diff --git a/GPT_SoVITS/prepare_datasets/1-get-text.py b/GPT_SoVITS/prepare_datasets/1-get-text.py index e01a63b9..7be2508c 100644 --- a/GPT_SoVITS/prepare_datasets/1-get-text.py +++ b/GPT_SoVITS/prepare_datasets/1-get-text.py @@ -10,13 +10,13 @@ all_parts = os.environ.get("all_parts") os.environ["CUDA_VISIBLE_DEVICES"] = os.environ.get("_CUDA_VISIBLE_DEVICES") opt_dir = os.environ.get("opt_dir") bert_pretrained_dir = os.environ.get("bert_pretrained_dir") -is_half = eval(os.environ.get("is_half", "True")) +import torch +is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() import sys, numpy as np, traceback, pdb import os.path from glob import glob from tqdm import tqdm from text.cleaner import clean_text -import torch from transformers import AutoModelForMaskedLM, AutoTokenizer import numpy as np diff --git a/GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py b/GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py index 17394ee4..00ebab56 100644 --- a/GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py +++ b/GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py @@ -10,11 +10,12 @@ os.environ["CUDA_VISIBLE_DEVICES"]= os.environ.get("_CUDA_VISIBLE_DEVICES") from feature_extractor import cnhubert opt_dir= os.environ.get("opt_dir") cnhubert.cnhubert_base_path= os.environ.get("cnhubert_base_dir") -is_half=eval(os.environ.get("is_half","True")) +import torch +is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() import pdb,traceback,numpy as np,logging from scipy.io import wavfile -import librosa,torch +import librosa now_dir = os.getcwd() sys.path.append(now_dir) from tools.my_utils import load_audio diff --git a/GPT_SoVITS/prepare_datasets/3-get-semantic.py b/GPT_SoVITS/prepare_datasets/3-get-semantic.py index 3448a580..64e640e7 100644 --- a/GPT_SoVITS/prepare_datasets/3-get-semantic.py +++ b/GPT_SoVITS/prepare_datasets/3-get-semantic.py @@ -8,7 +8,8 @@ os.environ["CUDA_VISIBLE_DEVICES"] = os.environ.get("_CUDA_VISIBLE_DEVICES") opt_dir = os.environ.get("opt_dir") pretrained_s2G = os.environ.get("pretrained_s2G") s2config_path = os.environ.get("s2config_path") -is_half = eval(os.environ.get("is_half", "True")) +import torch +is_half = eval(os.environ.get("is_half", "True")) and torch.cuda.is_available() import math, traceback import multiprocessing import sys, pdb @@ -19,7 +20,7 @@ from random import shuffle import torch.multiprocessing as mp from glob import glob from tqdm import tqdm -import logging, librosa, utils, torch +import logging, librosa, utils from module.models import SynthesizerTrn logging.getLogger("numba").setLevel(logging.WARNING) diff --git a/README.md b/README.md index c6b774e9..9b3d583a 100644 --- a/README.md +++ b/README.md @@ -106,7 +106,7 @@ conda install -c conda-forge 'ffmpeg<7' Download and place [ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) and [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) in the GPT-SoVITS root. -##### Mac Users +##### MacOS Users ```bash brew install ffmpeg ``` @@ -156,7 +156,7 @@ For English or Japanese ASR (additionally), download models from [Faster Whisper Users in the China region can download this model by entering the links below -- [Faster Whisper Large V3](https://www.icloud.com/iclouddrive/0c4pQxFs7oWyVU1iMTq2DbmLA#faster-whisper-large-v3) (Click "Download a copy", log out if you encounter errors while downloading.) +- [Faster Whisper Large V3](https://www.icloud.com/iclouddrive/00bUEp9_mcjMq_dhHu_vrAFDQ#faster-whisper-large-v3) (Click "Download a copy", log out if you encounter errors while downloading.) - [Faster Whisper Large V3](https://hf-mirror.com/Systran/faster-whisper-large-v3) (HuggingFace mirror site) @@ -227,7 +227,7 @@ ASR processing is performed through Faster_Whisper(ASR marking except Chinese) (No progress bars, GPU performance may cause time delays) ``` -python ./tools/asr/fasterwhisper_asr.py -i -o -l +python ./tools/asr/fasterwhisper_asr.py -i -o -l -p ``` A custom list save path is enabled diff --git a/docs/cn/README.md b/docs/cn/README.md index e46dce3a..0074cdce 100644 --- a/docs/cn/README.md +++ b/docs/cn/README.md @@ -106,7 +106,7 @@ conda install -c conda-forge 'ffmpeg<7' 下载并将 [ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) 和 [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) 放置在 GPT-SoVITS 根目录下。 -##### Mac 用户 +##### MacOS 用户 ```bash brew install ffmpeg ``` @@ -155,7 +155,7 @@ docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-Docker 对于英语与日语自动语音识别（附加）,从 [Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) 下载模型，并将它们放置在 `tools/asr/models` 中。 此外，[其他模型](https://huggingface.co/Systran)可能具有类似效果，但占用更小的磁盘空间。 中国地区用户可以通过以下链接下载： -- [Faster Whisper Large V3](https://www.icloud.com/iclouddrive/0c4pQxFs7oWyVU1iMTq2DbmLA#faster-whisper-large-v3)（点击“下载副本”，如果下载时遇到错误，请退出登录） +- [Faster Whisper Large V3](https://www.icloud.com/iclouddrive/00bUEp9_mcjMq_dhHu_vrAFDQ#faster-whisper-large-v3)(点击“下载副本”，如果下载时遇到错误，请退出登录） - [Faster Whisper Large V3](https://hf-mirror.com/Systran/faster-whisper-large-v3)(Hugging Face镜像站) @@ -185,7 +185,7 @@ D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin. - [ ] **高优先级：** - [x] 日语和英语的本地化。 - - [ ] 用户指南。 + - [x] 用户指南。 - [x] 日语和英语数据集微调训练。 - [ ] **功能:** @@ -226,9 +226,9 @@ python tools/asr/funasr_asr.py -i -o 通过Faster_Whisper进行ASR处理（除中文之外的ASR标记） （没有进度条，GPU性能可能会导致时间延迟） -```` -python ./tools/asr/fasterwhisper_asr.py -i -o -l -```` +``` +python ./tools/asr/fasterwhisper_asr.py -i -o -l -p +``` 启用自定义列表保存路径 ## 致谢 diff --git a/docs/ja/README.md b/docs/ja/README.md index 92ff5614..e3a9f003 100644 --- a/docs/ja/README.md +++ b/docs/ja/README.md @@ -102,7 +102,7 @@ conda install -c conda-forge 'ffmpeg<7' [ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) と [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) をダウンロードし、GPT-SoVITS のルートディレクトリに置きます。 -##### Mac ユーザー +##### MacOS ユーザー ```bash brew install ffmpeg ``` @@ -209,7 +209,7 @@ ASR処理はFaster_Whisperを通じて実行されます(中国語を除くASR (進行状況バーは表示されません。GPU のパフォーマンスにより時間遅延が発生する可能性があります) ``` -python ./tools/asr/fasterwhisper_asr.py -i -o -l +python ./tools/asr/fasterwhisper_asr.py -i -o -l -p ``` カスタムリストの保存パスが有効になっています diff --git a/docs/ko/README.md b/docs/ko/README.md index 80b68484..4deb2c4e 100644 --- a/docs/ko/README.md +++ b/docs/ko/README.md @@ -102,7 +102,7 @@ conda install -c conda-forge 'ffmpeg<7' [ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe)와 [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe)를 GPT-SoVITS root 디렉토리에 넣습니다. -##### Mac 사용자 +##### MacOS 사용자 ```bash brew install ffmpeg ``` @@ -213,7 +213,7 @@ ASR 처리는 Faster_Whisper(중국어를 제외한 ASR 마킹)를 통해 수행 (진행률 표시줄 없음, GPU 성능으로 인해 시간 지연이 발생할 수 있음) ``` -python ./tools/asr/fasterwhisper_asr.py -i -o -l +python ./tools/asr/fasterwhisper_asr.py -i -o -l -p ``` 사용자 정의 목록 저장 경로가 활성화되었습니다. diff --git a/docs/tr/README.md b/docs/tr/README.md index d0936f13..5b9a103f 100644 --- a/docs/tr/README.md +++ b/docs/tr/README.md @@ -102,7 +102,7 @@ conda install -c conda-forge 'ffmpeg<7' [ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) ve [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) dosyalarını indirin ve GPT-SoVITS kök dizinine yerleştirin. -##### Mac Kullanıcıları +##### MacOS Kullanıcıları ```bash brew install ffmpeg ``` diff --git a/tools/asr/config.py b/tools/asr/config.py index 8fe68388..d7e184c1 100644 --- a/tools/asr/config.py +++ b/tools/asr/config.py @@ -21,11 +21,13 @@ asr_dict = { 'lang': ['zh'], 'size': ['large'], 'path': 'funasr_asr.py', + 'precision': ['float32'] }, "Faster Whisper (多语种)": { 'lang': ['auto', 'zh', 'en', 'ja'], 'size': check_fw_local_models(), - 'path': 'fasterwhisper_asr.py' - } + 'path': 'fasterwhisper_asr.py', + 'precision': ['float32', 'float16', 'int8'] + }, } diff --git a/tools/asr/fasterwhisper_asr.py b/tools/asr/fasterwhisper_asr.py index e9fc6a47..da8eadfb 100644 --- a/tools/asr/fasterwhisper_asr.py +++ b/tools/asr/fasterwhisper_asr.py @@ -101,8 +101,8 @@ if __name__ == '__main__': parser.add_argument("-l", "--language", type=str, default='ja', choices=language_code_list, help="Language of the audio files.") - parser.add_argument("-p", "--precision", type=str, default='float16', choices=['float16','float32'], - help="fp16 or fp32") + parser.add_argument("-p", "--precision", type=str, default='float16', choices=['float16','float32','int8'], + help="fp16, int8 or fp32") cmd = parser.parse_args() output_file_path = execute_asr( diff --git a/tools/asr/funasr_asr.py b/tools/asr/funasr_asr.py index 831da6c1..ec786782 100644 --- a/tools/asr/funasr_asr.py +++ b/tools/asr/funasr_asr.py @@ -4,7 +4,8 @@ import argparse import os import traceback from tqdm import tqdm - +# from funasr.utils import version_checker +# version_checker.check_for_update = lambda: None from funasr import AutoModel path_asr = 'tools/asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch' @@ -14,6 +15,7 @@ path_asr = path_asr if os.path.exists(path_asr) else "iic/speech_paraformer-l path_vad = path_vad if os.path.exists(path_vad) else "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch" path_punc = path_punc if os.path.exists(path_punc) else "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch" + model = AutoModel( model = path_asr, model_revision = "v2.0.4", diff --git a/tools/i18n/locale/en_US.json b/tools/i18n/locale/en_US.json index 6936fb91..3d06e2d2 100644 --- a/tools/i18n/locale/en_US.json +++ b/tools/i18n/locale/en_US.