├── Eng_docs.md ├── LICENSE ├── README.md ├── add_speaker.py ├── attentions.py ├── commons.py ├── configs └── config.json ├── data_utils.py ├── dataset_raw └── wav_structure.txt ├── filelists ├── test.txt ├── train.txt └── val.txt ├── flask_api.py ├── hubert ├── __init__.py ├── hubert_model.py ├── hubert_model_onnx.py └── put_hubert_ckpt_here ├── inference ├── __init__.py ├── infer_tool.py ├── infer_tool_grad.py └── slicer.py ├── inference_main.py ├── logs └── 32k │ └── put_pretrained_model_here ├── losses.py ├── mel_processing.py ├── model_onnx.py ├── model_onnx_48k.py ├── models.py ├── modules.py ├── onnx_export.py ├── onnx_export_48k.py ├── preprocess_flist_config.py ├── preprocess_hubert_f0.py ├── raw └── put_raw_wav_here ├── requirements.txt ├── resample.py ├── sovits_gradio.py ├── spec_gen.py ├── train.py ├── utils.py └── vdecoder ├── __init__.py └── hifigan ├── env.py ├── models.py ├── nvSTFT.py └── utils.py /Eng_docs.md: -------------------------------------------------------------------------------- 1 | # SoftVC VITS Singing Voice Conversion 2 | 3 | ## Updates 4 | > According to incomplete statistics, it seems that training with multiple speakers may lead to **worsened leaking of voice timbre**. It is not recommended to train models with more than 5 speakers. The current suggestion is to try to train models with only a single speaker if you want to achieve a voice timbre that is more similar to the target. 5 | > Fixed the issue with unwanted staccato, improving audio quality by a decent amount.\ 6 | > The 2.0 version has been moved to the 2.0 branch.\ 7 | > Version 3.0 uses the code structure of FreeVC, which isn't compatible with older versions.\ 8 | > Compared to [DiffSVC](https://github.com/prophesier/diff-svc) , diffsvc performs much better when the training data is of extremely high quality, but this repository may perform better on datasets with lower quality. Additionally, this repository is much faster in terms of inference speed compared to diffsvc. 9 | 10 | ## Model Overview 11 | A singing voice conversion (SVC) model, using the SoftVC encoder to extract features from the input audio, sent into VITS along with the F0 to replace the original input to achieve a voice conversion effect. Additionally, changing the vocoder to [NSF HiFiGAN](https://github.com/openvpi/DiffSinger/tree/refactor/modules/nsf_hifigan) to fix the issue with unwanted staccato. 12 | 13 | ## Notice 14 | + The current branch is the 32kHz version, which requires less vram during inferencing, as well as faster inferencing speeds, and datasets for said branch take up less disk space. Thus the 32 kHz branch is recommended for use. 15 | + If you want to train 48 kHz variant models, switch to the [main branch](https://github.com/innnky/so-vits-svc/tree/main). 16 | 17 | 18 | ## Required models 19 | + soft vc hubert：[hubert-soft-0d54a1f4.pt](https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt) 20 | + Place under `hubert`. 21 | + Pretrained models [G_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth) and [D_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth) 22 | + Place under `logs/32k`. 23 | + Pretrained models are required, because from experiments, training from scratch can be rather unpredictable to say the least, and training with a pretrained model can greatly improve training speeds. 24 | + The pretrained model includes云灏, 即霜, 辉宇·星AI, 派蒙, and 绫地宁宁, covering the common ranges of both male and female voices, and so it can be seen as a rather universal pretrained model. 25 | + The pretrained model exludes the `optimizer speaker_embedding` section, rendering it only usable for pretraining and incapable of inferencing with. 26 | ```shell 27 | # For simple downloading. 28 | # hubert 29 | wget -P hubert/ https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt 30 | # G&D pretrained models 31 | wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth 32 | wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth 33 | 34 | ``` 35 | 36 | ## Colab notebook script for dataset creation and training. 37 | [colab training notebook](https://colab.research.google.com/drive/1rCUOOVG7-XQlVZuWRAj5IpGrMM8t07pE?usp=sharing) 38 | 39 | ## Dataset preparation 40 | All that is required is that the data be put under the `dataset_raw` folder in the structure format provided below. 41 | ```shell 42 | dataset_raw 43 | ├───speaker0 44 | │ ├───xxx1-xxx1.wav 45 | │ ├───... 46 | │ └───Lxx-0xx8.wav 47 | └───speaker1 48 | ├───xx2-0xxx2.wav 49 | ├───... 50 | └───xxx7-xxx007.wav 51 | ``` 52 | 53 | ## Data pre-processing. 54 | 1. Resample to 32khz 55 | 56 | ```shell 57 | python resample.py 58 | ``` 59 | 2. Automatically sort out training set, validation set, test set, and automatically generate configuration files. 60 | ```shell 61 | python preprocess_flist_config.py 62 | # Notice. 63 | # The n_speakers value in the config will be set automatically according to the amount of speakers in the dataset. 64 | # To reserve space for additionally added speakers in the dataset, the n_speakers value will be be set to twice the actual amount. 65 | # If you want even more space for adding more data, you can edit the n_speakers value in the config after runing this step. 66 | # This can not be changed after training starts. 67 | ``` 68 | 3. Generate hubert and F0 features/ 69 | ```shell 70 | python preprocess_hubert_f0.py 71 | ``` 72 | After running the step above, the `dataset` folder will contain all the pre-processed data, you can delete the `dataset_raw` folder after that. 73 | 74 | ## Training. 75 | ```shell 76 | python train.py -c configs/config.json -m 32k 77 | ``` 78 | 79 | ## Inferencing. 80 | 81 | Use [inference_main.py](inference_main.py) 82 | + Edit `model_path` to your newest checkpoint. 83 | + Place the input audio under the `raw` folder. 84 | + Change `clean_names` to the output file name. 85 | + Use `trans` to edit the pitch shifting amount (semitones). 86 | + Change `spk_list` to the speaker name. 87 | 88 | ## Onnx Exporting. 89 | ### **When exporting Onnx, please make sure you re-clone the whole repository!!!** 90 | Use [onnx_export.py](onnx_export.py) 91 | + Create a new folder called `checkpoints`. 92 | + Create a project folder in `checkpoints` folder with the desired name for your project, let's use `myproject` as example. Folder structure looks like `./checkpoints/myproject`. 93 | + Rename your model to `model.pth`, rename your config file to `config.json` then move them into `myproject` folder. 94 | + Modify [onnx_export.py](onnx_export.py) where `path = "NyaruTaffy"`, change `NyaruTaffy` to your project name, here it will be `path = "myproject"`. 95 | + Run [onnx_export.py](onnx_export.py) 96 | + Once it finished, a `model.onnx` will be generated in `myproject` folder, that's the model you just exported. 97 | + Notice: if you want to export a 48K model, please follow the instruction below or use `model_onnx_48k.py` directly. 98 | + Open [model_onnx.py](model_onnx.py) and change `hps={"sampling_rate": 32000...}` to `hps={"sampling_rate": 48000}` in class `SynthesizerTrn`. 99 | + Open [nvSTFT](/vdecoder/hifigan/nvSTFT.py) and replace all `32000` with `48000` 100 | ### Onnx Model UI Support 101 | + [MoeSS](https://github.com/NaruseMioShirakana/MoeSS) 102 | + All training function and transformation are removed, only if they are all removed you are actually using Onnx. 103 | 104 | ## Gradio (WebUI) 105 | Use [sovits_gradio.py](sovits_gradio.py) to run Gradio WebUI 106 | + Create a new folder called `checkpoints`. 107 | + Create a project folder in `checkpoints` folder with the desired name for your project, let's use `myproject` as example. Folder structure looks like `./checkpoints/myproject`. 108 | + Rename your model to `model.pth`, rename your config file to `config.json` then move them into `myproject` folder. 109 | + Run [sovits_gradio.py](sovits_gradio.py) 110 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # SoftVC VITS Singing Voice Conversion 2 | 3 | ## SoVITS已停止更新并正式Archive 4 | 后续维护等工作的一些其他相关仓库： 5 | + [so-vits-svc](https://github.com/svc-develop-team/so-vits-svc) ：svc社区准备接手维护sovits，如果有pr和issue可以往这里提交 6 | + [SoftVitsResearch](https://github.com/NaruseMioShirakana/SoftVitsResearch) ：用于制作一些比较花里胡哨的功能（主要面向Onnx端-MoeSS） 7 | 8 | ## 声明 9 | 10 | 本项目为开源、离线的项目，SvcDevelopTeam 的所有成员与本项目的所有开发者以及维护者（以下简称贡献者）对本项目没有控制力。本项目的贡献者从未向任何组织或个人提供包括但不限于数据集提取、数据集加工、算力支持、训练支持、推理等一切形式的帮助；本项目的贡献者不知晓也无法知晓使用者使用该项目的用途。故一切基于本项目训练的 AI 模型和合成的音频都与本项目贡献者无关。一切由此造成的问题由使用者自行承担。 11 | 12 | 此项目完全离线运行，不能收集任何用户信息或获取用户输入数据。因此，这个项目的贡献者不知道所有的用户输入和模型，因此不负责任何用户输入。 13 | 14 | 本项目只是一个框架项目，本身并没有语音合成的功能，所有的功能都需要用户自己训练模型。同时，这个项目没有任何模型，任何二次分发的项目都与这个项目的贡献者无关。 15 | 16 | ## 📏 使用规约 17 | 18 | # 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 无关！ 19 | 20 | 1. 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。 21 | 2. 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。 22 | 3. 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！ 23 | 4. 禁止使用该项目从事违法行为与宗教、政治等活动，该项目维护者坚决抵制上述行为，不同意此条则禁止使用该项目。 24 | 5. 继续使用视为已同意本仓库 README 所述相关条例，本仓库 README 已进行劝导义务，不对后续可能存在问题负责。 25 | 6. 如果将此项目用于任何其他企划，请提前联系并告知本仓库作者，十分感谢。 26 | 27 | ## English docs 28 | [Check here](Eng_docs.md) 29 | 30 | 31 | ## Updates 32 | > 诈尸一下，4.0-v2不少人炼出来效果很烂，那我就把这分支直接删了 \ 33 | > 更新了4.0-v2模型，全部流程同4.0，在[4.0-v2分支](https://github.com/innnky/so-vits-svc/tree/4.0-v2) 这是sovits最后一次更新 \ 34 | > **4.0模型及colab脚本已更新**：在[4.0分支](https://github.com/innnky/so-vits-svc/tree/4.0) 统一采样率使用44100hz（但推理显存占用比3.0的32khz还小），更换特征提取为contentvec， 目前稳定性还没有经过广泛测试 35 | > 36 | > 据不完全统计，多说话人似乎会导致**音色泄漏加重**，不建议训练超过5人的模型，目前的建议是如果想炼出来更像目标音色，**尽可能炼单说话人的**\ 37 | > 断音问题已解决，音质提升了不少\ 38 | > 2.0版本已经移至 sovits_2.0分支\ 39 | > 3.0版本使用FreeVC的代码结构，与旧版本不通用\ 40 | > 与[DiffSVC](https://github.com/prophesier/diff-svc) 相比，在训练数据质量非常高时diffsvc有着更好的表现，对于质量差一些的数据集，本仓库可能会有更好的表现，此外，本仓库推理速度上比diffsvc快很多 41 | 42 | 43 | ## 模型简介 44 | 歌声音色转换模型，通过SoftVC内容编码器提取源音频语音特征，与F0同时输入VITS替换原本的文本输入达到歌声转换的效果。同时，更换声码器为 [NSF HiFiGAN](https://github.com/openvpi/DiffSinger/tree/refactor/modules/nsf_hifigan) 解决断音问题 45 | 46 | 47 | ## 注意 48 | + 当前分支是32khz版本的分支，32khz模型推理更快，显存占用大幅减小，数据集所占硬盘空间也大幅降低，推荐训练该版本模型 49 | + 如果要训练48khz的模型请切换到[main分支](https://github.com/innnky/so-vits-svc/tree/main) 50 | 51 | 52 | ## 预先下载的模型文件 53 | + soft vc hubert：[hubert-soft-0d54a1f4.pt](https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt) 54 | + 放在`hubert`目录下 55 | + 预训练底模文件 [G_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth) 与 [D_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth) 56 | + 放在`logs/32k`目录下 57 | + 预训练底模为必选项，因为据测试从零开始训练有概率不收敛，同时底模也能加快训练速度 58 | + 预训练底模训练数据集包含云灏 即霜 辉宇·星AI 派蒙 绫地宁宁，覆盖男女生常见音域，可以认为是相对通用的底模 59 | + 底模删除了`optimizer speaker_embedding`等无关权重, 只可以用于初始化训练，无法用于推理 60 | + 该底模和48khz底模通用 61 | ```shell 62 | # 一键下载 63 | # hubert 64 | wget -P hubert/ https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt 65 | # G与D预训练模型 66 | wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth 67 | wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth 68 | 69 | ``` 70 | 71 | 72 | ## colab一键数据集制作、训练脚本 73 | [一键colab](https://colab.research.google.com/drive/1_-gh9i-wCPNlRZw6pYF-9UufetcVrGBX?usp=sharing) 74 | 75 | 76 | ## 数据集准备 77 | 仅需要以以下文件结构将数据集放入dataset_raw目录即可 78 | ```shell 79 | dataset_raw 80 | ├───speaker0 81 | │ ├───xxx1-xxx1.wav 82 | │ ├───... 83 | │ └───Lxx-0xx8.wav 84 | └───speaker1 85 | ├───xx2-0xxx2.wav 86 | ├───... 87 | └───xxx7-xxx007.wav 88 | ``` 89 | 90 | 91 | ## 数据预处理 92 | 1. 重采样至 32khz 93 | 94 | ```shell 95 | python resample.py 96 | ``` 97 | 2. 自动划分训练集 验证集 测试集 以及自动生成配置文件 98 | ```shell 99 | python preprocess_flist_config.py 100 | # 注意 101 | # 自动生成的配置文件中，说话人数量n_speakers会自动按照数据集中的人数而定 102 | # 为了给之后添加说话人留下一定空间，n_speakers自动设置为 当前数据集人数乘2 103 | # 如果想多留一些空位可以在此步骤后 自行修改生成的config.json中n_speakers数量 104 | # 一旦模型开始训练后此项不可再更改 105 | ``` 106 | 3. 生成hubert与f0 107 | ```shell 108 | python preprocess_hubert_f0.py 109 | ``` 110 | 执行完以上步骤后 dataset 目录便是预处理完成的数据，可以删除dataset_raw文件夹了 111 | 112 | 113 | ## 训练 114 | ```shell 115 | python train.py -c configs/config.json -m 32k 116 | ``` 117 | 118 | 119 | ## 推理 120 | 121 | 使用 [inference_main.py](inference_main.py) 122 | + 更改`model_path`为你自己训练的最新模型记录点 123 | + 将待转换的音频放在`raw`文件夹下 124 | + `clean_names` 写待转换的音频名称 125 | + `trans` 填写变调半音数量 126 | + `spk_list` 填写合成的说话人名称 127 | 128 | 129 | ## Onnx导出 130 | ### 重要的事情说三遍：导出Onnx时，请重新克隆整个仓库！！！导出Onnx时，请重新克隆整个仓库！！！导出Onnx时，请重新克隆整个仓库！！！ 131 | 使用 [onnx_export.py](onnx_export.py) 132 | + 新建文件夹：`checkpoints` 并打开 133 | + 在`checkpoints`文件夹中新建一个文件夹作为项目文件夹，文件夹名为你的项目名称，比如`aziplayer` 134 | + 将你的模型更名为`model.pth`，配置文件更名为`config.json`，并放置到刚才创建的`aziplayer`文件夹下 135 | + 将 [onnx_export.py](onnx_export.py) 中`path = "NyaruTaffy"` 的 `"NyaruTaffy"` 修改为你的项目名称，`path = "aziplayer"` 136 | + 运行 [onnx_export.py](onnx_export.py) 137 | + 等待执行完毕，在你的项目文件夹下会生成一个`model.onnx`，即为导出的模型 138 | + 注意：若想导出48K模型，请按照以下步骤修改文件，或者直接使用`model_onnx_48k.py` 139 | + 请打开[model_onnx.py](model_onnx.py)，将其中最后一个class`SynthesizerTrn`的hps中`sampling_rate`32000改为48000 140 | + 请打开[nvSTFT](/vdecoder/hifigan/nvSTFT.py)，将其中所有32000改为48000 141 | ### Onnx模型支持的UI 142 | + [MoeSS](https://github.com/NaruseMioShirakana/MoeSS) 143 | + 我去除了所有的训练用函数和一切复杂的转置，一行都没有保留，因为我认为只有去除了这些东西，才知道你用的是Onnx 144 | 145 | ## Gradio（WebUI） 146 | 使用 [sovits_gradio.py](sovits_gradio.py) 147 | + 新建文件夹：checkpoints 并打开 148 | + 在checkpoints文件夹中新建一个文件夹作为项目文件夹，文件夹名为你的项目名称 149 | + 将你的模型更名为model.pth，配置文件更名为config.json，并放置到刚才创建的文件夹下 150 | + 运行 [sovits_gradio.py](sovits_gradio.py) 151 | 152 | ## 📚 一些法律条例参考 153 | 154 | #### 任何国家，地区，组织和个人使用此项目必须遵守以下法律 155 | 156 | #### 《民法典》 157 | 158 | ##### 第一千零一十九条 159 | 160 | 任何组织或者个人不得以丑化、污损，或者利用信息技术手段伪造等方式侵害他人的肖像权。未经肖像权人同意，不得制作、使用、公开肖像权人的肖像，但是法律另有规定的除外。未经肖像权人同意，肖像作品权利人不得以发表、复制、发行、出租、展览等方式使用或者公开肖像权人的肖像。对自然人声音的保护，参照适用肖像权保护的有关规定。 161 | 162 | ##### 第一千零二十四条 163 | 164 | 【名誉权】民事主体享有名誉权。任何组织或者个人不得以侮辱、诽谤等方式侵害他人的名誉权。 165 | 166 | ##### 第一千零二十七条 167 | 168 | 【作品侵害名誉权】行为人发表的文学、艺术作品以真人真事或者特定人为描述对象，含有侮辱、诽谤内容，侵害他人名誉权的，受害人有权依法请求该行为人承担民事责任。行为人发表的文学、艺术作品不以特定人为描述对象，仅其中的情节与该特定人的情况相似的，不承担民事责任。 169 | 170 | #### 《[中华人民共和国宪法](http://www.gov.cn/guoqing/2018-03/22/content_5276318.htm)》 171 | 172 | #### 《[中华人民共和国刑法](http://gongbao.court.gov.cn/Details/f8e30d0689b23f57bfc782d21035c3.html?sw=中华人民共和国刑法)》 173 | 174 | #### 《[中华人民共和国民法典](http://gongbao.court.gov.cn/Details/51eb6750b8361f79be8f90d09bc202.html)》 175 | 176 | #### 《[中华人民共和国合同法](http://www.npc.gov.cn/zgrdw/npc/lfzt/rlyw/2016-07/01/content_1992739.htm)》 177 | -------------------------------------------------------------------------------- /add_speaker.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | from tqdm import tqdm 4 | from random import shuffle 5 | import json 6 | 7 | 8 | if __name__ == "__main__": 9 | parser = argparse.ArgumentParser() 10 | parser.add_argument("--train_list", type=str, default="./filelists/train.txt", help="path to train list") 11 | parser.add_argument("--val_list", type=str, default="./filelists/val.txt", help="path to val list") 12 | parser.add_argument("--test_list", type=str, default="./filelists/test.txt", help="path to test list") 13 | parser.add_argument("--source_dir", type=str, default="./dataset/32k", help="path to source dir") 14 | args = parser.parse_args() 15 | 16 | previous_config = json.load(open("configs/config.json", "rb")) 17 | 18 | train = [] 19 | val = [] 20 | test = [] 21 | idx = 0 22 | spk_dict = previous_config["spk"] 23 | spk_id = max([i for i in spk_dict.values()]) + 1 24 | for speaker in tqdm(os.listdir(args.source_dir)): 25 | if speaker not in spk_dict.keys(): 26 | spk_dict[speaker] = spk_id 27 | spk_id += 1 28 | wavs = [os.path.join(args.source_dir, speaker, i)for i in os.listdir(os.path.join(args.source_dir, speaker))] 29 | wavs = [i for i in wavs if i.endswith("wav")] 30 | shuffle(wavs) 31 | train += wavs[2:-10] 32 | val += wavs[:2] 33 | test += wavs[-10:] 34 | 35 | assert previous_config["model"]["n_speakers"] > len(spk_dict.keys()) 36 | shuffle(train) 37 | shuffle(val) 38 | shuffle(test) 39 | 40 | print("Writing", args.train_list) 41 | with open(args.train_list, "w") as f: 42 | for fname in tqdm(train): 43 | wavpath = fname 44 | f.write(wavpath + "\n") 45 | 46 | print("Writing", args.val_list) 47 | with open(args.val_list, "w") as f: 48 | for fname in tqdm(val): 49 | wavpath = fname 50 | f.write(wavpath + "\n") 51 | 52 | print("Writing", args.test_list) 53 | with open(args.test_list, "w") as f: 54 | for fname in tqdm(test): 55 | wavpath = fname 56 | f.write(wavpath + "\n") 57 | 58 | previous_config["spk"] = spk_dict 59 | 60 | print("Writing configs/config.json") 61 | with open("configs/config.json", "w") as f: 62 | json.dump(previous_config, f, indent=2) 63 | -------------------------------------------------------------------------------- /attentions.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import numpy as np 4 | import torch 5 | from torch import nn 6 | from torch.nn import functional as F 7 | 8 | import commons 9 | import modules 10 | from modules import LayerNorm 11 | 12 | 13 | class Encoder(nn.Module): 14 | def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs): 15 | super().__init__() 16 | self.hidden_channels = hidden_channels 17 | self.filter_channels = filter_channels 18 | self.n_heads = n_heads 19 | self.n_layers = n_layers 20 | self.kernel_size = kernel_size 21 | self.p_dropout = p_dropout 22 | self.window_size = window_size 23 | 24 | self.drop = nn.Dropout(p_dropout) 25 | self.attn_layers = nn.ModuleList() 26 | self.norm_layers_1 = nn.ModuleList() 27 | self.ffn_layers = nn.ModuleList() 28 | self.norm_layers_2 = nn.ModuleList() 29 | for i in range(self.n_layers): 30 | self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size)) 31 | self.norm_layers_1.append(LayerNorm(hidden_channels)) 32 | self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout)) 33 | self.norm_layers_2.append(LayerNorm(hidden_channels)) 34 | 35 | def forward(self, x, x_mask): 36 | attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1) 37 | x = x * x_mask 38 | for i in range(self.n_layers): 39 | y = self.attn_layers[i](x, x, attn_mask) 40 | y = self.drop(y) 41 | x = self.norm_layers_1[i](x + y) 42 | 43 | y = self.ffn_layers[i](x, x_mask) 44 | y = self.drop(y) 45 | x = self.norm_layers_2[i](x + y) 46 | x = x * x_mask 47 | return x 48 | 49 | 50 | class Decoder(nn.Module): 51 | def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs): 52 | super().__init__() 53 | self.hidden_channels = hidden_channels 54 | self.filter_channels = filter_channels 55 | self.n_heads = n_heads 56 | self.n_layers = n_layers 57 | self.kernel_size = kernel_size 58 | self.p_dropout = p_dropout 59 | self.proximal_bias = proximal_bias 60 | self.proximal_init = proximal_init 61 | 62 | self.drop = nn.Dropout(p_dropout) 63 | self.self_attn_layers = nn.ModuleList() 64 | self.norm_layers_0 = nn.ModuleList() 65 | self.encdec_attn_layers = nn.ModuleList() 66 | self.norm_layers_1 = nn.ModuleList() 67 | self.ffn_layers = nn.ModuleList() 68 | self.norm_layers_2 = nn.ModuleList() 69 | for i in range(self.n_layers): 70 | self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init)) 71 | self.norm_layers_0.append(LayerNorm(hidden_channels)) 72 | self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout)) 73 | self.norm_layers_1.append(LayerNorm(hidden_channels)) 74 | self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True)) 75 | self.norm_layers_2.append(LayerNorm(hidden_channels)) 76 | 77 | def forward(self, x, x_mask, h, h_mask): 78 | """ 79 | x: decoder input 80 | h: encoder output 81 | """ 82 | self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype) 83 | encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1) 84 | x = x * x_mask 85 | for i in range(self.n_layers): 86 | y = self.self_attn_layers[i](x, x, self_attn_mask) 87 | y = self.drop(y) 88 | x = self.norm_layers_0[i](x + y) 89 | 90 | y = self.encdec_attn_layers[i](x, h, encdec_attn_mask) 91 | y = self.drop(y) 92 | x = self.norm_layers_1[i](x + y) 93 | 94 | y = self.ffn_layers[i](x, x_mask) 95 | y = self.drop(y) 96 | x = self.norm_layers_2[i](x + y) 97 | x = x * x_mask 98 | return x 99 | 100 | 101 | class MultiHeadAttention(nn.Module): 102 | def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False): 103 | super().__init__() 104 | assert channels % n_heads == 0 105 | 106 | self.channels = channels 107 | self.out_channels = out_channels 108 | self.n_heads = n_heads 109 | self.p_dropout = p_dropout 110 | self.window_size = window_size 111 | self.heads_share = heads_share 112 | self.block_length = block_length 113 | self.proximal_bias = proximal_bias 114 | self.proximal_init = proximal_init 115 | self.attn = None 116 | 117 | self.k_channels = channels // n_heads 118 | self.conv_q = nn.Conv1d(channels, channels, 1) 119 | self.conv_k = nn.Conv1d(channels, channels, 1) 120 | self.conv_v = nn.Conv1d(channels, channels, 1) 121 | self.conv_o = nn.Conv1d(channels, out_channels, 1) 122 | self.drop = nn.Dropout(p_dropout) 123 | 124 | if window_size is not None: 125 | n_heads_rel = 1 if heads_share else n_heads 126 | rel_stddev = self.k_channels**-0.5 127 | self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev) 128 | self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev) 129 | 130 | nn.init.xavier_uniform_(self.conv_q.weight) 131 | nn.init.xavier_uniform_(self.conv_k.weight) 132 | nn.init.xavier_uniform_(self.conv_v.weight) 133 | if proximal_init: 134 | with torch.no_grad(): 135 | self.conv_k.weight.copy_(self.conv_q.weight) 136 | self.conv_k.bias.copy_(self.conv_q.bias) 137 | 138 | def forward(self, x, c, attn_mask=None): 139 | q = self.conv_q(x) 140 | k = self.conv_k(c) 141 | v = self.conv_v(c) 142 | 143 | x, self.attn = self.attention(q, k, v, mask=attn_mask) 144 | 145 | x = self.conv_o(x) 146 | return x 147 | 148 | def attention(self, query, key, value, mask=None): 149 | # reshape [b, d, t] -> [b, n_h, t, d_k] 150 | b, d, t_s, t_t = (*key.size(), query.size(2)) 151 | query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3) 152 | key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3) 153 | value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3) 154 | 155 | scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1)) 156 | if self.window_size is not None: 157 | assert t_s == t_t, "Relative attention is only available for self-attention." 158 | key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s) 159 | rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings) 160 | scores_local = self._relative_position_to_absolute_position(rel_logits) 161 | scores = scores + scores_local 162 | if self.proximal_bias: 163 | assert t_s == t_t, "Proximal bias is only available for self-attention." 164 | scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype) 165 | if mask is not None: 166 | scores = scores.masked_fill(mask == 0, -1e4) 167 | if self.block_length is not None: 168 | assert t_s == t_t, "Local attention is only available for self-attention." 169 | block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length) 170 | scores = scores.masked_fill(block_mask == 0, -1e4) 171 | p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s] 172 | p_attn = self.drop(p_attn) 173 | output = torch.matmul(p_attn, value) 174 | if self.window_size is not None: 175 | relative_weights = self._absolute_position_to_relative_position(p_attn) 176 | value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s) 177 | output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings) 178 | output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t] 179 | return output, p_attn 180 | 181 | def _matmul_with_relative_values(self, x, y): 182 | """ 183 | x: [b, h, l, m] 184 | y: [h or 1, m, d] 185 | ret: [b, h, l, d] 186 | """ 187 | ret = torch.matmul(x, y.unsqueeze(0)) 188 | return ret 189 | 190 | def _matmul_with_relative_keys(self, x, y): 191 | """ 192 | x: [b, h, l, d] 193 | y: [h or 1, m, d] 194 | ret: [b, h, l, m] 195 | """ 196 | ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1)) 197 | return ret 198 | 199 | def _get_relative_embeddings(self, relative_embeddings, length): 200 | max_relative_position = 2 * self.window_size + 1 201 | # Pad first before slice to avoid using cond ops. 202 | pad_length = max(length - (self.window_size + 1), 0) 203 | slice_start_position = max((self.window_size + 1) - length, 0) 204 | slice_end_position = slice_start_position + 2 * length - 1 205 | if pad_length > 0: 206 | padded_relative_embeddings = F.pad( 207 | relative_embeddings, 208 | commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]])) 209 | else: 210 | padded_relative_embeddings = relative_embeddings 211 | used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position] 212 | return used_relative_embeddings 213 | 214 | def _relative_position_to_absolute_position(self, x): 215 | """ 216 | x: [b, h, l, 2*l-1] 217 | ret: [b, h, l, l] 218 | """ 219 | batch, heads, length, _ = x.size() 220 | # Concat columns of pad to shift from relative to absolute indexing. 