import glob import json import os import platform import shutil import signal import subprocess import webbrowser import GPUtil import gradio as gr import psutil import torch import yaml from config import yml_config from tools.log import logger bert_model_paths = [ "./bert/chinese-roberta-wwm-ext-large/pytorch_model.bin", "./bert/deberta-v2-large-japanese-char-wwm/pytorch_model.bin", "./bert/deberta-v3-large/pytorch_model.bin", "./bert/deberta-v3-large/spm.model", ] emo_model_paths = [ "./emotional/clap-htsat-fused/pytorch_model.bin" ] train_base_model_paths = ["D_0.pth", "G_0.pth", "DUR_0.pth"] default_yaml_path = "default_config.yml" default_config_path = "configs/config.json" def load_yaml_data_in_raw(yml_path=yml_config): with open(yml_path, "r", encoding="utf-8") as file: # data = yaml.safe_load(file) data = file.read() return str(data) def load_json_data_in_raw(json_path): with open(json_path, "r", encoding="utf-8") as file: json_data = json.load(file) formatted_json_data = json.dumps(json_data, ensure_ascii=False, indent=2) return formatted_json_data def load_json_data_in_fact(json_path): with open(json_path, "r", encoding="utf-8") as file: json_data = json.load(file) return json_data def load_yaml_data_in_fact(yml_path=yml_config): with open(yml_path, "r", encoding="utf-8") as file: yml = yaml.safe_load(file) # data = file.read() return yml def fill_openi_token(token: str): yml = load_yaml_data_in_fact() yml["mirror"] = "openi" yml["openi_token"] = token write_yaml_data_in_fact(yml) msg = "openi 令牌已填写完成" logger.info(msg) return gr.Textbox(value=msg), gr.Code(value=load_yaml_data_in_raw()) def load_train_param(cfg_path): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] train_json_path = os.path.join(data_path, cfg_path).replace("\\", "/") json_data = load_json_data_in_fact(train_json_path) bs = json_data["train"]["batch_size"] nc = json_data["train"].get("keep_ckpts", 5) li = json_data["train"]["log_interval"] ei = json_data["train"]["eval_interval"] ep = json_data["train"]["epochs"] lr = json_data["train"]["learning_rate"] ver = json_data["version"] msg = f"加载训练配置文件: {train_json_path}" logger.info(msg) return ( gr.Textbox(value=msg), gr.Code(label=train_json_path, value=load_yaml_data_in_raw(train_json_path)), gr.Slider(value=bs), gr.Slider(value=nc), gr.Slider(value=li), gr.Slider(value=ei), gr.Slider(value=ep), gr.Slider(value=lr), gr.Dropdown(value=ver), ) def write_yaml_data_in_fact(yml, yml_path=yml_config): with open(yml_path, "w", encoding="utf-8") as file: yaml.safe_dump(yml, file, allow_unicode=True) # data = file.read() return yml def write_json_data_in_fact(json_path, json_data): with open(json_path, "w", encoding="utf-8") as file: json.dump(json_data, file, ensure_ascii=False, indent=2) def check_if_exists_model(paths: list[str]): check_results = { path: os.path.exists(path) and os.path.isfile(path) for path in paths } val = [path for path, exists in check_results.items() if exists] return val def check_bert_models(): return gr.CheckboxGroup(value=check_if_exists_model(bert_model_paths)) def check_emo_models(): return gr.CheckboxGroup(value=check_if_exists_model(emo_model_paths)) def check_base_models(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] models_dir = yml["train_ms"]["model"] model_paths = [ os.path.join(data_path, models_dir, p).replace("\\", "/") for p in train_base_model_paths ] return gr.CheckboxGroup( label="检测底模状态", info="最好去下载底模进行训练", choices=model_paths, value=check_if_exists_model(model_paths), interactive=False, ) def modify_data_path(data_path): yml = load_yaml_data_in_fact() yml["dataset_path"] = data_path write_yaml_data_in_fact(yml) txt_box = gr.Textbox(value=data_path) return ( gr.Dropdown(value=data_path), txt_box, txt_box, txt_box, gr.Code(value=load_yaml_data_in_raw()), check_base_models(), ) def modify_preprocess_param(trans_path, cfg_path, val_per_lang, max_val_total): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] yml["preprocess_text"]["transcription_path"] = trans_path yml["preprocess_text"]["config_path"] = cfg_path yml["preprocess_text"]["val_per_lang"] = val_per_lang yml["preprocess_text"]["max_val_total"] = max_val_total write_yaml_data_in_fact(yml) whole_path = os.path.join(data_path, cfg_path).replace("\\", "/") logger.info("预处理配置: ", whole_path) if not os.path.exists(whole_path): os.makedirs(os.path.dirname(whole_path), exist_ok=True) shutil.copy(default_config_path, os.path.dirname(whole_path)) return gr.Dropdown(value=trans_path), gr.Code(value=load_yaml_data_in_raw()) def modify_resample_path(in_dir, out_dir, sr): yml = load_yaml_data_in_fact() yml["resample"]["in_dir"] = in_dir yml["resample"]["out_dir"] = out_dir