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Reverberation and delay removal model(by FoxJoy):", "ASR 模型": "ASR model", "ASR 模型尺寸": "ASR model size", + "数据类型精度": "Computing precision", "ASR 语言设置": "ASR language", "ASR进程输出信息": "ASR output log", "GPT模型列表": "GPT weight list", diff --git a/tools/i18n/locale/es_ES.json b/tools/i18n/locale/es_ES.json index dcf17f8e..9d4e44ba 100644 --- a/tools/i18n/locale/es_ES.json +++ b/tools/i18n/locale/es_ES.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Modelos de eliminación de reverberación y retardo (por FoxJoy)", "ASR 模型": "Modelo ASR", "ASR 模型尺寸": "Tamaño del modelo ASR", + "数据类型精度": "precisión del tipo de datos", "ASR 语言设置": "Configuración del idioma ASR", "ASR进程输出信息": "Información de salida del proceso ASR", "GPT模型列表": "Lista de modelos GPT", diff --git a/tools/i18n/locale/fr_FR.json b/tools/i18n/locale/fr_FR.json index 194ff0ea..516593a9 100644 --- a/tools/i18n/locale/fr_FR.json +++ b/tools/i18n/locale/fr_FR.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Modèle de suppression de réverbération et de retard (par FoxJoy) :", "ASR 模型": "Modèle ASR", "ASR 模型尺寸": "Taille du modèle ASR", + "数据类型精度": "précision du type de données", "ASR 语言设置": "Paramètres de langue ASR", "ASR进程输出信息": "Informations de processus ASR", "GPT模型列表": "Liste des modèles GPT", diff --git a/tools/i18n/locale/it_IT.json b/tools/i18n/locale/it_IT.json index 71edbe95..a1a7f88d 100644 --- a/tools/i18n/locale/it_IT.json +++ b/tools/i18n/locale/it_IT.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Modello per rimuovere la riverberazione e il ritardo (by FoxJoy):", "ASR 模型": "Modello ASR", "ASR 模型尺寸": "Dimensioni del modello ASR", + "数据类型精度": "precisione del tipo di dati", "ASR 语言设置": "Impostazioni linguistiche ASR", "ASR进程输出信息": "Informazioni sull'output del processo ASR", "GPT模型列表": "Elenco dei modelli GPT", diff --git a/tools/i18n/locale/ja_JP.json b/tools/i18n/locale/ja_JP.json index b8ddb1a8..0f467ffd 100644 --- a/tools/i18n/locale/ja_JP.json +++ b/tools/i18n/locale/ja_JP.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3、リバーブ除去と遅延除去モデル（by FoxJoy）：", "ASR 模型": "ASR モデル", "ASR 模型尺寸": "ASRモデルサイズ", + "数据类型精度": "データ型の精度", "ASR 语言设置": "ASR 言語設定", "ASR进程输出信息": "ASRプロセスの出力情報", "GPT模型列表": "GPTモデルリスト", diff --git a/tools/i18n/locale/ko_KR.json b/tools/i18n/locale/ko_KR.json index 22ae5ce0..53a9d108 100644 --- a/tools/i18n/locale/ko_KR.json +++ b/tools/i18n/locale/ko_KR.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. 잔향 제거 및 지연 제거 모델 (by FoxJoy):", "ASR 模型": "ASR 모델", "ASR 模型尺寸": "ASR 모델 크기", + "数据类型精度": "데이터 유형 정밀도", "ASR 语言设置": "ASR 언어 설정", "ASR进程输出信息": "ASR 프로세스 출력 정보", "GPT模型列表": "GPT 모델 목록", diff --git a/tools/i18n/locale/pt_BR.json b/tools/i18n/locale/pt_BR.json index 0454364f..95e3e332 100644 --- a/tools/i18n/locale/pt_BR.json +++ b/tools/i18n/locale/pt_BR.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Modelo de remoção de reverberação e atraso (por FoxJoy):", "ASR 模型": "Modelo ASR", "ASR 模型尺寸": "Tamanho do modelo ASR", + "数据类型精度": "precisão do tipo de dado", "ASR 语言设置": "Configurações de idioma do ASR", "ASR进程输出信息": "Informações de saída do processo ASR", "GPT模型列表": "Lista de modelos GPT", diff --git a/tools/i18n/locale/ru_RU.json b/tools/i18n/locale/ru_RU.json index a5a7df2d..42ca5913 100644 --- a/tools/i18n/locale/ru_RU.json +++ b/tools/i18n/locale/ru_RU.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Модель удаления реверберации и задержек (от FoxJoy):", "ASR 模型": "Модель ASR", "ASR 模型尺寸": "Размер модели ASR", + "数据类型精度": "точность типа данных", "ASR 语言设置": "Настройки языка ASR", "ASR进程输出信息": "Информация о процессе ASR", "GPT模型列表": "Список моделей GPT", diff --git a/tools/i18n/locale/tr_TR.json b/tools/i18n/locale/tr_TR.json index 8469237a..179d8548 100644 --- a/tools/i18n/locale/tr_TR.json +++ b/tools/i18n/locale/tr_TR.