221 | x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]])) 222 | 223 | # Concat extra elements so to add up to shape (len+1, 2*len-1). 224 | x_flat = x.view([batch, heads, length * 2 * length]) 225 | x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]])) 226 | 227 | # Reshape and slice out the padded elements. 228 | x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:] 229 | return x_final 230 | 231 | def _absolute_position_to_relative_position(self, x): 232 | """ 233 | x: [b, h, l, l] 234 | ret: [b, h, l, 2*l-1] 235 | """ 236 | batch, heads, length, _ = x.size() 237 | # padd along column 238 | x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]])) 239 | x_flat = x.view([batch, heads, length**2 + length*(length -1)]) 240 | # add 0's in the beginning that will skew the elements after reshape 241 | x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]])) 242 | x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:] 243 | return x_final 244 | 245 | def _attention_bias_proximal(self, length): 246 | """Bias for self-attention to encourage attention to close positions. 247 | Args: 248 | length: an integer scalar. 249 | Returns: 250 | a Tensor with shape [1, 1, length, length] 251 | """ 252 | r = torch.arange(length, dtype=torch.float32) 253 | diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1) 254 | return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0) 255 | 256 | 257 | class FFN(nn.Module): 258 | def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False): 259 | super().__init__() 260 | self.in_channels = in_channels 261 | self.out_channels = out_channels 262 | self.filter_channels = filter_channels 263 | self.kernel_size = kernel_size 264 | self.p_dropout = p_dropout 265 | self.activation = activation 266 | self.causal = causal 267 | 268 | if causal: 269 | self.padding = self._causal_padding 270 | else: 271 | self.padding = self._same_padding 272 | 273 | self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size) 274 | self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size) 275 | self.drop = nn.Dropout(p_dropout) 276 | 277 | def forward(self, x, x_mask): 278 | x = self.conv_1(self.padding(x * x_mask)) 279 | if self.activation == "gelu": 280 | x = x * torch.sigmoid(1.702 * x) 281 | else: 282 | x = torch.relu(x) 283 | x = self.drop(x) 284 | x = self.conv_2(self.padding(x * x_mask)) 285 | return x * x_mask 286 | 287 | def _causal_padding(self, x): 288 | if self.kernel_size == 1: 289 | return x 290 | pad_l = self.kernel_size - 1 291 | pad_r = 0 292 | padding = [[0, 0], [0, 0], [pad_l, pad_r]] 293 | x = F.pad(x, commons.convert_pad_shape(padding)) 294 | return x 295 | 296 | def _same_padding(self, x): 297 | if self.kernel_size == 1: 298 | return x 299 | pad_l = (self.kernel_size - 1) // 2 300 | pad_r = self.kernel_size // 2 301 | padding = [[0, 0], [0, 0], [pad_l, pad_r]] 302 | x = F.pad(x, commons.convert_pad_shape(padding)) 303 | return x 304 | -------------------------------------------------------------------------------- /commons.py: -------------------------------------------------------------------------------- 1 | import math 2 | import numpy as np 3 | import torch 4 | from torch import nn 5 | from torch.nn import functional as F 6 | 7 | def slice_pitch_segments(x, ids_str, segment_size=4): 8 | ret = torch.zeros_like(x[:, :segment_size]) 9 | for i in range(x.size(0)): 10 | idx_str = ids_str[i] 11 | idx_end = idx_str + segment_size 12 | ret[i] = x[i, idx_str:idx_end] 13 | return ret 14 | 15 | def rand_slice_segments_with_pitch(x, pitch, x_lengths=None, segment_size=4): 16 | b, d, t = x.size() 17 | if x_lengths is None: 18 | x_lengths = t 19 | ids_str_max = x_lengths - segment_size + 1 20 | ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long) 21 | ret = slice_segments(x, ids_str, segment_size) 22 | ret_pitch = slice_pitch_segments(pitch, ids_str, segment_size) 23 | return ret, ret_pitch, ids_str 24 | 25 | def init_weights(m, mean=0.0, std=0.01): 26 | classname = m.__class__.__name__ 27 | if classname.find("Conv") != -1: 28 | m.weight.data.normal_(mean, std) 29 | 30 | 31 | def get_padding(kernel_size, dilation=1): 32 | return int((kernel_size*dilation - dilation)/2) 33 | 34 | 35 | def convert_pad_shape(pad_shape): 36 | l = pad_shape[::-1] 37 | pad_shape = [item for sublist in l for item in sublist] 38 | return pad_shape 39 | 40 | 41 | def intersperse(lst, item): 42 | result = [item] * (len(lst) * 2 + 1) 43 | result[1::2] = lst 44 | return result 45 | 46 | 47 | def kl_divergence(m_p, logs_p, m_q, logs_q): 48 | """KL(P||Q)""" 49 | kl = (logs_q - logs_p) - 0.5 50 | kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q) 51 | return kl 52 | 53 | 54 | def rand_gumbel(shape): 55 | """Sample from the Gumbel distribution, protect from overflows.""" 56 | uniform_samples = torch.rand(shape) * 0.99998 + 0.00001 57 | return -torch.log(-torch.log(uniform_samples)) 58 | 59 | 60 | def rand_gumbel_like(x): 61 | g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device) 62 | return g 63 | 64 | 65 | def slice_segments(x, ids_str, segment_size=4): 66 | ret = torch.zeros_like(x[:, :, :segment_size]) 67 | for i in range(x.size(0)): 68 | idx_str = ids_str[i] 69 | idx_end = idx_str + segment_size 70 | ret[i] = x[i, :, idx_str:idx_end] 71 | return ret 72 | 73 | 74 | def rand_slice_segments(x, x_lengths=None, segment_size=4): 75 | b, d, t = x.size() 76 | if x_lengths is None: 77 | x_lengths = t 78 | ids_str_max = x_lengths - segment_size + 1 79 | ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long) 80 | ret = slice_segments(x, ids_str, segment_size) 81 | return ret, ids_str 82 | 83 | 84 | def rand_spec_segments(x, x_lengths=None, segment_size=4): 85 | b, d, t = x.size() 86 | if x_lengths is None: 87 | x_lengths = t 88 | ids_str_max = x_lengths - segment_size 89 | ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long) 90 | ret = slice_segments(x, ids_str, segment_size) 91 | return ret, ids_str 92 | 93 | 94 | def get_timing_signal_1d( 95 | length, channels, min_timescale=1.0, max_timescale=1.0e4): 96 | position = torch.arange(length, dtype=torch.float) 97 | num_timescales = channels // 2 98 | log_timescale_increment = ( 99 | math.log(float(max_timescale) / float(min_timescale)) / 100 | (num_timescales - 1)) 101 | inv_timescales = min_timescale * torch.exp( 102 | torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment) 103 | scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1) 104 | signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0) 105 | signal = F.pad(signal, [0, 0, 0, channels % 2]) 106 | signal = signal.view(1, channels, length) 107 | return signal 108 | 109 | 110 | def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4): 111 | b, channels, length = x.size() 112 | signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale) 113 | return x + signal.to(dtype=x.dtype, device=x.device) 114 | 115 | 116 | def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1): 117 | b, channels, length = x.size() 118 | signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale) 119 | return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis) 120 | 121 | 122 | def subsequent_mask(length): 123 | mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0) 124 | return mask 125 | 126 | 127 | @torch.jit.script 128 | def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels): 129 | n_channels_int = n_channels[0] 130 | in_act = input_a + input_b 131 | t_act = torch.tanh(in_act[:, :n_channels_int, :]) 132 | s_act = torch.sigmoid(in_act[:, n_channels_int:, :]) 133 | acts = t_act * s_act 134 | return acts 135 | 136 | 137 | def convert_pad_shape(pad_shape): 138 | l = pad_shape[::-1] 139 | pad_shape = [item for sublist in l for item in sublist] 140 | return pad_shape 141 | 142 | 143 | def shift_1d(x): 144 | x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1] 145 | return x 146 | 147 | 148 | def sequence_mask(length, max_length=None): 149 | if max_length is None: 150 | max_length = length.max() 151 | x = torch.arange(max_length, dtype=length.dtype, device=length.device) 152 | return x.unsqueeze(0) < length.unsqueeze(1) 153 | 154 | 155 | def generate_path(duration, mask): 156 | """ 157 | duration: [b, 1, t_x] 158 | mask: [b, 1, t_y, t_x] 159 | """ 160 | device = duration.device 161 | 162 | b, _, t_y, t_x = mask.shape 163 | cum_duration = torch.cumsum(duration, -1) 164 | 165 | cum_duration_flat = cum_duration.view(b * t_x) 166 | path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype) 167 | path = path.view(b, t_x, t_y) 168 | path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1] 169 | path = path.unsqueeze(1).transpose(2,3) * mask 170 | return path 171 | 172 | 173 | def clip_grad_value_(parameters, clip_value, norm_type=2): 174 | if isinstance(parameters, torch.Tensor): 175 | parameters = [parameters] 176 | parameters = list(filter(lambda p: p.grad is not None, parameters)) 177 | norm_type = float(norm_type) 178 | if clip_value is not None: 179 | clip_value = float(clip_value) 180 | 181 | total_norm = 0 182 | for p in parameters: 183 | param_norm = p.grad.data.norm(norm_type) 184 | total_norm += param_norm.item() ** norm_type 185 | if clip_value is not None: 186 | p.grad.data.clamp_(min=-clip_value, max=clip_value) 187 | total_norm = total_norm ** (1. / norm_type) 188 | return total_norm 189 | -------------------------------------------------------------------------------- /configs/config.json: -------------------------------------------------------------------------------- 1 | 请使用生成的config文件 -------------------------------------------------------------------------------- /data_utils.py: -------------------------------------------------------------------------------- 1 | import time 2 | import os 3 | import random 4 | import numpy as np 5 | import torch 6 | import torch.utils.data 7 | 8 | import commons 9 | from mel_processing import spectrogram_torch, spec_to_mel_torch 10 | from utils import load_wav_to_torch, load_filepaths_and_text, transform 11 | 12 | # import h5py 13 | 14 | 15 | """Multi speaker version""" 16 | 17 | 18 | class TextAudioSpeakerLoader(torch.utils.data.Dataset): 19 | """ 20 | 1) loads audio, speaker_id, text pairs 21 | 2) normalizes text and converts them to sequences of integers 22 | 3) computes spectrograms from audio files. 23 | """ 24 | 25 | def __init__(self, audiopaths, hparams): 26 | self.audiopaths = load_filepaths_and_text(audiopaths) 27 | self.max_wav_value = hparams.data.max_wav_value 28 | self.sampling_rate = hparams.data.sampling_rate 29 | self.filter_length = hparams.data.filter_length 30 | self.hop_length = hparams.data.hop_length 31 | self.win_length = hparams.data.win_length 32 | self.sampling_rate = hparams.data.sampling_rate 33 | self.use_sr = hparams.train.use_sr 34 | self.spec_len = hparams.train.max_speclen 35 | self.spk_map = hparams.spk 36 | 37 | random.seed(1234) 38 | random.shuffle(self.audiopaths) 39 | 40 | def get_audio(self, filename): 41 | filename = filename.replace("\\", "/") 42 | audio, sampling_rate = load_wav_to_torch(filename) 43 | if sampling_rate != self.sampling_rate: 44 | raise ValueError("{} SR doesn't match target {} SR".format( 45 | sampling_rate, self.sampling_rate)) 46 | audio_norm = audio / self.max_wav_value 47 | audio_norm = audio_norm.unsqueeze(0) 48 | spec_filename = filename.replace(".wav", ".spec.pt") 49 | if os.path.exists(spec_filename): 50 | spec = torch.load(spec_filename) 51 | else: 52 | spec = spectrogram_torch(audio_norm, self.filter_length, 53 | self.sampling_rate, self.hop_length, self.win_length, 54 | center=False) 55 | spec = torch.squeeze(spec, 0) 56 | torch.save(spec, spec_filename) 57 | 58 | spk = filename.split("/")[-2] 59 | spk = torch.LongTensor([self.spk_map[spk]]) 60 | 61 | c = torch.load(filename + ".soft.pt").squeeze(0) 62 | c = torch.repeat_interleave(c, repeats=2, dim=1) 63 | 64 | f0 = np.load(filename + ".f0.npy") 65 | f0 = torch.FloatTensor(f0) 66 | lmin = min(c.size(-1), spec.size(-1), f0.shape[0]) 67 | assert abs(c.size(-1) - spec.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape, filename) 68 | assert abs(lmin - spec.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape) 69 | assert abs(lmin - c.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape) 70 | spec, c, f0 = spec[:, :lmin], c[:, :lmin], f0[:lmin] 71 | audio_norm = audio_norm[:, :lmin * self.hop_length] 72 | _spec, _c, _audio_norm, _f0 = spec, c, audio_norm, f0 73 | while spec.size(-1) < self.spec_len: 74 | spec = torch.cat((spec, _spec), -1) 75 | c = torch.cat((c, _c), -1) 76 | f0 = torch.cat((f0, _f0), -1) 77 | audio_norm = torch.cat((audio_norm, _audio_norm), -1) 78 | start = random.randint(0, spec.size(-1) - self.spec_len) 79 | end = start + self.spec_len 80 | spec = spec[:, start:end] 81 | c = c[:, start:end] 82 | f0 = f0[start:end] 83 | audio_norm = audio_norm[:, start * self.hop_length:end * self.hop_length] 84 | 85 | return c, f0, spec, audio_norm, spk 86 | 87 | def __getitem__(self, index): 88 | return self.get_audio(self.audiopaths[index][0]) 89 | 90 | def __len__(self): 91 | return len(self.audiopaths) 92 | 93 | 94 | class EvalDataLoader(torch.utils.data.Dataset): 95 | """ 96 | 1) loads audio, speaker_id, text pairs 97 | 2) normalizes text and converts them to sequences of integers 98 | 3) computes spectrograms from audio files. 99 | """ 100 | 101 | def __init__(self, audiopaths, hparams): 102 | self.audiopaths = load_filepaths_and_text(audiopaths) 103 | self.max_wav_value = hparams.data.max_wav_value 104 | self.sampling_rate = hparams.data.sampling_rate 105 | self.filter_length = hparams.data.filter_length 106 | self.hop_length = hparams.data.hop_length 107 | self.win_length = hparams.data.win_length 108 | self.sampling_rate = hparams.data.sampling_rate 109 | self.use_sr = hparams.train.use_sr 110 | self.audiopaths = self.audiopaths[:5] 111 | self.spk_map = hparams.spk 112 | 113 | 114 | def get_audio(self, filename): 115 | filename = filename.replace("\\", "/") 116 | audio, sampling_rate = load_wav_to_torch(filename) 117 | if sampling_rate != self.sampling_rate: 118 | raise ValueError("{} SR doesn't match target {} SR".format( 119 | sampling_rate, self.sampling_rate)) 120 | audio_norm = audio / self.max_wav_value 121 | audio_norm = audio_norm.unsqueeze(0) 122 | spec_filename = filename.replace(".wav", ".spec.pt") 123 | if os.path.exists(spec_filename): 124 | spec = torch.load(spec_filename) 125 | else: 126 | spec = spectrogram_torch(audio_norm, self.filter_length, 127 | self.sampling_rate, self.hop_length, self.win_length, 128 | center=False) 129 | spec = torch.squeeze(spec, 0) 130 | torch.save(spec, spec_filename) 131 | 132 | spk = filename.split("/")[-2] 133 | spk = torch.LongTensor([self.spk_map[spk]]) 134 | 135 | c = torch.load(filename + ".soft.pt").squeeze(0) 136 | 137 | c = torch.repeat_interleave(c, repeats=2, dim=1) 138 | 139 | f0 = np.load(filename + ".f0.npy") 140 | f0 = torch.FloatTensor(f0) 141 | lmin = min(c.size(-1), spec.size(-1), f0.shape[0]) 142 | assert abs(c.size(-1) - spec.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape) 143 | assert abs(f0.shape[0] - spec.shape[-1]) < 4, (c.size(-1), spec.size(-1), f0.shape) 144 | spec, c, f0 = spec[:, :lmin], c[:, :lmin], f0[:lmin] 145 | audio_norm = audio_norm[:, :lmin * self.hop_length] 146 | 147 | return c, f0, spec, audio_norm, spk 148 | 149 | def __getitem__(self, index): 150 | return self.get_audio(self.audiopaths[index][0]) 151 | 152 | def __len__(self): 153 | return len(self.audiopaths) 154 | 155 | -------------------------------------------------------------------------------- /dataset_raw/wav_structure.txt: -------------------------------------------------------------------------------- 1 | 数据集准备 2 | 3 | raw 4 | ├───speaker0 5 | │ ├───xxx1-xxx1.wav 6 | │ ├───... 7 | │ └───Lxx-0xx8.wav 8 | └───speaker1 9 | ├───xx2-0xxx2.wav 10 | ├───... 11 | └───xxx7-xxx007.wav 12 | 13 | 此外还需要编辑config.json 14 | 15 | "n_speakers": 10 16 | 17 | "spk":{ 18 | "speaker0": 0, 19 | "speaker1": 1, 20 | } 21 | -------------------------------------------------------------------------------- /filelists/test.txt: -------------------------------------------------------------------------------- 1 | ./dataset/32k/yunhao/001829.wav 2 | ./dataset/32k/yunhao/001827.wav 3 | ./dataset/32k/jishuang/000104.wav 4 | ./dataset/32k/nen/kne110_005.wav 5 | ./dataset/32k/nen/kne110_004.wav 6 | ./dataset/32k/jishuang/000223.wav 7 | ./dataset/32k/yunhao/001828.wav 8 | -------------------------------------------------------------------------------- /filelists/train.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/filelists/train.txt -------------------------------------------------------------------------------- /filelists/val.txt: -------------------------------------------------------------------------------- 1 | ./dataset/32k/nen/kne110_005.wav 2 | ./dataset/32k/yunhao/001827.wav 3 | ./dataset/32k/jishuang/000104.wav 4 | ./dataset/32k/jishuang/000223.wav 5 | ./dataset/32k/nen/kne110_004.wav 6 | ./dataset/32k/yunhao/001828.wav 7 | -------------------------------------------------------------------------------- /flask_api.py: -------------------------------------------------------------------------------- 1 | import io 2 | import logging 3 | 4 | import soundfile 5 | import torch 6 | import torchaudio 7 | from flask import Flask, request, send_file 8 | from flask_cors import CORS 9 | 10 | from inference.infer_tool import Svc, RealTimeVC 11 | 12 | app = Flask(__name__) 13 | 14 | CORS(app) 15 | 16 | logging.getLogger('numba').setLevel(logging.WARNING) 17 | 18 | 19 | @app.route("/voiceChangeModel", methods=["POST"]) 20 | def voice_change_model(): 21 | request_form = request.form 22 | wave_file = request.files.get("sample", None) 23 | # 变调信息 24 | f_pitch_change = float(request_form.get("fPitchChange", 0)) 25 | # DAW所需的采样率 26 | daw_sample = int(float(request_form.get("sampleRate", 0))) 27 | speaker_id = int(float(request_form.get("sSpeakId", 0))) 28 | # http获得wav文件并转换 29 | input_wav_path = io.BytesIO(wave_file.read()) 30 | 31 | # 模型推理 32 | if raw_infer: 33 | out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) 34 | tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample) 35 | else: 36 | out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path) 37 | tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample) 38 | # 返回音频 39 | out_wav_path = io.BytesIO() 40 | soundfile.write(out_wav_path, tar_audio.cpu().numpy(), daw_sample, format="wav") 41 | out_wav_path.seek(0) 42 | return send_file(out_wav_path, download_name="temp.wav", as_attachment=True) 43 | 44 | 45 | if __name__ == '__main__': 46 | # 启用则为直接切片合成，False为交叉淡化方式 47 | # vst插件调整0.3-0.5s切片时间可以降低延迟，直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音 48 | # 自行选择能接受的方法，或将vst最大切片时间调整为1s，此处设为Ture，延迟大音质稳定一些 49 | raw_infer = True 50 | # 每个模型和config是唯一对应的 51 | model_name = "logs/32k/G_174000-Copy1.pth" 52 | config_name = "configs/config.json" 53 | svc_model = Svc(model_name, config_name) 54 | svc = RealTimeVC() 55 | # 此处与vst插件对应，不建议更改 56 | app.run(port=6842, host="0.0.0.0", debug=False, threaded=False) 57 | -------------------------------------------------------------------------------- /hubert/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/hubert/__init__.py -------------------------------------------------------------------------------- /hubert/hubert_model.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import random 3 | from typing import Optional, Tuple 4 | 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as t_func 8 | from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present 9 | 10 | 11 | class Hubert(nn.Module): 12 | def __init__(self, num_label_embeddings: int = 100, mask: bool = True): 13 | super().__init__() 14 | self._mask = mask 15 | self.feature_extractor = FeatureExtractor() 16 | self.feature_projection = FeatureProjection() 17 | self.positional_embedding = PositionalConvEmbedding() 18 | self.norm = nn.LayerNorm(768) 19 | self.dropout = nn.Dropout(0.1) 20 | self.encoder = TransformerEncoder( 21 | nn.TransformerEncoderLayer( 22 | 768, 12, 3072, activation="gelu", batch_first=True 23 | ), 24 | 12, 25 | ) 26 | self.proj = nn.Linear(768, 256) 27 | 28 | self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_()) 29 | self.label_embedding = nn.Embedding(num_label_embeddings, 256) 30 | 31 | def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: 32 | mask = None 33 | if self.training and self._mask: 34 | mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2) 35 | x[mask] = self.masked_spec_embed.to(x.dtype) 36 | return x, mask 37 | 38 | def encode( 39 | self, x: torch.Tensor, layer: Optional[int] = None 40 | ) -> Tuple[torch.Tensor, torch.Tensor]: 41 | x = self.feature_extractor(x) 42 | x = self.feature_projection(x.transpose(1, 2)) 43 | x, mask = self.mask(x) 44 | x = x + self.positional_embedding(x) 45 | x = self.dropout(self.norm(x)) 46 | x = self.encoder(x, output_layer=layer) 47 | return x, mask 48 | 49 | def logits(self, x: torch.Tensor) -> torch.Tensor: 50 | logits = torch.cosine_similarity( 51 | x.unsqueeze(2), 52 | self.label_embedding.weight.unsqueeze(0).unsqueeze(0), 53 | dim=-1, 54 | ) 55 | return logits / 0.1 56 | 57 | def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: 58 | x, mask = self.encode(x) 59 | x = self.proj(x) 60 | logits = self.logits(x) 61 | return logits, mask 62 | 63 | 64 | class HubertSoft(Hubert): 65 | def __init__(self): 66 | super().__init__() 67 | 68 | @torch.inference_mode() 69 | def units(self, wav: torch.Tensor) -> torch.Tensor: 70 | wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2)) 71 | x, _ = self.encode(wav) 72 | return self.proj(x) 73 | 74 | 75 | class FeatureExtractor(nn.Module): 76 | def __init__(self): 77 | super().__init__() 78 | self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False) 79 | self.norm0 = nn.GroupNorm(512, 512) 80 | self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False) 81 | self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False) 82 | self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False) 83 | self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False) 84 | self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False) 85 | self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False) 86 | 87 | def forward(self, x: torch.Tensor) -> torch.Tensor: 88 | x = t_func.gelu(self.norm0(self.conv0(x))) 89 | x = t_func.gelu(self.conv1(x)) 90 | x = t_func.gelu(self.conv2(x)) 91 | x = t_func.gelu(self.conv3(x)) 92 | x = t_func.gelu(self.conv4(x)) 93 | x = t_func.gelu(self.conv5(x)) 94 | x = t_func.gelu(self.conv6(x)) 95 | return x 96 | 97 | 98 | class FeatureProjection(nn.Module): 99 | def __init__(self): 100 | super().__init__() 101 | self.norm = nn.LayerNorm(512) 102 | self.projection = nn.Linear(512, 768) 103 | self.dropout = nn.Dropout(0.1) 104 | 105 | def forward(self, x: torch.Tensor) -> torch.Tensor: 106 | x = self.norm(x) 107 | x = self.projection(x) 108 | x = self.dropout(x) 109 | return x 110 | 111 | 112 | class PositionalConvEmbedding(nn.Module): 113 | def __init__(self): 114 | super().__init__() 115 | self.conv = nn.Conv1d( 116 | 768, 117 | 768, 118 | kernel_size=128, 119 | padding=128 // 2, 120 | groups=16, 121 | ) 122 | self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2) 123 | 124 | def forward(self, x: torch.Tensor) -> torch.Tensor: 125 | x = self.conv(x.transpose(1, 2)) 126 | x = t_func.gelu(x[:, :, :-1]) 127 | return x.transpose(1, 2) 128 | 129 | 130 | class TransformerEncoder(nn.Module): 131 | def __init__( 132 | self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int 133 | ) -> None: 134 | super(TransformerEncoder, self).