Upload so-vits-svc_for_aliyun.ipynb

so-vits-svc_for_aliyun.ipynb

@@ -0,0 +1,824 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "5cfafef2-66b0-449e-b3f5-734215fdb747",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "## Warning：请自行解决数据集授权问题，禁止使用非授权数据集进行训练！任何由于使用非授权数据集进行训练造成的问题，需自行承担全部责任和后果!与仓库、仓库维护者、svc develop team 、镜像作者无关！ \n",
+    "\n",
+    "    本项目是基于学术交流目的建立，仅供交流与学习使用，并非为生产环境准备。\n",
+    "\n",
+    "    任何发布到视频平台的基于 sovits 制作的视频，都必须要在简介明确指明用于变声器转换的输入源歌声、音频，例如：使用他人发布的视频音频，通过分离的人声作为输入源进行转换的，必须要给出明确的原视频、音乐链接；若使用是自己的人声，或是使用其他歌声合成引擎合成的声音作为输入源进行转换的，也必须在简介加以说明。\n",
+    "\n",
+    "    由输入源造成的侵权问题需自行承担全部责任和一切后果。使用其他商用歌声合成软件作为输入源时，请确保遵守该软件的使用条例，注意，许多歌声合成引擎使用条例中明确指明不可用于输入源进行转换！\n",
+    "\n",
+    "    禁止使用该项目从事违法行为与宗教、政治等活动，该项目维护者、镜像制作者坚决抵制上述行为，不同意此条则禁止使用该项目。\n",
+    "\n",
+    "    继续使用视为已同意上述条款\n",
+    "\n",
+    "### 所有由使用者非法使用所产生的任何后果，均与镜像制作者，项目维护着，开发者 无任何关系 "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "70a8b412-1c64-4006-b8d5-47c2a2ba6a52",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "## 这是一个简介，以及部分教程  \n",
+    "\n",
+    "    初次使用请耐心看完\n",
+    "\n",
+    "    作者:bilibili@kiss丿冷鸟鸟\n",
+    "    邮箱:2649406963@qq.com\n",
+    "    第一次制作镜像，如有bug还请多多理解(摊手)\n",
+    "\n",
+    "    本镜像基于so-vits-svc项目分支sovits4.1-stable（目前为主分支并推荐使用)\n",
+    "    \n",
+    "### 更新日志(可以看一下)\n",
+    "      2023.6.11\n",
+    "    1.缝合了羽毛布团大佬的webui，能用webui进行推理辣!(无法进行训练，但是可以预处理,等我之后想办法解决)\n",
+    "    2.修正了从v1到v5版本的命名错误(至今我才发现768打成了786,草)\n",
+    "    3.新增响度嵌入模型底模(仅768编码器可用)\n",
+    "    4.新增了部分编码器，但是没有底模所以暂时不提供使用\n",
+    "    5.修复了部分镜像v5产生的bug,修改了部分说明,小白务必记得仔细查看\n",
+    "      2023.6.3\n",
+    "    1.更新whisper-ppg编码器 (咬字更清晰) 不过浅扩散功能待更新\n",
+    "    2.增加静态/动态声线融合 (与云端训练基本无关)\n",
+    "    3.增加响度嵌入 (使用后训练出的模型将匹配到输入源响度，否则为训练集响度,开启后能更加匹配原曲的响度)\n",
+    "    4.增加特征检索,来自于RVC (跟聚类方案可以减小音色泄漏，咬字比聚类稍好，但会降低推理速度，俗称聚类模型plus)\n",
+    "    5.新增tensorboard (训练输出可视化)\n",
+    "    6.更新了整合包链接为最新版 (截至6.3的下一天)\n",
+    "    \n",
+    "      2023.05.27\n",
+    "    1.更新resample.py文件，修复了在重采样过程中产生的爆音问题\n",
+    "    2.更新768l12的底模以获得更好的效果\n",
+    "\n",
+    "      2023.05.24\n",
+    "    1.输入更换为 Content Vec 的第12层Transformer输出，并兼容4.0分支\n",
+    "    2.浅层扩散，可以使用浅层扩散模型提升音质\n",
+    "    3.新增vec768l12以及hubert编码器\n",
+    "\n",
+    "### 待更新(希望不会越加越长)\n",
+    "    2023.05.27\n",
+    "    1.onnx编码器(暂时不会更新这玩意，估计用不到，白白浪费储存空间，如果有人需要可以提，或者去看原项目)\n",
+    "     就是懒 (摊手)\n",
+    "    2.whisper-ppg编码器的浅扩散底模添加\n",
+    "    3.可能会试图添加一个建议的webui供推理,可能(√)\n",
+    "    4.很多编码器的底模添加\n",
+    "\n",
+    "项目地址：https://github.com/svc-develop-team/so-vits-svc\n",
+    "\n",
+    "本地建议搭配bilibili@羽毛布团 大佬的整合包使用\n",
+    "\n",
+    "百度网盘：https://pan.baidu.com/s/12u_LDyb5KSOfvjJ9LVwCIQ?pwd=g8n4 提取码：g8n4 \n",
+    "\n",
+    "    目前已经完成环境配置，已经装载用于预训练的底模，包括扩散模型以及主要模型的底模 开箱即用\n",
+    "    他们存放于pre_trained_model,扩散模型底模名为model_0.pt,主要模型底模名称为G_0.pt,D_0.pt\n",
+    "\n",
+    "                           \n",
+    "####                           注意 \n",
+    "###            sovits4.1-stable训练的模型与其他分支基本不兼容 \n",
+    "    但是其他分支的模型可修改配置文件，在这个分支上使用,当然我的推荐是重新训练一个模型\n",