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3. Yankı ve gecikme giderme modeli (FoxJoy tarafından):", "ASR 模型": "ASR modeli", "ASR 模型尺寸": "ASR model boyutu", + "数据类型精度": "veri türü doğruluğu", "ASR 语言设置": "ASR dil ayarları", "ASR进程输出信息": "ASR işlemi çıktı bilgisi", "GPT模型列表": "GPT model listesi", diff --git a/tools/i18n/locale/zh_CN.json b/tools/i18n/locale/zh_CN.json index 84cbca25..ddc2eda3 100644 --- a/tools/i18n/locale/zh_CN.json +++ b/tools/i18n/locale/zh_CN.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3、去混响、去延迟模型（by FoxJoy）：", "ASR 模型": "ASR 模型", "ASR 模型尺寸": "ASR 模型尺寸", + "数据类型精度": "数据类型精度", "ASR 语言设置": "ASR 语言设置", "ASR进程输出信息": "ASR进程输出信息", "GPT模型列表": "GPT模型列表", diff --git a/tools/i18n/locale/zh_HK.json b/tools/i18n/locale/zh_HK.json index c2297973..253243ec 100644 --- a/tools/i18n/locale/zh_HK.json +++ b/tools/i18n/locale/zh_HK.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3、去混響、去延遲模型（by FoxJoy）：", "ASR 模型": "ASR 模型", "ASR 模型尺寸": "ASR 模型尺寸", + "数据类型精度": "數據類型精度", "ASR 语言设置": "ASR 語言設置", "ASR进程输出信息": "ASR進程輸出信息", "GPT模型列表": "GPT模型列表", diff --git a/tools/i18n/locale/zh_SG.json b/tools/i18n/locale/zh_SG.json index 81c7ab37..5c90a364 100644 --- a/tools/i18n/locale/zh_SG.json +++ b/tools/i18n/locale/zh_SG.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3、去混響、去延遲模型（by FoxJoy）：", "ASR 模型": "ASR 模型", "ASR 模型尺寸": "ASR 模型尺寸", + "数据类型精度": "數據類型精度", "ASR 语言设置": "ASR 語言設定", "ASR进程输出信息": "ASR進程輸出資訊", "GPT模型列表": "GPT模型列表", diff --git a/tools/i18n/locale/zh_TW.json b/tools/i18n/locale/zh_TW.json index 014c16bd..d9191e50 100644 --- a/tools/i18n/locale/zh_TW.json +++ b/tools/i18n/locale/zh_TW.json @@ -34,6 +34,7 @@ "3、去混响、去延迟模型（by FoxJoy）：": "3、去混響、去延遲模型（by FoxJoy）：", "ASR 模型": "ASR 模型", "ASR 模型尺寸": "ASR 模型尺寸", + "数据类型精度": "數據類型精度", "ASR 语言设置": "ASR 語言設置", "ASR进程输出信息": "ASR進程輸出資訊", "GPT模型列表": "GPT模型列表", diff --git a/tools/uvr5/bs_roformer/bs_roformer.py b/tools/uvr5/bs_roformer/bs_roformer.py index a159211d..88af3caa 100644 --- a/tools/uvr5/bs_roformer/bs_roformer.py +++ b/tools/uvr5/bs_roformer/bs_roformer.py @@ -479,27 +479,31 @@ class BSRoformer(Module): 'b s f t c -> b (f s) t c') # merge stereo / mono into the frequency, with frequency leading dimension, for band splitting x = rearrange(stft_repr, 'b f t c -> b t (f c)') - + # print("460:", x.dtype)#fp32 x = self.band_split(x) # axial / hierarchical attention + # print("487:",x.dtype)#fp16 for transformer_block in self.layers: if len(transformer_block) == 3: linear_transformer, time_transformer, freq_transformer = transformer_block x, ft_ps = pack([x], 'b * d') + # print("494:", x.dtype)#fp16 x = linear_transformer(x) + # print("496:", x.dtype)#fp16 x, = unpack(x, ft_ps, 'b * d') else: time_transformer, freq_transformer = transformer_block + # print("501:", x.dtype)#fp16 x = rearrange(x, 'b t f d -> b f t d') x, ps = pack([x], '* t d') x = time_transformer(x) - + # print("505:", x.dtype)#fp16 x, = unpack(x, ps, '* t d') x = rearrange(x, 'b f t d -> b t f d') x, ps = pack([x], '* f d') @@ -508,10 +512,11 @@ class BSRoformer(Module): x, = unpack(x, ps, '* f d') + # print("515:", x.dtype)######fp16 x = self.final_norm(x) num_stems = len(self.mask_estimators) - + # print("519:", x.dtype)#fp32 mask = torch.stack([fn(x) for fn in self.mask_estimators], dim=1) mask = rearrange(mask, 'b n t (f c) -> b n f t c', c=2) diff --git a/tools/uvr5/bsroformer.py b/tools/uvr5/bsroformer.py index 5c01011d..d1620326 100644 --- a/tools/uvr5/bsroformer.py +++ b/tools/uvr5/bsroformer.py @@ -1,4 +1,5 @@ # This code is modified from https://github.com/ZFTurbo/ +import pdb import librosa from tqdm import tqdm @@ -10,6 +11,7 @@ import torch.nn as nn import warnings warnings.filterwarnings("ignore") +from bs_roformer.bs_roformer import BSRoformer class BsRoformer_Loader: def get_model_from_config(self): @@ -40,7 +42,7 @@ class BsRoformer_Loader: } - from bs_roformer.bs_roformer import BSRoformer + model = BSRoformer( **dict(config) ) @@ -95,6 +97,8 @@ class BsRoformer_Loader: part = nn.functional.pad(input=part, pad=(0, C - length), mode='reflect') else: part = nn.functional.pad(input=part, pad=(0, C - length, 0, 0), mode='constant', value=0) + if(self.is_half==True): + part=part.half() batch_data.append(part) batch_locations.append((i, length)) i += step @@ -102,6 +106,7 @@ class