yml["resample"]["sampling_rate"] = int(sr) write_yaml_data_in_fact(yml) msg = f"重采样参数已更改: [{in_dir}, {out_dir}, {sr}]\n" logger.info(msg) return ( gr.Textbox(value=in_dir), gr.Textbox(value=out_dir), gr.Textbox(value=msg), gr.Dropdown(value=sr), gr.Code(value=load_yaml_data_in_raw()), ) def modify_bert_config(cfg_path, nps, dev, multi): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] yml["bert_gen"]["config_path"] = cfg_path yml["bert_gen"]["num_processes"] = int(nps) yml["bert_gen"]["device"] = dev yml["bert_gen"]["use_multi_device"] = multi write_yaml_data_in_fact(yml) whole_path = os.path.join(data_path, cfg_path).replace("\\", "/") logger.info("bert配置路径: ", whole_path) if not os.path.exists(whole_path): os.makedirs(os.path.dirname(whole_path), exist_ok=True) shutil.copy(default_config_path, os.path.dirname(whole_path)) return ( gr.Textbox(value=cfg_path), gr.Slider(value=int(nps)), gr.Dropdown(value=dev), gr.Radio(value=multi), gr.Code(value=load_yaml_data_in_raw()), ) def modify_train_path(model, cfg_path): yml = load_yaml_data_in_fact() yml["train_ms"]["config_path"] = cfg_path yml["train_ms"]["model"] = model write_yaml_data_in_fact(yml) logger.info(f"训练配置文件路径: {cfg_path}\n") logger.info(f"训练模型文件夹路径: {model}") return ( gr.Textbox(value=model), gr.Textbox(value=cfg_path), gr.Code(value=load_yaml_data_in_raw()), check_base_models(), ) def modify_train_param(bs, nc, li, ei, ep, lr, ver): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] cfg_path = yml["train_ms"]["config_path"] ok = False whole_path = os.path.join(data_path, cfg_path).replace("\\", "/") logger.info("config_path: ", whole_path) if not os.path.exists(whole_path): os.makedirs(os.path.dirname(whole_path), exist_ok=True) shutil.copy(default_config_path, os.path.dirname(whole_path)) if os.path.exists(whole_path) and os.path.isfile(whole_path): ok = True with open(whole_path, "r", encoding="utf-8") as file: json_data = json.load(file) json_data["train"]["batch_size"] = bs json_data["train"]["keep_ckpts"] = nc json_data["train"]["log_interval"] = li json_data["train"]["eval_interval"] = ei json_data["train"]["epochs"] = ep json_data["train"]["learning_rate"] = lr json_data["version"] = ver with open(whole_path, "w", encoding="utf-8") as file: json.dump(json_data, file, ensure_ascii=False, indent=2) msg = f"成功更改训练参数! [{bs},{nc},{li},{ei},{ep},{lr}]" logger.info(msg) else: msg = f"打开训练配置文件时出现错误: {whole_path}\n" f"该文件不存在或损坏，现在打开默认配置文件" logger.error(msg) return gr.Textbox(value=msg), gr.Code( label=whole_path if ok else default_config_path, value=load_json_data_in_raw(whole_path) if ok else load_json_data_in_raw(default_config_path), ) def modify_infer_param(model_path, config_path, port, share, debug, ver): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] yml["webui"]["model"] = os.path.relpath(model_path, start=data_path) yml["webui"]["config_path"] = os.path.relpath(config_path, start=data_path) port = int(port) port = port if 0 <= port <= 65535 else 10086 yml["webui"]["port"] = port yml["webui"]["share"] = share yml["webui"]["debug"] = debug write_yaml_data_in_fact(yml) json_data = load_json_data_in_fact(config_path) json_data["version"] = ver write_json_data_in_fact(config_path, json_data) msg = f"修改推理配置文件成功: [{model_path}, {config_path}, {port}, {ver}]" logger.info(msg) return ( gr.Textbox(value=msg), gr.Code(value=load_yaml_data_in_raw()), gr.Code( label=config_path, value=load_json_data_in_raw(config_path) if os.path.exists(config_path) else load_json_data_in_raw(default_config_path), ), ) def get_status(): """获取电脑运行状态""" cpu_percent = psutil.cpu_percent(interval=1) memory_info = psutil.virtual_memory() memory_total = memory_info.total memory_available = memory_info.available memory_used = memory_info.used memory_percent = memory_info.percent gpuInfo = [] devices = ["cpu"] for i in range(torch.cuda.device_count()): devices.append(f"cuda:{i}") if torch.cuda.device_count() > 0: gpus = GPUtil.getGPUs() for gpu in gpus: gpuInfo.append( { "GPU编号": gpu.id, "GPU负载": f"{gpu.load} %", "专用GPU内存": { "总内存": f"{gpu.memoryTotal} MB", "已使用": f"{gpu.memoryUsed} MB", "空闲": f"{gpu.memoryFree} MB", }, } ) status_data = { "devices": devices, "CPU占用率": f"{cpu_percent} %", "总内存": f"{memory_total // (1024 * 1024)} MB", "可用内存": f"{memory_available // (1024 * 1024)} MB", "已使用内存": f"{memory_used // (1024 * 1024)} MB", "百分数": f"{memory_percent} %", "gpu信息": gpuInfo, } formatted_json_data = json.dumps(status_data, ensure_ascii=False, indent=2) logger.info(formatted_json_data) return str(formatted_json_data) def get_gpu_status(): return gr.Code(value=get_status()) def list_infer_models(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] inf_models, json_files = [], [] for root, dirs, files in