+    "    修改方法，添加代码\n",
+    "    例:\n",
+    "    ...\n",
+    "    \"ssl_dim\": 768,\n",
+    "    \"n_speakers\": 1\n",
+    "    修改为\n",
+    "    ...\n",
+    "    \"ssl_dim\": 768,\n",
+    "    \"n_speakers\": 1,\n",
+    "    \"speech_encoder\": \"vec768l12\",\n",
+    "    \"speaker_embedding\": false\n",
+    "\n",
+    "#### 下面是关于几个 编码器 以及 f0预测器 的区别说明(感谢bilibili@羽毛布団) \n",
+    "\n",
+    "    编码器,默认vec768l12,不推荐ver25619,不提供基于该模型训练,如需要可拉取其他镜像\n",
+    "    vec256l9: ContentVec(256Layer9)，旧版本叫v1，So-VITS-SVC 4.0的基础版本，暂不支持扩散模型(不推荐)\n",
+    "    vec768l12: 特征输入更换为ContentVec的第12层Transformer输出，更还原音色，支持响度嵌入(768特有口胡)\n",
+    "    hubertsoft: So-VITS-SVC 3.0使用的编码器，咬字更为准确，但可能存在音色泄露问题(一个模型只训练一哥声音应该能很好地避免这个问题(?)\n",
+    "    whipser: 咬字更为准确,但是音色还原度不如vec，并且更吃配置。请注意使用这个编码器的时候单个训练数据的长度需要小于30s(暂不支持浅扩散)\n",
+    "    \n",
+    "    f0预测器\n",
+    "    crepe: 抗噪能力最强，但预处理速度很慢(数据集嘈杂推荐使用)\n",
+    "    pm: 预处理速度快，但抗噪能力较弱 \n",
+    "    dio: 原来预处理使用的f0预测器 \n",
+    "    harvest: 有一定抗噪能力，预处理显存占用友好(显存小推荐使用)\n",
+    "\n",
+    "    其他参数我会在下面使用的时候注明\n",
+    "\n",
+    "#### 摸鱼交流群:829974025 \n",
+    "\n",
+    "heart heart heart heart heart heart heart ♥"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2a673177-d43e-4a79-8d66-f12f351801d3",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "#### 请先将数据集上传至 so-vits-svc/dataset_raw文件夹\n",
+    "    关于数据集的处理\n",
+    "    请观看bv:114514(待制作)\n",
+    "    数据集要求\n",
+    "    5s - 15s左右的纯人声切片，总时长建议30分钟到2小时，但 宁缺毋滥，不要往数据集里面丢垃圾，就算是只有5分钟的优秀数据集，也可能比30分钟塞满垃圾的片段效果好\n",
+    "    切片并不是强求5~15s,稍微大点可以安但是不要太大就行   注:使用whisper作为编码器,音频切片需要小于30秒\n",
+    "    数据集的切片尽量在本地进行，切片工具audio-slicer\n",
+    "    \n",
+    "工具链接:https://github.com/flutydeer/audio-slicer\n",
+    "#### 如何上传？\n",
+    "    建议打包成压缩文件，上传后再解压(不会解压直接一股脑全拖过来，只要不嫌慢，随意)\n",
+    "    尽量只训练单个说话人\n",
+    "    假设你训练的是miku模型\n",
+    "    /root/so-vits-svc/dataset_raw/miku\n",
+    "    此时miku文件夹内存放的是miku数据集\n",
+    "    miku为说话人,也就是下方的speaker0\n",
+    "    请尽量不要训练多个角色模型\n",
+    "    我的意思不推荐\n",
+    "上传文件也可参考平台的帮助文档https://www.autodl.com/docs/netdisk/"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8752163c-5d4e-45df-aa2b-cf385c4a6303",
+   "metadata": {},
+   "source": [
+    "#### 数据集文件结构\n",
+    "    dataset_raw\n",
+    "    ├───speaker0\n",
+    "    │   ├───xxx1-xxx1.wav\n",
+    "    │   ├───...\n",
+    "    │   └───Lxx-0xx8.wav\n",
+    "    └───speaker1\n",
+    "        ├───xx2-0xxx2.wav\n",
+    "        ├───...\n",
+    "        └───xxx7-xxx007.wav"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "13bab566-0be4-42c6-b31b-7ab75bd704d5",
+   "metadata": {},
+   "source": [
+    "### 如果你在重新进入笔记本，运行命令时发生报错:can't open file 'xxxxx': [Errno 2] No such file or directory\n",
+    "    请执行下面的进入项目文件夹命令"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5556f776-8343-4376-80bc-9f2f89692e1d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#克隆sovits仓库\n",
+    "!git clone https://ghproxy.com/https://github.com/svc-develop-team/so-vits-svc"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d6a59e50-22ae-40d3-8277-8e217329b2c3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#创建环境\n",