BsRoformer_Loader: if len(batch_data) >= batch_size or (i >= mix.shape[1]): arr = torch.stack(batch_data, dim=0) + # print(23333333,arr.dtype) x = model(arr) window = window_middle @@ -192,14 +197,18 @@ class BsRoformer_Loader: # print("Elapsed time: {:.2f} sec".format(time.time() - start_time)) - def __init__(self, model_path, device): + def __init__(self, model_path, device,is_half): self.device = device self.extract_instrumental=True model = self.get_model_from_config() - state_dict = torch.load(model_path) + state_dict = torch.load(model_path,map_location="cpu") model.load_state_dict(state_dict) - self.model = model.to(device) + self.is_half=is_half + if(is_half==False): + self.model = model.to(device) + else: + self.model = model.half().to(device) def _path_audio_(self, input, others_root, vocal_root, format, is_hp3=False): diff --git a/tools/uvr5/webui.py b/tools/uvr5/webui.py index 712c892e..60dfdaa7 100644 --- a/tools/uvr5/webui.py +++ b/tools/uvr5/webui.py @@ -17,8 +17,8 @@ from bsroformer import BsRoformer_Loader weight_uvr5_root = "tools/uvr5/uvr5_weights" uvr5_names = [] for name in os.listdir(weight_uvr5_root): - if name.endswith(".pth") or "onnx" in name: - uvr5_names.append(name.replace(".pth", "")) + if name.endswith(".pth") or name.endswith(".ckpt") or "onnx" in name: + uvr5_names.append(name.replace(".pth", "").replace(".ckpt", "")) device=sys.argv[1] is_half=eval(sys.argv[2]) @@ -37,8 +37,9 @@ def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format elif model_name == "Bs_Roformer" or "bs_roformer" in model_name.lower(): func = BsRoformer_Loader pre_fun = func( - model_path = os.path.join(weight_uvr5_root, model_name + ".pth"), + model_path = os.path.join(weight_uvr5_root, model_name + ".ckpt"), device = device, + is_half=is_half ) else: func = AudioPre if "DeEcho" not in model_name else AudioPreDeEcho diff --git a/webui.py b/webui.py index 71d12465..82fef695 100644 --- a/webui.py +++ b/webui.py @@ -192,7 +192,7 @@ def change_tts_inference(if_tts,bert_path,cnhubert_base_path,gpu_number,gpt_path yield i18n("TTS推理进程已关闭") from tools.asr.config import asr_dict -def open_asr(asr_inp_dir, asr_opt_dir, asr_model, asr_model_size, asr_lang): +def open_asr(asr_inp_dir, asr_opt_dir, asr_model, asr_model_size, asr_lang, asr_precision): global p_asr if(p_asr==None): asr_inp_dir=my_utils.clean_path(asr_inp_dir) @@ -202,16 +202,18 @@ def open_asr(asr_inp_dir, asr_opt_dir, asr_model, asr_model_size, asr_lang): cmd += f' -o "{asr_opt_dir}"' cmd += f' -s {asr_model_size}' cmd += f' -l {asr_lang}' - cmd += " -p %s"%("float16"if is_half==True else "float32") - - yield "ASR任务开启：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True} + cmd += f" -p {asr_precision}" + output_file_name = os.path.basename(asr_inp_dir) + output_folder = asr_opt_dir or "output/asr_opt" + output_file_path = os.path.abspath(f'{output_folder}/{output_file_name}.list') + yield "ASR任务开启：%s"%cmd, {"__type__":"update","visible":False}, {"__type__":"update","visible":True}, {"__type__":"update"} print(cmd) p_asr = Popen(cmd, shell=True) p_asr.wait() p_asr=None - yield f"ASR任务完成, 查看终端进行下一步",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield f"ASR任务完成, 查看终端进行下一步", {"__type__":"update","visible":True}, {"__type__":"update","visible":False}, {"__type__":"update","value":output_file_path} else: - yield "已有正在进行的ASR任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "已有正在进行的ASR任务，需先终止才能开启下一次任务", {"__type__":"update","visible":False}, {"__type__":"update","visible":True}, {"__type__":"update"} # return None def close_asr(): @@ -219,7 +221,7 @@ def close_asr(): if(p_asr!=None): kill_process(p_asr.pid) p_asr=None - return "已终止ASR进程",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + return "已终止ASR进程", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} def open_denoise(denoise_inp_dir, denoise_opt_dir): global p_denoise if(p_denoise==None): @@ -227,14 +229,14 @@ def open_denoise(denoise_inp_dir, denoise_opt_dir): denoise_opt_dir=my_utils.clean_path(denoise_opt_dir) cmd = '"%s" tools/cmd-denoise.py -i "%s" -o "%s" -p %s'%(python_exec,denoise_inp_dir,denoise_opt_dir,"float16"if is_half==True else "float32") - yield "语音降噪任务开启：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "语音降噪任务开启：%s"%cmd, {"__type__":"update","visible":False}, {"__type__":"update","visible":True}, {"__type__":"update"} print(cmd) p_denoise = Popen(cmd, shell=True) p_denoise.wait() p_denoise=None - yield f"语音降噪任务完成, 查看终端进行下一步",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield f"语音降噪任务完成, 查看终端进行下一步", {"__type__":"update","visible":True}, {"__type__":"update","visible":False}, {"__type__":"update","value":denoise_opt_dir} else: - yield "已有正在进行的语音降噪任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "已有正在进行的语音降噪任务，需先终止才能开启下一次任务", {"__type__":"update","visible":False}, {"__type__":"update","visible":True}, {"__type__":"update"} # return None def close_denoise(): @@ -242,7 +244,7 @@ def close_denoise(): if(p_denoise!