__init__() 135 | self.layers = nn.ModuleList( 136 | [copy.deepcopy(encoder_layer) for _ in range(num_layers)] 137 | ) 138 | self.num_layers = num_layers 139 | 140 | def forward( 141 | self, 142 | src: torch.Tensor, 143 | mask: torch.Tensor = None, 144 | src_key_padding_mask: torch.Tensor = None, 145 | output_layer: Optional[int] = None, 146 | ) -> torch.Tensor: 147 | output = src 148 | for layer in self.layers[:output_layer]: 149 | output = layer( 150 | output, src_mask=mask, src_key_padding_mask=src_key_padding_mask 151 | ) 152 | return output 153 | 154 | 155 | def _compute_mask( 156 | shape: Tuple[int, int], 157 | mask_prob: float, 158 | mask_length: int, 159 | device: torch.device, 160 | min_masks: int = 0, 161 | ) -> torch.Tensor: 162 | batch_size, sequence_length = shape 163 | 164 | if mask_length < 1: 165 | raise ValueError("`mask_length` has to be bigger than 0.") 166 | 167 | if mask_length > sequence_length: 168 | raise ValueError( 169 | f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`" 170 | ) 171 | 172 | # compute number of masked spans in batch 173 | num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random()) 174 | num_masked_spans = max(num_masked_spans, min_masks) 175 | 176 | # make sure num masked indices <= sequence_length 177 | if num_masked_spans * mask_length > sequence_length: 178 | num_masked_spans = sequence_length // mask_length 179 | 180 | # SpecAugment mask to fill 181 | mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool) 182 | 183 | # uniform distribution to sample from, make sure that offset samples are < sequence_length 184 | uniform_dist = torch.ones( 185 | (batch_size, sequence_length - (mask_length - 1)), device=device 186 | ) 187 | 188 | # get random indices to mask 189 | mask_indices = torch.multinomial(uniform_dist, num_masked_spans) 190 | 191 | # expand masked indices to masked spans 192 | mask_indices = ( 193 | mask_indices.unsqueeze(dim=-1) 194 | .expand((batch_size, num_masked_spans, mask_length)) 195 | .reshape(batch_size, num_masked_spans * mask_length) 196 | ) 197 | offsets = ( 198 | torch.arange(mask_length, device=device)[None, None, :] 199 | .expand((batch_size, num_masked_spans, mask_length)) 200 | .reshape(batch_size, num_masked_spans * mask_length) 201 | ) 202 | mask_idxs = mask_indices + offsets 203 | 204 | # scatter indices to mask 205 | mask = mask.scatter(1, mask_idxs, True) 206 | 207 | return mask 208 | 209 | 210 | def hubert_soft( 211 | path: str, 212 | ) -> HubertSoft: 213 | r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`. 214 | Args: 215 | path (str): path of a pretrained model 216 | """ 217 | hubert = HubertSoft() 218 | checkpoint = torch.load(path) 219 | consume_prefix_in_state_dict_if_present(checkpoint, "module.") 220 | hubert.load_state_dict(checkpoint) 221 | hubert.eval() 222 | return hubert 223 | -------------------------------------------------------------------------------- /hubert/hubert_model_onnx.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import random 3 | from typing import Optional, Tuple 4 | 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.functional as t_func 8 | from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present 9 | 10 | 11 | class Hubert(nn.Module): 12 | def __init__(self, num_label_embeddings: int = 100, mask: bool = True): 13 | super().__init__() 14 | self._mask = mask 15 | self.feature_extractor = FeatureExtractor() 16 | self.feature_projection = FeatureProjection() 17 | self.positional_embedding = PositionalConvEmbedding() 18 | self.norm = nn.LayerNorm(768) 19 | self.dropout = nn.Dropout(0.1) 20 | self.encoder = TransformerEncoder( 21 | nn.TransformerEncoderLayer( 22 | 768, 12, 3072, activation="gelu", batch_first=True 23 | ), 24 | 12, 25 | ) 26 | self.proj = nn.Linear(768, 256) 27 | 28 | self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_()) 29 | self.label_embedding = nn.Embedding(num_label_embeddings, 256) 30 | 31 | def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: 32 | mask = None 33 | if self.training and self._mask: 34 | mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2) 35 | x[mask] = self.masked_spec_embed.to(x.dtype) 36 | return x, mask 37 | 38 | def encode( 39 | self, x: torch.Tensor, layer: Optional[int] = None 40 | ) -> Tuple[torch.Tensor, torch.Tensor]: 41 | x = self.feature_extractor(x) 42 | x = self.feature_projection(x.transpose(1, 2)) 43 | x, mask = self.mask(x) 44 | x = x + self.positional_embedding(x) 45 | x = self.dropout(self.norm(x)) 46 | x = self.encoder(x, output_layer=layer) 47 | return x, mask 48 | 49 | def logits(self, x: torch.Tensor) -> torch.Tensor: 50 | logits = torch.cosine_similarity( 51 | x.unsqueeze(2), 52 | self.label_embedding.weight.unsqueeze(0).unsqueeze(0), 53 | dim=-1, 54 | ) 55 | return logits / 0.1 56 | 57 | 58 | class HubertSoft(Hubert): 59 | def __init__(self): 60 | super().__init__() 61 | 62 | def units(self, wav: torch.Tensor) -> torch.Tensor: 63 | wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2)) 64 | x, _ = self.encode(wav) 65 | return self.proj(x) 66 | 67 | def forward(self, x): 68 | return self.units(x) 69 | 70 | class FeatureExtractor(nn.Module): 71 | def __init__(self): 72 | super().__init__() 73 | self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False) 74 | self.norm0 = nn.GroupNorm(512, 512) 75 | self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False) 76 | self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False) 77 | self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False) 78 | self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False) 79 | self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False) 80 | self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False) 81 | 82 | def forward(self, x: torch.Tensor) -> torch.Tensor: 83 | x = t_func.gelu(self.norm0(self.conv0(x))) 84 | x = t_func.gelu(self.conv1(x)) 85 | x = t_func.gelu(self.conv2(x)) 86 | x = t_func.gelu(self.conv3(x)) 87 | x = t_func.gelu(self.conv4(x)) 88 | x = t_func.gelu(self.conv5(x)) 89 | x = t_func.gelu(self.conv6(x)) 90 | return x 91 | 92 | 93 | class FeatureProjection(nn.Module): 94 | def __init__(self): 95 | super().__init__() 96 | self.norm = nn.LayerNorm(512) 97 | self.projection = nn.Linear(512, 768) 98 | self.dropout = nn.Dropout(0.1) 99 | 100 | def forward(self, x: torch.Tensor) -> torch.Tensor: 101 | x = self.norm(x) 102 | x = self.projection(x) 103 | x = self.dropout(x) 104 | return x 105 | 106 | 107 | class PositionalConvEmbedding(nn.Module): 108 | def __init__(self): 109 | super().__init__() 110 | self.conv = nn.Conv1d( 111 | 768, 112 | 768, 113 | kernel_size=128, 114 | padding=128 // 2, 115 | groups=16, 116 | ) 117 | self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2) 118 | 119 | def forward(self, x: torch.Tensor) -> torch.Tensor: 120 | x = self.conv(x.transpose(1, 2)) 121 | x = t_func.gelu(x[:, :, :-1]) 122 | return x.transpose(1, 2) 123 | 124 | 125 | class TransformerEncoder(nn.Module): 126 | def __init__( 127 | self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int 128 | ) -> None: 129 | super(TransformerEncoder, self).__init__() 130 | self.layers = nn.ModuleList( 131 | [copy.deepcopy(encoder_layer) for _ in range(num_layers)] 132 | ) 133 | self.num_layers = num_layers 134 | 135 | def forward( 136 | self, 137 | src: torch.Tensor, 138 | mask: torch.Tensor = None, 139 | src_key_padding_mask: torch.Tensor = None, 140 | output_layer: Optional[int] = None, 141 | ) -> torch.Tensor: 142 | output = src 143 | for layer in self.layers[:output_layer]: 144 | output = layer( 145 | output, src_mask=mask, src_key_padding_mask=src_key_padding_mask 146 | ) 147 | return output 148 | 149 | 150 | def _compute_mask( 151 | shape: Tuple[int, int], 152 | mask_prob: float, 153 | mask_length: int, 154 | device: torch.device, 155 | min_masks: int = 0, 156 | ) -> torch.Tensor: 157 | batch_size, sequence_length = shape 158 | 159 | if mask_length < 1: 160 | raise ValueError("`mask_length` has to be bigger than 0.") 161 | 162 | if mask_length > sequence_length: 163 | raise ValueError( 164 | f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`" 165 | ) 166 | 167 | # compute number of masked spans in batch 168 | num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random()) 169 | num_masked_spans = max(num_masked_spans, min_masks) 170 | 171 | # make sure num masked indices <= sequence_length 172 | if num_masked_spans * mask_length > sequence_length: 173 | num_masked_spans = sequence_length // mask_length 174 | 175 | # SpecAugment mask to fill 176 | mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool) 177 | 178 | # uniform distribution to sample from, make sure that offset samples are < sequence_length 179 | uniform_dist = torch.ones( 180 | (batch_size, sequence_length - (mask_length - 1)), device=device 181 | ) 182 | 183 | # get random indices to mask 184 | mask_indices = torch.multinomial(uniform_dist, num_masked_spans) 185 | 186 | # expand masked indices to masked spans 187 | mask_indices = ( 188 | mask_indices.unsqueeze(dim=-1) 189 | .expand((batch_size, num_masked_spans, mask_length)) 190 | .reshape(batch_size, num_masked_spans * mask_length) 191 | ) 192 | offsets = ( 193 | torch.arange(mask_length, device=device)[None, None, :] 194 | .expand((batch_size, num_masked_spans, mask_length)) 195 | .reshape(batch_size, num_masked_spans * mask_length) 196 | ) 197 | mask_idxs = mask_indices + offsets 198 | 199 | # scatter indices to mask 200 | mask = mask.scatter(1, mask_idxs, True) 201 | 202 | return mask 203 | 204 | 205 | def hubert_soft( 206 | path: str, 207 | ) -> HubertSoft: 208 | r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`. 209 | Args: 210 | path (str): path of a pretrained model 211 | """ 212 | hubert = HubertSoft() 213 | checkpoint = torch.load(path) 214 | consume_prefix_in_state_dict_if_present(checkpoint, "module.") 215 | hubert.load_state_dict(checkpoint) 216 | hubert.eval() 217 | return hubert 218 | -------------------------------------------------------------------------------- /hubert/put_hubert_ckpt_here: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/hubert/put_hubert_ckpt_here -------------------------------------------------------------------------------- /inference/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/inference/__init__.py -------------------------------------------------------------------------------- /inference/infer_tool.py: -------------------------------------------------------------------------------- 1 | import hashlib 2 | import json 3 | import logging 4 | import os 5 | import time 6 | from pathlib import Path 7 | 8 | import librosa 9 | import maad 10 | import numpy as np 11 | # import onnxruntime 12 | import parselmouth 13 | import soundfile 14 | import torch 15 | import torchaudio 16 | 17 | from hubert import hubert_model 18 | import utils 19 | from models import SynthesizerTrn 20 | 21 | logging.getLogger('matplotlib').setLevel(logging.WARNING) 22 | 23 | 24 | def read_temp(file_name): 25 | if not os.path.exists(file_name): 26 | with open(file_name, "w") as f: 27 | f.write(json.dumps({"info": "temp_dict"})) 28 | return {} 29 | else: 30 | try: 31 | with open(file_name, "r") as f: 32 | data = f.read() 33 | data_dict = json.loads(data) 34 | if os.path.getsize(file_name) > 50 * 1024 * 1024: 35 | f_name = file_name.replace("\\", "/").split("/")[-1] 36 | print(f"clean {f_name}") 37 | for wav_hash in list(data_dict.keys()): 38 | if int(time.time()) - int(data_dict[wav_hash]["time"]) > 14 * 24 * 3600: 39 | del data_dict[wav_hash] 40 | except Exception as e: 41 | print(e) 42 | print(f"{file_name} error,auto rebuild file") 43 | data_dict = {"info": "temp_dict"} 44 | return data_dict 45 | 46 | 47 | def write_temp(file_name, data): 48 | with open(file_name, "w") as f: 49 | f.write(json.dumps(data)) 50 | 51 | 52 | def timeit(func): 53 | def run(*args, **kwargs): 54 | t = time.time() 55 | res = func(*args, **kwargs) 56 | print('executing \'%s\' costed %.3fs' % (func.__name__, time.time() - t)) 57 | return res 58 | 59 | return run 60 | 61 | 62 | def format_wav(audio_path): 63 | if Path(audio_path).suffix == '.wav': 64 | return 65 | raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True, sr=None) 66 | soundfile.write(Path(audio_path).with_suffix(".wav"), raw_audio, raw_sample_rate) 67 | 68 | 69 | def get_end_file(dir_path, end): 70 | file_lists = [] 71 | for root, dirs, files in os.walk(dir_path): 72 | files = [f for f in files if f[0] != '.'] 73 | dirs[:] = [d for d in dirs if d[0] != '.'] 74 | for f_file in files: 75 | if f_file.endswith(end): 76 | file_lists.append(os.path.join(root, f_file).replace("\\", "/")) 77 | return file_lists 78 | 79 | 80 | def get_md5(content): 81 | return hashlib.new("md5", content).hexdigest() 82 | 83 | 84 | def resize2d_f0(x, target_len): 85 | source = np.array(x) 86 | source[source < 0.001] = np.nan 87 | target = np.interp(np.arange(0, len(source) * target_len, len(source)) / target_len, np.arange(0, len(source)), 88 | source) 89 | res = np.nan_to_num(target) 90 | return res 91 | 92 | def get_f0(x, p_len,f0_up_key=0): 93 | 94 | time_step = 160 / 16000 * 1000 95 | f0_min = 50 96 | f0_max = 1100 97 | f0_mel_min = 1127 * np.log(1 + f0_min / 700) 98 | f0_mel_max = 1127 * np.log(1 + f0_max / 700) 99 | 100 | f0 = parselmouth.Sound(x, 16000).to_pitch_ac( 101 | time_step=time_step / 1000, voicing_threshold=0.6, 102 | pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency'] 103 | if len(f0) > p_len: 104 | f0 = f0[:p_len] 105 | pad_size=(p_len - len(f0) + 1) // 2 106 | if(pad_size>0 or p_len - len(f0) - pad_size>0): 107 | f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant') 108 | 109 | f0 *= pow(2, f0_up_key / 12) 110 | f0_mel = 1127 * np.log(1 + f0 / 700) 111 | f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1 112 | f0_mel[f0_mel <= 1] = 1 113 | f0_mel[f0_mel > 255] = 255 114 | f0_coarse = np.rint(f0_mel).astype(np.int) 115 | return f0_coarse, f0 116 | 117 | def clean_pitch(input_pitch): 118 | num_nan = np.sum(input_pitch == 1) 119 | if num_nan / len(input_pitch) > 0.9: 120 | input_pitch[input_pitch != 1] = 1 121 | return input_pitch 122 | 123 | 124 | def plt_pitch(input_pitch): 125 | input_pitch = input_pitch.astype(float) 126 | input_pitch[input_pitch == 1] = np.nan 127 | return input_pitch 128 | 129 | 130 | def f0_to_pitch(ff): 131 | f0_pitch = 69 + 12 * np.log2(ff / 440) 132 | return f0_pitch 133 | 134 | 135 | def fill_a_to_b(a, b): 136 | if len(a) < len(b): 137 | for _ in range(0, len(b) - len(a)): 138 | a.append(a[0]) 139 | 140 | 141 | def mkdir(paths: list): 142 | for path in paths: 143 | if not os.path.exists(path): 144 | os.mkdir(path) 145 | 146 | 147 | class Svc(object): 148 | def __init__(self, net_g_path, config_path, hubert_path="hubert/hubert-soft-0d54a1f4.pt", 149 | onnx=False): 150 | self.onnx = onnx 151 | self.net_g_path = net_g_path 152 | self.hubert_path = hubert_path 153 | self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu") 154 | self.net_g_ms = None 155 | self.hps_ms = utils.get_hparams_from_file(config_path) 156 | self.target_sample = self.hps_ms.data.sampling_rate 157 | self.hop_size = self.hps_ms.data.hop_length 158 | self.speakers = {} 159 | for spk, sid in self.hps_ms.spk.items(): 160 | self.speakers[sid] = spk 161 | self.spk2id = self.hps_ms.spk 162 | # 加载hubert 163 | self.hubert_soft = hubert_model.hubert_soft(hubert_path) 164 | if torch.cuda.is_available(): 165 | self.hubert_soft = self.hubert_soft.cuda() 166 | self.load_model() 167 | 168 | def load_model(self): 169 | # 获取模型配置 170 | if self.onnx: 171 | raise NotImplementedError 172 | # self.net_g_ms = SynthesizerTrnForONNX( 173 | # 178, 174 | # self.hps_ms.data.filter_length // 2 + 1, 175 | # self.hps_ms.train.segment_size // self.hps_ms.data.hop_length, 176 | # n_speakers=self.hps_ms.data.n_speakers, 177 | # **self.hps_ms.model) 178 | # _ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None) 179 | else: 180 | self.net_g_ms = SynthesizerTrn( 181 | self.hps_ms.data.filter_length // 2 + 1, 182 | self.hps_ms.train.segment_size // self.hps_ms.data.hop_length, 183 | **self.hps_ms.model) 184 | _ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None) 185 | if "half" in self.net_g_path and torch.cuda.is_available(): 186 | _ = self.net_g_ms.half().eval().to(self.dev) 187 | else: 188 | _ = self.net_g_ms.eval().to(self.dev) 189 | 190 | def get_units(self, source, sr): 191 | 192 | source = source.unsqueeze(0).to(self.dev) 193 | with torch.inference_mode(): 194 | start = time.time() 195 | units = self.hubert_soft.units(source) 196 | use_time = time.time() - start 197 | print("hubert use time:{}".format(use_time)) 198 | return units 199 | 200 | 201 | def get_unit_pitch(self, in_path, tran): 202 | source, sr = torchaudio.load(in_path) 203 | source = torchaudio.functional.resample(source, sr, 16000) 204 | if len(source.shape) == 2 and source.shape[1] >= 2: 205 | source = torch.mean(source, dim=0).unsqueeze(0) 206 | soft = self.get_units(source, sr).squeeze(0).cpu().numpy() 207 | f0_coarse, f0 = get_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran) 208 | return soft, f0 209 | 210 | def infer(self, speaker_id, tran, raw_path): 211 | if type(speaker_id) == str: 212 | speaker_id = self.spk2id[speaker_id] 213 | sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0) 214 | soft, pitch = self.get_unit_pitch(raw_path, tran) 215 | f0 = torch.FloatTensor(clean_pitch(pitch)).unsqueeze(0).to(self.dev) 216 | if "half" in self.net_g_path and torch.cuda.is_available(): 217 | stn_tst = torch.HalfTensor(soft) 218 | else: 219 | stn_tst = torch.FloatTensor(soft) 220 | with torch.no_grad(): 221 | x_tst = stn_tst.unsqueeze(0).to(self.dev) 222 | start = time.time() 223 | x_tst = torch.repeat_interleave(x_tst, repeats=2, dim=1).transpose(1, 2) 224 | audio = self.net_g_ms.infer(x_tst, f0=f0, g=sid)[0,0].data.float() 225 | use_time = time.time() - start 226 | print("vits use time:{}".format(use_time)) 227 | return audio, audio.shape[-1] 228 | 229 | 230 | # class SvcONNXInferModel(object): 231 | # def __init__(self, hubert_onnx, vits_onnx, config_path): 232 | # self.config_path = config_path 233 | # self.vits_onnx = vits_onnx 234 | # self.hubert_onnx = hubert_onnx 235 | # self.hubert_onnx_session = onnxruntime.InferenceSession(hubert_onnx, providers=['CUDAExecutionProvider', ]) 236 | # self.inspect_onnx(self.hubert_onnx_session) 237 | # self.vits_onnx_session = onnxruntime.InferenceSession(vits_onnx, providers=['CUDAExecutionProvider', ]) 238 | # self.inspect_onnx(self.vits_onnx_session) 239 | # self.hps_ms = utils.get_hparams_from_file(self.config_path) 240 | # self.target_sample = self.hps_ms.data.sampling_rate 241 | # self.feature_input = FeatureInput(self.hps_ms.data.sampling_rate, self.hps_ms.data.hop_length) 242 | # 243 | # @staticmethod 244 | # def inspect_onnx(session): 245 | # for i in session.get_inputs(): 246 | # print("name:{}\tshape:{}\tdtype:{}".format(i.name, i.shape, i.type)) 247 | # for i in session.get_outputs(): 248 | # print("name:{}\tshape:{}\tdtype:{}".format(i.name, i.shape, i.type)) 249 | # 250 | # def infer(self, speaker_id, tran, raw_path): 251 | # sid = np.array([int(speaker_id)], dtype=np.int64) 252 | # soft, pitch = self.get_unit_pitch(raw_path, tran) 253 | # pitch = np.expand_dims(pitch, axis=0).astype(np.int64) 254 | # stn_tst = soft 255 | # x_tst = np.expand_dims(stn_tst, axis=0) 256 | # x_tst_lengths = np.array([stn_tst.shape[0]], dtype=np.int64) 257 | # # 使用ONNX Runtime进行推理 258 | # start = time.time() 259 | # audio = self.vits_onnx_session.run(output_names=["audio"], 260 | # input_feed={ 261 | # "hidden_unit": x_tst, 262 | # "lengths": x_tst_lengths, 263 | # "pitch": pitch, 264 | # "sid": sid, 265 | # })[0][0, 0] 266 | # use_time = time.time() - start 267 | # print("vits_onnx_session.run time:{}".format(use_time)) 268 | # audio = torch.from_numpy(audio) 269 | # return audio, audio.shape[-1] 270 | # 271 | # def get_units(self, source, sr): 272 | # source = torchaudio.functional.resample(source, sr, 16000) 273 | # if len(source.shape) == 2 and source.shape[1] >= 2: 274 | # source = torch.mean(source, dim=0).unsqueeze(0) 275 | # source = source.unsqueeze(0) 276 | # # 使用ONNX Runtime进行推理 277 | # start = time.time() 278 | # units = self.hubert_onnx_session.run(output_names=["embed"], 279 | # input_feed={"source": source.numpy()})[0] 280 | # use_time = time.time() - start 281 | # print("hubert_onnx_session.run time:{}".format(use_time)) 282 | # return units 283 | # 284 | # def transcribe(self, source, sr, length, transform): 285 | # feature_pit = self.feature_input.compute_f0(source, sr) 286 | # feature_pit = feature_pit * 2 ** (transform / 12) 287 | # feature_pit = resize2d_f0(feature_pit, length) 288 | # coarse_pit = self.feature_input.coarse_f0(feature_pit) 289 | # return coarse_pit 290 | # 291 | # def get_unit_pitch(self, in_path, tran): 292 | # source, sr = torchaudio.load(in_path) 293 | # soft = self.get_units(source, sr).squeeze(0) 294 | # input_pitch = self.transcribe(source.numpy()[0], sr, soft.shape[0], tran) 295 | # return soft, input_pitch 296 | 297 | 298 | class RealTimeVC: 299 | def __init__(self): 300 | self.last_chunk = None 301 | self.last_o = None 302 | self.chunk_len = 16000 # 区块长度 303 | self.pre_len = 3840 # 交叉淡化长度，640的倍数 304 | 305 | """输入输出都是1维numpy 音频波形数组""" 306 | 307 | def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path): 308 | audio, sr = torchaudio.load(input_wav_path) 309 | audio = audio.cpu().numpy()[0] 310 | temp_wav = io.BytesIO() 311 | if self.last_chunk is None: 312 | input_wav_path.seek(0) 313 | audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path) 314 | audio = audio.cpu().numpy() 315 | self.last_chunk = audio[-self.pre_len:] 316 | self.last_o = audio 317 | return audio[-self.chunk_len:] 318 | else: 319 | audio = np.concatenate([self.last_chunk, audio]) 320 | soundfile.write(temp_wav, audio, sr, format="wav") 321 | temp_wav.seek(0) 322 | audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav) 323 | audio = audio.cpu().numpy() 324 | ret = maad.util.crossfade(self.last_o, audio, self.pre_len) 325 | self.last_chunk = audio[-self.pre_len:] 326 | self.last_o = audio 327 | return ret[self.chunk_len:2 * self.chunk_len] 328 | -------------------------------------------------------------------------------- /inference/infer_tool_grad.py: -------------------------------------------------------------------------------- 1 | import hashlib 2 | import json 3 | import logging 4 | import os 5 | import time 6 | from pathlib import Path 7 | import io 8 | import librosa 9 | import maad 10 | import numpy as np 11 | from inference import slicer 12 | import parselmouth 13 | import soundfile 14 | import torch 15 | import torchaudio 16 | 17 | from hubert import hubert_model 18 | import utils 19 | from models import SynthesizerTrn 20 | logging.getLogger('numba').setLevel(logging.WARNING) 21 | logging.getLogger('matplotlib').setLevel(logging.WARNING) 22 | 23 | def resize2d_f0(x, target_len): 24 | source = np.array(x) 25 | source[source < 0.001] = np.nan 26 | target = np.interp(np.arange(0, len(source) * target_len, len(source)) / target_len, np.arange(0, len(source)), 27 | source) 28 | res = np.nan_to_num(target) 29 | return res 30 | 31 | def get_f0(x, p_len,f0_up_key=0): 32 | 33 | time_step = 160 / 16000 * 1000 34 | f0_min = 50 35 | f0_max = 1100 36 | f0_mel_min = 1127 * np.log(1 + f0_min / 700) 37 | f0_mel_max = 1127 * np.log(1 + f0_max / 700) 38 | 39 | f0 = parselmouth.Sound(x, 16000).to_pitch_ac( 40 | time_step=time_step / 1000, voicing_threshold=0.6, 41 | pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency'] 42 | 43 | pad_size=(p_len - len(f0) + 1) // 2 44 | if(pad_size>0 or p_len - len(f0) - pad_size>0): 45 | f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant') 46 | 47 | f0 *= pow(2, f0_up_key / 12) 48 | f0_mel = 1127 * np.log(1 + f0 / 700) 49 | f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1 50 | f0_mel[f0_mel <= 1] = 1 51 | f0_mel[f0_mel > 255] = 255 52 | f0_coarse = np.rint(f0_mel).astype(np.int) 53 | return f0_coarse, f0 54 | 55 | def clean_pitch(input_pitch): 56 | num_nan = np.sum(input_pitch == 1) 57 | if num_nan / len(input_pitch) > 0.9: 58 | input_pitch[input_pitch != 1] = 1 59 | return input_pitch 60 | 61 | 62 | def plt_pitch(input_pitch): 