+    "!conda create -n sovits python=3.8\n",
+    "#初始化\n",
+    "!conda init\n",
+    "#激活\n",
+    "!conda activate sovits\n",
+    "#安装torch\n",
+    "!conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7bf0e3bf-caa6-4c9d-95eb-26d0342558a4",
+   "metadata": {
+    "ExecutionIndicator": {
+     "show": true
+    },
+    "execution": {
+     "iopub.execute_input": "2023-07-19T03:33:13.532566Z",
+     "iopub.status.busy": "2023-07-19T03:33:13.531535Z",
+     "iopub.status.idle": "2023-07-19T03:33:13.543838Z",
+     "shell.execute_reply": "2023-07-19T03:33:13.542961Z",
+     "shell.execute_reply.started": "2023-07-19T03:33:13.532519Z"
+    },
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#进入项目文件夹\n",
+    "%cd /mnt/workspace/so-vits-svc"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a8d486d9-eda0-47d7-8715-3898cd2b7a1f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#安装依赖\n",
+    "!pip install pyworld==0.2.12\n",
+    "!pip install -r requirements.txt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0baf0de1-9ee4-43da-8bec-62c18734c63f",
+   "metadata": {
+    "ExecutionIndicator": {
+     "show": true
+    },
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#下载contentvec编码器\n",
+    "!wget -cP /mnt/workspace/so-vits-svc/pretrain https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt -O checkpoint_best_legacy_500.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "09bf5744-5315-4235-8129-32977c509f3a",
+   "metadata": {
+    "ExecutionIndicator": {
+     "show": true
+    },
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#下载hubertsoft编码器\n",
+    "!wget -P /mnt/workspace/so-vits-svc/pretrain https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d53a38a6-f4de-40f6-aa1a-d9d5b022f617",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#下载Whisper-ppg编码器\n",
+    "!wget -P /mnt/workspace/so-vits-svc/pretrain https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5da911c3-561e-43b5-848e-86f4262e1141",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#下载rmvpe编码器\n",
+    "!wget -P /mnt/workspace/so-vits-svc/pretrain https://huggingface.co/datasets/Jinbiii/Jinbi_s_projects/resolve/main/rmvpe.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b93256be-5cdf-40b7-a873-216a26458f61",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#下载cnhubertlarge编码器\n",
+    "!wget -P /mnt/workspace/so-vits-svc/pretrain https://huggingface.co/TencentGameMate/chinese-hubert-large/resolve/main/chinese-hubert-large-fairseq-ckpt.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "81ba4373-605c-4aba-ab88-da832e06177c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#下载dphubert编码器\n",
+    "!wget -P /mnt/workspace/so-vits-svc/pretrain https://huggingface.co/pyf98/DPHuBERT/resolve/main/DPHuBERT-sp0.75.pth"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2490bf4e-5d7b-4f35-847c-a7eba071489c",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "!wget -P /mnt/workspace/so-vits-svc/pretrain https://valle.blob.core.windows.net/share/wavlm/WavLM-Base+.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "59d97962-a8ed-46d7-a98e-2411da8e11a5",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#下载nsf_hifigan\n",