os.walk(data_path): for file in files: filepath = os.path.join(root, file).replace("\\", "/") if file.startswith("G_") and file.lower().endswith(".pth"): inf_models.append(filepath) elif file.lower().endswith(".json"): json_files.append(filepath) logger.info("找到推理模型文件: " + str(inf_models)) logger.info("找到推理配置文件: " + str(json_files)) return gr.Dropdown(choices=inf_models), gr.Dropdown(choices=json_files) def do_resample(nps): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] in_dir = yml["resample"]["in_dir"] comp_in_dir = os.path.join(os.path.abspath(data_path), in_dir).replace("\\", "/") logger.info(f"\n重采样路径: {comp_in_dir}") cmd = f"python resample.py --processes {nps}" logger.info(cmd) subprocess.run(cmd, shell=True) return gr.Textbox(value="重采样完成!") def do_transcript(lang, workers): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] in_dir = yml["resample"]["in_dir"] comp_in_dir = os.path.join(os.path.abspath(data_path), in_dir).replace("\\", "/") logger.info(f"\n转写文件夹路径: {comp_in_dir}") cmd = f'python asr_transcript.py -f "{comp_in_dir}" -l {lang} -w {workers}' logger.info(cmd) subprocess.run(cmd, shell=True) return gr.Textbox(value=f"\n转写文件夹路径: {comp_in_dir}\n转写到.lab完成!") def do_extract(raw_path, lang, unclean, char_name): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] lab_path = os.path.join(os.path.abspath(data_path), raw_path).replace("\\", "/") unclean_path = os.path.join( data_path, os.path.splitext(unclean)[0] + ".txt" ).replace("\\", "/") logger.info(f"\n提取转写文本路径: {lab_path}") lab_ok = False for root, _, files in os.walk(lab_path): for f_name in files: if str(f_name).lower().endswith(".lab"): lab_ok = True break if lab_ok: break if os.path.exists(lab_path) and os.path.isdir(lab_path): if lab_ok: cmd = f'python extract_list.py -f "{lab_path}" -l {lang} -n "{char_name}" -o "{unclean_path}"' logger.info(cmd) subprocess.run(cmd, shell=True) msg = f"提取完成!生成如下文件: {unclean_path}" logger.info(msg) else: msg = "未找到提取转写文本路径下的.lab文件!" logger.warning(msg) else: msg = "路径未选择正确!" logger.error(msg) return gr.Textbox(value=msg) def do_clean_list(ban_chars, unclean, clean): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] unclean_path = os.path.join(data_path, unclean) clean_path = os.path.join(data_path, clean) if os.path.exists(unclean_path) and os.path.isfile(unclean_path): cmd = f'python clean_list.py -c "{ban_chars}" -i "{unclean_path}" -o "{clean_path}"' logger.info(cmd) subprocess.run(cmd, shell=True) msg = "清洗标注文本完成!" logger.info(msg) else: msg = "未找到可清洗标注文本，请到2.2节重新生成!" logger.warning(msg) return gr.Textbox(value=msg) def do_preprocess_text(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] trans_path = yml["preprocess_text"]["transcription_path"] comp_trans_path = os.path.join(os.path.abspath(data_path), trans_path).replace( "\\", "/" ) logger.info(f"\n清洗后标注文本文件路径: {comp_trans_path}") if os.path.exists(comp_trans_path) and os.path.isfile(comp_trans_path): cmd = "python preprocess_text.py" logger.info(cmd) subprocess.run(cmd, shell=True) msg = "文本预处理完成!" else: msg = "\n清洗后标注文本文件不存在或失效！" logger.info(msg) return gr.Textbox(value=msg) def do_bert_gen(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] train_list_path = yml["preprocess_text"]["train_path"] val_list_path = yml["preprocess_text"]["val_path"] comp_t_path = os.path.join(os.path.abspath(data_path), train_list_path).replace( "\\", "/" ) comp_v_path = os.path.join(os.path.abspath(data_path), val_list_path).replace( "\\", "/" ) if os.path.exists(comp_t_path) and os.path.isfile(comp_t_path): subprocess.run("python bert_gen.py", shell=True) msg = "bert文件生成完成!" logger.info(msg) else: msg = f"未找到训练集和验证集文本!\ntrain: {comp_t_path}\nval:{comp_v_path}" logger.error(msg) return gr.Textbox(value=msg) def modify_emo_gen(emo_cfg, emo_nps, emo_device): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] yml["emo_gen"]["config_path"] = emo_cfg yml["emo_gen"]["num_processes"] = emo_nps yml["emo_gen"]["device"] = emo_device write_yaml_data_in_fact(yml) comp_emo_cfg = os.path.join(os.path.abspath(data_path), emo_cfg).replace("\\", "/") if not os.path.exists(comp_emo_cfg): os.makedirs(os.path.dirname(comp_emo_cfg), exist_ok=True) shutil.copy(default_config_path, os.path.dirname(comp_emo_cfg)) msg = f"修改emo配置参数: [配置路径:{comp_emo_cfg}, 处理数:{emo_nps}, 设备:{emo_device}]" logger.info(msg) return gr.Textbox(value=msg), gr.Code(value=load_yaml_data_in_raw()) def do_emo_gen(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] emo_config_path = yml["emo_gen"]["config_path"] comp_emo_path = os.path.join(os.path.abspath(data_path), emo_config_path).replace( "\\", "/" ) if os.path.exists(comp_emo_path) and os.path.isfile(comp_emo_path): subprocess.run("python emo_gen.py", shell=True) msg = "emo.npy文件生成完成!" logger.info(msg) else: msg = f"选定路径下未找到配置文件!