63 | input_pitch = input_pitch.astype(float) 64 | input_pitch[input_pitch == 1] = np.nan 65 | return input_pitch 66 | 67 | 68 | def f0_to_pitch(ff): 69 | f0_pitch = 69 + 12 * np.log2(ff / 440) 70 | return f0_pitch 71 | 72 | 73 | def fill_a_to_b(a, b): 74 | if len(a) < len(b): 75 | for _ in range(0, len(b) - len(a)): 76 | a.append(a[0]) 77 | 78 | 79 | def mkdir(paths: list): 80 | for path in paths: 81 | if not os.path.exists(path): 82 | os.mkdir(path) 83 | 84 | 85 | class VitsSvc(object): 86 | def __init__(self): 87 | self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 88 | self.SVCVITS = None 89 | self.hps = None 90 | self.speakers = None 91 | self.hubert_soft = utils.get_hubert_model() 92 | 93 | def set_device(self, device): 94 | self.device = torch.device(device) 95 | self.hubert_soft.to(self.device) 96 | if self.SVCVITS != None: 97 | self.SVCVITS.to(self.device) 98 | 99 | def loadCheckpoint(self, path): 100 | self.hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json") 101 | self.SVCVITS = SynthesizerTrn( 102 | self.hps.data.filter_length // 2 + 1, 103 | self.hps.train.segment_size // self.hps.data.hop_length, 104 | **self.hps.model) 105 | _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", self.SVCVITS, None) 106 | _ = self.SVCVITS.eval().to(self.device) 107 | self.speakers = self.hps.spk 108 | 109 | def get_units(self, source, sr): 110 | source = source.unsqueeze(0).to(self.device) 111 | with torch.inference_mode(): 112 | units = self.hubert_soft.units(source) 113 | return units 114 | 115 | 116 | def get_unit_pitch(self, in_path, tran): 117 | source, sr = torchaudio.load(in_path) 118 | source = torchaudio.functional.resample(source, sr, 16000) 119 | if len(source.shape) == 2 and source.shape[1] >= 2: 120 | source = torch.mean(source, dim=0).unsqueeze(0) 121 | soft = self.get_units(source, sr).squeeze(0).cpu().numpy() 122 | f0_coarse, f0 = get_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran) 123 | return soft, f0 124 | 125 | def infer(self, speaker_id, tran, raw_path): 126 | speaker_id = self.speakers[speaker_id] 127 | sid = torch.LongTensor([int(speaker_id)]).to(self.device).unsqueeze(0) 128 | soft, pitch = self.get_unit_pitch(raw_path, tran) 129 | f0 = torch.FloatTensor(clean_pitch(pitch)).unsqueeze(0).to(self.device) 130 | stn_tst = torch.FloatTensor(soft) 131 | with torch.no_grad(): 132 | x_tst = stn_tst.unsqueeze(0).to(self.device) 133 | x_tst = torch.repeat_interleave(x_tst, repeats=2, dim=1).transpose(1, 2) 134 | audio = self.SVCVITS.infer(x_tst, f0=f0, g=sid)[0,0].data.float() 135 | return audio, audio.shape[-1] 136 | 137 | def inference(self,srcaudio,chara,tran,slice_db): 138 | sampling_rate, audio = srcaudio 139 | audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32) 140 | if len(audio.shape) > 1: 141 | audio = librosa.to_mono(audio.transpose(1, 0)) 142 | if sampling_rate != 16000: 143 | audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000) 144 | soundfile.write("tmpwav.wav", audio, 16000, format="wav") 145 | chunks = slicer.cut("tmpwav.wav", db_thresh=slice_db) 146 | audio_data, audio_sr = slicer.chunks2audio("tmpwav.wav", chunks) 147 | audio = [] 148 | for (slice_tag, data) in audio_data: 149 | length = int(np.ceil(len(data) / audio_sr * self.hps.data.sampling_rate)) 150 | raw_path = io.BytesIO() 151 | soundfile.write(raw_path, data, audio_sr, format="wav") 152 | raw_path.seek(0) 153 | if slice_tag: 154 | _audio = np.zeros(length) 155 | else: 156 | out_audio, out_sr = self.infer(chara, tran, raw_path) 157 | _audio = out_audio.cpu().numpy() 158 | audio.extend(list(_audio)) 159 | audio = (np.array(audio) * 32768.0).astype('int16') 160 | return (self.hps.data.sampling_rate,audio) 161 | -------------------------------------------------------------------------------- /inference/slicer.py: -------------------------------------------------------------------------------- 1 | import librosa 2 | import torch 3 | import torchaudio 4 | 5 | 6 | class Slicer: 7 | def __init__(self, 8 | sr: int, 9 | threshold: float = -40., 10 | min_length: int = 5000, 11 | min_interval: int = 300, 12 | hop_size: int = 20, 13 | max_sil_kept: int = 5000): 14 | if not min_length >= min_interval >= hop_size: 15 | raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size') 16 | if not max_sil_kept >= hop_size: 17 | raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size') 18 | min_interval = sr * min_interval / 1000 19 | self.threshold = 10 ** (threshold / 20.) 20 | self.hop_size = round(sr * hop_size / 1000) 21 | self.win_size = min(round(min_interval), 4 * self.hop_size) 22 | self.min_length = round(sr * min_length / 1000 / self.hop_size) 23 | self.min_interval = round(min_interval / self.hop_size) 24 | self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size) 25 | 26 | def _apply_slice(self, waveform, begin, end): 27 | if len(waveform.shape) > 1: 28 | return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)] 29 | else: 30 | return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)] 31 | 32 | # @timeit 33 | def slice(self, waveform): 34 | if len(waveform.shape) > 1: 35 | samples = librosa.to_mono(waveform) 36 | else: 37 | samples = waveform 38 | if samples.shape[0] <= self.min_length: 39 | return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}} 40 | rms_list = librosa.feature.rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0) 41 | sil_tags = [] 42 | silence_start = None 43 | clip_start = 0 44 | for i, rms in enumerate(rms_list): 45 | # Keep looping while frame is silent. 46 | if rms < self.threshold: 47 | # Record start of silent frames. 48 | if silence_start is None: 49 | silence_start = i 50 | continue 51 | # Keep looping while frame is not silent and silence start has not been recorded. 52 | if silence_start is None: 53 | continue 54 | # Clear recorded silence start if interval is not enough or clip is too short 55 | is_leading_silence = silence_start == 0 and i > self.max_sil_kept 56 | need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length 57 | if not is_leading_silence and not need_slice_middle: 58 | silence_start = None 59 | continue 60 | # Need slicing. Record the range of silent frames to be removed. 61 | if i - silence_start <= self.max_sil_kept: 62 | pos = rms_list[silence_start: i + 1].argmin() + silence_start 63 | if silence_start == 0: 64 | sil_tags.append((0, pos)) 65 | else: 66 | sil_tags.append((pos, pos)) 67 | clip_start = pos 68 | elif i - silence_start <= self.max_sil_kept * 2: 69 | pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin() 70 | pos += i - self.max_sil_kept 71 | pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start 72 | pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept 73 | if silence_start == 0: 74 | sil_tags.append((0, pos_r)) 75 | clip_start = pos_r 76 | else: 77 | sil_tags.append((min(pos_l, pos), max(pos_r, pos))) 78 | clip_start = max(pos_r, pos) 79 | else: 80 | pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start 81 | pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept 82 | if silence_start == 0: 83 | sil_tags.append((0, pos_r)) 84 | else: 85 | sil_tags.append((pos_l, pos_r)) 86 | clip_start = pos_r 87 | silence_start = None 88 | # Deal with trailing silence. 89 | total_frames = rms_list.shape[0] 90 | if silence_start is not None and total_frames - silence_start >= self.min_interval: 91 | silence_end = min(total_frames, silence_start + self.max_sil_kept) 92 | pos = rms_list[silence_start: silence_end + 1].argmin() + silence_start 93 | sil_tags.append((pos, total_frames + 1)) 94 | # Apply and return slices. 95 | if len(sil_tags) == 0: 96 | return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}} 97 | else: 98 | chunks = [] 99 | # 第一段静音并非从头开始，补上有声片段 100 | if sil_tags[0][0]: 101 | chunks.append( 102 | {"slice": False, "split_time": f"0,{min(waveform.shape[0], sil_tags[0][0] * self.hop_size)}"}) 103 | for i in range(0, len(sil_tags)): 104 | # 标识有声片段（跳过第一段） 105 | if i: 106 | chunks.append({"slice": False, 107 | "split_time": f"{sil_tags[i - 1][1] * self.hop_size},{min(waveform.shape[0], sil_tags[i][0] * self.hop_size)}"}) 108 | # 标识所有静音片段 109 | chunks.append({"slice": True, 110 | "split_time": f"{sil_tags[i][0] * self.hop_size},{min(waveform.shape[0], sil_tags[i][1] * self.hop_size)}"}) 111 | # 最后一段静音并非结尾，补上结尾片段 112 | if sil_tags[-1][1] * self.hop_size < len(waveform): 113 | chunks.append({"slice": False, "split_time": f"{sil_tags[-1][1] * self.hop_size},{len(waveform)}"}) 114 | chunk_dict = {} 115 | for i in range(len(chunks)): 116 | chunk_dict[str(i)] = chunks[i] 117 | return chunk_dict 118 | 119 | 120 | def cut(audio_path, db_thresh=-30, min_len=5000): 121 | audio, sr = librosa.load(audio_path, sr=None) 122 | slicer = Slicer( 123 | sr=sr, 124 | threshold=db_thresh, 125 | min_length=min_len 126 | ) 127 | chunks = slicer.slice(audio) 128 | return chunks 129 | 130 | 131 | def chunks2audio(audio_path, chunks): 132 | chunks = dict(chunks) 133 | audio, sr = torchaudio.load(audio_path) 134 | if len(audio.shape) == 2 and audio.shape[1] >= 2: 135 | audio = torch.mean(audio, dim=0).unsqueeze(0) 136 | audio = audio.cpu().numpy()[0] 137 | result = [] 138 | for k, v in chunks.items(): 139 | tag = v["split_time"].split(",") 140 | if tag[0] != tag[1]: 141 | result.append((v["slice"], audio[int(tag[0]):int(tag[1])])) 142 | return result, sr 143 | -------------------------------------------------------------------------------- /inference_main.py: -------------------------------------------------------------------------------- 1 | import io 2 | import logging 3 | import time 4 | from pathlib import Path 5 | 6 | import librosa 7 | import numpy as np 8 | import soundfile 9 | 10 | from inference import infer_tool 11 | from inference import slicer 12 | from inference.infer_tool import Svc 13 | 14 | logging.getLogger('numba').setLevel(logging.WARNING) 15 | chunks_dict = infer_tool.read_temp("inference/chunks_temp.json") 16 | 17 | model_path = "logs/32k/G_174000-Copy1.pth" 18 | config_path = "configs/config.json" 19 | svc_model = Svc(model_path, config_path) 20 | infer_tool.mkdir(["raw", "results"]) 21 | 22 | # 支持多个wav文件，放在raw文件夹下 23 | clean_names = ["君の知らない物語-src"] 24 | trans = [-5] # 音高调整，支持正负（半音） 25 | spk_list = ['yunhao'] # 每次同时合成多语者音色 26 | slice_db = -40 # 默认-40，嘈杂的音频可以-30，干声保留呼吸可以-50 27 | wav_format = 'flac' # 音频输出格式 28 | 29 | infer_tool.fill_a_to_b(trans, clean_names) 30 | for clean_name, tran in zip(clean_names, trans): 31 | raw_audio_path = f"raw/{clean_name}" 32 | if "." not in raw_audio_path: 33 | raw_audio_path += ".wav" 34 | infer_tool.format_wav(raw_audio_path) 35 | wav_path = Path(raw_audio_path).with_suffix('.wav') 36 | chunks = slicer.cut(wav_path, db_thresh=slice_db) 37 | audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks) 38 | 39 | for spk in spk_list: 40 | audio = [] 41 | for (slice_tag, data) in audio_data: 42 | print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======') 43 | length = int(np.ceil(len(data) / audio_sr * svc_model.target_sample)) 44 | raw_path = io.BytesIO() 45 | soundfile.write(raw_path, data, audio_sr, format="wav") 46 | raw_path.seek(0) 47 | if slice_tag: 48 | print('jump empty segment') 49 | _audio = np.zeros(length) 50 | else: 51 | out_audio, out_sr = svc_model.infer(spk, tran, raw_path) 52 | _audio = out_audio.cpu().numpy() 53 | audio.extend(list(_audio)) 54 | 55 | res_path = f'./results/{clean_name}_{tran}key_{spk}.{wav_format}' 56 | soundfile.write(res_path, audio, svc_model.target_sample, format=wav_format) 57 | -------------------------------------------------------------------------------- /logs/32k/put_pretrained_model_here: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/logs/32k/put_pretrained_model_here -------------------------------------------------------------------------------- /losses.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch.nn import functional as F 3 | 4 | import commons 5 | 6 | 7 | def feature_loss(fmap_r, fmap_g): 8 | loss = 0 9 | for dr, dg in zip(fmap_r, fmap_g): 10 | for rl, gl in zip(dr, dg): 11 | rl = rl.float().detach() 12 | gl = gl.float() 13 | loss += torch.mean(torch.abs(rl - gl)) 14 | 15 | return loss * 2 16 | 17 | 18 | def discriminator_loss(disc_real_outputs, disc_generated_outputs): 19 | loss = 0 20 | r_losses = [] 21 | g_losses = [] 22 | for dr, dg in zip(disc_real_outputs, disc_generated_outputs): 23 | dr = dr.float() 24 | dg = dg.float() 25 | r_loss = torch.mean((1-dr)**2) 26 | g_loss = torch.mean(dg**2) 27 | loss += (r_loss + g_loss) 28 | r_losses.append(r_loss.item()) 29 | g_losses.append(g_loss.item()) 30 | 31 | return loss, r_losses, g_losses 32 | 33 | 34 | def generator_loss(disc_outputs): 35 | loss = 0 36 | gen_losses = [] 37 | for dg in disc_outputs: 38 | dg = dg.float() 39 | l = torch.mean((1-dg)**2) 40 | gen_losses.append(l) 41 | loss += l 42 | 43 | return loss, gen_losses 44 | 45 | 46 | def kl_loss(z_p, logs_q, m_p, logs_p, z_mask): 47 | """ 48 | z_p, logs_q: [b, h, t_t] 49 | m_p, logs_p: [b, h, t_t] 50 | """ 51 | z_p = z_p.float() 52 | logs_q = logs_q.float() 53 | m_p = m_p.float() 54 | logs_p = logs_p.float() 55 | z_mask = z_mask.float() 56 | #print(logs_p) 57 | kl = logs_p - logs_q - 0.5 58 | kl += 0.5 * ((z_p - m_p)**2) * torch.exp(-2. * logs_p) 59 | kl = torch.sum(kl * z_mask) 60 | l = kl / torch.sum(z_mask) 61 | return l 62 | -------------------------------------------------------------------------------- /mel_processing.py: -------------------------------------------------------------------------------- 1 | import math 2 | import os 3 | import random 4 | import torch 5 | from torch import nn 6 | import torch.nn.functional as F 7 | import torch.utils.data 8 | import numpy as np 9 | import librosa 10 | import librosa.util as librosa_util 11 | from librosa.util import normalize, pad_center, tiny 12 | from scipy.signal import get_window 13 | from scipy.io.wavfile import read 14 | from librosa.filters import mel as librosa_mel_fn 15 | 16 | MAX_WAV_VALUE = 32768.0 17 | 18 | 19 | def dynamic_range_compression_torch(x, C=1, clip_val=1e-5): 20 | """ 21 | PARAMS 22 | ------ 23 | C: compression factor 24 | """ 25 | return torch.log(torch.clamp(x, min=clip_val) * C) 26 | 27 | 28 | def dynamic_range_decompression_torch(x, C=1): 29 | """ 30 | PARAMS 31 | ------ 32 | C: compression factor used to compress 33 | """ 34 | return torch.exp(x) / C 35 | 36 | 37 | def spectral_normalize_torch(magnitudes): 38 | output = dynamic_range_compression_torch(magnitudes) 39 | return output 40 | 41 | 42 | def spectral_de_normalize_torch(magnitudes): 43 | output = dynamic_range_decompression_torch(magnitudes) 44 | return output 45 | 46 | 47 | mel_basis = {} 48 | hann_window = {} 49 | 50 | 51 | def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False): 52 | if torch.min(y) < -1.: 53 | print('min value is ', torch.min(y)) 54 | if torch.max(y) > 1.: 55 | print('max value is ', torch.max(y)) 56 | 57 | global hann_window 58 | dtype_device = str(y.dtype) + '_' + str(y.device) 59 | wnsize_dtype_device = str(win_size) + '_' + dtype_device 60 | if wnsize_dtype_device not in hann_window: 61 | hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device) 62 | 63 | y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect') 64 | y = y.squeeze(1) 65 | 66 | spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device], 67 | center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False) 68 | 69 | spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) 70 | return spec 71 | 72 | 73 | def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax): 74 | global mel_basis 75 | dtype_device = str(spec.dtype) + '_' + str(spec.device) 76 | fmax_dtype_device = str(fmax) + '_' + dtype_device 77 | if fmax_dtype_device not in mel_basis: 78 | mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) 79 | mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device) 80 | spec = torch.matmul(mel_basis[fmax_dtype_device], spec) 81 | spec = spectral_normalize_torch(spec) 82 | return spec 83 | 84 | 85 | def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False): 86 | if torch.min(y) < -1.: 87 | print('min value is ', torch.min(y)) 88 | if torch.max(y) > 1.: 89 | print('max value is ', torch.max(y)) 90 | 91 | global mel_basis, hann_window 92 | dtype_device = str(y.dtype) + '_' + str(y.device) 93 | fmax_dtype_device = str(fmax) + '_' + dtype_device 94 | wnsize_dtype_device = str(win_size) + '_' + dtype_device 95 | if fmax_dtype_device not in mel_basis: 96 | mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax) 97 | mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device) 98 | if wnsize_dtype_device not in hann_window: 99 | hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device) 100 | 101 | y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect') 102 | y = y.squeeze(1) 103 | 104 | spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device], 105 | center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False) 106 | 107 | spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) 108 | 109 | spec = torch.matmul(mel_basis[fmax_dtype_device], spec) 110 | spec = spectral_normalize_torch(spec) 111 | 112 | return spec 113 | -------------------------------------------------------------------------------- /model_onnx.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | from torch import nn 5 | from torch.nn import functional as F 6 | 7 | import attentions 8 | import commons 9 | import modules 10 | 11 | from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d 12 | from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm 13 | from commons import init_weights, get_padding 14 | from vdecoder.hifigan.models import Generator 15 | from utils import f0_to_coarse 16 | 17 | class ResidualCouplingBlock(nn.Module): 18 | def __init__(self, 19 | channels, 20 | hidden_channels, 21 | kernel_size, 22 | dilation_rate, 23 | n_layers, 24 | n_flows=4, 25 | gin_channels=0): 26 | super().__init__() 27 | self.channels = channels 28 | self.hidden_channels = hidden_channels 29 | self.kernel_size = kernel_size 30 | self.dilation_rate = dilation_rate 31 | self.n_layers = n_layers 32 | self.n_flows = n_flows 33 | self.gin_channels = gin_channels 34 | 35 | self.flows = nn.ModuleList() 36 | for i in range(n_flows): 37 | self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True)) 38 | self.flows.append(modules.Flip()) 39 | 40 | def forward(self, x, x_mask, g=None, reverse=False): 41 | if not reverse: 42 | for flow in self.flows: 43 | x, _ = flow(x, x_mask, g=g, reverse=reverse) 44 | else: 45 | for flow in reversed(self.flows): 46 | x = flow(x, x_mask, g=g, reverse=reverse) 47 | return x 48 | 49 | 50 | class Encoder(nn.Module): 51 | def __init__(self, 52 | in_channels, 53 | out_channels, 54 | hidden_channels, 55 | kernel_size, 56 | dilation_rate, 57 | n_layers, 58 | gin_channels=0): 59 | super().__init__() 60 | self.in_channels = in_channels 61 | self.out_channels = out_channels 62 | self.hidden_channels = hidden_channels 63 | self.kernel_size = kernel_size 64 | self.dilation_rate = dilation_rate 65 | self.n_layers = n_layers 66 | self.gin_channels = gin_channels 67 | 68 | self.pre = nn.Conv1d(in_channels, hidden_channels, 1) 69 | self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels) 70 | self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) 71 | 72 | def forward(self, x, x_lengths, g=None): 73 | # print(x.shape,x_lengths.shape) 74 | x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) 75 | x = self.pre(x) * x_mask 76 | x = self.enc(x, x_mask, g=g) 77 | stats = self.proj(x) * x_mask 78 | m, logs = torch.split(stats, self.out_channels, dim=1) 79 | z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask 80 | return z, m, logs, x_mask 81 | 82 | 83 | class TextEncoder(nn.Module): 84 | def __init__(self, 85 | in_channels, 86 | out_channels, 87 | hidden_channels, 88 | kernel_size, 89 | dilation_rate, 90 | n_layers, 91 | gin_channels=0, 92 | filter_channels=None, 93 | n_heads=None, 94 | p_dropout=None): 95 | super().__init__() 96 | self.in_channels = in_channels 97 | self.out_channels = out_channels 98 | self.hidden_channels = hidden_channels 99 | self.kernel_size = kernel_size 100 | self.dilation_rate = dilation_rate 101 | self.n_layers = n_layers 102 | self.gin_channels = gin_channels 103 | self.pre = nn.Conv1d(in_channels, hidden_channels, 1) 104 | self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) 105 | self.f0_emb = nn.Embedding(256, hidden_channels) 106 | 107 | self.enc_ = attentions.Encoder( 108 | hidden_channels, 109 | filter_channels, 110 | n_heads, 111 | n_layers, 112 | kernel_size, 113 | p_dropout) 114 | 115 | def forward(self, x, x_lengths, f0=None): 116 | x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) 117 | x = self.pre(x) * x_mask 118 | x = x + self.f0_emb(f0.long()).transpose(1,2) 119 | x = self.enc_(x * x_mask, x_mask) 120 | stats = self.proj(x) * x_mask 121 | m, logs = torch.split(stats, self.out_channels, dim=1) 122 | z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask 123 | 124 | return z, m, logs, x_mask 125 | 126 | 127 | 128 | class DiscriminatorP(torch.nn.Module): 129 | def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): 130 | super(DiscriminatorP, self).__init__() 131 | self.period = period 132 | self.use_spectral_norm = use_spectral_norm 133 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 134 | self.convs = nn.ModuleList([ 135 | norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 136 | norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 137 | norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 138 | norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 139 | norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))), 140 | ]) 141 | self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0))) 142 | 143 | def forward(self, x): 144 | fmap = [] 145 | 146 | # 1d to 2d 147 | b, c, t = x.shape 148 | if t % self.period != 0: # pad first 149 | n_pad = self.period - (t % self.period) 150 | x = F.pad(x, (0, n_pad), "reflect") 151 | t = t + n_pad 152 | x = x.view(b, c, t // self.period, self.period) 153 | 154 | for l in self.convs: 155 | x = l(x) 156 | x = F.leaky_relu(x, modules.LRELU_SLOPE) 157 | fmap.append(x) 158 | x = self.conv_post(x) 159 | fmap.append(x) 160 | x = torch.flatten(x, 1, -1) 161 | 162 | return x, fmap 163 | 164 | 165 | class DiscriminatorS(torch.nn.Module): 166 | def __init__(self, use_spectral_norm=False): 167 | super(DiscriminatorS, self).__init__() 168 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 169 | self.convs = nn.ModuleList([ 170 | norm_f(Conv1d(1, 16, 15, 1, padding=7)), 171 | norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), 172 | norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), 173 | norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), 174 | norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), 175 | norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), 176 | ]) 177 | self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) 178 | 179 | def forward(self, x): 180 | fmap = [] 181 | 182 | for l in self.convs: 183 | x = l(x) 184 | x = F.leaky_relu(x, modules.LRELU_SLOPE) 185 | fmap.append(x) 186 | x = self.conv_post(x) 187 | fmap.append(x) 188 | x = torch.flatten(x, 1, -1) 189 | 190 | return x, fmap 191 | 192 | 193 | class MultiPeriodDiscriminator(torch.nn.Module): 194 | def __init__(self, use_spectral_norm=False): 195 | super(MultiPeriodDiscriminator, self).__init__() 196 | periods = [2,3,5,7,11] 197 | 198 | discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)] 199 | discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods] 200 | self.discriminators = nn.ModuleList(discs) 201 | 202 | def forward(self, y, y_hat): 203 | y_d_rs = [] 204 | y_d_gs = [] 205 | fmap_rs = [] 206 | fmap_gs = [] 207 | for i, d in enumerate(self.discriminators): 208 | y_d_r, fmap_r = d(y) 209 | y_d_g, fmap_g = d(y_hat) 210 | y_d_rs.append(y_d_r) 211 | y_d_gs.append(y_d_g) 212 | fmap_rs.append(fmap_r) 213 | fmap_gs.append(fmap_g) 214 | 215 | return y_d_rs, y_d_gs, fmap_rs, fmap_gs 216 | 217 | 218 | class SpeakerEncoder(torch.nn.Module): 219 | def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256): 220 | super(SpeakerEncoder, self).