+    "!wget -cP /mnt/workspace/pretrain https://huggingface.co/datasets/Jinbiii/Jinbi_s_projects/resolve/main/nsf_hifigan_20221211.zip\n",
+    "#解压\n",
+    "%cd /mnt/workspace/so-vits-svc/pretrain\n",
+    "!unzip nsf_hifigan_20221211\n",
+    "%cd /mnt/workspace/so-vits-svc/"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "254c6415-7999-4228-9b9f-ddb027ef7283",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#下载扩散模型底模\n",
+    "!wget -P logs/44k/diffusion https://huggingface.co/Kakaru/sovits-whisper-pretrain/resolve/main/diffusion/model_0.pt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7189ee02-7c9d-4b6d-8a36-6e73d8b0a361",
+   "metadata": {
+    "ExecutionIndicator": {
+     "show": true
+    },
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#下载预训练底模\n",
+    "!wget -cP /mnt/workspace/so-vits-svc https://huggingface.co/datasets/Jinbiii/Jinbi_s_projects/resolve/main/pre_trained_model.zip"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "034c9b94-7828-4420-9038-3b1f2a6641e3",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#解压预训练��模\n",
+    "!unzip pre_trained_model"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "62a0d38c-b580-427f-855e-060e819d3915",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "请先将数据集上传至Hugging Face上的Dataset处，以节省时间\n",
+    "下载时注意将链接中的blob改为resolve\n",
+    "例如：https://huggingface.co/datasets/xxx/xxxxx/blob/main/xxx.zip 改为 https://huggingface.co/datasets/xxx/xxxxx/resolve/main/xxx.zip改为"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "913e816f-8545-4fa8-96dd-990aff067b3d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#下载数据集\n",
+    "!wget -cP /mnt/workspace/so-vits-svc/dataset_raw https://huggingface.co/datasets/Jinbiii/Jinbi_s_projects/resolve/main/Geping.zip"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d5656920-2ad0-4177-98a0-3f287505b95e",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#解压数据集\n",
+    "%cd /mnt/workspace/so-vits-svc/dataset_raw\n",
+    "!unzip Geping\n",
+    "%cd /mnt/workspace/so-vits-svc"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b95bfbc4-048f-4acd-a035-16f2fb366b66",
+   "metadata": {},
+   "source": [
+    "开始预处理"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ef0ad7e1-44ba-474f-beae-2a2b56481795",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#进行数据集的重采样(如果报错，请检查是否有数据集太长/短)\n",
+    "!python resample.py "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b316b1e7-5b2f-4edb-8f2c-06c57dc4b3fe",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "#### 选择编码器，并生成配置文件  \n",
+    "\n",
+    "    配置文件存放在config文件夹，名为config.json\n",
+    "    扩散模型的配置文件默认就行,如需查看参数，请见/config_template/diffusion_template.yaml\n",
+    "    (懒得去翻译了)\n",
+    "    \n",
+    "    all_in_mem,cache_all_data：加载所有数据集到内存中，硬盘IO过于低下、同时内存容量 远大于 数据集体积时可以启用(可以较大的提升训练速度)\n",
+    "    \n",
+    "    epoch 总最大训练轮数,中途可中断训练，继续训练会从保存的最后一个模型处开始训练,不一定需要全部训练完，一般3000，4000就可以试试效果，大概1w,2w就能比较好的拟合了\n",