\n需要的config路径 : {comp_emo_path}" logger.error(msg) return gr.Textbox(value=msg) def do_clap_gen(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] train_list_path = yml["preprocess_text"]["train_path"] val_list_path = yml["preprocess_text"]["val_path"] comp_t_path = os.path.join(os.path.abspath(data_path), train_list_path).replace( "\\", "/" ) comp_v_path = os.path.join(os.path.abspath(data_path), val_list_path).replace( "\\", "/" ) msg = f"确保生成了train.list和val.list在对应目录下(由.list通过预处理得到)" logger.warning(msg) msg = f"train: {comp_t_path}" logger.warning(msg) msg = f"val: {comp_v_path}" logger.warning(msg) if os.path.exists(comp_t_path) and os.path.isfile(comp_t_path): subprocess.Popen("python clap_gen.py", shell=True) msg = "clap文件生成完成!" logger.info(msg) else: msg = f"未找到训练集和验证集文本!\ntrain: {comp_t_path}\nval:{comp_v_path}" logger.error(msg) return gr.Textbox(value=msg) def do_my_train(): yml = load_yaml_data_in_fact() n_gpus = torch.cuda.device_count() # subprocess.run(f'python train_ms.py', shell=True) if os.path.exists(r"..\vits\python.exe") and os.path.isfile(r"..\vits\python.exe"): cmd = ( r"..\vits\python ..\vits\Scripts\torchrun.exe " f"--nproc_per_node={n_gpus} train_ms.py" ) else: cmd = f"torchrun --nproc_per_node={n_gpus} train_ms.py" subprocess.Popen(cmd, shell=True) train_port = yml["train_ms"]["env"]["MASTER_PORT"] train_addr = yml["train_ms"]["env"]["MASTER_ADDR"] url = f"env://{train_addr}:{train_port}" msg = f"训练开始!\nMASTER_URL: {url}\n使用gpu数:{n_gpus}\n推荐按下终止训练按钮来结束!" logger.info(msg) return gr.Textbox(value=msg) def do_tensorboard(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] train_model_dir = yml["train_ms"]["model"] whole_dir = os.path.join(data_path, train_model_dir).replace("\\", "/") if os.path.exists(r"..\vits\python.exe") and os.path.isfile(r"..\vits\python.exe"): first_cmd = r"..\vits\python ..\vits\Scripts\tensorboard.exe " else: first_cmd = "tensorboard " tb_cmd = ( first_cmd + f"--logdir={whole_dir} " f"--port={11451} " f'--window_title="训练情况一览" ' f"--reload_interval={120}" ) subprocess.Popen(tb_cmd, shell=True) url = f"http://localhost:{11451}" webbrowser.open(url=url) msg = tb_cmd + "\n" + url logger.info(msg) return gr.Textbox(value=msg) def do_webui_infer(): yml = load_yaml_data_in_fact() data_path = yml["dataset_path"] model_path = yml["webui"]["model"] config_path = yml["webui"]["config_path"] comp_m_path = os.path.join(os.path.abspath(data_path), model_path) comp_c_path = os.path.join(os.path.abspath(data_path), config_path) if os.path.exists(comp_c_path) and os.path.exists(comp_m_path): webui_port = yml["webui"]["port"] subprocess.Popen("python app.py", shell=True) url = f"http://localhost:{webui_port} | http://127.0.0.1:{webui_port}" msg = f"推理端已开启, 到控制台中复制网址打开页面\n{url}\n选择的模型:{model_path}" logger.info(msg) else: msg = f"未找到有效的模型或配置文件!\n模型路径:{comp_m_path}\n配置路径:{comp_c_path}" logger.error(msg) return gr.Textbox(value=msg) def compress_model(cfg_path, in_path, out_path): subprocess.Popen( "python compress_model.py" f" -c {cfg_path}" f" -i {in_path}", shell=True ) msg = "到控制台中查看压缩结果" logger.info(msg) return gr.Textbox(value=msg) def kill_specific_process_linux(cmd): try: output = subprocess.check_output(["pgrep", "-f", cmd], text=True) pids = output.strip().split("\n") for pid in pids: if pid: logger.critical(f"终止进程: {pid}") os.kill(int(pid), signal.SIGTERM) # os.kill(int(pid), signal.SIGKILL) except subprocess.CalledProcessError: logger.error("没有找到匹配的进程。") except Exception as e: logger.error(f"发生错误: {e}") def kill_specific_process_windows(cmd): try: # 使用tasklist和findstr来找到匹配特定命令行模式的进程 output = subprocess.check_output( f'tasklist /FO CSV /V | findstr /C:"{cmd}"', shell=True, text=True ) lines = output.strip().split("\n") for line in lines: if line: pid = line.split(",")[1].strip('"') logger.critical(f"终止进程: {pid}") subprocess.run(["taskkill", "/PID", pid, "/F"], shell=True) # 强制终止 except subprocess.CalledProcessError: logger.error(f"没有找到匹配的{cmd}进程。") except Exception as e: logger.error(f"发生错误: {e}") def stop_train_ms(): yml = load_yaml_data_in_fact() train_port = yml["train_ms"]["env"]["MASTER_PORT"] train_addr = yml["train_ms"]["env"]["MASTER_ADDR"] if platform.system() == "Windows": kill_specific_process_windows("torchrun") else: kill_specific_process_linux("torchrun") url = f"env://{train_addr}:{train_port}" msg = f"训练结束!\nMASTER_URL: {url}" logger.critical(msg) return gr.Textbox(value=msg) def stop_tensorboard(): if platform.system() == "Windows": kill_specific_process_windows("tensorboard") else: kill_specific_process_linux("tensorboard") msg = "关闭tensorboard!