__init__() 221 | self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True) 222 | self.linear = nn.Linear(model_hidden_size, model_embedding_size) 223 | self.relu = nn.ReLU() 224 | 225 | def forward(self, mels): 226 | self.lstm.flatten_parameters() 227 | _, (hidden, _) = self.lstm(mels) 228 | embeds_raw = self.relu(self.linear(hidden[-1])) 229 | return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True) 230 | 231 | def compute_partial_slices(self, total_frames, partial_frames, partial_hop): 232 | mel_slices = [] 233 | for i in range(0, total_frames-partial_frames, partial_hop): 234 | mel_range = torch.arange(i, i+partial_frames) 235 | mel_slices.append(mel_range) 236 | 237 | return mel_slices 238 | 239 | def embed_utterance(self, mel, partial_frames=128, partial_hop=64): 240 | mel_len = mel.size(1) 241 | last_mel = mel[:,-partial_frames:] 242 | 243 | if mel_len > partial_frames: 244 | mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop) 245 | mels = list(mel[:,s] for s in mel_slices) 246 | mels.append(last_mel) 247 | mels = torch.stack(tuple(mels), 0).squeeze(1) 248 | 249 | with torch.no_grad(): 250 | partial_embeds = self(mels) 251 | embed = torch.mean(partial_embeds, axis=0).unsqueeze(0) 252 | #embed = embed / torch.linalg.norm(embed, 2) 253 | else: 254 | with torch.no_grad(): 255 | embed = self(last_mel) 256 | 257 | return embed 258 | 259 | 260 | class SynthesizerTrn(nn.Module): 261 | """ 262 | Synthesizer for Training 263 | """ 264 | 265 | def __init__(self, 266 | spec_channels, 267 | segment_size, 268 | inter_channels, 269 | hidden_channels, 270 | filter_channels, 271 | n_heads, 272 | n_layers, 273 | kernel_size, 274 | p_dropout, 275 | resblock, 276 | resblock_kernel_sizes, 277 | resblock_dilation_sizes, 278 | upsample_rates, 279 | upsample_initial_channel, 280 | upsample_kernel_sizes, 281 | gin_channels, 282 | ssl_dim, 283 | n_speakers, 284 | **kwargs): 285 | 286 | super().__init__() 287 | self.spec_channels = spec_channels 288 | self.inter_channels = inter_channels 289 | self.hidden_channels = hidden_channels 290 | self.filter_channels = filter_channels 291 | self.n_heads = n_heads 292 | self.n_layers = n_layers 293 | self.kernel_size = kernel_size 294 | self.p_dropout = p_dropout 295 | self.resblock = resblock 296 | self.resblock_kernel_sizes = resblock_kernel_sizes 297 | self.resblock_dilation_sizes = resblock_dilation_sizes 298 | self.upsample_rates = upsample_rates 299 | self.upsample_initial_channel = upsample_initial_channel 300 | self.upsample_kernel_sizes = upsample_kernel_sizes 301 | self.segment_size = segment_size 302 | self.gin_channels = gin_channels 303 | self.ssl_dim = ssl_dim 304 | self.emb_g = nn.Embedding(n_speakers, gin_channels) 305 | 306 | self.enc_p_ = TextEncoder(ssl_dim, inter_channels, hidden_channels, 5, 1, 16,0, filter_channels, n_heads, p_dropout) 307 | hps = { 308 | "sampling_rate": 32000, 309 | "inter_channels": 192, 310 | "resblock": "1", 311 | "resblock_kernel_sizes": [3, 7, 11], 312 | "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]], 313 | "upsample_rates": [10, 8, 2, 2], 314 | "upsample_initial_channel": 512, 315 | "upsample_kernel_sizes": [16, 16, 4, 4], 316 | "gin_channels": 256, 317 | } 318 | self.dec = Generator(h=hps) 319 | self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels) 320 | self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels) 321 | 322 | def forward(self, c, c_lengths, f0, g=None): 323 | g = self.emb_g(g.unsqueeze(0)).transpose(1,2) 324 | z_p, m_p, logs_p, c_mask = self.enc_p_(c.transpose(1,2), c_lengths, f0=f0_to_coarse(f0)) 325 | z = self.flow(z_p, c_mask, g=g, reverse=True) 326 | o = self.dec(z * c_mask, g=g, f0=f0.float()) 327 | return o 328 | 329 | -------------------------------------------------------------------------------- /model_onnx_48k.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | from torch import nn 5 | from torch.nn import functional as F 6 | 7 | import attentions 8 | import commons 9 | import modules 10 | 11 | from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d 12 | from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm 13 | from commons import init_weights, get_padding 14 | from vdecoder.hifigan.models import Generator 15 | from utils import f0_to_coarse 16 | 17 | class ResidualCouplingBlock(nn.Module): 18 | def __init__(self, 19 | channels, 20 | hidden_channels, 21 | kernel_size, 22 | dilation_rate, 23 | n_layers, 24 | n_flows=4, 25 | gin_channels=0): 26 | super().__init__() 27 | self.channels = channels 28 | self.hidden_channels = hidden_channels 29 | self.kernel_size = kernel_size 30 | self.dilation_rate = dilation_rate 31 | self.n_layers = n_layers 32 | self.n_flows = n_flows 33 | self.gin_channels = gin_channels 34 | 35 | self.flows = nn.ModuleList() 36 | for i in range(n_flows): 37 | self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True)) 38 | self.flows.append(modules.Flip()) 39 | 40 | def forward(self, x, x_mask, g=None, reverse=False): 41 | if not reverse: 42 | for flow in self.flows: 43 | x, _ = flow(x, x_mask, g=g, reverse=reverse) 44 | else: 45 | for flow in reversed(self.flows): 46 | x = flow(x, x_mask, g=g, reverse=reverse) 47 | return x 48 | 49 | 50 | class Encoder(nn.Module): 51 | def __init__(self, 52 | in_channels, 53 | out_channels, 54 | hidden_channels, 55 | kernel_size, 56 | dilation_rate, 57 | n_layers, 58 | gin_channels=0): 59 | super().__init__() 60 | self.in_channels = in_channels 61 | self.out_channels = out_channels 62 | self.hidden_channels = hidden_channels 63 | self.kernel_size = kernel_size 64 | self.dilation_rate = dilation_rate 65 | self.n_layers = n_layers 66 | self.gin_channels = gin_channels 67 | 68 | self.pre = nn.Conv1d(in_channels, hidden_channels, 1) 69 | self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels) 70 | self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) 71 | 72 | def forward(self, x, x_lengths, g=None): 73 | # print(x.shape,x_lengths.shape) 74 | x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) 75 | x = self.pre(x) * x_mask 76 | x = self.enc(x, x_mask, g=g) 77 | stats = self.proj(x) * x_mask 78 | m, logs = torch.split(stats, self.out_channels, dim=1) 79 | z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask 80 | return z, m, logs, x_mask 81 | 82 | 83 | class TextEncoder(nn.Module): 84 | def __init__(self, 85 | in_channels, 86 | out_channels, 87 | hidden_channels, 88 | kernel_size, 89 | dilation_rate, 90 | n_layers, 91 | gin_channels=0, 92 | filter_channels=None, 93 | n_heads=None, 94 | p_dropout=None): 95 | super().__init__() 96 | self.in_channels = in_channels 97 | self.out_channels = out_channels 98 | self.hidden_channels = hidden_channels 99 | self.kernel_size = kernel_size 100 | self.dilation_rate = dilation_rate 101 | self.n_layers = n_layers 102 | self.gin_channels = gin_channels 103 | self.pre = nn.Conv1d(in_channels, hidden_channels, 1) 104 | self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) 105 | self.f0_emb = nn.Embedding(256, hidden_channels) 106 | 107 | self.enc_ = attentions.Encoder( 108 | hidden_channels, 109 | filter_channels, 110 | n_heads, 111 | n_layers, 112 | kernel_size, 113 | p_dropout) 114 | 115 | def forward(self, x, x_lengths, f0=None): 116 | x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) 117 | x = self.pre(x) * x_mask 118 | x = x + self.f0_emb(f0.long()).transpose(1,2) 119 | x = self.enc_(x * x_mask, x_mask) 120 | stats = self.proj(x) * x_mask 121 | m, logs = torch.split(stats, self.out_channels, dim=1) 122 | z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask 123 | 124 | return z, m, logs, x_mask 125 | 126 | 127 | 128 | class DiscriminatorP(torch.nn.Module): 129 | def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): 130 | super(DiscriminatorP, self).__init__() 131 | self.period = period 132 | self.use_spectral_norm = use_spectral_norm 133 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 134 | self.convs = nn.ModuleList([ 135 | norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 136 | norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 137 | norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 138 | norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 139 | norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))), 140 | ]) 141 | self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0))) 142 | 143 | def forward(self, x): 144 | fmap = [] 145 | 146 | # 1d to 2d 147 | b, c, t = x.shape 148 | if t % self.period != 0: # pad first 149 | n_pad = self.period - (t % self.period) 150 | x = F.pad(x, (0, n_pad), "reflect") 151 | t = t + n_pad 152 | x = x.view(b, c, t // self.period, self.period) 153 | 154 | for l in self.convs: 155 | x = l(x) 156 | x = F.leaky_relu(x, modules.LRELU_SLOPE) 157 | fmap.append(x) 158 | x = self.conv_post(x) 159 | fmap.append(x) 160 | x = torch.flatten(x, 1, -1) 161 | 162 | return x, fmap 163 | 164 | 165 | class DiscriminatorS(torch.nn.Module): 166 | def __init__(self, use_spectral_norm=False): 167 | super(DiscriminatorS, self).__init__() 168 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 169 | self.convs = nn.ModuleList([ 170 | norm_f(Conv1d(1, 16, 15, 1, padding=7)), 171 | norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), 172 | norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), 173 | norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), 174 | norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), 175 | norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), 176 | ]) 177 | self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) 178 | 179 | def forward(self, x): 180 | fmap = [] 181 | 182 | for l in self.convs: 183 | x = l(x) 184 | x = F.leaky_relu(x, modules.LRELU_SLOPE) 185 | fmap.append(x) 186 | x = self.conv_post(x) 187 | fmap.append(x) 188 | x = torch.flatten(x, 1, -1) 189 | 190 | return x, fmap 191 | 192 | 193 | class MultiPeriodDiscriminator(torch.nn.Module): 194 | def __init__(self, use_spectral_norm=False): 195 | super(MultiPeriodDiscriminator, self).__init__() 196 | periods = [2,3,5,7,11] 197 | 198 | discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)] 199 | discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods] 200 | self.discriminators = nn.ModuleList(discs) 201 | 202 | def forward(self, y, y_hat): 203 | y_d_rs = [] 204 | y_d_gs = [] 205 | fmap_rs = [] 206 | fmap_gs = [] 207 | for i, d in enumerate(self.discriminators): 208 | y_d_r, fmap_r = d(y) 209 | y_d_g, fmap_g = d(y_hat) 210 | y_d_rs.append(y_d_r) 211 | y_d_gs.append(y_d_g) 212 | fmap_rs.append(fmap_r) 213 | fmap_gs.append(fmap_g) 214 | 215 | return y_d_rs, y_d_gs, fmap_rs, fmap_gs 216 | 217 | 218 | class SpeakerEncoder(torch.nn.Module): 219 | def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256): 220 | super(SpeakerEncoder, self).__init__() 221 | self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True) 222 | self.linear = nn.Linear(model_hidden_size, model_embedding_size) 223 | self.relu = nn.ReLU() 224 | 225 | def forward(self, mels): 226 | self.lstm.flatten_parameters() 227 | _, (hidden, _) = self.lstm(mels) 228 | embeds_raw = self.relu(self.linear(hidden[-1])) 229 | return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True) 230 | 231 | def compute_partial_slices(self, total_frames, partial_frames, partial_hop): 232 | mel_slices = [] 233 | for i in range(0, total_frames-partial_frames, partial_hop): 234 | mel_range = torch.arange(i, i+partial_frames) 235 | mel_slices.append(mel_range) 236 | 237 | return mel_slices 238 | 239 | def embed_utterance(self, mel, partial_frames=128, partial_hop=64): 240 | mel_len = mel.size(1) 241 | last_mel = mel[:,-partial_frames:] 242 | 243 | if mel_len > partial_frames: 244 | mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop) 245 | mels = list(mel[:,s] for s in mel_slices) 246 | mels.append(last_mel) 247 | mels = torch.stack(tuple(mels), 0).squeeze(1) 248 | 249 | with torch.no_grad(): 250 | partial_embeds = self(mels) 251 | embed = torch.mean(partial_embeds, axis=0).unsqueeze(0) 252 | #embed = embed / torch.linalg.norm(embed, 2) 253 | else: 254 | with torch.no_grad(): 255 | embed = self(last_mel) 256 | 257 | return embed 258 | 259 | 260 | class SynthesizerTrn(nn.Module): 261 | """ 262 | Synthesizer for Training 263 | """ 264 | 265 | def __init__(self, 266 | spec_channels, 267 | segment_size, 268 | inter_channels, 269 | hidden_channels, 270 | filter_channels, 271 | n_heads, 272 | n_layers, 273 | kernel_size, 274 | p_dropout, 275 | resblock, 276 | resblock_kernel_sizes, 277 | resblock_dilation_sizes, 278 | upsample_rates, 279 | upsample_initial_channel, 280 | upsample_kernel_sizes, 281 | gin_channels, 282 | ssl_dim, 283 | n_speakers, 284 | **kwargs): 285 | 286 | super().__init__() 287 | self.spec_channels = spec_channels 288 | self.inter_channels = inter_channels 289 | self.hidden_channels = hidden_channels 290 | self.filter_channels = filter_channels 291 | self.n_heads = n_heads 292 | self.n_layers = n_layers 293 | self.kernel_size = kernel_size 294 | self.p_dropout = p_dropout 295 | self.resblock = resblock 296 | self.resblock_kernel_sizes = resblock_kernel_sizes 297 | self.resblock_dilation_sizes = resblock_dilation_sizes 298 | self.upsample_rates = upsample_rates 299 | self.upsample_initial_channel = upsample_initial_channel 300 | self.upsample_kernel_sizes = upsample_kernel_sizes 301 | self.segment_size = segment_size 302 | self.gin_channels = gin_channels 303 | self.ssl_dim = ssl_dim 304 | self.emb_g = nn.Embedding(n_speakers, gin_channels) 305 | 306 | self.enc_p_ = TextEncoder(ssl_dim, inter_channels, hidden_channels, 5, 1, 16,0, filter_channels, n_heads, p_dropout) 307 | hps = { 308 | "sampling_rate": 48000, 309 | "inter_channels": 192, 310 | "resblock": "1", 311 | "resblock_kernel_sizes": [3, 7, 11], 312 | "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]], 313 | "upsample_rates": [10, 8, 2, 2], 314 | "upsample_initial_channel": 512, 315 | "upsample_kernel_sizes": [16, 16, 4, 4], 316 | "gin_channels": 256, 317 | } 318 | self.dec = Generator(h=hps) 319 | self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels) 320 | self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels) 321 | 322 | def forward(self, c, c_lengths, f0, g=None): 323 | g = self.emb_g(g.unsqueeze(0)).transpose(1,2) 324 | z_p, m_p, logs_p, c_mask = self.enc_p_(c.transpose(1,2), c_lengths, f0=f0_to_coarse(f0)) 325 | z = self.flow(z_p, c_mask, g=g, reverse=True) 326 | o = self.dec(z * c_mask, g=g, f0=f0.float()) 327 | return o 328 | 329 | -------------------------------------------------------------------------------- /models.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | from torch import nn 5 | from torch.nn import functional as F 6 | 7 | import attentions 8 | import commons 9 | import modules 10 | 11 | from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d 12 | from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm 13 | from commons import init_weights, get_padding 14 | from vdecoder.hifigan.models import Generator 15 | from utils import f0_to_coarse 16 | 17 | class ResidualCouplingBlock(nn.Module): 18 | def __init__(self, 19 | channels, 20 | hidden_channels, 21 | kernel_size, 22 | dilation_rate, 23 | n_layers, 24 | n_flows=4, 25 | gin_channels=0): 26 | super().__init__() 27 | self.channels = channels 28 | self.hidden_channels = hidden_channels 29 | self.kernel_size = kernel_size 30 | self.dilation_rate = dilation_rate 31 | self.n_layers = n_layers 32 | self.n_flows = n_flows 33 | self.gin_channels = gin_channels 34 | 35 | self.flows = nn.ModuleList() 36 | for i in range(n_flows): 37 | self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True)) 38 | self.flows.append(modules.Flip()) 39 | 40 | def forward(self, x, x_mask, g=None, reverse=False): 41 | if not reverse: 42 | for flow in self.flows: 43 | x, _ = flow(x, x_mask, g=g, reverse=reverse) 44 | else: 45 | for flow in reversed(self.flows): 46 | x = flow(x, x_mask, g=g, reverse=reverse) 47 | return x 48 | 49 | 50 | class Encoder(nn.Module): 51 | def __init__(self, 52 | in_channels, 53 | out_channels, 54 | hidden_channels, 55 | kernel_size, 56 | dilation_rate, 57 | n_layers, 58 | gin_channels=0): 59 | super().__init__() 60 | self.in_channels = in_channels 61 | self.out_channels = out_channels 62 | self.hidden_channels = hidden_channels 63 | self.kernel_size = kernel_size 64 | self.dilation_rate = dilation_rate 65 | self.n_layers = n_layers 66 | self.gin_channels = gin_channels 67 | 68 | self.pre = nn.Conv1d(in_channels, hidden_channels, 1) 69 | self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels) 70 | self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) 71 | 72 | def forward(self, x, x_lengths, g=None): 73 | # print(x.shape,x_lengths.shape) 74 | x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) 75 | x = self.pre(x) * x_mask 76 | x = self.enc(x, x_mask, g=g) 77 | stats = self.proj(x) * x_mask 78 | m, logs = torch.split(stats, self.out_channels, dim=1) 79 | z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask 80 | return z, m, logs, x_mask 81 | 82 | 83 | class TextEncoder(nn.Module): 84 | def __init__(self, 85 | in_channels, 86 | out_channels, 87 | hidden_channels, 88 | kernel_size, 89 | dilation_rate, 90 | n_layers, 91 | gin_channels=0, 92 | filter_channels=None, 93 | n_heads=None, 94 | p_dropout=None): 95 | super().__init__() 96 | self.in_channels = in_channels 97 | self.out_channels = out_channels 98 | self.hidden_channels = hidden_channels 99 | self.kernel_size = kernel_size 100 | self.dilation_rate = dilation_rate 101 | self.n_layers = n_layers 102 | self.gin_channels = gin_channels 103 | self.pre = nn.Conv1d(in_channels, hidden_channels, 1) 104 | self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) 105 | self.f0_emb = nn.Embedding(256, hidden_channels) 106 | 107 | self.enc_ = attentions.Encoder( 108 | hidden_channels, 109 | filter_channels, 110 | n_heads, 111 | n_layers, 112 | kernel_size, 113 | p_dropout) 114 | 115 | def forward(self, x, x_lengths, f0=None): 116 | x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) 117 | x = self.pre(x) * x_mask 118 | x = x + self.f0_emb(f0).transpose(1,2) 119 | x = self.enc_(x * x_mask, x_mask) 120 | stats = self.proj(x) * x_mask 121 | m, logs = torch.split(stats, self.out_channels, dim=1) 122 | z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask 123 | 124 | return z, m, logs, x_mask 125 | 126 | 127 | 128 | class DiscriminatorP(torch.nn.Module): 129 | def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): 130 | super(DiscriminatorP, self).__init__() 131 | self.period = period 132 | self.use_spectral_norm = use_spectral_norm 133 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 134 | self.convs = nn.ModuleList([ 135 | norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 136 | norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 137 | norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 138 | norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), 139 | norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))), 140 | ]) 141 | self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0))) 142 | 143 | def forward(self, x): 144 | fmap = [] 145 | 146 | # 1d to 2d 147 | b, c, t = x.shape 148 | if t % self.period != 0: # pad first 149 | n_pad = self.period - (t % self.period) 150 | x = F.pad(x, (0, n_pad), "reflect") 151 | t = t + n_pad 152 | x = x.view(b, c, t // self.period, self.period) 153 | 154 | for l in self.convs: 155 | x = l(x) 156 | x = F.leaky_relu(x, modules.LRELU_SLOPE) 157 | fmap.append(x) 158 | x = self.conv_post(x) 159 | fmap.append(x) 160 | x = torch.flatten(x, 1, -1) 161 | 162 | return x, fmap 163 | 164 | 165 | class DiscriminatorS(torch.nn.Module): 166 | def __init__(self, use_spectral_norm=False): 167 | super(DiscriminatorS, self).__init__() 168 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 169 | self.convs = nn.ModuleList([ 170 | norm_f(Conv1d(1, 16, 15, 1, padding=7)), 171 | norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), 172 | norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), 173 | norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), 174 | norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), 175 | norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), 176 | ]) 177 | self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) 178 | 179 | def forward(self, x): 180 | fmap = [] 181 | 182 | for l in self.convs: 183 | x = l(x) 184 | x = F.leaky_relu(x, modules.LRELU_SLOPE) 185 | fmap.append(x) 186 | x = self.conv_post(x) 187 | fmap.append(x) 188 | x = torch.flatten(x, 1, -1) 189 | 190 | return x, fmap 191 | 192 | 193 | class MultiPeriodDiscriminator(torch.nn.Module): 194 | def __init__(self, use_spectral_norm=False): 195 | super(MultiPeriodDiscriminator, self).__init__() 196 | periods = [2,3,5,7,11] 197 | 198 | discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)] 199 | discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods] 200 | self.discriminators = nn.ModuleList(discs) 201 | 202 | def forward(self, y, y_hat): 203 | y_d_rs = [] 204 | y_d_gs = [] 205 | fmap_rs = [] 206 | fmap_gs = [] 207 | for i, d in enumerate(self.discriminators): 208 | y_d_r, fmap_r = d(y) 209 | y_d_g, fmap_g = d(y_hat) 210 | y_d_rs.append(y_d_r) 211 | y_d_gs.append(y_d_g) 212 | fmap_rs.append(fmap_r) 213 | fmap_gs.append(fmap_g) 214 | 215 | return y_d_rs, y_d_gs, fmap_rs, fmap_gs 216 | 217 | 218 | class SpeakerEncoder(torch.nn.Module): 219 | def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256): 220 | super(SpeakerEncoder, self).__init__() 221 | self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True) 222 | self.linear = nn.Linear(model_hidden_size, model_embedding_size) 223 | self.relu = nn.ReLU() 224 | 225 | def forward(self, mels): 226 | self.lstm.flatten_parameters() 227 | _, (hidden, _) = self.lstm(mels) 228 | embeds_raw = self.relu(self.linear(hidden[-1])) 229 | return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True) 230 | 231 | def compute_partial_slices(self, total_frames, partial_frames, partial_hop): 232 | mel_slices = [] 233 | for i in range(0, total_frames-partial_frames, partial_hop): 234 | mel_range = torch.arange(i, i+partial_frames) 235 | mel_slices.append(mel_range) 236 | 237 | return mel_slices 238 | 239 | def embed_utterance(self, mel, partial_frames=128, partial_hop=64): 240 | mel_len = mel.size(1) 241 | last_mel = mel[:,-partial_frames:] 242 | 243 | if mel_len > partial_frames: 244 | mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop) 245 | mels = list(mel[:,s] for s in mel_slices) 246 | mels.append(last_mel) 247 | mels = torch.stack(tuple(mels), 0).squeeze(1) 248 | 249 | with torch.no_grad(): 250 | partial_embeds = self(mels) 251 | embed = torch.mean(partial_embeds, axis=0).unsqueeze(0) 252 | #embed = embed / torch.linalg.norm(embed, 2) 253 | else: 254 | with torch.no_grad(): 255 | embed = self(last_mel) 256 | 257 | return embed 258 | 259 | 260 | class SynthesizerTrn(nn.Module): 261 | """ 262 | Synthesizer for Training 263 | """ 264 | 265 | def __init__(self, 266 | spec_channels, 267 | segment_size, 268 | inter_channels, 269 | hidden_channels, 270 | filter_channels, 271 | n_heads, 272 | n_layers, 273 | kernel_size, 274 | p_dropout, 275 | resblock, 276 | resblock_kernel_sizes, 277 | resblock_dilation_sizes, 278 | upsample_rates, 279 | upsample_initial_channel, 280 | upsample_kernel_sizes, 281 | gin_channels, 282 | ssl_dim, 283 | n_speakers, 284 | **kwargs): 285 | 286 | super().