     with open(config_path, "w", encoding="utf-8") as f:
         json.dump(json_config, f, indent=2, ensure_ascii=False)
     if error_count > 0:
-        logger.error(
-            f"An error occurred in {error_count} lines. Please check {error_log_path} for details. You can proceed with lines that do not have errors."
-        )
-        raise Exception(
-            f"An error occurred in {error_count} lines. Please check {error_log_path} for details. You can proceed with lines that do not have errors."
-        )
+        if yomi_error == "skip":
+            logger.warning(
+                f"An error occurred in {error_count} lines. Proceed with lines without errors. Please check {error_log_path} for details."
+            )
+        else:
+            # yom_error == "raise"と"use"の場合。
+            # "use"の場合は、そもそもyomi_error = Falseで処理しているので、
+            # ここが実行されるのは他の例外のときなので、エラーをraiseする。
+            logger.error(
+                f"An error occurred in {error_count} lines. Please check {error_log_path} for details."
+            )
+            raise Exception(
+                f"An error occurred in {error_count} lines. Please check {error_log_path} for details."
+            )
     else:
         logger.info(
             "Training set and validation set generation from texts is complete!"

server_editor.py

@@ -214,7 +214,7 @@ async def read_item(item: TextRequest):
     try:
         # 最初に正規化しないと整合性がとれない
         text = text_normalize(item.text)
-        kata_tone_list = g2kata_tone(text, ignore_unknown=True)
+        kata_tone_list = g2kata_tone(text)
     except Exception as e:
         raise HTTPException(
             status_code=400,
@@ -289,7 +289,6 @@ def synthesis(request: SynthesisRequest):
         assist_text_weight=request.assistTextWeight,
         use_assist_text=bool(request.assistText),
         line_split=False,
-        ignore_unknown=True,
         pitch_scale=request.pitchScale,
         intonation_scale=request.intonationScale,
     )
@@ -348,7 +347,6 @@ def multi_synthesis(request: MultiSynthesisRequest):
             assist_text_weight=req.assistTextWeight,
             use_assist_text=bool(req.assistText),
             line_split=False,
-            ignore_unknown=True,
             pitch_scale=req.pitchScale,
             intonation_scale=req.intonationScale,
         )

text/__init__.py

@@ -18,28 +18,14 @@ def cleaned_text_to_sequence(cleaned_text, tones, language):
     return phones, tones, lang_ids
 
 
-def get_bert(
-    text,
-    word2ph,
-    language,
-    device,
-    assist_text=None,
-    assist_text_weight=0.7,
-    ignore_unknown=False,
-):
+def get_bert(text, word2ph, language, device, assist_text=None, assist_text_weight=0.7):
     if language == "ZH":
-        from .chinese_bert import get_bert_feature as zh_bert
-
-        return zh_bert(text, word2ph, device, assist_text, assist_text_weight)
+        from .chinese_bert import get_bert_feature
     elif language == "EN":
-        from .english_bert_mock import get_bert_feature as en_bert
-
-        return en_bert(text, word2ph, device, assist_text, assist_text_weight)
+        from .english_bert_mock import get_bert_feature
     elif language == "JP":
-        from .japanese_bert import get_bert_feature as jp_bert
-
-        return jp_bert(
-            text, word2ph, device, assist_text, assist_text_weight, ignore_unknown
-        )
+        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/cleaner.py

@@ -1,4 +1,4 @@
-def clean_text(text, language, use_jp_extra=True, ignore_unknown=False):
+def clean_text(text, language, use_jp_extra=True, raise_yomi_error=False):
     # Changed to import inside if condition to avoid unnecessary import
     if language == "ZH":
         from . import chinese as language_module
@@ -15,7 +15,7 @@ def clean_text(text, language, use_jp_extra=True, ignore_unknown=False):
 
         norm_text = language_module.text_normalize(text)
         phones, tones, word2ph = language_module.g2p(
-            norm_text, use_jp_extra, ignore_unknown=ignore_unknown
+            norm_text, use_jp_extra, raise_yomi_error=raise_yomi_error
         )
     else:
         raise ValueError(f"Language {language} not supported")