\n" logger.critical(msg) return gr.Textbox(value=msg) def stop_webui_infer(): yml = load_yaml_data_in_fact() webui_port = yml["webui"]["port"] if platform.system() == "Linux": kill_specific_process_linux("python app.py") else: kill_specific_process_windows("python app.py") msg = f"尝试终止推理进程，请到控制台查看情况\nport={webui_port}" logger.critical(msg) return gr.Textbox(value=msg) def get_dataset_folders() -> str: os.makedirs('Data', exist_ok=True) glob_list = glob.glob('Data/**/', recursive=False) glob_str = '{' + ",".join(glob_list).replace('\\', '/') + '}' logger.info(glob_str) return glob_str def do_all_process(selected_folders): msg = "\n".join(selected_folders) logger.info(msg) return gr.Textbox(value=msg) def update_dataset_folders(): return gr.FileExplorer(glob=get_dataset_folders()) def fn_create_folder(folder_name): new_path = os.path.join(init_yml['dataset_path'], folder_name) os.makedirs(os.path.join(new_path, "audios/raw"), exist_ok=True) os.makedirs(os.path.join(new_path, "filelists"), exist_ok=True) os.makedirs(os.path.join(new_path, "models"), exist_ok=True) msg = "创建了新的文件夹: " + new_path return gr.Textbox(value=msg), update_dataset_folders() def fn_delete_folder(selected_folders): for path in selected_folders: try: shutil.rmtree(path) except Exception as e: logger.error("删除文件夹发生错误：" + str(e)) msg = "删除了以下文件夹及其子目录\n" + "\n".join(selected_folders) logger.info(msg) return gr.Textbox(value=msg), update_dataset_folders() if __name__ == "__main__": init_yml = load_yaml_data_in_fact() with gr.Blocks( title="Bert-VITS-2-v2.0-管理器", theme=gr.themes.Soft(), css=os.path.abspath("./css/custom.css"), ) as app: with gr.Row(): with gr.Tabs(): with gr.TabItem("首页"): gr.Markdown( """ ## Bert-VITS2-v2.0 可视化界面 #### Copyright/Powered by 怕吃辣滴辣子酱 #### 许可： [AGPL 3.0 Licence](https://github.com/AnyaCoder/Bert-VITS2/blob/master/LICENSE) #### 请订阅我的频道: 1. Bilibili： [spicysama](https://space.bilibili.com/47278440) 2. github： [AnyaCoder](https://github.com/AnyaCoder) ### 严禁将此项目用于一切违反《中华人民共和国宪法》，《中华人民共和国刑法》，《中华人民共和国治安管理处罚法》和《中华人民共和国民法典》之用途。 ### 严禁用于任何政治相关用途。 ## References + [anyvoiceai/MassTTS](https://github.com/anyvoiceai/MassTTS) + [jaywalnut310/vits](https://github.com/jaywalnut310/vits) + [p0p4k/vits2_pytorch](https://github.com/p0p4k/vits2_pytorch) + [svc-develop-team/so-vits-svc](https://github.com/svc-develop-team/so-vits-svc) + [PaddlePaddle/PaddleSpeech](https://github.com/PaddlePaddle/PaddleSpeech) ## 感谢所有贡献者作出的努力 Made with [contrib.rocks](https://contrib.rocks). """ ) with gr.TabItem("填入openi token"): with gr.Row(): gr.Markdown( """ ### 为了后续步骤中能够方便地自动下载模型(bert/emo_gen阶段)，强烈推荐完成这一步骤! ### 去openi官网注册并登录后: ### [点击此处跳转到openi官网](https://openi.pcl.ac.cn/) ### , 点击右上角`个人头像`-> `设置` -> `应用`, 生成令牌(token) ### 复制token, 粘贴到下面的框框, 点击确认 """ ) with gr.Row(): openi_token_box = gr.Textbox( label="填入openi token", value=init_yml["openi_token"] ) with gr.Row(): openi_token_btn = gr.Button(value="确认填写", variant="primary") with gr.Row(): openi_token_status = gr.Textbox(label="状态信息") with gr.TabItem("模型检测"): CheckboxGroup_bert_models = gr.CheckboxGroup( label="检测bert模型状态", info="对应文件夹下必须有对应的模型文件(填入openi token后，则后续步骤中会自动下载)", choices=bert_model_paths, value=check_if_exists_model(bert_model_paths), interactive=False, ) check_pth_btn1 = gr.Button(value="检查bert模型状态") CheckboxGroup_emo_models = gr.CheckboxGroup( label="检测emo模型状态", info="对应文件夹下必须有对应的模型文件", choices=emo_model_paths, value=check_if_exists_model(emo_model_paths), interactive=False, ) check_pth_btn2 = gr.Button(value="检查emo模型状态") with gr.TabItem("数据处理"): with gr.Row(): dropdown_data_path = gr.Dropdown( label="选择数据集存放路径 (右侧的dataset_path)", info="详细说明可见右侧带注释的yaml文件", interactive=True, allow_custom_value=True, choices=[init_yml["dataset_path"]], value=init_yml["dataset_path"], ) with gr.Row(): data_path_btn = gr.Button(value="确认更改存放路径", variant="primary") with gr.Tabs(): with gr.TabItem("1. 音频重采样"): with gr.Row(): resample_in_box = gr.Textbox( label="输入音频文件夹in_dir", value=init_yml["resample"]["in_dir"], lines=1, interactive=True, ) resample_out_box = gr.Textbox( label="输出音频文件夹out_dir", lines=1, value=init_yml["resample"]["out_dir"], interactive=True, ) with gr.Row(): dropdown_resample_sr = gr.Dropdown( label="输出采样率(Hz)", choices=["16000", "22050", "44100", "48000"], value="44100", ) slider_resample_nps = gr.Slider( label="采样用的CPU核心数", minimum=1, maximum=64, step=1, value=2, ) with gr.Row(): resample_config_btn = gr.Button( value="确认重采样配置", variant="secondary", ) resample_btn = gr.Button( value="1. 音频重采样", variant="primary", ) with gr.Row(): resample_status = gr.Textbox( label="重采样结果", placeholder="执行重采样后可查看", lines=3, interactive=False, ) with gr.TabItem("2. 