__init__() 287 | self.spec_channels = spec_channels 288 | self.inter_channels = inter_channels 289 | self.hidden_channels = hidden_channels 290 | self.filter_channels = filter_channels 291 | self.n_heads = n_heads 292 | self.n_layers = n_layers 293 | self.kernel_size = kernel_size 294 | self.p_dropout = p_dropout 295 | self.resblock = resblock 296 | self.resblock_kernel_sizes = resblock_kernel_sizes 297 | self.resblock_dilation_sizes = resblock_dilation_sizes 298 | self.upsample_rates = upsample_rates 299 | self.upsample_initial_channel = upsample_initial_channel 300 | self.upsample_kernel_sizes = upsample_kernel_sizes 301 | self.segment_size = segment_size 302 | self.gin_channels = gin_channels 303 | self.ssl_dim = ssl_dim 304 | self.emb_g = nn.Embedding(n_speakers, gin_channels) 305 | 306 | self.enc_p_ = TextEncoder(ssl_dim, inter_channels, hidden_channels, 5, 1, 16,0, filter_channels, n_heads, p_dropout) 307 | hps = { 308 | "sampling_rate": 32000, 309 | "inter_channels": 192, 310 | "resblock": "1", 311 | "resblock_kernel_sizes": [3, 7, 11], 312 | "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]], 313 | "upsample_rates": [10, 8, 2, 2], 314 | "upsample_initial_channel": 512, 315 | "upsample_kernel_sizes": [16, 16, 4, 4], 316 | "gin_channels": 256, 317 | } 318 | self.dec = Generator(h=hps) 319 | self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels) 320 | self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels) 321 | 322 | def forward(self, c, f0, spec, g=None, mel=None, c_lengths=None, spec_lengths=None): 323 | if c_lengths == None: 324 | c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device) 325 | if spec_lengths == None: 326 | spec_lengths = (torch.ones(spec.size(0)) * spec.size(-1)).to(spec.device) 327 | 328 | g = self.emb_g(g).transpose(1,2) 329 | 330 | z_ptemp, m_p, logs_p, _ = self.enc_p_(c, c_lengths, f0=f0_to_coarse(f0)) 331 | z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g) 332 | 333 | z_p = self.flow(z, spec_mask, g=g) 334 | z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(z, f0, spec_lengths, self.segment_size) 335 | 336 | # o = self.dec(z_slice, g=g) 337 | o = self.dec(z_slice, g=g, f0=pitch_slice) 338 | 339 | return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q) 340 | 341 | def infer(self, c, f0, g=None, mel=None, c_lengths=None): 342 | if c_lengths == None: 343 | c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device) 344 | g = self.emb_g(g).transpose(1,2) 345 | 346 | z_p, m_p, logs_p, c_mask = self.enc_p_(c, c_lengths, f0=f0_to_coarse(f0)) 347 | z = self.flow(z_p, c_mask, g=g, reverse=True) 348 | 349 | o = self.dec(z * c_mask, g=g, f0=f0) 350 | 351 | return o 352 | -------------------------------------------------------------------------------- /modules.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import numpy as np 4 | import scipy 5 | import torch 6 | from torch import nn 7 | from torch.nn import functional as F 8 | 9 | from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d 10 | from torch.nn.utils import weight_norm, remove_weight_norm 11 | 12 | import commons 13 | from commons import init_weights, get_padding 14 | 15 | 16 | LRELU_SLOPE = 0.1 17 | 18 | 19 | class LayerNorm(nn.Module): 20 | def __init__(self, channels, eps=1e-5): 21 | super().__init__() 22 | self.channels = channels 23 | self.eps = eps 24 | 25 | self.gamma = nn.Parameter(torch.ones(channels)) 26 | self.beta = nn.Parameter(torch.zeros(channels)) 27 | 28 | def forward(self, x): 29 | x = x.transpose(1, -1) 30 | x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps) 31 | return x.transpose(1, -1) 32 | 33 | 34 | class ConvReluNorm(nn.Module): 35 | def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout): 36 | super().__init__() 37 | self.in_channels = in_channels 38 | self.hidden_channels = hidden_channels 39 | self.out_channels = out_channels 40 | self.kernel_size = kernel_size 41 | self.n_layers = n_layers 42 | self.p_dropout = p_dropout 43 | assert n_layers > 1, "Number of layers should be larger than 0." 44 | 45 | self.conv_layers = nn.ModuleList() 46 | self.norm_layers = nn.ModuleList() 47 | self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2)) 48 | self.norm_layers.append(LayerNorm(hidden_channels)) 49 | self.relu_drop = nn.Sequential( 50 | nn.ReLU(), 51 | nn.Dropout(p_dropout)) 52 | for _ in range(n_layers-1): 53 | self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2)) 54 | self.norm_layers.append(LayerNorm(hidden_channels)) 55 | self.proj = nn.Conv1d(hidden_channels, out_channels, 1) 56 | self.proj.weight.data.zero_() 57 | self.proj.bias.data.zero_() 58 | 59 | def forward(self, x, x_mask): 60 | x_org = x 61 | for i in range(self.n_layers): 62 | x = self.conv_layers[i](x * x_mask) 63 | x = self.norm_layers[i](x) 64 | x = self.relu_drop(x) 65 | x = x_org + self.proj(x) 66 | return x * x_mask 67 | 68 | 69 | class DDSConv(nn.Module): 70 | """ 71 | Dialted and Depth-Separable Convolution 72 | """ 73 | def __init__(self, channels, kernel_size, n_layers, p_dropout=0.): 74 | super().__init__() 75 | self.channels = channels 76 | self.kernel_size = kernel_size 77 | self.n_layers = n_layers 78 | self.p_dropout = p_dropout 79 | 80 | self.drop = nn.Dropout(p_dropout) 81 | self.convs_sep = nn.ModuleList() 82 | self.convs_1x1 = nn.ModuleList() 83 | self.norms_1 = nn.ModuleList() 84 | self.norms_2 = nn.ModuleList() 85 | for i in range(n_layers): 86 | dilation = kernel_size ** i 87 | padding = (kernel_size * dilation - dilation) // 2 88 | self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size, 89 | groups=channels, dilation=dilation, padding=padding 90 | )) 91 | self.convs_1x1.append(nn.Conv1d(channels, channels, 1)) 92 | self.norms_1.append(LayerNorm(channels)) 93 | self.norms_2.append(LayerNorm(channels)) 94 | 95 | def forward(self, x, x_mask, g=None): 96 | if g is not None: 97 | x = x + g 98 | for i in range(self.n_layers): 99 | y = self.convs_sep[i](x * x_mask) 100 | y = self.norms_1[i](y) 101 | y = F.gelu(y) 102 | y = self.convs_1x1[i](y) 103 | y = self.norms_2[i](y) 104 | y = F.gelu(y) 105 | y = self.drop(y) 106 | x = x + y 107 | return x * x_mask 108 | 109 | 110 | class WN(torch.nn.Module): 111 | def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0): 112 | super(WN, self).__init__() 113 | assert(kernel_size % 2 == 1) 114 | self.hidden_channels =hidden_channels 115 | self.kernel_size = kernel_size, 116 | self.dilation_rate = dilation_rate 117 | self.n_layers = n_layers 118 | self.gin_channels = gin_channels 119 | self.p_dropout = p_dropout 120 | 121 | self.in_layers = torch.nn.ModuleList() 122 | self.res_skip_layers = torch.nn.ModuleList() 123 | self.drop = nn.Dropout(p_dropout) 124 | 125 | if gin_channels != 0: 126 | cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1) 127 | self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight') 128 | 129 | for i in range(n_layers): 130 | dilation = dilation_rate ** i 131 | padding = int((kernel_size * dilation - dilation) / 2) 132 | in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size, 133 | dilation=dilation, padding=padding) 134 | in_layer = torch.nn.utils.weight_norm(in_layer, name='weight') 135 | self.in_layers.append(in_layer) 136 | 137 | # last one is not necessary 138 | if i < n_layers - 1: 139 | res_skip_channels = 2 * hidden_channels 140 | else: 141 | res_skip_channels = hidden_channels 142 | 143 | res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1) 144 | res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight') 145 | self.res_skip_layers.append(res_skip_layer) 146 | 147 | def forward(self, x, x_mask, g=None, **kwargs): 148 | output = torch.zeros_like(x) 149 | n_channels_tensor = torch.IntTensor([self.hidden_channels]) 150 | 151 | if g is not None: 152 | g = self.cond_layer(g) 153 | 154 | for i in range(self.n_layers): 155 | x_in = self.in_layers[i](x) 156 | if g is not None: 157 | cond_offset = i * 2 * self.hidden_channels 158 | g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:] 159 | else: 160 | g_l = torch.zeros_like(x_in) 161 | 162 | acts = commons.fused_add_tanh_sigmoid_multiply( 163 | x_in, 164 | g_l, 165 | n_channels_tensor) 166 | acts = self.drop(acts) 167 | 168 | res_skip_acts = self.res_skip_layers[i](acts) 169 | if i < self.n_layers - 1: 170 | res_acts = res_skip_acts[:,:self.hidden_channels,:] 171 | x = (x + res_acts) * x_mask 172 | output = output + res_skip_acts[:,self.hidden_channels:,:] 173 | else: 174 | output = output + res_skip_acts 175 | return output * x_mask 176 | 177 | def remove_weight_norm(self): 178 | if self.gin_channels != 0: 179 | torch.nn.utils.remove_weight_norm(self.cond_layer) 180 | for l in self.in_layers: 181 | torch.nn.utils.remove_weight_norm(l) 182 | for l in self.res_skip_layers: 183 | torch.nn.utils.remove_weight_norm(l) 184 | 185 | 186 | class ResBlock1(torch.nn.Module): 187 | def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)): 188 | super(ResBlock1, self).__init__() 189 | self.convs1 = nn.ModuleList([ 190 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0], 191 | padding=get_padding(kernel_size, dilation[0]))), 192 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1], 193 | padding=get_padding(kernel_size, dilation[1]))), 194 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2], 195 | padding=get_padding(kernel_size, dilation[2]))) 196 | ]) 197 | self.convs1.apply(init_weights) 198 | 199 | self.convs2 = nn.ModuleList([ 200 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, 201 | padding=get_padding(kernel_size, 1))), 202 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, 203 | padding=get_padding(kernel_size, 1))), 204 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, 205 | padding=get_padding(kernel_size, 1))) 206 | ]) 207 | self.convs2.apply(init_weights) 208 | 209 | def forward(self, x, x_mask=None): 210 | for c1, c2 in zip(self.convs1, self.convs2): 211 | xt = F.leaky_relu(x, LRELU_SLOPE) 212 | if x_mask is not None: 213 | xt = xt * x_mask 214 | xt = c1(xt) 215 | xt = F.leaky_relu(xt, LRELU_SLOPE) 216 | if x_mask is not None: 217 | xt = xt * x_mask 218 | xt = c2(xt) 219 | x = xt + x 220 | if x_mask is not None: 221 | x = x * x_mask 222 | return x 223 | 224 | def remove_weight_norm(self): 225 | for l in self.convs1: 226 | remove_weight_norm(l) 227 | for l in self.convs2: 228 | remove_weight_norm(l) 229 | 230 | 231 | class ResBlock2(torch.nn.Module): 232 | def __init__(self, channels, kernel_size=3, dilation=(1, 3)): 233 | super(ResBlock2, self).__init__() 234 | self.convs = nn.ModuleList([ 235 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0], 236 | padding=get_padding(kernel_size, dilation[0]))), 237 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1], 238 | padding=get_padding(kernel_size, dilation[1]))) 239 | ]) 240 | self.convs.apply(init_weights) 241 | 242 | def forward(self, x, x_mask=None): 243 | for c in self.convs: 244 | xt = F.leaky_relu(x, LRELU_SLOPE) 245 | if x_mask is not None: 246 | xt = xt * x_mask 247 | xt = c(xt) 248 | x = xt + x 249 | if x_mask is not None: 250 | x = x * x_mask 251 | return x 252 | 253 | def remove_weight_norm(self): 254 | for l in self.convs: 255 | remove_weight_norm(l) 256 | 257 | 258 | class Log(nn.Module): 259 | def forward(self, x, x_mask, reverse=False, **kwargs): 260 | if not reverse: 261 | y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask 262 | logdet = torch.sum(-y, [1, 2]) 263 | return y, logdet 264 | else: 265 | x = torch.exp(x) * x_mask 266 | return x 267 | 268 | 269 | class Flip(nn.Module): 270 | def forward(self, x, *args, reverse=False, **kwargs): 271 | x = torch.flip(x, [1]) 272 | if not reverse: 273 | logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device) 274 | return x, logdet 275 | else: 276 | return x 277 | 278 | 279 | class ElementwiseAffine(nn.Module): 280 | def __init__(self, channels): 281 | super().__init__() 282 | self.channels = channels 283 | self.m = nn.Parameter(torch.zeros(channels,1)) 284 | self.logs = nn.Parameter(torch.zeros(channels,1)) 285 | 286 | def forward(self, x, x_mask, reverse=False, **kwargs): 287 | if not reverse: 288 | y = self.m + torch.exp(self.logs) * x 289 | y = y * x_mask 290 | logdet = torch.sum(self.logs * x_mask, [1,2]) 291 | return y, logdet 292 | else: 293 | x = (x - self.m) * torch.exp(-self.logs) * x_mask 294 | return x 295 | 296 | 297 | class ResidualCouplingLayer(nn.Module): 298 | def __init__(self, 299 | channels, 300 | hidden_channels, 301 | kernel_size, 302 | dilation_rate, 303 | n_layers, 304 | p_dropout=0, 305 | gin_channels=0, 306 | mean_only=False): 307 | assert channels % 2 == 0, "channels should be divisible by 2" 308 | super().__init__() 309 | self.channels = channels 310 | self.hidden_channels = hidden_channels 311 | self.kernel_size = kernel_size 312 | self.dilation_rate = dilation_rate 313 | self.n_layers = n_layers 314 | self.half_channels = channels // 2 315 | self.mean_only = mean_only 316 | 317 | self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1) 318 | self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels) 319 | self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1) 320 | self.post.weight.data.zero_() 321 | self.post.bias.data.zero_() 322 | 323 | def forward(self, x, x_mask, g=None, reverse=False): 324 | x0, x1 = torch.split(x, [self.half_channels]*2, 1) 325 | h = self.pre(x0) * x_mask 326 | h = self.enc(h, x_mask, g=g) 327 | stats = self.post(h) * x_mask 328 | if not self.mean_only: 329 | m, logs = torch.split(stats, [self.half_channels]*2, 1) 330 | else: 331 | m = stats 332 | logs = torch.zeros_like(m) 333 | 334 | if not reverse: 335 | x1 = m + x1 * torch.exp(logs) * x_mask 336 | x = torch.cat([x0, x1], 1) 337 | logdet = torch.sum(logs, [1,2]) 338 | return x, logdet 339 | else: 340 | x1 = (x1 - m) * torch.exp(-logs) * x_mask 341 | x = torch.cat([x0, x1], 1) 342 | return x 343 | -------------------------------------------------------------------------------- /onnx_export.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | import numpy as np 4 | import onnx 5 | from onnxsim import simplify 6 | import onnxruntime as ort 7 | import onnxoptimizer 8 | import torch 9 | from model_onnx import SynthesizerTrn 10 | import utils 11 | from hubert import hubert_model_onnx 12 | 13 | def main(HubertExport,NetExport): 14 | 15 | path = "NyaruTaffy" 16 | 17 | if(HubertExport): 18 | device = torch.device("cuda") 19 | hubert_soft = utils.get_hubert_model() 20 | test_input = torch.rand(1, 1, 16000) 21 | input_names = ["source"] 22 | output_names = ["embed"] 23 | torch.onnx.export(hubert_soft.to(device), 24 | test_input.to(device), 25 | "hubert3.0.onnx", 26 | dynamic_axes={ 27 | "source": { 28 | 2: "sample_length" 29 | } 30 | }, 31 | verbose=False, 32 | opset_version=13, 33 | input_names=input_names, 34 | output_names=output_names) 35 | if(NetExport): 36 | device = torch.device("cuda") 37 | hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json") 38 | SVCVITS = SynthesizerTrn( 39 | hps.data.filter_length // 2 + 1, 40 | hps.train.segment_size // hps.data.hop_length, 41 | **hps.model) 42 | _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None) 43 | _ = SVCVITS.eval().to(device) 44 | for i in SVCVITS.parameters(): 45 | i.requires_grad = False 46 | test_hidden_unit = torch.rand(1, 50, 256) 47 | test_lengths = torch.LongTensor([50]) 48 | test_pitch = torch.rand(1, 50) 49 | test_sid = torch.LongTensor([0]) 50 | input_names = ["hidden_unit", "lengths", "pitch", "sid"] 51 | output_names = ["audio", ] 52 | SVCVITS.eval() 53 | torch.onnx.export(SVCVITS, 54 | ( 55 | test_hidden_unit.to(device), 56 | test_lengths.to(device), 57 | test_pitch.to(device), 58 | test_sid.to(device) 59 | ), 60 | f"checkpoints/{path}/model.onnx", 61 | dynamic_axes={ 62 | "hidden_unit": [0, 1], 63 | "pitch": [1] 64 | }, 65 | do_constant_folding=False, 66 | opset_version=16, 67 | verbose=False, 68 | input_names=input_names, 69 | output_names=output_names) 70 | 71 | 72 | if __name__ == '__main__': 73 | main(False,True) 74 | -------------------------------------------------------------------------------- /onnx_export_48k.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | import numpy as np 4 | import onnx 5 | from onnxsim import simplify 6 | import onnxruntime as ort 7 | import onnxoptimizer 8 | import torch 9 | from model_onnx_48k import SynthesizerTrn 10 | import utils 11 | from hubert import hubert_model_onnx 12 | 13 | def main(HubertExport,NetExport): 14 | 15 | path = "NyaruTaffy" 16 | 17 | if(HubertExport): 18 | device = torch.device("cuda") 19 | hubert_soft = hubert_model_onnx.hubert_soft("hubert/model.pt") 20 | test_input = torch.rand(1, 1, 16000) 21 | input_names = ["source"] 22 | output_names = ["embed"] 23 | torch.onnx.export(hubert_soft.to(device), 24 | test_input.to(device), 25 | "hubert3.0.onnx", 26 | dynamic_axes={ 27 | "source": { 28 | 2: "sample_length" 29 | } 30 | }, 31 | verbose=False, 32 | opset_version=13, 33 | input_names=input_names, 34 | output_names=output_names) 35 | if(NetExport): 36 | device = torch.device("cuda") 37 | hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json") 38 | SVCVITS = SynthesizerTrn( 39 | hps.data.filter_length // 2 + 1, 40 | hps.train.segment_size // hps.data.hop_length, 41 | **hps.model) 42 | _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None) 43 | _ = SVCVITS.eval().to(device) 44 | for i in SVCVITS.parameters(): 45 | i.requires_grad = False 46 | test_hidden_unit = torch.rand(1, 50, 256) 47 | test_lengths = torch.LongTensor([50]) 48 | test_pitch = torch.rand(1, 50) 49 | test_sid = torch.LongTensor([0]) 50 | input_names = ["hidden_unit", "lengths", "pitch", "sid"] 51 | output_names = ["audio", ] 52 | SVCVITS.eval() 53 | torch.onnx.export(SVCVITS, 54 | ( 55 | test_hidden_unit.to(device), 56 | test_lengths.to(device), 57 | test_pitch.to(device), 58 | test_sid.to(device) 59 | ), 60 | f"checkpoints/{path}/model.onnx", 61 | dynamic_axes={ 62 | "hidden_unit": [0, 1], 63 | "pitch": [1] 64 | }, 65 | do_constant_folding=False, 66 | opset_version=16, 67 | verbose=False, 68 | input_names=input_names, 69 | output_names=output_names) 70 | 71 | 72 | if __name__ == '__main__': 73 | main(False,True) 74 | -------------------------------------------------------------------------------- /preprocess_flist_config.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | import re 4 | 5 | from tqdm import tqdm 6 | from random import shuffle 7 | import json 8 | config_template = { 9 | "train": { 10 | "log_interval": 200, 11 | "eval_interval": 1000, 12 | "seed": 1234, 13 | "epochs": 10000, 14 | "learning_rate": 1e-4, 15 | "betas": [0.8, 0.99], 16 | "eps": 1e-9, 17 | "batch_size": 12, 18 | "fp16_run": False, 19 | "lr_decay": 0.999875, 20 | "segment_size": 17920, 21 | "init_lr_ratio": 1, 22 | "warmup_epochs": 0, 23 | "c_mel": 45, 24 | "c_kl": 1.0, 25 | "use_sr": True, 26 | "max_speclen": 384, 27 | "port": "8001", 28 | "keep_ckpts": 5, # Number of ckpts to keep, if 0, keep all 29 | }, 30 | "data": { 31 | "training_files":"filelists/train.txt", 32 | "validation_files":"filelists/val.txt", 33 | "max_wav_value": 32768.0, 34 | "sampling_rate": 32000, 35 | "filter_length": 1280, 36 | "hop_length": 320, 37 | "win_length": 1280, 38 | "n_mel_channels": 80, 39 | "mel_fmin": 0.0, 40 | "mel_fmax": None 41 | }, 42 | "model": { 43 | "inter_channels": 192, 44 | "hidden_channels": 192, 45 | "filter_channels": 768, 46 | "n_heads": 2, 47 | "n_layers": 6, 48 | "kernel_size": 3, 49 | "p_dropout": 0.1, 50 | "resblock": "1", 51 | "resblock_kernel_sizes": [3,7,11], 52 | "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], 53 | "upsample_rates": [10,8,2,2], 54 | "upsample_initial_channel": 512, 55 | "upsample_kernel_sizes": [16,16,4,4], 56 | "n_layers_q": 3, 57 | "use_spectral_norm": False, 58 | "gin_channels": 256, 59 | "ssl_dim": 256, 60 | "n_speakers": 0, 61 | }, 62 | "spk":{ 63 | "nen": 0, 64 | "paimon": 1, 65 | "yunhao": 2 66 | } 67 | } 68 | 69 | pattern = re.compile(r'^[\.a-zA-Z0-9_\/]+$') 70 | 71 | if __name__ == "__main__": 72 | parser = argparse.ArgumentParser() 73 | parser.add_argument("--train_list", type=str, default="./filelists/train.txt", help="path to train list") 74 | parser.add_argument("--val_list", type=str, default="./filelists/val.txt", help="path to val list") 75 | parser.add_argument("--test_list", type=str, default="./filelists/test.txt", help="path to test list") 76 | parser.add_argument("--source_dir", type=str, default="./dataset/32k", help="path to source dir") 77 | args = parser.parse_args() 78 | 79 | train = [] 80 | val = [] 81 | test = [] 82 | idx = 0 83 | spk_dict = {} 84 | spk_id = 0 85 | for speaker in tqdm(os.listdir(args.source_dir)): 86 | spk_dict[speaker] = spk_id 87 | spk_id += 1 88 | wavs = ["/".join([args.source_dir, speaker, i]) for i in os.listdir(os.path.join(args.source_dir, speaker))] 89 | for wavpath in wavs: 90 | if not pattern.match(wavpath): 91 | print(f"warning：文件名{wavpath}中包含非字母数字下划线，可能会导致错误。（也可能不会）") 92 | if len(wavs) < 10: 93 | print(f"warning：{speaker}数据集数量小于10条，请补充数据") 94 | wavs = [i for i in wavs if i.endswith("wav")] 95 | shuffle(wavs) 96 | train += wavs[2:-2] 97 | val += wavs[:2] 98 | test += wavs[-2:] 99 | n_speakers = len(spk_dict.keys())*2 100 | shuffle(train) 101 | shuffle(val) 102 | shuffle(test) 103 | 104 | print("Writing", args.train_list) 105 | with open(args.train_list, "w") as f: 106 | for fname in tqdm(train): 107 | wavpath = fname 108 | f.write(wavpath + "\n") 109 | 110 | print("Writing", args.val_list) 111 | with open(args.val_list, "w") as f: 112 | for fname in tqdm(val): 113 | wavpath = fname 114 | f.write(wavpath + "\n") 115 | 116 | print("Writing", args.test_list) 117 | with open(args.test_list, "w") as f: 118 | for fname in tqdm(test): 119 | wavpath = fname 120 | f.write(wavpath + "\n") 121 | 122 | config_template["model"]["n_speakers"] = n_speakers 123 | config_template["spk"] = spk_dict 124 | print("Writing configs/config.json") 125 | with open("configs/config.json", "w") as f: 126 | json.dump(config_template, f, indent=2) 127 | -------------------------------------------------------------------------------- /preprocess_hubert_f0.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | 4 | import torch 5 | import json 6 | from glob import glob 7 | 8 | from pyworld import pyworld 9 | from tqdm import tqdm 10 | from scipy.io import wavfile 11 | 12 | import utils 13 | from mel_processing import mel_spectrogram_torch 14 | #import h5py 15 | import logging 16 | logging.getLogger('numba').setLevel(logging.WARNING) 17 | 18 | import parselmouth 19 | import librosa 20 | import numpy as np 21 | 22 | 23 | def get_f0(path,p_len=None, f0_up_key=0): 24 | x, _ = librosa.load(path, 32000) 25 | if p_len is None: 26 | p_len = x.shape[0]//320 27 | else: 28 | assert abs(p_len-x.shape[0]//320) < 3, (path, p_len, x.shape) 29 | time_step = 320 / 32000 * 1000 30 | f0_min = 50 31 | f0_max = 1100 32 | f0_mel_min = 1127 * np.log(1 + f0_min / 700) 33 | f0_mel_max = 1127 * np.log(1 + f0_max / 700) 34 | 35 | f0 = parselmouth.Sound(x, 32000).to_pitch_ac( 36 | time_step=time_step / 1000, voicing_threshold=0.6, 37 | pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency'] 38 | 39 | pad_size=(p_len - len(f0) + 1) // 2 40 | if(pad_size>0 or p_len - len(f0) - pad_size>0): 41 | f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant') 42 | 43 | f0bak = f0.copy() 44 | f0 *= pow(2, f0_up_key / 12) 45 | f0_mel = 1127 * np.log(1 + f0 / 700) 46 | f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1 47 | f0_mel[f0_mel <= 1] = 1 48 | f0_mel[f0_mel > 255] = 255 49 | f0_coarse = np.rint(f0_mel).astype(np.int) 50 | return f0_coarse, f0bak 51 | 52 | def resize2d(x, target_len): 53 | source = np.array(x) 54 | source[source<0.001] = np.nan 55 | target = np.interp(np.arange(0, len(source)*target_len, len(source))/ target_len, np.arange(0, len(source)), source) 56 | res = np.nan_to_num(target) 57 | return res 58 | 59 | def compute_f0(path, c_len): 60 | x, sr = librosa.load(path, sr=32000) 61 | f0, t = pyworld.dio( 62 | x.astype(np.double), 63 | fs=sr, 64 | f0_ceil=800, 65 | frame_period=1000 * 320 / sr, 66 | ) 67 | f0 = pyworld.stonemask(x.astype(np.double), f0, t, 32000) 68 | for index, pitch in enumerate(f0): 69 | f0[index] = round(pitch, 1) 70 | assert abs(c_len - x.shape[0]//320) < 3, (c_len, f0.shape) 71 | 72 | return None, resize2d(f0, c_len) 73 | 74 | 75 | def process(filename): 76 | print(filename) 77 | save_name = filename+".soft.pt" 78 | if not os.path.exists(save_name): 79 | devive = torch.device("cuda" if torch.cuda.is_available() else "cpu") 80 | wav, _ = librosa.load(filename, sr=16000) 81 | wav = torch.from_numpy(wav).unsqueeze(0).to(devive) 82 | c = utils.get_hubert_content(hmodel, wav) 83 | torch.save(c.cpu(), save_name) 84 | else: 85 | c = torch.load(save_name) 86 | f0path = filename+".f0.npy" 87 | if not os.path.exists(f0path): 88 | cf0, f0 = compute_f0(filename, c.shape[-1] * 2) 89 | np.save(f0path, f0) 90 | 91 | 92 | 93 | if __name__ == "__main__": 94 | parser = argparse.ArgumentParser() 95 | parser.add_argument("--in_dir", type=str, default="dataset/32k", help="path to input dir") 96 | args = parser.parse_args() 97 | 98 | print("Loading hubert for content...") 