+    "    batch_size 根据显存大小和数据集大小调节,默认值6,显存够可选12，但是请不要太大\n",
+    "    learning_rat 学习率，请和你的batch_size同步,默认为0.0001.当batch_size改为12时，学习率应该改为0.0002\n",
+    "\n",
+    "    eval_interval 每隔多少步保存一次模型,会自动清除老的模型。默认800steps保存一次,请配合keep_ckpt参数设置\n",
+    "    keep_ckpts 一共需要保存多少个模型，默认3，一组模型1G左右，看你硬盘大小    记得备份优秀模型防止被删除\n",
+    "    log_interval 每个多少步输出一次训练日志，\n",
+    "    loss 损失值 一般越低越好.一般在28左右,如果特别大请检查是否加载底模/底模是否对应编码器/数据集是否有问题\n",
+    "    warmup_epochs 预热轮数，此时学习率不会衰减\n",
+    "    seed 模型初始化种子.如果训练效果不理想，可以换一个种子重新训练\n",
+    "\n",
+    "    其他参数默认即可\n",
+    "#### 训练途中请不要修改配置文件"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6d426aea-8b0e-424d-8313-12b957283921",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#选择vec768l12编码器执行这条(带响度嵌入)\n",
+    "!python preprocess_flist_config.py --speech_encoder vec768l12 --vol_aug\n",
+    "%cp pre_trained_model/vol_emb/768l12/* logs/44k"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "05f27476-a4b6-4d89-93e7-e24ecbec69d2",
+   "metadata": {
+    "ExecutionIndicator": {
+     "show": true
+    },
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#选择vec768l12编码器执行这条(不带响度嵌入)\n",
+    "!python preprocess_flist_config.py --speech_encoder vec768l12\n",
+    "%cp -r pre_trained_model/768l12/* logs/44k"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "bc2224e1-5dc2-4569-8261-217af904d318",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#选择hubertsoft编码器执行这条\n",
+    "!python preprocess_flist_config.py --speech_encoder hubertsoft\n",
+    "%cp pre_trained_model/hubertsoft/* logs/44k"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4c3aacdf-11b5-43e7-8340-24ef4d5d9e36",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#选择了whisper-ppg编码器执行这条（暂无扩散模型底模,不支持浅扩散）\n",
+    "!python preprocess_flist_config.py --speech_encoder whisper-ppg\n",
+    "%cp pre_trained_model/whisper-ppg/* logs/44k"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "95242345-25d1-4fbe-b5b5-de25a7e5f645",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "#### 请选择f0预测器\n",
+    "    选择哪个预测器就删掉哪一个前面的的注释符号#，然后运行(不要删掉感叹号!)\n",
+    "    如果需要用浅扩散模型，需要在代码最后面增加 --use_diff 参数\n",
+    "    比如\n",
+    "    python preprocess_hubert_f0.py --f0_predictor dio --use_diff"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9594cf8a-3ce5-49f5-88c9-ce323f35ab67",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "!python preprocess_hubert_f0.py --f0_predictor crepe --use_diff\n",
+    "#!python preprocess_hubert_f0.py --f0_predictor dio\n",
+    "#!python preprocess_hubert_f0.py --f0_predictor pm\n",
+    "#!python preprocess_hubert_f0.py --f0_predictor harvest"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7cc95c5a-5612-4f30-8f28-6c1c9dd57ac8",
+   "metadata": {},
+   "source": [
+    "#### 如果你上一条选择了使用浅扩散模型，则根据你选择编码器，执行下面的命令 将对应的扩散模型放入/logs/44k/diffusion\n",
+    "如没有，则请跳过这部分"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ead62598-cbf4-4198-8fcd-a71e07484cb7",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#选择了vec768l12编码器执行这条\n",