转写文本生成"): with gr.Row(): dropdown_lang = gr.Dropdown( label="选择语言", info="ZH中文，JP日语，EN英语", choices=["ZH", "JP", "EN"], value="ZH", ) slider_transcribe = gr.Slider( label="转写进程数", info="目的路径与前一节一致\n 重采样的输入路径", minimum=1, maximum=10, step=1, value=1, interactive=True, ) clean_txt_box = gr.Textbox( label="非法字符集", info="在此文本框内出现的字符都会被整行删除", lines=1, value="{}<>", interactive=True, ) with gr.Row(): unclean_box = gr.Textbox( label="未清洗的文本", info="仅将.lab提取到这个文件里, 请保持txt格式", lines=1, value=os.path.splitext( init_yml["preprocess_text"][ "transcription_path" ] )[0] + ".txt", interactive=True, ) clean_box = gr.Textbox( label="已清洗的文本", info="将未清洗的文本做去除非法字符集处理后的文本", lines=1, value=init_yml["preprocess_text"][ "transcription_path" ], interactive=True, ) char_name_box = gr.Textbox( label="输入角色名", info="区分说话人用", lines=1, placeholder="填入一个名称", interactive=True, ) with gr.Row(): transcribe_btn = gr.Button( value="2.1 转写文本", interactive=True ) extract_list_btn = gr.Button( value="2.2 合成filelist", ) clean_trans_btn = gr.Button(value="2.3 清洗标注") with gr.Row(): preprocess_status_box = gr.Textbox(label="标注状态") with gr.TabItem("3. 文本预处理"): with gr.Row(): slider_val_per_spk = gr.Slider( label="每种语言的验证集条数", info="TensorBoard里的每种语言eval音频展示条目", minimum=1, maximum=20, step=1, value=init_yml["preprocess_text"]["val_per_lang"], ) slider_max_val_total = gr.Slider( label="验证集最大条数", info="多于此项的会被截断并放到训练集中", minimum=8, maximum=160, step=8, value=init_yml["preprocess_text"]["max_val_total"], ) with gr.Row(): dropdown_filelist_path = gr.Dropdown( interactive=True, label="输入filelist路径", allow_custom_value=True, choices=[ init_yml["preprocess_text"][ "transcription_path" ] ], value=init_yml["preprocess_text"][ "transcription_path" ], ) preprocess_config_box = gr.Textbox( label="预处理配置文件路径", value=init_yml["preprocess_text"]["config_path"], ) with gr.Row(): preprocess_config_btn = gr.Button(value="更新预处理配置文件") preprocess_text_btn = gr.Button( value="标注文本预处理", variant="primary" ) with gr.Row(): label_status = gr.Textbox(label="转写状态") with gr.TabItem("4. bert_gen"): with gr.Row(): bert_dataset_box = gr.Textbox( label="数据集存放路径", text_align="right", value=str(init_yml["dataset_path"]).rstrip("/"), lines=1, interactive=False, scale=10, ) gr.Markdown( """

## + """ ) bert_config_box = gr.Textbox( label="bert_gen配置文件路径", text_align="left", value=init_yml["bert_gen"]["config_path"], lines=1, interactive=True, scale=10, ) with gr.Row(): slider_bert_nps = gr.Slider( label="bert_gen并行处理数", minimum=1, maximum=12, step=1, value=init_yml["bert_gen"]["num_processes"], ) dropdown_bert_dev = gr.Dropdown( label="bert_gen处理设备", choices=["cuda", "cpu"], value=init_yml["bert_gen"]["device"], ) radio_bert_multi = gr.Radio( label="使用多卡推理", choices=[True, False], value=False ) with gr.Row(): bert_config_btn = gr.Button(value="确认更改bert配置项") bert_gen_btn = gr.Button( value="Go! Bert Gen!", variant="primary" ) with gr.Row(): bert_status = gr.Textbox(label="状态信息") with gr.TabItem("5. clap_gen"): with gr.Row(): gr.Markdown(""" ### 和第4步差不多，点就完了 ### 作用：提取情绪特征，生成`.emo.npy`文件训练使用 """) with gr.Row(): slider_clap_nps = gr.Slider( label="clap_gen并行处理数", minimum=1, maximum=12, step=1, value=init_yml["emo_gen"]["num_processes"], ) dropdown_clap_dev = gr.Dropdown( label="clap_gen处理设备", choices=["cuda", "cpu"], value=init_yml["emo_gen"]["device"], ) clap_config_box = gr.Textbox( label="clap_gen配置文件路径", value=init_yml["emo_gen"]["config_path"] ) with gr.Row(): clap_conf_btn = gr.Button( value="确认更改clap配置", variant="secondary" ) clap_gen_btn = gr.Button( value="Clap Gen!", variant="primary" ) with gr.Row(): clap_status = gr.Textbox(label="状态信息") with gr.TabItem("训练界面"): with gr.Tabs(): with gr.TabItem("训练配置文件路径"): with gr.Row(): train_dataset_box_1 = gr.Textbox( label="数据集存放路径", text_align="right", value=str(init_yml["dataset_path"]).rstrip("/"), lines=1, interactive=False, scale=20, ) gr.Markdown( """

## + """ ) train_config_box = gr.Textbox( label="train_ms配置文件路径", text_align="left", value=init_yml["train_ms"]["config_path"], lines=1, interactive=True, scale=20, ) with gr.Row(): train_dataset_box_2 = gr.Textbox( label="数据集存放路径", text_align="right", value=str(init_yml["dataset_path"]).rstrip("/"), lines=1, interactive=False, scale=20, ) gr.Markdown( """

## + """ ) train_model_box = gr.Textbox( label="train_ms模型文件夹路径", value=init_yml["train_ms"]["model"], lines=1, interactive=True, scale=20, ) with gr.Row(): train_ms_path_btn = gr.Button(value="更改训练路径配置") CheckboxGroup_train_models = check_base_models() check_pth_btn3 = gr.Button(value="检查训练底模状态") with gr.TabItem("训练参数设置"): with gr.Row(): slider_batch_size = gr.Slider( minimum=1, maximum=40, value=4, step=1, label="batch_size 批处理大小", ) slider_keep_ckpts = gr.Slider( minimum=1, maximum=20, value=5, step=1, label="keep_ckpts 