=None): kill_process(p_denoise.pid) p_denoise=None - return "已终止语音降噪进程",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + return "已终止语音降噪进程", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} p_train_SoVITS=None def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers1Ba,pretrained_s2G,pretrained_s2D): @@ -272,21 +274,21 @@ def open1Ba(batch_size,total_epoch,exp_name,text_low_lr_rate,if_save_latest,if_s with open(tmp_config_path,"w")as f:f.write(json.dumps(data)) cmd = '"%s" GPT_SoVITS/s2_train.py --config "%s"'%(python_exec,tmp_config_path) - yield "SoVITS训练开始：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "SoVITS训练开始：%s"%cmd, {"__type__":"update","visible":False}, {"__type__":"update","visible":True} print(cmd) p_train_SoVITS = Popen(cmd, shell=True) p_train_SoVITS.wait() p_train_SoVITS=None - yield "SoVITS训练完成",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "SoVITS训练完成", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} else: - yield "已有正在进行的SoVITS训练任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "已有正在进行的SoVITS训练任务，需先终止才能开启下一次任务", {"__type__":"update","visible":False}, {"__type__":"update","visible":True} def close1Ba(): global p_train_SoVITS if(p_train_SoVITS!=None): kill_process(p_train_SoVITS.pid) p_train_SoVITS=None - return "已终止SoVITS训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + return "已终止SoVITS训练", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} p_train_GPT=None def open1Bb(batch_size,total_epoch,exp_name,if_dpo,if_save_latest,if_save_every_weights,save_every_epoch,gpu_numbers,pretrained_s1): @@ -319,21 +321,21 @@ def open1Bb(batch_size,total_epoch,exp_name,if_dpo,if_save_latest,if_save_every_ with open(tmp_config_path, "w") as f:f.write(yaml.dump(data, default_flow_style=False)) # cmd = '"%s" GPT_SoVITS/s1_train.py --config_file "%s" --train_semantic_path "%s/6-name2semantic.tsv" --train_phoneme_path "%s/2-name2text.txt" --output_dir "%s/logs_s1"'%(python_exec,tmp_config_path,s1_dir,s1_dir,s1_dir) cmd = '"%s" GPT_SoVITS/s1_train.py --config_file "%s" '%(python_exec,tmp_config_path) - yield "GPT训练开始：%s"%cmd,{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "GPT训练开始：%s"%cmd, {"__type__":"update","visible":False}, {"__type__":"update","visible":True} print(cmd) p_train_GPT = Popen(cmd, shell=True) p_train_GPT.wait() p_train_GPT=None - yield "GPT训练完成",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "GPT训练完成", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} else: - yield "已有正在进行的GPT训练任务，需先终止才能开启下一次任务",{"__type__":"update","visible":False},{"__type__":"update","visible":True} + yield "已有正在进行的GPT训练任务，需先终止才能开启下一次任务", {"__type__":"update","visible":False}, {"__type__":"update","visible":True} def close1Bb(): global p_train_GPT if(p_train_GPT!=None): kill_process(p_train_GPT.pid) p_train_GPT=None - return "已终止GPT训练",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + return "已终止GPT训练", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} ps_slice=[] def open_slice(inp,opt_root,threshold,min_length,min_interval,hop_size,max_sil_kept,_max,alpha,n_parts): @@ -341,12 +343,12 @@ def open_slice(inp,opt_root,threshold,min_length,min_interval,hop_size,max_sil_k inp = my_utils.clean_path(inp) opt_root = my_utils.clean_path(opt_root) if(os.path.exists(inp)==False): - yield "输入路径不存在",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "输入路径不存在", {"__type__":"update","visible":True}, {"__type__":"update","visible":False}, {"__type__": "update"}, {"__type__": "update"} return if os.path.isfile(inp):n_parts=1 elif os.path.isdir(inp):pass else: - yield "输入路径存在但既不是文件也不是文件夹",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "输入路径存在但既不是文件也不是文件夹", {"__type__":"update","visible":True}, {"__type__":"update","visible":False}, {"__type__": "update"}, {"__type__": "update"} return