99 | hmodel = utils.get_hubert_model(0 if torch.cuda.is_available() else None) 100 | print("Loaded hubert.") 101 | 102 | filenames = glob(f'{args.in_dir}/*/*.wav', recursive=True)#[:10] 103 | 104 | for filename in tqdm(filenames): 105 | process(filename) 106 | -------------------------------------------------------------------------------- /raw/put_raw_wav_here: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/raw/put_raw_wav_here -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | Flask==2.1.2 2 | Flask_Cors==3.0.10 3 | gradio==3.4.1 4 | numpy==1.19.2 5 | playsound==1.3.0 6 | PyAudio==0.2.12 7 | pydub==0.25.1 8 | pyworld==0.3.0 9 | requests==2.28.1 10 | scipy==1.7.3 11 | sounddevice==0.4.5 12 | SoundFile==0.10.3.post1 13 | starlette==0.19.1 14 | torch==1.10.0+cu113 15 | torchaudio==0.10.0+cu113 16 | tqdm==4.63.0 17 | scikit-maad 18 | praat-parselmouth 19 | onnx 20 | onnxsim 21 | onnxoptimizer 22 | -------------------------------------------------------------------------------- /resample.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | import librosa 4 | import numpy as np 5 | from multiprocessing import Pool, cpu_count 6 | from scipy.io import wavfile 7 | from tqdm import tqdm 8 | 9 | 10 | def process(item): 11 | spkdir, wav_name, args = item 12 | # speaker 's5', 'p280', 'p315' are excluded, 13 | speaker = spkdir.replace("\\", "/").split("/")[-1] 14 | wav_path = os.path.join(args.in_dir, speaker, wav_name) 15 | if os.path.exists(wav_path) and '.wav' in wav_path: 16 | os.makedirs(os.path.join(args.out_dir2, speaker), exist_ok=True) 17 | wav, sr = librosa.load(wav_path, None) 18 | wav, _ = librosa.effects.trim(wav, top_db=20) 19 | peak = np.abs(wav).max() 20 | if peak > 1.0: 21 | wav = 0.98 * wav / peak 22 | wav2 = librosa.resample(wav, orig_sr=sr, target_sr=args.sr2) 23 | wav2 /= max(wav2.max(), -wav2.min()) 24 | save_name = wav_name 25 | save_path2 = os.path.join(args.out_dir2, speaker, save_name) 26 | wavfile.write( 27 | save_path2, 28 | args.sr2, 29 | (wav2 * np.iinfo(np.int16).max).astype(np.int16) 30 | ) 31 | 32 | 33 | 34 | if __name__ == "__main__": 35 | parser = argparse.ArgumentParser() 36 | parser.add_argument("--sr2", type=int, default=32000, help="sampling rate") 37 | parser.add_argument("--in_dir", type=str, default="./dataset_raw", help="path to source dir") 38 | parser.add_argument("--out_dir2", type=str, default="./dataset/32k", help="path to target dir") 39 | args = parser.parse_args() 40 | processs = cpu_count()-2 if cpu_count() >4 else 1 41 | pool = Pool(processes=processs) 42 | 43 | for speaker in os.listdir(args.in_dir): 44 | spk_dir = os.path.join(args.in_dir, speaker) 45 | if os.path.isdir(spk_dir): 46 | print(spk_dir) 47 | for _ in tqdm(pool.imap_unordered(process, [(spk_dir, i, args) for i in os.listdir(spk_dir) if i.endswith("wav")])): 48 | pass 49 | -------------------------------------------------------------------------------- /sovits_gradio.py: -------------------------------------------------------------------------------- 1 | from inference.infer_tool_grad import VitsSvc 2 | import gradio as gr 3 | import os 4 | 5 | class VitsGradio: 6 | def __init__(self): 7 | self.so = VitsSvc() 8 | self.lspk = [] 9 | self.modelPaths = [] 10 | for root,dirs,files in os.walk("checkpoints"): 11 | for dir in dirs: 12 | self.modelPaths.append(dir) 13 | with gr.Blocks() as self.Vits: 14 | with gr.Tab("VoiceConversion"): 15 | with gr.Row(visible=False) as self.VoiceConversion: 16 | with gr.Column(): 17 | with gr.Row(): 18 | with gr.Column(): 19 | self.srcaudio = gr.Audio(label = "输入音频") 20 | self.btnVC = gr.Button("说话人转换") 21 | with gr.Column(): 22 | self.dsid = gr.Dropdown(label = "目标角色", choices = self.lspk) 23 | self.tran = gr.Slider(label = "升降调", maximum = 60, minimum = -60, step = 1, value = 0) 24 | self.th = gr.Slider(label = "切片阈值", maximum = 32767, minimum = -32768, step = 0.1, value = -40) 25 | with gr.Row(): 26 | self.VCOutputs = gr.Audio() 27 | self.btnVC.click(self.so.inference, inputs=[self.srcaudio,self.dsid,self.tran,self.th], outputs=[self.VCOutputs]) 28 | with gr.Tab("SelectModel"): 29 | with gr.Column(): 30 | modelstrs = gr.Dropdown(label = "模型", choices = self.modelPaths, value = self.modelPaths[0], type = "value") 31 | devicestrs = gr.Dropdown(label = "设备", choices = ["cpu","cuda"], value = "cpu", type = "value") 32 | btnMod = gr.Button("载入模型") 33 | btnMod.click(self.loadModel, inputs=[modelstrs,devicestrs], outputs = [self.dsid,self.VoiceConversion]) 34 | 35 | def loadModel(self, path, device): 36 | self.lspk = [] 37 | self.so.set_device(device) 38 | self.so.loadCheckpoint(path) 39 | for spk, sid in self.so.hps.spk.items(): 40 | self.lspk.append(spk) 41 | VChange = gr.update(visible = True) 42 | SDChange = gr.update(choices = self.lspk, value = self.lspk[0]) 43 | return [SDChange,VChange] 44 | 45 | grVits = VitsGradio() 46 | 47 | grVits.Vits.launch() -------------------------------------------------------------------------------- /spec_gen.py: -------------------------------------------------------------------------------- 1 | from data_utils import TextAudioSpeakerLoader, EvalDataLoader 2 | import json 3 | from tqdm import tqdm 4 | 5 | from utils import HParams 6 | 7 | config_path = 'configs/config.json' 8 | with open(config_path, "r") as f: 9 | data = f.read() 10 | config = json.loads(data) 11 | hps = HParams(**config) 12 | 13 | train_dataset = TextAudioSpeakerLoader("filelists/train.txt", hps) 14 | test_dataset = TextAudioSpeakerLoader("filelists/test.txt", hps) 15 | eval_dataset = TextAudioSpeakerLoader("filelists/val.txt", hps) 16 | 17 | for _ in tqdm(train_dataset): 18 | pass 19 | for _ in tqdm(eval_dataset): 20 | pass 21 | for _ in tqdm(test_dataset): 22 | pass -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import logging 2 | logging.getLogger('matplotlib').setLevel(logging.WARNING) 3 | import os 4 | import json 5 | import argparse 6 | import itertools 7 | import math 8 | import torch 9 | from torch import nn, optim 10 | from torch.nn import functional as F 11 | from torch.utils.data import DataLoader 12 | from torch.utils.tensorboard import SummaryWriter 13 | import torch.multiprocessing as mp 14 | import torch.distributed as dist 15 | from torch.nn.parallel import DistributedDataParallel as DDP 16 | from torch.cuda.amp import autocast, GradScaler 17 | 18 | import commons 19 | import utils 20 | from data_utils import TextAudioSpeakerLoader, EvalDataLoader 21 | from models import ( 22 | SynthesizerTrn, 23 | MultiPeriodDiscriminator, 24 | ) 25 | from losses import ( 26 | kl_loss, 27 | generator_loss, discriminator_loss, feature_loss 28 | ) 29 | 30 | from mel_processing import mel_spectrogram_torch, spec_to_mel_torch 31 | 32 | torch.backends.cudnn.benchmark = True 33 | global_step = 0 34 | 35 | 36 | # os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO' 37 | 38 | 39 | def main(): 40 | """Assume Single Node Multi GPUs Training Only""" 41 | assert torch.cuda.is_available(), "CPU training is not allowed." 42 | hps = utils.get_hparams() 43 | 44 | n_gpus = torch.cuda.device_count() 45 | os.environ['MASTER_ADDR'] = 'localhost' 46 | os.environ['MASTER_PORT'] = hps.train.port 47 | 48 | mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,)) 49 | 50 | 51 | def run(rank, n_gpus, hps): 52 | global global_step 53 | if rank == 0: 54 | logger = utils.get_logger(hps.model_dir) 55 | logger.info(hps) 56 | utils.check_git_hash(hps.model_dir) 57 | writer = SummaryWriter(log_dir=hps.model_dir) 58 | writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval")) 59 | 60 | # for pytorch on win, backend use gloo 61 | dist.init_process_group(backend= 'gloo' if os.name == 'nt' else 'nccl', init_method='env://', world_size=n_gpus, rank=rank) 62 | torch.manual_seed(hps.train.seed) 63 | torch.cuda.set_device(rank) 64 | 65 | train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps) 66 | train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True, 67 | batch_size=hps.train.batch_size) 68 | if rank == 0: 69 | eval_dataset = EvalDataLoader(hps.data.validation_files, hps) 70 | eval_loader = DataLoader(eval_dataset, num_workers=1, shuffle=False, 71 | batch_size=1, pin_memory=False, 72 | drop_last=False) 73 | 74 | net_g = SynthesizerTrn( 75 | hps.data.filter_length // 2 + 1, 76 | hps.train.segment_size // hps.data.hop_length, 77 | **hps.model).cuda(rank) 78 | net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank) 79 | optim_g = torch.optim.AdamW( 80 | net_g.parameters(), 81 | hps.train.learning_rate, 82 | betas=hps.train.betas, 83 | eps=hps.train.eps) 84 | optim_d = torch.optim.AdamW( 85 | net_d.parameters(), 86 | hps.train.learning_rate, 87 | betas=hps.train.betas, 88 | eps=hps.train.eps) 89 | net_g = DDP(net_g, device_ids=[rank]) # , find_unused_parameters=True) 90 | net_d = DDP(net_d, device_ids=[rank]) 91 | 92 | try: 93 | _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, 94 | optim_g) 95 | _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, 96 | optim_d) 97 | global_step = (epoch_str - 1) * len(train_loader) 98 | except: 99 | epoch_str = 1 100 | global_step = 0 101 | 102 | scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2) 103 | scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2) 104 | 105 | scaler = GradScaler(enabled=hps.train.fp16_run) 106 | 107 | for epoch in range(epoch_str, hps.train.epochs + 1): 108 | if rank == 0: 109 | train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, 110 | [train_loader, eval_loader], logger, [writer, writer_eval]) 111 | else: 112 | train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, 113 | [train_loader, None], None, None) 114 | scheduler_g.step() 115 | scheduler_d.step() 116 | 117 | 118 | def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers): 119 | net_g, net_d = nets 120 | optim_g, optim_d = optims 121 | scheduler_g, scheduler_d = schedulers 122 | train_loader, eval_loader = loaders 123 | if writers is not None: 124 | writer, writer_eval = writers 125 | 126 | # train_loader.batch_sampler.set_epoch(epoch) 127 | global global_step 128 | 129 | net_g.train() 130 | net_d.train() 131 | for batch_idx, items in enumerate(train_loader): 132 | c, f0, spec, y, spk = items 133 | g = spk.cuda(rank, non_blocking=True) 134 | spec, y = spec.cuda(rank, non_blocking=True), y.cuda(rank, non_blocking=True) 135 | c = c.cuda(rank, non_blocking=True) 136 | f0 = f0.cuda(rank, non_blocking=True) 137 | mel = spec_to_mel_torch( 138 | spec, 139 | hps.data.filter_length, 140 | hps.data.n_mel_channels, 141 | hps.data.sampling_rate, 142 | hps.data.mel_fmin, 143 | hps.data.mel_fmax) 144 | 145 | with autocast(enabled=hps.train.fp16_run): 146 | y_hat, ids_slice, z_mask, \ 147 | (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(c, f0, spec, g=g, mel=mel) 148 | 149 | y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length) 150 | y_hat_mel = mel_spectrogram_torch( 151 | y_hat.squeeze(1), 152 | hps.data.filter_length, 153 | hps.data.n_mel_channels, 154 | hps.data.sampling_rate, 155 | hps.data.hop_length, 156 | hps.data.win_length, 157 | hps.data.mel_fmin, 158 | hps.data.mel_fmax 159 | ) 160 | y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice 161 | 162 | # Discriminator 163 | y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach()) 164 | 165 | with autocast(enabled=False): 166 | loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g) 167 | loss_disc_all = loss_disc 168 | 169 | optim_d.zero_grad() 170 | scaler.scale(loss_disc_all).backward() 171 | scaler.unscale_(optim_d) 172 | grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None) 173 | scaler.step(optim_d) 174 | 175 | with autocast(enabled=hps.train.fp16_run): 176 | # Generator 177 | y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat) 178 | with autocast(enabled=False): 179 | loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel 180 | loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl 181 | loss_fm = feature_loss(fmap_r, fmap_g) 182 | loss_gen, losses_gen = generator_loss(y_d_hat_g) 183 | loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl 184 | optim_g.zero_grad() 185 | scaler.scale(loss_gen_all).backward() 186 | scaler.unscale_(optim_g) 187 | grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None) 188 | scaler.step(optim_g) 189 | scaler.update() 190 | 191 | if rank == 0: 192 | if global_step % hps.train.log_interval == 0: 193 | lr = optim_g.param_groups[0]['lr'] 194 | losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_kl] 195 | logger.info('Train Epoch: {} [{:.0f}%]'.format( 196 | epoch, 197 | 100. * batch_idx / len(train_loader))) 198 | logger.info([x.item() for x in losses] + [global_step, lr]) 199 | 200 | scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr, 201 | "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g} 202 | scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/kl": loss_kl}) 203 | 204 | scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}) 205 | scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}) 206 | scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}) 207 | image_dict = { 208 | "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()), 209 | "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()), 210 | "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()), 211 | } 212 | 213 | utils.summarize( 214 | writer=writer, 215 | global_step=global_step, 216 | images=image_dict, 217 | scalars=scalar_dict 218 | ) 219 | 220 | if global_step % hps.train.eval_interval == 0: 221 | evaluate(hps, net_g, eval_loader, writer_eval) 222 | utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch, 223 | os.path.join(hps.model_dir, "G_{}.pth".format(global_step))) 224 | utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch, 225 | os.path.join(hps.model_dir, "D_{}.pth".format(global_step))) 226 | 227 | keep_ckpts = getattr(hps.train, 'keep_ckpts', 0) 228 | if keep_ckpts > 0: 229 | utils.clean_checkpoints(path_to_models='logs/32k/', n_ckpts_to_keep=keep_ckpts, sort_by_time=True) 230 | 231 | 232 | global_step += 1 233 | 234 | if rank == 0: 235 | logger.info('====> Epoch: {}'.format(epoch)) 236 | 237 | 238 | def evaluate(hps, generator, eval_loader, writer_eval): 239 | generator.eval() 240 | image_dict = {} 241 | audio_dict = {} 242 | with torch.no_grad(): 243 | for batch_idx, items in enumerate(eval_loader): 244 | c, f0, spec, y, spk = items 245 | g = spk[:1].cuda(0) 246 | spec, y = spec[:1].cuda(0), y[:1].cuda(0) 247 | c = c[:1].cuda(0) 248 | f0 = f0[:1].cuda(0) 249 | mel = spec_to_mel_torch( 250 | spec, 251 | hps.data.filter_length, 252 | hps.data.n_mel_channels, 253 | hps.data.sampling_rate, 254 | hps.data.mel_fmin, 255 | hps.data.mel_fmax) 256 | y_hat = generator.module.infer(c, f0, g=g, mel=mel) 257 | 258 | y_hat_mel = mel_spectrogram_torch( 259 | y_hat.squeeze(1).float(), 260 | hps.data.filter_length, 261 | hps.data.n_mel_channels, 262 | hps.data.sampling_rate, 263 | hps.data.hop_length, 264 | hps.data.win_length, 265 | hps.data.mel_fmin, 266 | hps.data.mel_fmax 267 | ) 268 | 269 | audio_dict.update({ 270 | f"gen/audio_{batch_idx}": y_hat[0], 271 | f"gt/audio_{batch_idx}": y[0] 272 | }) 273 | image_dict.update({ 274 | f"gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy()), 275 | "gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy()) 276 | }) 277 | utils.summarize( 278 | writer=writer_eval, 279 | global_step=global_step, 280 | images=image_dict, 281 | audios=audio_dict, 282 | audio_sampling_rate=hps.data.sampling_rate 283 | ) 284 | generator.train() 285 | 286 | 287 | if __name__ == "__main__": 288 | main() 289 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import glob 3 | import re 4 | import sys 5 | import argparse 6 | import logging 7 | import json 8 | import subprocess 9 | 10 | import librosa 11 | import numpy as np 12 | import torchaudio 13 | from scipy.io.wavfile import read 14 | import torch 15 | import torchvision 16 | from torch.nn import functional as F 17 | from commons import sequence_mask 18 | from hubert import hubert_model 19 | MATPLOTLIB_FLAG = False 20 | 21 | logging.basicConfig(stream=sys.stdout, level=logging.DEBUG) 22 | logger = logging 23 | 24 | f0_bin = 256 25 | f0_max = 1100.0 26 | f0_min = 50.0 27 | f0_mel_min = 1127 * np.log(1 + f0_min / 700) 28 | f0_mel_max = 1127 * np.log(1 + f0_max / 700) 29 | 30 | def f0_to_coarse(f0): 31 | is_torch = isinstance(f0, torch.Tensor) 32 | f0_mel = 1127 * (1 + f0 / 700).log() if is_torch else 1127 * np.log(1 + f0 / 700) 33 | f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * (f0_bin - 2) / (f0_mel_max - f0_mel_min) + 1 34 | 35 | f0_mel[f0_mel <= 1] = 1 36 | f0_mel[f0_mel > f0_bin - 1] = f0_bin - 1 37 | f0_coarse = (f0_mel + 0.5).long() if is_torch else np.rint(f0_mel).astype(np.int) 38 | assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (f0_coarse.max(), f0_coarse.min()) 39 | return f0_coarse 40 | 41 | 42 | def get_hubert_model(rank=None): 43 | 44 | hubert_soft = hubert_model.hubert_soft("hubert/hubert-soft-0d54a1f4.pt") 45 | if rank is not None: 46 | hubert_soft = hubert_soft.cuda(rank) 47 | return hubert_soft 48 | 49 | def get_hubert_content(hmodel, y=None, path=None): 50 | if path is not None: 51 | source, sr = torchaudio.load(path) 52 | source = torchaudio.functional.resample(source, sr, 16000) 53 | if len(source.shape) == 2 and source.shape[1] >= 2: 54 | source = torch.mean(source, dim=0).unsqueeze(0) 55 | else: 56 | source = y 57 | source = source.unsqueeze(0) 58 | with torch.inference_mode(): 59 | units = hmodel.units(source) 60 | return units.transpose(1,2) 61 | 62 | 63 | def get_content(cmodel, y): 64 | with torch.no_grad(): 65 | c = cmodel.extract_features(y.squeeze(1))[0] 66 | c = c.transpose(1, 2) 67 | return c 68 | 69 | 70 | 71 | def transform(mel, height): # 68-92 72 | #r = np.random.random() 73 | #rate = r * 0.3 + 0.85 # 0.85-1.15 74 | #height = int(mel.size(-2) * rate) 75 | tgt = torchvision.transforms.functional.resize(mel, (height, mel.size(-1))) 76 | if height >= mel.size(-2): 77 | return tgt[:, :mel.size(-2), :] 78 | else: 79 | silence = tgt[:,-1:,:].repeat(1,mel.size(-2)-height,1) 80 | silence += torch.randn_like(silence) / 10 81 | return torch.cat((tgt, silence), 1) 82 | 83 | 84 | def stretch(mel, width): # 0.5-2 85 | return torchvision.transforms.functional.resize(mel, (mel.size(-2), width)) 86 | 87 | 88 | def load_checkpoint(checkpoint_path, model, optimizer=None): 89 | assert os.path.isfile(checkpoint_path) 90 | checkpoint_dict = torch.load(checkpoint_path, map_location='cpu') 91 | iteration = checkpoint_dict['iteration'] 92 | learning_rate = checkpoint_dict['learning_rate'] 93 | if iteration is None: 94 | iteration = 1 95 | if learning_rate is None: 96 | learning_rate = 0.0002 97 | if optimizer is not None and checkpoint_dict['optimizer'] is not None: 98 | optimizer.load_state_dict(checkpoint_dict['optimizer']) 99 | saved_state_dict = checkpoint_dict['model'] 100 | if hasattr(model, 'module'): 101 | state_dict = model.module.state_dict() 102 | else: 103 | state_dict = model.state_dict() 104 | new_state_dict= {} 105 | for k, v in state_dict.items(): 106 | try: 107 | new_state_dict[k] = saved_state_dict[k] 108 | except: 109 | logger.info("%s is not in the checkpoint" % k) 110 | new_state_dict[k] = v 111 | if hasattr(model, 'module'): 112 | model.module.load_state_dict(new_state_dict) 113 | else: 114 | model.load_state_dict(new_state_dict) 115 | logger.info("Loaded checkpoint '{}' (iteration {})" .format( 116 | checkpoint_path, iteration)) 117 | return model, optimizer, learning_rate, iteration 118 | 119 | 120 | def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path): 121 | logger.info("Saving model and optimizer state at iteration {} to {}".format( 122 | iteration, checkpoint_path)) 123 | if hasattr(model, 'module'): 124 | state_dict = model.module.state_dict() 125 | else: 126 | state_dict = model.state_dict() 127 | torch.save({'model': state_dict, 128 | 'iteration': iteration, 129 | 'optimizer': optimizer.state_dict(), 130 | 'learning_rate': learning_rate}, checkpoint_path) 131 | # clean_ckpt = False 132 | # if clean_ckpt: 133 | # clean_checkpoints(path_to_models='logs/32k/', n_ckpts_to_keep=3, sort_by_time=True) 134 | 135 | def clean_checkpoints(path_to_models='logs/48k/', n_ckpts_to_keep=2, sort_by_time=True): 136 | """Freeing up space by deleting saved ckpts 137 | 138 | Arguments: 139 | path_to_models -- Path to the model directory 140 | n_ckpts_to_keep -- Number of ckpts to keep, excluding G_0.pth and D_0.pth 141 | sort_by_time -- True -> chronologically delete ckpts 142 | False -> lexicographically delete ckpts 143 | """ 144 | ckpts_files = [f for f in os.listdir(path_to_models) if os.path.isfile(os.path.join(path_to_models, f))] 145 | name_key = (lambda _f: int(re.compile('._(\d+)\.pth').match(_f).group(1))) 146 | time_key = (lambda _f: os.path.getmtime(os.path.join(path_to_models, _f))) 147 | sort_key = time_key if sort_by_time else name_key 148 | x_sorted = lambda _x: sorted([f for f in ckpts_files if f.startswith(_x) and not f.endswith('_0.pth')], key=sort_key) 149 | to_del = [os.path.join(path_to_models, fn) for fn in 150 | (x_sorted('G')[:-n_ckpts_to_keep] + x_sorted('D')[:-n_ckpts_to_keep])] 151 | del_info = lambda fn: logger.info(f".. Free up space by deleting ckpt {fn}") 152 | del_routine = lambda x: [os.remove(x), del_info(x)] 153 | rs = [del_routine(fn) for fn in to_del] 154 | 155 | def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050): 156 | for k, v in scalars.items(): 157 | writer.add_scalar(k, v, global_step) 158 | for k, v in histograms.items(): 159 | writer.add_histogram(k, v, global_step) 160 | for k, v in images.items(): 161 | writer.add_image(k, v, global_step, dataformats='HWC') 162 | for k, v in audios.items(): 163 | writer.add_audio(k, v, global_step, audio_sampling_rate) 164 | 165 | 166 | def latest_checkpoint_path(dir_path, regex="G_*.pth"): 167 | f_list = glob.glob(os.path.join(dir_path, regex)) 168 | f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f)))) 169 | x = f_list[-1] 170 | print(x) 171 | return x 172 | 173 | 174 | def plot_spectrogram_to_numpy(spectrogram): 175 | global MATPLOTLIB_FLAG 176 | if not MATPLOTLIB_FLAG: 177 | import matplotlib 178 | matplotlib.use("Agg") 179 | MATPLOTLIB_FLAG = True 180 | mpl_logger = logging.getLogger('matplotlib') 181 | mpl_logger.setLevel(logging.WARNING) 182 | import matplotlib.pylab as plt 183 | import numpy as np 184 | 185 | fig, ax = plt.subplots(figsize=(10,2)) 186 | im = ax.imshow(spectrogram, aspect="auto", origin="lower", 187 | interpolation='none') 188 | plt.colorbar(im, ax=ax) 189 | plt.xlabel("Frames") 190 | plt.ylabel("Channels") 191 | plt.tight_layout() 192 | 193 | fig.canvas.draw() 194 | data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='') 195 | data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,)) 196 | plt.close() 197 | return data 198 | 199 | 200 | def plot_alignment_to_numpy(alignment, info=None): 201 | global MATPLOTLIB_FLAG 202 | if not MATPLOTLIB_FLAG: 203 | import matplotlib 204 | matplotlib.use("Agg") 205 | MATPLOTLIB_FLAG = True 206 | mpl_logger = logging.getLogger('matplotlib') 207 | mpl_logger.setLevel(logging.WARNING) 208 | import matplotlib.pylab as plt 209 | import numpy as np 210 | 211 | fig, ax = plt.subplots(figsize=(6, 4)) 212 | im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower', 213 | interpolation='none') 214 | fig.colorbar(im, ax=ax) 215 | xlabel = 'Decoder timestep' 216 | if info is not None: 217 | xlabel += '\n\n' + info 218 | plt.xlabel(xlabel) 219 | plt.ylabel('Encoder timestep') 220 | plt.tight_layout() 221 | 222 | fig.canvas.draw() 223 | data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='') 224 | data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,)) 225 | plt.close() 226 | return data 227 | 228 | 229 | def load_wav_to_torch(full_path): 230 | sampling_rate, data = read(full_path) 231 | return torch.FloatTensor(data.astype(np.float32)), sampling_rate 232 | 233 | 234 | def load_filepaths_and_text(filename, split="|"): 235 | with open(filename, encoding='utf-8') as f: 236 | filepaths_and_text = [line.strip().split(split) for line in f] 237 | return filepaths_and_text 238 | 239 | 240 | def get_hparams(init=True): 241 | parser = argparse.ArgumentParser() 242 | parser.add_argument('-c', '--config', type=str, default="./configs/base.json", 243 | help='JSON file for configuration') 244 | parser.add_argument('-m', '--model', type=str, required=True, 245 | help='Model name') 246 | 247 | args = parser.parse_args() 248 | model_dir = os.path.join("./logs", args.model) 249 | 250 | if not os.path.exists(model_dir): 251 | os.makedirs(model_dir) 252 | 253 | config_path = args.config 254 | config_save_path = os.path.join(model_dir, "config.json") 255 | if init: 256 | with open(config_path, "r") as f: 257 | data = f.read() 258 | with open(config_save_path, "w") as f: 259 | f.write(data) 260 | else: 261 | with open(config_save_path, "r") as f: 262 | data = f.read() 263 | config = json.loads(data) 264 | 265 | hparams = HParams(**config) 266 | hparams.model_dir = model_dir 267 | return hparams 268 | 269 | 270 | def get_hparams_from_dir(model_dir): 271 | config_save_path = os.path.join(model_dir, "config.json") 272 | with open(config_save_path, "r") as f: 273 | data = f.read() 274 | config = json.loads(data) 275 | 276 | hparams =HParams(**config) 277 | hparams.model_dir = model_dir 278 | return hparams 279 | 280 | 281 | def get_hparams_from_file(config_path): 282 | with open(config_path, "r") as f: 283 | data = f.read() 284 | config = json.loads(data) 285 | 286 | hparams =HParams(**config) 287 | return hparams 288 | 289 | 290 | def check_git_hash(model_dir): 291 | source_dir = os.path.dirname(os.path.realpath(__file__)) 292 | if not os.path.exists(os.path.join(source_dir, ".git")): 293 | logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format( 294 | source_dir 295 | )) 296 | return 297 | 298 | cur_hash = subprocess.getoutput("git rev-parse HEAD") 299 | 300 | path = os.path.join(model_dir, "githash") 301 | if os.path.exists(path): 302 | saved_hash = open(path).read() 303 | if saved_hash != cur_hash: 304 | logger.warn("git hash values are different. {}(saved) != {}(current)".format( 305 | saved_hash[:8], cur_hash[:8])) 306 | else: 307 | open(path, "w").write(cur_hash) 308 | 309 | 310 | def get_logger(model_dir, filename="train.log"): 311 | global logger 312 | logger = logging.getLogger(os.path.basename(model_dir)) 313 | logger.setLevel(logging.DEBUG) 314 | 315 | formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s") 316 | if not os.path.exists(model_dir): 317 | os.makedirs(model_dir) 318 | h = logging.FileHandler(os.path.join(model_dir, filename)) 319 | h.setLevel(logging.DEBUG) 320 | h.setFormatter(formatter) 321 | logger.addHandler(h) 322 | return logger 323 | 324 | 325 | class HParams(): 326 | def __init__(self, **kwargs): 327 | for k, v in kwargs.items(): 328 | if type(v) == dict: 329 | v = HParams(**v) 330 | self[k] = v 331 | 332 | def keys(self): 333 | return self.