最多保存n个最新模型", info="若超过，则删除最早的" ) with gr.Row(): slider_log_interval = gr.Slider( minimum=50, maximum=3000, value=200, step=50, label="log_interval 打印日志步数间隔", ) slider_eval_interval = gr.Slider( minimum=100, maximum=5000, value=1000, step=50, label="eval_interval 保存模型步数间隔", ) with gr.Row(): slider_epochs = gr.Slider( minimum=50, maximum=2000, value=100, step=50, label="epochs 训练轮数", ) slider_lr = gr.Slider( minimum=0.0001, maximum=0.0010, value=0.0003, step=0.0001, label="learning_rate 初始学习率", ) with gr.Row(): dropdown_version = gr.Dropdown( label="模型版本选择", info="推荐使用最新版底模和版本训练", choices=["2.2", "2.1", "2.0.2", "2.0.1", "2.0", "1.1.1", "1.1.0", "1.0.1"], value="2.2", ) with gr.Row(): train_ms_load_btn = gr.Button( value="加载训练参数配置", variant="primary" ) train_ms_param_btn = gr.Button( value="更改训练参数配置", variant="primary" ) with gr.Row(): train_btn = gr.Button( value="3.1 点击开始训练", variant="primary" ) train_btn_2 = gr.Button( value="3.2 继续训练", variant="primary" ) stop_train_btn = gr.Button( value="终止训练", variant="secondary" ) with gr.Row(): train_output_box = gr.Textbox( label="状态信息", lines=1, autoscroll=True ) with gr.TabItem("TensorBoard"): with gr.Row(): gr.Markdown( """ ### Tensorboard的logdir 默认为训练的models路径 ### 请在前一节 `训练配置文件路径` 查看 """ ) with gr.Row(): open_tb_btn = gr.Button("开启Tensorboard") stop_tb_btn = gr.Button("关闭Tensorboard") with gr.Row(): tb_output_box = gr.Textbox( label="状态信息", lines=1, autoscroll=True ) with gr.TabItem("推理界面"): with gr.Tabs(): with gr.TabItem("模型选择"): with gr.Row(): dropdown_infer_model = gr.Dropdown( label="选择推理模型", info="默认选择预处理阶段配置的文件夹内容; 也可以自己输入路径。", interactive=True, allow_custom_value=True, ) dropdown_infer_config = gr.Dropdown( label="选择配置文件", info="默认选择预处理阶段配置的文件夹内容; 也可以自己输入路径。", interactive=True, allow_custom_value=True, ) with gr.Row(): dropdown_model_fresh_btn = gr.Button(value="刷新推理模型列表") with gr.Row(): webui_port_box = gr.Textbox( label="WebUI推理的端口号", placeholder="范围:[0, 65535]", max_lines=1, lines=1, value=init_yml["webui"]["port"], interactive=True, ) infer_ver_box = gr.Dropdown( label="更改推理版本", info="已经实现兼容推理，请选择合适的版本", choices=["2.1", "2.0.2", "2.0.1", "2.0", "1.1.1", "1.1.0", "1.0.1"], value="2.1", ) with gr.Row(): radio_webui_share = gr.Radio( label="公开", info="是否公开部署，对外网开放", choices=[True, False], value=init_yml["webui"]["share"], ) radio_webui_debug = gr.Radio( label="调试模式", info="是否开启debug模式", choices=[True, False], value=init_yml["webui"]["debug"], ) with gr.Row(): infer_config_btn = gr.Button(value="更新推理配置文件") stop_infer_btn = gr.Button(value="结束WebUI推理") with gr.Row(): infer_webui_btn = gr.Button( value="开启WebUI推理", variant="primary" ) with gr.Row(): infer_webui_box = gr.Textbox( label="提示信息", interactive=False ) with gr.TabItem("模型压缩"): with gr.Row(): compress_config = gr.Textbox( label="压缩配置文件", info="模型对应的config.json" ) with gr.Row(): compress_input_path = gr.Textbox( label="待压缩模型路径", info="所谓的模型是：G_{步数}.pth" ) with gr.Row(): compress_output_path = gr.Textbox( label="输出模型路径", info="输出为：G_{步数}_release.pth", value="在待压缩模型路径的同一文件夹下", interactive=False, ) with gr.Row(): compress_btn = gr.Button( value="压缩模型", variant="primary" ) with gr.Row(): compress_status = gr.Textbox(label="状态信息") with gr.TabItem("一键训练(前瞻试验版本)"): with gr.Row(): gr.Markdown(""" ### 把所有的音频放入到Data下的每个文件夹的audios/raw中 ### 这里只展示了Data下的子一级文件夹，方便选择数据集 ### - 创建文件夹指的是会把所需的文件夹下所有子目录都建立好 ### - 目前创建与删除重启之后才能生效 """) with gr.Row(): selected_folders = gr.FileExplorer( label="选择数据集文件夹", glob=get_dataset_folders(), height=150, root=".", file_count="multiple" ) with gr.Row(): create_folder_box = gr.Textbox(label="输入要创建的文件夹名称", placeholder="可以含有中文，但不要有空格或特殊符号") final_folder_box = gr.Textbox(label="输入最终输出的文件夹名称", placeholder="可以含有中文，但不要有空格或特殊符号") with gr.Row(): create_folder_btn = gr.Button(variant="primary", value="点击创建所需文件夹") delete_folder_btn = gr.Button(variant="stop", value="点击删除选定文件夹") with gr.Row(): all_process_btn = gr.Button( value="开始一键训练", variant="primary" ) with gr.Row(): all_process_status = gr.Textbox(label="状态信息") with gr.Tabs(): with gr.TabItem("yaml配置文件状态"): code_config_yml = gr.Code( interactive=False, label=yml_config, value=load_yaml_data_in_raw(), language="yaml", elem_id="yml_code", ) with gr.TabItem("带注释的yaml配置文件"): code_default_yml = gr.Code( interactive=False, label=default_yaml_path, value=load_yaml_data_in_raw(default_yaml_path), language="yaml", elem_id="yml_code", ) with gr.TabItem("训练的json配置文件"): code_train_config_json = gr.Code( interactive=False, label=default_config_path, value=load_json_data_in_raw(default_config_path), language="json", elem_id="json_code", ) with gr.TabItem("推理的json配置文件"): code_infer_config_json = gr.Code( interactive=False, label=default_config_path, value=load_json_data_in_raw(default_config_path), language="json", elem_id="json_code", ) with