if (ps_slice == []): for i_part in range(n_parts): @@ -354,13 +356,13 @@ def open_slice(inp,opt_root,threshold,min_length,min_interval,hop_size,max_sil_k print(cmd) p = Popen(cmd, shell=True) ps_slice.append(p) - yield "切割执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True} + yield "切割执行中", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}, {"__type__": "update"}, {"__type__": "update"} for p in ps_slice: p.wait() ps_slice=[] - yield "切割结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "切割结束", {"__type__":"update","visible":True}, {"__type__":"update","visible":False}, {"__type__": "update", "value":opt_root}, {"__type__": "update", "value":opt_root} else: - yield "已有正在进行的切割任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True} + yield "已有正在进行的切割任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True}, {"__type__": "update"}, {"__type__": "update"} def close_slice(): global ps_slice @@ -467,7 +469,7 @@ def open1b(inp_text,inp_wav_dir,exp_name,gpu_numbers,ssl_pretrained_dir): for p in ps1b: p.wait() ps1b=[] - yield "SSL提取进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "SSL提取进程结束", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} else: yield "已有正在进行的SSL提取任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True} @@ -524,7 +526,7 @@ def open1c(inp_text,exp_name,gpu_numbers,pretrained_s2G_path): with open(path_semantic, "w", encoding="utf8") as f: f.write("\n".join(opt) + "\n") ps1c=[] - yield "语义token提取进程结束",{"__type__":"update","visible":True},{"__type__":"update","visible":False} + yield "语义token提取进程结束", {"__type__":"update","visible":True}, {"__type__":"update","visible":False} else: yield "已有正在进行的语义token提取任务，需先终止才能开启下一次任务", {"__type__": "update", "visible": False}, {"__type__": "update", "visible": True} @@ -689,33 +691,37 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app: uvr5_info = gr.Textbox(label=i18n("UVR5进程输出信息")) gr.Markdown(value=i18n("0b-语音切分工具")) with gr.Row(): - with gr.Row(): - slice_inp_path=gr.Textbox(label=i18n("音频自动切分输入路径，可文件可文件夹"),value="") - slice_opt_root=gr.Textbox(label=i18n("切分后的子音频的输出根目录"),value="output/slicer_opt") - threshold=gr.Textbox(label=i18n("threshold:音量小于这个值视作静音的备选切割点"),value="-34") - min_length=gr.Textbox(label=i18n("min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值"),value="4000") - min_interval=gr.Textbox(label=i18n("min_interval:最短切割间隔"),value="300") - hop_size=gr.Textbox(label=i18n("hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）"),value="10") - max_sil_kept=gr.Textbox(label=i18n("max_sil_kept:切完后静音最多留多长"),value="500") - with gr.Row(): - open_slicer_button=gr.Button(i18n("开启语音切割"), variant="primary",visible=True) - close_slicer_button=gr.Button(i18n("终止语音切割"), variant="primary",visible=False) - _max=gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("max:归一化后最大值多少"),value=0.9,interactive=True) - alpha=gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("alpha_mix:混多少比例归一化后音频进来"),value=0.25,interactive=True) - n_process=gr.Slider(minimum=1,maximum=n_cpu,step=1,label=i18n("切割使用的进程数"),value=4,interactive=True) - slicer_info = gr.Textbox(label=i18n("语音切割进程输出信息")) + with gr.Column(scale=3): + with gr.Row(): + slice_inp_path=gr.Textbox(label=i18n("音频自动切分输入路径，可文件可文件夹"),value="") + slice_opt_root=gr.Textbox(label=i18n("切分后的子音频的输出根目录"),value="output/slicer_opt") + with gr.Row(): + threshold=gr.Textbox(label=i18n("threshold:音量小于这个值视作静音的备选切割点"),value="-34") + min_length=gr.Textbox(label=i18n("min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值"),value="4000") + min_interval=gr.Textbox(label=i18n("min_interval:最短切割间隔"),value="300") + hop_size=gr.Textbox(label=i18n("hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）"),value="10") + max_sil_kept=gr.Textbox(label=i18n("max_sil_kept:切完后静音最多留多长"),value="500") + with gr.Row(): + _max=gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("max:归一化后最大值多少"),value=0.9,interactive=True) + alpha=gr.Slider(minimum=0,maximum=1,step=0.05,label=i18n("alpha_mix:混多少比例归一化后音频进来"),value=0.25,interactive=True) + n_process=gr.Slider(minimum=1,maximum=n_cpu,step=1,label=i18n("切割使用的进程数"),value=4,interactive=True) + with gr.Row(): + slicer_info = gr.Textbox(label=i18n("语音切割进程输出信息")) + open_slicer_button=gr.Button(i18n("开启语音切割"), variant="primary",visible=True) + close_slicer_button=gr.Button(i18n("终止语音切割"), variant="primary",visible=False) gr.Markdown(value=i18n("0bb-语音降噪工具")) with