+    "%cp -r pre_trained_model/diffusion/768l12/* logs/44k/diffusion"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "eee66f94-0997-435e-95f5-37d3d025eb84",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#选择了hubertsoft编码器执行这条\n",
+    "%cp pre_trained_model/diffusion/hubertsoft/* logs/44k/diffusion"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6da4706f-e862-41ba-9b7d-5fab7c97d620",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "#### 下面进行训练,请注意你是否选择了浅扩散模型\n",
+    "    注意:只要jupyterlab不出现重启（几乎不会），jupyterlab的终端就会一直运行，训练也不会中断，无论是本地主机断网还是关机\n",
+    "    如果想要中途停止训练，只需在终端中按下ctrl + c,或者点击上面的中断内核\n",
+    "    继续训练也是执行以下命令，从最后保存的节点继续训练"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ddee8b5d-24e5-485a-af6f-941610b356a9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#如果你在上面选择了浅扩散模型，需要训练扩散模型，扩散模型训练方法为:\n",
+    "#如果没有选择，请跳过此步骤\n",
+    "!python train_diff.py -c configs/diffusion.yaml "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ebb3516e-534a-44af-be51-211d23583063",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#主模型正式训练（继续训练也是点这个）\n",
+    "!python train.py -c configs/config.json -m 44k"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f87c6c55-3914-4088-96b2-686e854af6a3",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "#### 输出日志参数说明\n",
+    "    loss 损失值，看loss的收敛程度(看下面的tensorboard)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e7051279-ccbe-49df-a866-7bf2723812e6",
+   "metadata": {
+    "ExecutionIndicator": {
+     "show": false
+    },
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#启用tensorboard\n",
+    "#运行之后去实例监控里面查看，实例监控入口位于快捷工具里面,jupyterlab你怎么开的就在哪里找\n",
+    "!ps -ef | grep tensorboard | awk '{print $2}' | xargs kill -9\n",
+    "!tensorboard --port 6007 --logdir /mnt/workspace/so-vits-svc/logs"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "1ae3f472-ea5d-4908-bd84-82962650f189",
+   "metadata": {},
+   "source": [
+    "#### 聚类模型和特征检索模型"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "29c5119f-ea50-44c3-bd45-a2af7b1a5437",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#聚类模型训练(个人不太喜欢所以不太推荐) \n",
+    "#模型的输出会在logs/44k/kmeans_10000.pt，默认使用gpu训练，所需要的时间大大缩短\n",
+    "!python cluster/train_cluster.py --gpu"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dc6d0b09-a541-484c-938c-e29e49d5b16e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#特征检错训练（聚类模型plus），很快\n",
+    "#生成的模型会保存在 logs/44k/feature_and_index.pkl\n",
+    "!python train_index.py -c configs/config.json"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6f8a53fb-f9c8-4ccb-8bf5-57f926918234",
+   "metadata": {},
+   "source": [
+    "#### 模型压缩(非必要)\n",
+    "生成的模型含有继续训练所需的信息。如果确认不再训练，可以移除模型中此部分信息，得到约 1/3 大小的最终模型。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5d28e48d-1aa1-450c-bb78-5ea2dd8823fb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#请把***G_***.pth改为你自己的模型名称,release.pth为压缩后的最终模型\n",
+    "!python compress_model.py -c=\"configs/config.json\" -i=\"logs/44k/G_.pth\" -o=\"logs/44k/release.pth\""