gr.TabItem("其他状态"): code_gpu_json = gr.Code( label="本机资源使用情况", interactive=False, value=get_status(), language="json", elem_id="gpu_code", ) gpu_json_btn = gr.Button(value="刷新本机状态") openi_token_btn.click( fn=fill_openi_token, inputs=[openi_token_box], outputs=[openi_token_status, code_config_yml], ) check_pth_btn1.click( fn=check_bert_models, inputs=[], outputs=[CheckboxGroup_bert_models] ) check_pth_btn2.click( fn=check_emo_models, inputs=[], outputs=[CheckboxGroup_emo_models] ) check_pth_btn3.click( fn=check_base_models, inputs=[], outputs=[CheckboxGroup_train_models] ) data_path_btn.click( fn=modify_data_path, inputs=[dropdown_data_path], outputs=[ dropdown_data_path, bert_dataset_box, train_dataset_box_1, train_dataset_box_2, code_config_yml, CheckboxGroup_train_models, ], ) preprocess_config_btn.click( fn=modify_preprocess_param, inputs=[ dropdown_filelist_path, preprocess_config_box, slider_val_per_spk, slider_max_val_total, ], outputs=[dropdown_filelist_path, code_config_yml], ) preprocess_text_btn.click( fn=do_preprocess_text, inputs=[], outputs=[label_status] ) resample_config_btn.click( fn=modify_resample_path, inputs=[resample_in_box, resample_out_box, dropdown_resample_sr], outputs=[ resample_in_box, resample_out_box, resample_status, dropdown_resample_sr, code_config_yml, ], ) resample_btn.click( fn=do_resample, inputs=[slider_resample_nps], outputs=[resample_status] ) transcribe_btn.click( fn=do_transcript, inputs=[dropdown_lang, slider_transcribe], outputs=[preprocess_status_box], ) extract_list_btn.click( fn=do_extract, inputs=[resample_in_box, dropdown_lang, unclean_box, char_name_box], outputs=[preprocess_status_box], ) clean_trans_btn.click( fn=do_clean_list, inputs=[clean_txt_box, unclean_box, clean_box], outputs=[preprocess_status_box], ) bert_config_btn.click( fn=modify_bert_config, inputs=[ bert_config_box, slider_bert_nps, dropdown_bert_dev, radio_bert_multi, ], outputs=[ bert_config_box, slider_bert_nps, dropdown_bert_dev, radio_bert_multi, code_config_yml, ], ) bert_gen_btn.click(fn=do_bert_gen, inputs=[], outputs=[bert_status]) train_ms_load_btn.click( fn=load_train_param, inputs=[train_config_box], outputs=[ train_output_box, code_train_config_json, slider_batch_size, slider_keep_ckpts, slider_log_interval, slider_eval_interval, slider_epochs, slider_lr, dropdown_version, ], ) train_ms_path_btn.click( fn=modify_train_path, inputs=[train_model_box, train_config_box], outputs=[ train_model_box, train_config_box, code_config_yml, CheckboxGroup_train_models, ], ) train_ms_param_btn.click( fn=modify_train_param, inputs=[ slider_batch_size, slider_keep_ckpts, slider_log_interval, slider_eval_interval, slider_epochs, slider_lr, dropdown_version, ], outputs=[train_output_box, code_train_config_json], ) train_btn.click(fn=do_my_train, inputs=[], outputs=[train_output_box]) train_btn_2.click(fn=do_my_train, inputs=[], outputs=[train_output_box]) stop_train_btn.click(fn=stop_train_ms, inputs=[], outputs=[train_output_box]) open_tb_btn.click(fn=do_tensorboard, inputs=[], outputs=[tb_output_box]) stop_tb_btn.click(fn=stop_tensorboard, inputs=[], outputs=[tb_output_box]) dropdown_model_fresh_btn.click( fn=list_infer_models, inputs=[], outputs=[dropdown_infer_model, dropdown_infer_config], ) infer_config_btn.click( fn=modify_infer_param, inputs=[ dropdown_infer_model, dropdown_infer_config, webui_port_box, radio_webui_share, radio_webui_debug, infer_ver_box, ], outputs=[infer_webui_box, code_config_yml, code_infer_config_json], ) infer_webui_btn.click(fn=do_webui_infer, inputs=[], outputs=[infer_webui_box]) compress_btn.click( fn=compress_model, inputs=[compress_config, compress_input_path, compress_output_path], outputs=[compress_status], ) stop_infer_btn.click(fn=stop_webui_infer, inputs=[], outputs=[infer_webui_box]) gpu_json_btn.click(fn=get_gpu_status, inputs=[], outputs=[code_gpu_json]) clap_gen_btn.click(fn=do_clap_gen, inputs=[], outputs=[clap_status]) clap_conf_btn.click(fn=modify_emo_gen, inputs=[clap_config_box, slider_clap_nps, dropdown_clap_dev], outputs=[clap_status, code_config_yml]) # selected_folders.change(fn=update_dataset_folders, inputs=[], outputs=[selected_folders]) all_process_btn.click(fn=do_all_process, inputs=[selected_folders], outputs=[all_process_status]) create_folder_btn.click(fn=fn_create_folder, inputs=[create_folder_box], outputs=[all_process_status, selected_folders]) delete_folder_btn.click(fn=fn_delete_folder, inputs=[selected_folders], outputs=[all_process_status, selected_folders]) os.environ["no_proxy"] = "localhost,127.0.0.1,0.0.0.0" webbrowser.open("http://127.0.0.1:6007") app.launch(share=False, server_port=6007)