gr.Row(): + with gr.Column(scale=3): + with gr.Row(): + denoise_input_dir=gr.Textbox(label=i18n("降噪音频文件输入文件夹"),value="") + denoise_output_dir=gr.Textbox(label=i18n("降噪结果输出文件夹"),value="output/denoise_opt") + with gr.Row(): + denoise_info = gr.Textbox(label=i18n("语音降噪进程输出信息")) open_denoise_button = gr.Button(i18n("开启语音降噪"), variant="primary",visible=True) close_denoise_button = gr.Button(i18n("终止语音降噪进程"), variant="primary",visible=False) - denoise_input_dir=gr.Textbox(label=i18n("降噪音频文件输入文件夹"),value="") - denoise_output_dir=gr.Textbox(label=i18n("降噪结果输出文件夹"),value="output/denoise_opt") - denoise_info = gr.Textbox(label=i18n("语音降噪进程输出信息")) gr.Markdown(value=i18n("0c-中文批量离线ASR工具")) with gr.Row(): - open_asr_button = gr.Button(i18n("开启离线批量ASR"), variant="primary",visible=True) - close_asr_button = gr.Button(i18n("终止ASR进程"), variant="primary",visible=False) - with gr.Column(): + with gr.Column(scale=3): with gr.Row(): asr_inp_dir = gr.Textbox( label=i18n("输入文件夹路径"), @@ -746,17 +752,39 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app: interactive = True, value="zh" ) + asr_precision = gr.Dropdown( + label = i18n("数据类型精度"), + choices = ["float32"], + interactive = True, + value="float32" + ) with gr.Row(): - asr_info = gr.Textbox(label=i18n("ASR进程输出信息")) + asr_info = gr.Textbox(label=i18n("ASR进程输出信息")) + open_asr_button = gr.Button(i18n("开启离线批量ASR"), variant="primary",visible=True) + close_asr_button = gr.Button(i18n("终止ASR进程"), variant="primary",visible=False) def change_lang_choices(key): #根据选择的模型修改可选的语言 # return gr.Dropdown(choices=asr_dict[key]['lang']) return {"__type__": "update", "choices": asr_dict[key]['lang'],"value":asr_dict[key]['lang'][0]} def change_size_choices(key): # 根据选择的模型修改可选的模型尺寸 # return gr.Dropdown(choices=asr_dict[key]['size']) - return {"__type__": "update", "choices": asr_dict[key]['size']} + return {"__type__": "update", "choices": asr_dict[key]['size'],"value":asr_dict[key]['size'][-1]} + def change_precision_choices(key): #根据选择的模型修改可选的语言 + if key =="Faster Whisper (多语种)": + if default_batch_size <= 4: + precision = 'int8' + elif is_half: + precision = 'float16' + else: + precision = 'float32' + else: + precision = 'float32' + # return gr.Dropdown(choices=asr_dict[key]['precision']) + return {"__type__": "update", "choices": asr_dict[key]['precision'],"value":precision} asr_model.change(change_lang_choices, [asr_model], [asr_lang]) asr_model.change(change_size_choices, [asr_model], [asr_size]) + asr_model.change(change_precision_choices, [asr_model], [asr_precision]) + gr.Markdown(value=i18n("0d-语音文本校对标注工具")) with gr.Row(): @@ -769,11 +797,11 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app: label_info = gr.Textbox(label=i18n("打标工具进程输出信息")) if_label.change(change_label, [if_label,path_list], [label_info]) if_uvr5.change(change_uvr5, [if_uvr5], [uvr5_info]) - open_asr_button.click(open_asr, [asr_inp_dir, asr_opt_dir, asr_model, asr_size, asr_lang], [asr_info,open_asr_button,close_asr_button]) + open_asr_button.click(open_asr, [asr_inp_dir, asr_opt_dir, asr_model, asr_size, asr_lang, asr_precision], [asr_info,open_asr_button,close_asr_button,path_list]) close_asr_button.click(close_asr, [], [asr_info,open_asr_button,close_asr_button]) - open_slicer_button.click(open_slice, [slice_inp_path,slice_opt_root,threshold,min_length,min_interval,hop_size,max_sil_kept,_max,alpha,n_process], [slicer_info,open_slicer_button,close_slicer_button]) + open_slicer_button.click(open_slice, [slice_inp_path,slice_opt_root,threshold,min_length,min_interval,hop_size,max_sil_kept,_max,alpha,n_process], [slicer_info,open_slicer_button,close_slicer_button,asr_inp_dir,denoise_input_dir]) close_slicer_button.click(close_slice, [], [slicer_info,open_slicer_button,close_slicer_button]) - open_denoise_button.click(open_denoise, [denoise_input_dir,denoise_output_dir], [denoise_info,open_denoise_button,close_denoise_button]) + open_denoise_button.click(open_denoise, [denoise_input_dir,denoise_output_dir], [denoise_info,open_denoise_button,close_denoise_button,asr_inp_dir]) close_denoise_button.click(close_denoise, [], [denoise_info,open_denoise_button,close_denoise_button]) with gr.TabItem(i18n("1-GPT-SoVITS-TTS")): @@ -876,4 +904,4 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app: share=is_share, server_port=webui_port_main, quiet=True, - ) + ) \ No newline at end of file