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "129e43a2-402e-4b64-be1e-9f878c014727",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "#### 推理(建议在本地进行)，当然云端也行\n",
+    "    本地整合包链接在最上方，不要漏看了,下面是webui"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5fe77fbb-e883-414f-a8b1-4d1df564101e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#使用webui进行推理,由于技术问题无法用于训练，但是可以用来推理和预处理,推理得到的音频文件存放于results文件夹\n",
+    "#运行后生成两条链接，可以通过自定义服务打开，也可以直接点击第二条打开（前者比较稳定，后者很不推荐,音频稍微长一点点,及其容易转换到一半超时,除非走投无路否则别用）\n",
+    "!python app.py"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "442429a0-23a3-4cc1-a44a-0d5e3d198bac",
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "#使用 inference_main.py\n",
+    "#具体推理请参考源项目地址,这里不列举其他参数\n",
+    "# 例\n",
+    "!python inference_main.py -m \"logs/44k/G_350400.pth\" -c \"configs/config.json\" -n \"祝福.wav\" -t -12 -s \"Geping\" -f0p \"crepe\" -dm \"logs/44k/diffusion/model_6000.pt\" -dc \"logs/44k/diffusion/configs.yaml\" -ks 150\n",
+    "!python inference_main.py -m \"logs/44k/G_330400.pth\" -c \"configs/config.json\" -n \"向云端.wav\" -t -12 -s \"Geping\" -f0p \"crepe\" -dm \"logs/44k/diffusion/model_6000.pt\" -dc \"logs/44k/diffusion/configs.yaml\" -ks 150\n",
+    "!python inference_main.py -m \"logs/44k/G_192800.pth\" -c \"configs/config.json\" -n \"杀死那个石家庄人.wav\" -t 0 -s \"Geping\" -f0p \"crepe\" -dm \"logs/44k/diffusion/model_6000.pt\" -dc \"logs/44k/diffusion/configs.yaml\" -ks 200\n",
+    "!python inference_main.py -m \"logs/44k/G_192800.pth\" -c \"configs/config.json\" -n \"单相思.wav\" -t -12 -s \"Geping\" -f0p \"crepe\" -cm \"logs/44k/kmeans_10000.pt\" -cr 0.3 -dm \"logs/44k/diffusion/model_6000.pt\" -dc \"logs/44k/diffusion/configs.yaml\" -ks 320\n",
+    "!python inference_main.py -m \"logs/44k/G_192800.pth\" -c \"configs/config.json\" -n \"idol1.wav\" -t -12 -s \"Geping\" -f0p \"rmvpe\" -cm \"logs/44k/kmeans_10000.pt\" -cr 0.3\n",
+    "#将G_30400.pth改为你的模型名称\n",
+    "#config.json改为你的配置文件名称(默认为config.json)\n",
+    "#君の知らない物語-src.wav改为你用于转换的歌曲名称，请将用于转换的歌曲存放于  raw  文件夹  内\n",
+    "#nen改为说话人名称，也就是你的数据集文件夹名，在配置文件最下方spk内可看见\n",
+    "#音高部分随缘改"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5ac6e67e-33df-424d-b56c-37b327d85024",
+   "metadata": {},
+   "source": [
+    "下载模型的时候记得同时下载配置文件，该镜像如果出现问题，请联系镜像作者。\n",
+    "先暂时用着"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2a5024b8-7144-44ab-a077-fb5ddd1783b8",
+   "metadata": {
+    "tags": []
+   },
+   "source": [
+    "## 请及时备份\n",
+    "重新训练模型需要删除数据集，配置文件，以及已经训练好的模型。\n",
+    "删除完后重复你第一次训练的步骤即可"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4e5ab7a0-3666-4673-a796-2a20e0efbb71",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#删除数据集，配置文件，已经训练好的模型\n",
+    "#慎用\n",
+    "%rm -rf logs/44k/*\n",
+    "%rm -rf dataset_raw/*\n",
+    "%rm -rf configs/*\n",
+    "%rm -rf dataset/*\n",
+    "%cd logs/44k/\n",
+    "%mkdir diffusion"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "90b4ad4d-6f4d-4501-b0f5-44cb3dabae04",
+   "metadata": {},
+   "source": [
+    "碰见bug请与作者联系,能修就修,不能修就摇人修("
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.6.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}