__dict__.keys() 334 | 335 | def items(self): 336 | return self.__dict__.items() 337 | 338 | def values(self): 339 | return self.__dict__.values() 340 | 341 | def __len__(self): 342 | return len(self.__dict__) 343 | 344 | def __getitem__(self, key): 345 | return getattr(self, key) 346 | 347 | def __setitem__(self, key, value): 348 | return setattr(self, key, value) 349 | 350 | def __contains__(self, key): 351 | return key in self.__dict__ 352 | 353 | def __repr__(self): 354 | return self.__dict__.__repr__() 355 | 356 | -------------------------------------------------------------------------------- /vdecoder/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/innnky/so-vits-svc/87939fb057b6b6231c465d59433fa9da0b45243c/vdecoder/__init__.py -------------------------------------------------------------------------------- /vdecoder/hifigan/env.py: -------------------------------------------------------------------------------- 1 | import os 2 | import shutil 3 | 4 | 5 | class AttrDict(dict): 6 | def __init__(self, *args, **kwargs): 7 | super(AttrDict, self).__init__(*args, **kwargs) 8 | self.__dict__ = self 9 | 10 | 11 | def build_env(config, config_name, path): 12 | t_path = os.path.join(path, config_name) 13 | if config != t_path: 14 | os.makedirs(path, exist_ok=True) 15 | shutil.copyfile(config, os.path.join(path, config_name)) 16 | -------------------------------------------------------------------------------- /vdecoder/hifigan/models.py: -------------------------------------------------------------------------------- 1 | import os 2 | import json 3 | from .env import AttrDict 4 | import numpy as np 5 | import torch 6 | import torch.nn.functional as F 7 | import torch.nn as nn 8 | from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d 9 | from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm 10 | from .utils import init_weights, get_padding 11 | 12 | LRELU_SLOPE = 0.1 13 | 14 | 15 | def load_model(model_path, device='cuda'): 16 | config_file = os.path.join(os.path.split(model_path)[0], 'config.json') 17 | with open(config_file) as f: 18 | data = f.read() 19 | 20 | global h 21 | json_config = json.loads(data) 22 | h = AttrDict(json_config) 23 | 24 | generator = Generator(h).to(device) 25 | 26 | cp_dict = torch.load(model_path) 27 | generator.load_state_dict(cp_dict['generator']) 28 | generator.eval() 29 | generator.remove_weight_norm() 30 | del cp_dict 31 | return generator, h 32 | 33 | 34 | class ResBlock1(torch.nn.Module): 35 | def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)): 36 | super(ResBlock1, self).__init__() 37 | self.h = h 38 | self.convs1 = nn.ModuleList([ 39 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0], 40 | padding=get_padding(kernel_size, dilation[0]))), 41 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1], 42 | padding=get_padding(kernel_size, dilation[1]))), 43 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2], 44 | padding=get_padding(kernel_size, dilation[2]))) 45 | ]) 46 | self.convs1.apply(init_weights) 47 | 48 | self.convs2 = nn.ModuleList([ 49 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, 50 | padding=get_padding(kernel_size, 1))), 51 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, 52 | padding=get_padding(kernel_size, 1))), 53 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, 54 | padding=get_padding(kernel_size, 1))) 55 | ]) 56 | self.convs2.apply(init_weights) 57 | 58 | def forward(self, x): 59 | for c1, c2 in zip(self.convs1, self.convs2): 60 | xt = F.leaky_relu(x, LRELU_SLOPE) 61 | xt = c1(xt) 62 | xt = F.leaky_relu(xt, LRELU_SLOPE) 63 | xt = c2(xt) 64 | x = xt + x 65 | return x 66 | 67 | def remove_weight_norm(self): 68 | for l in self.convs1: 69 | remove_weight_norm(l) 70 | for l in self.convs2: 71 | remove_weight_norm(l) 72 | 73 | 74 | class ResBlock2(torch.nn.Module): 75 | def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)): 76 | super(ResBlock2, self).__init__() 77 | self.h = h 78 | self.convs = nn.ModuleList([ 79 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0], 80 | padding=get_padding(kernel_size, dilation[0]))), 81 | weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1], 82 | padding=get_padding(kernel_size, dilation[1]))) 83 | ]) 84 | self.convs.apply(init_weights) 85 | 86 | def forward(self, x): 87 | for c in self.convs: 88 | xt = F.leaky_relu(x, LRELU_SLOPE) 89 | xt = c(xt) 90 | x = xt + x 91 | return x 92 | 93 | def remove_weight_norm(self): 94 | for l in self.convs: 95 | remove_weight_norm(l) 96 | 97 | 98 | def padDiff(x): 99 | return F.pad(F.pad(x, (0,0,-1,1), 'constant', 0) - x, (0,0,0,-1), 'constant', 0) 100 | 101 | class SineGen(torch.nn.Module): 102 | """ Definition of sine generator 103 | SineGen(samp_rate, harmonic_num = 0, 104 | sine_amp = 0.1, noise_std = 0.003, 105 | voiced_threshold = 0, 106 | flag_for_pulse=False) 107 | samp_rate: sampling rate in Hz 108 | harmonic_num: number of harmonic overtones (default 0) 109 | sine_amp: amplitude of sine-wavefrom (default 0.1) 110 | noise_std: std of Gaussian noise (default 0.003) 111 | voiced_thoreshold: F0 threshold for U/V classification (default 0) 112 | flag_for_pulse: this SinGen is used inside PulseGen (default False) 113 | Note: when flag_for_pulse is True, the first time step of a voiced 114 | segment is always sin(np.pi) or cos(0) 115 | """ 116 | 117 | def __init__(self, samp_rate, harmonic_num=0, 118 | sine_amp=0.1, noise_std=0.003, 119 | voiced_threshold=0, 120 | flag_for_pulse=False): 121 | super(SineGen, self).__init__() 122 | self.sine_amp = sine_amp 123 | self.noise_std = noise_std 124 | self.harmonic_num = harmonic_num 125 | self.dim = self.harmonic_num + 1 126 | self.sampling_rate = samp_rate 127 | self.voiced_threshold = voiced_threshold 128 | self.flag_for_pulse = flag_for_pulse 129 | 130 | def _f02uv(self, f0): 131 | # generate uv signal 132 | uv = (f0 > self.voiced_threshold).type(torch.float32) 133 | return uv 134 | 135 | def _f02sine(self, f0_values): 136 | """ f0_values: (batchsize, length, dim) 137 | where dim indicates fundamental tone and overtones 138 | """ 139 | # convert to F0 in rad. The interger part n can be ignored 140 | # because 2 * np.pi * n doesn't affect phase 141 | rad_values = (f0_values / self.sampling_rate) % 1 142 | 143 | # initial phase noise (no noise for fundamental component) 144 | rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], \ 145 | device=f0_values.device) 146 | rand_ini[:, 0] = 0 147 | rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini 148 | 149 | # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad) 150 | if not self.flag_for_pulse: 151 | # for normal case 152 | 153 | # To prevent torch.cumsum numerical overflow, 154 | # it is necessary to add -1 whenever \sum_k=1^n rad_value_k > 1. 155 | # Buffer tmp_over_one_idx indicates the time step to add -1. 156 | # This will not change F0 of sine because (x-1) * 2*pi = x * 2*pi 157 | tmp_over_one = torch.cumsum(rad_values, 1) % 1 158 | tmp_over_one_idx = (padDiff(tmp_over_one)) < 0 159 | cumsum_shift = torch.zeros_like(rad_values) 160 | cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0 161 | 162 | sines = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) 163 | * 2 * np.pi) 164 | else: 165 | # If necessary, make sure that the first time step of every 166 | # voiced segments is sin(pi) or cos(0) 167 | # This is used for pulse-train generation 168 | 169 | # identify the last time step in unvoiced segments 170 | uv = self._f02uv(f0_values) 171 | uv_1 = torch.roll(uv, shifts=-1, dims=1) 172 | uv_1[:, -1, :] = 1 173 | u_loc = (uv < 1) * (uv_1 > 0) 174 | 175 | # get the instantanouse phase 176 | tmp_cumsum = torch.cumsum(rad_values, dim=1) 177 | # different batch needs to be processed differently 178 | for idx in range(f0_values.shape[0]): 179 | temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :] 180 | temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :] 181 | # stores the accumulation of i.phase within 182 | # each voiced segments 183 | tmp_cumsum[idx, :, :] = 0 184 | tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum 185 | 186 | # rad_values - tmp_cumsum: remove the accumulation of i.phase 187 | # within the previous voiced segment. 188 | i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1) 189 | 190 | # get the sines 191 | sines = torch.cos(i_phase * 2 * np.pi) 192 | return sines 193 | 194 | def forward(self, f0): 195 | """ sine_tensor, uv = forward(f0) 196 | input F0: tensor(batchsize=1, length, dim=1) 197 | f0 for unvoiced steps should be 0 198 | output sine_tensor: tensor(batchsize=1, length, dim) 199 | output uv: tensor(batchsize=1, length, 1) 200 | """ 201 | with torch.no_grad(): 202 | f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, 203 | device=f0.device) 204 | # fundamental component 205 | fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device)) 206 | 207 | # generate sine waveforms 208 | sine_waves = self._f02sine(fn) * self.sine_amp 209 | 210 | # generate uv signal 211 | # uv = torch.ones(f0.shape) 212 | # uv = uv * (f0 > self.voiced_threshold) 213 | uv = self._f02uv(f0) 214 | 215 | # noise: for unvoiced should be similar to sine_amp 216 | # std = self.sine_amp/3 -> max value ~ self.sine_amp 217 | # . for voiced regions is self.noise_std 218 | noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3 219 | noise = noise_amp * torch.randn_like(sine_waves) 220 | 221 | # first: set the unvoiced part to 0 by uv 222 | # then: additive noise 223 | sine_waves = sine_waves * uv + noise 224 | return sine_waves, uv, noise 225 | 226 | 227 | class SourceModuleHnNSF(torch.nn.Module): 228 | """ SourceModule for hn-nsf 229 | SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1, 230 | add_noise_std=0.003, voiced_threshod=0) 231 | sampling_rate: sampling_rate in Hz 232 | harmonic_num: number of harmonic above F0 (default: 0) 233 | sine_amp: amplitude of sine source signal (default: 0.1) 234 | add_noise_std: std of additive Gaussian noise (default: 0.003) 235 | note that amplitude of noise in unvoiced is decided 236 | by sine_amp 237 | voiced_threshold: threhold to set U/V given F0 (default: 0) 238 | Sine_source, noise_source = SourceModuleHnNSF(F0_sampled) 239 | F0_sampled (batchsize, length, 1) 240 | Sine_source (batchsize, length, 1) 241 | noise_source (batchsize, length 1) 242 | uv (batchsize, length, 1) 243 | """ 244 | 245 | def __init__(self, sampling_rate, harmonic_num=0, sine_amp=0.1, 246 | add_noise_std=0.003, voiced_threshod=0): 247 | super(SourceModuleHnNSF, self).__init__() 248 | 249 | self.sine_amp = sine_amp 250 | self.noise_std = add_noise_std 251 | 252 | # to produce sine waveforms 253 | self.l_sin_gen = SineGen(sampling_rate, harmonic_num, 254 | sine_amp, add_noise_std, voiced_threshod) 255 | 256 | # to merge source harmonics into a single excitation 257 | self.l_linear = torch.nn.Linear(harmonic_num + 1, 1) 258 | self.l_tanh = torch.nn.Tanh() 259 | 260 | def forward(self, x): 261 | """ 262 | Sine_source, noise_source = SourceModuleHnNSF(F0_sampled) 263 | F0_sampled (batchsize, length, 1) 264 | Sine_source (batchsize, length, 1) 265 | noise_source (batchsize, length 1) 266 | """ 267 | # source for harmonic branch 268 | sine_wavs, uv, _ = self.l_sin_gen(x) 269 | sine_merge = self.l_tanh(self.l_linear(sine_wavs)) 270 | 271 | # source for noise branch, in the same shape as uv 272 | noise = torch.randn_like(uv) * self.sine_amp / 3 273 | return sine_merge, noise, uv 274 | 275 | 276 | class Generator(torch.nn.Module): 277 | def __init__(self, h): 278 | super(Generator, self).__init__() 279 | self.h = h 280 | 281 | self.num_kernels = len(h["resblock_kernel_sizes"]) 282 | self.num_upsamples = len(h["upsample_rates"]) 283 | self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(h["upsample_rates"])) 284 | self.m_source = SourceModuleHnNSF( 285 | sampling_rate=h["sampling_rate"], 286 | harmonic_num=8) 287 | self.noise_convs = nn.ModuleList() 288 | self.conv_pre = weight_norm(Conv1d(h["inter_channels"], h["upsample_initial_channel"], 7, 1, padding=3)) 289 | resblock = ResBlock1 if h["resblock"] == '1' else ResBlock2 290 | self.ups = nn.ModuleList() 291 | for i, (u, k) in enumerate(zip(h["upsample_rates"], h["upsample_kernel_sizes"])): 292 | c_cur = h["upsample_initial_channel"] // (2 ** (i + 1)) 293 | self.ups.append(weight_norm( 294 | ConvTranspose1d(h["upsample_initial_channel"] // (2 ** i), h["upsample_initial_channel"] // (2 ** (i + 1)), 295 | k, u, padding=(k - u) // 2))) 296 | if i + 1 < len(h["upsample_rates"]): # 297 | stride_f0 = np.prod(h["upsample_rates"][i + 1:]) 298 | self.noise_convs.append(Conv1d( 299 | 1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=stride_f0 // 2)) 300 | else: 301 | self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1)) 302 | self.resblocks = nn.ModuleList() 303 | for i in range(len(self.ups)): 304 | ch = h["upsample_initial_channel"] // (2 ** (i + 1)) 305 | for j, (k, d) in enumerate(zip(h["resblock_kernel_sizes"], h["resblock_dilation_sizes"])): 306 | self.resblocks.append(resblock(h, ch, k, d)) 307 | 308 | self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3)) 309 | self.ups.apply(init_weights) 310 | self.conv_post.apply(init_weights) 311 | self.cond = nn.Conv1d(h['gin_channels'], h['upsample_initial_channel'], 1) 312 | 313 | def forward(self, x, f0, g=None): 314 | # print(1,x.shape,f0.shape,f0[:, None].shape) 315 | f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2) # bs,n,t 316 | # print(2,f0.shape) 317 | har_source, noi_source, uv = self.m_source(f0) 318 | har_source = har_source.transpose(1, 2) 319 | x = self.conv_pre(x) 320 | x = x + self.cond(g) 321 | # print(124,x.shape,har_source.shape) 322 | for i in range(self.num_upsamples): 323 | x = F.leaky_relu(x, LRELU_SLOPE) 324 | # print(3,x.shape) 325 | x = self.ups[i](x) 326 | x_source = self.noise_convs[i](har_source) 327 | # print(4,x_source.shape,har_source.shape,x.shape) 328 | x = x + x_source 329 | xs = None 330 | for j in range(self.num_kernels): 331 | if xs is None: 332 | xs = self.resblocks[i * self.num_kernels + j](x) 333 | else: 334 | xs += self.resblocks[i * self.num_kernels + j](x) 335 | x = xs / self.num_kernels 336 | x = F.leaky_relu(x) 337 | x = self.conv_post(x) 338 | x = torch.tanh(x) 339 | 340 | return x 341 | 342 | def remove_weight_norm(self): 343 | print('Removing weight norm...') 344 | for l in self.ups: 345 | remove_weight_norm(l) 346 | for l in self.resblocks: 347 | l.remove_weight_norm() 348 | remove_weight_norm(self.conv_pre) 349 | remove_weight_norm(self.conv_post) 350 | 351 | 352 | class DiscriminatorP(torch.nn.Module): 353 | def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): 354 | super(DiscriminatorP, self).__init__() 355 | self.period = period 356 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 357 | self.convs = nn.ModuleList([ 358 | norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))), 359 | norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))), 360 | norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))), 361 | norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))), 362 | norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))), 363 | ]) 364 | self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0))) 365 | 366 | def forward(self, x): 367 | fmap = [] 368 | 369 | # 1d to 2d 370 | b, c, t = x.shape 371 | if t % self.period != 0: # pad first 372 | n_pad = self.period - (t % self.period) 373 | x = F.pad(x, (0, n_pad), "reflect") 374 | t = t + n_pad 375 | x = x.view(b, c, t // self.period, self.period) 376 | 377 | for l in self.convs: 378 | x = l(x) 379 | x = F.leaky_relu(x, LRELU_SLOPE) 380 | fmap.append(x) 381 | x = self.conv_post(x) 382 | fmap.append(x) 383 | x = torch.flatten(x, 1, -1) 384 | 385 | return x, fmap 386 | 387 | 388 | class MultiPeriodDiscriminator(torch.nn.Module): 389 | def __init__(self, periods=None): 390 | super(MultiPeriodDiscriminator, self).__init__() 391 | self.periods = periods if periods is not None else [2, 3, 5, 7, 11] 392 | self.discriminators = nn.ModuleList() 393 | for period in self.periods: 394 | self.discriminators.append(DiscriminatorP(period)) 395 | 396 | def forward(self, y, y_hat): 397 | y_d_rs = [] 398 | y_d_gs = [] 399 | fmap_rs = [] 400 | fmap_gs = [] 401 | for i, d in enumerate(self.discriminators): 402 | y_d_r, fmap_r = d(y) 403 | y_d_g, fmap_g = d(y_hat) 404 | y_d_rs.append(y_d_r) 405 | fmap_rs.append(fmap_r) 406 | y_d_gs.append(y_d_g) 407 | fmap_gs.append(fmap_g) 408 | 409 | return y_d_rs, y_d_gs, fmap_rs, fmap_gs 410 | 411 | 412 | class DiscriminatorS(torch.nn.Module): 413 | def __init__(self, use_spectral_norm=False): 414 | super(DiscriminatorS, self).__init__() 415 | norm_f = weight_norm if use_spectral_norm == False else spectral_norm 416 | self.convs = nn.ModuleList([ 417 | norm_f(Conv1d(1, 128, 15, 1, padding=7)), 418 | norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)), 419 | norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)), 420 | norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)), 421 | norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)), 422 | norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)), 423 | norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), 424 | ]) 425 | self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) 426 | 427 | def forward(self, x): 428 | fmap = [] 429 | for l in self.convs: 430 | x = l(x) 431 | x = F.leaky_relu(x, LRELU_SLOPE) 432 | fmap.append(x) 433 | x = self.conv_post(x) 434 | fmap.append(x) 435 | x = torch.flatten(x, 1, -1) 436 | 437 | return x, fmap 438 | 439 | 440 | class MultiScaleDiscriminator(torch.nn.Module): 441 | def __init__(self): 442 | super(MultiScaleDiscriminator, self).__init__() 443 | self.discriminators = nn.ModuleList([ 444 | DiscriminatorS(use_spectral_norm=True), 445 | DiscriminatorS(), 446 | DiscriminatorS(), 447 | ]) 448 | self.meanpools = nn.ModuleList([ 449 | AvgPool1d(4, 2, padding=2), 450 | AvgPool1d(4, 2, padding=2) 451 | ]) 452 | 453 | def forward(self, y, y_hat): 454 | y_d_rs = [] 455 | y_d_gs = [] 456 | fmap_rs = [] 457 | fmap_gs = [] 458 | for i, d in enumerate(self.discriminators): 459 | if i != 0: 460 | y = self.meanpools[i - 1](y) 461 | y_hat = self.meanpools[i - 1](y_hat) 462 | y_d_r, fmap_r = d(y) 463 | y_d_g, fmap_g = d(y_hat) 464 | y_d_rs.append(y_d_r) 465 | fmap_rs.append(fmap_r) 466 | y_d_gs.append(y_d_g) 467 | fmap_gs.append(fmap_g) 468 | 469 | return y_d_rs, y_d_gs, fmap_rs, fmap_gs 470 | 471 | 472 | def feature_loss(fmap_r, fmap_g): 473 | loss = 0 474 | for dr, dg in zip(fmap_r, fmap_g): 475 | for rl, gl in zip(dr, dg): 476 | loss += torch.mean(torch.abs(rl - gl)) 477 | 478 | return loss * 2 479 | 480 | 481 | def discriminator_loss(disc_real_outputs, disc_generated_outputs): 482 | loss = 0 483 | r_losses = [] 484 | g_losses = [] 485 | for dr, dg in zip(disc_real_outputs, disc_generated_outputs): 486 | r_loss = torch.mean((1 - dr) ** 2) 487 | g_loss = torch.mean(dg ** 2) 488 | loss += (r_loss + g_loss) 489 | r_losses.append(r_loss.item()) 490 | g_losses.append(g_loss.item()) 491 | 492 | return loss, r_losses, g_losses 493 | 494 | 495 | def generator_loss(disc_outputs): 496 | loss = 0 497 | gen_losses = [] 498 | for dg in disc_outputs: 499 | l = torch.mean((1 - dg) ** 2) 500 | gen_losses.append(l) 501 | loss += l 502 | 503 | return loss, gen_losses 504 | -------------------------------------------------------------------------------- /vdecoder/hifigan/nvSTFT.py: -------------------------------------------------------------------------------- 1 | import math 2 | import os 3 | os.environ["LRU_CACHE_CAPACITY"] = "3" 4 | import random 5 | import torch 6 | import torch.utils.data 7 | import numpy as np 8 | import librosa 9 | from librosa.util import normalize 10 | from librosa.filters import mel as librosa_mel_fn 11 | from scipy.io.wavfile import read 12 | import soundfile as sf 13 | 14 | def load_wav_to_torch(full_path, target_sr=None, return_empty_on_exception=False): 15 | sampling_rate = None 16 | try: 17 | data, sampling_rate = sf.read(full_path, always_2d=True)# than soundfile. 18 | except Exception as ex: 19 | print(f"'{full_path}' failed to load.\nException:") 20 | print(ex) 21 | if return_empty_on_exception: 22 | return [], sampling_rate or target_sr or 32000 23 | else: 24 | raise Exception(ex) 25 | 26 | if len(data.shape) > 1: 27 | data = data[:, 0] 28 | assert len(data) > 2# check duration of audio file is > 2 samples (because otherwise the slice operation was on the wrong dimension) 29 | 30 | if np.issubdtype(data.dtype, np.integer): # if audio data is type int 31 | max_mag = -np.iinfo(data.dtype).min # maximum magnitude = min possible value of intXX 32 | else: # if audio data is type fp32 33 | max_mag = max(np.amax(data), -np.amin(data)) 34 | max_mag = (2**31)+1 if max_mag > (2**15) else ((2**15)+1 if max_mag > 1.01 else 1.0) # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32 35 | 36 | data = torch.FloatTensor(data.astype(np.float32))/max_mag 37 | 38 | if (torch.isinf(data) | torch.isnan(data)).any() and return_empty_on_exception:# resample will crash with inf/NaN inputs. return_empty_on_exception will return empty arr instead of except 39 | return [], sampling_rate or target_sr or 32000 40 | if target_sr is not None and sampling_rate != target_sr: 41 | data = torch.from_numpy(librosa.core.resample(data.numpy(), orig_sr=sampling_rate, target_sr=target_sr)) 42 | sampling_rate = target_sr 43 | 44 | return data, sampling_rate 45 | 46 | def dynamic_range_compression(x, C=1, clip_val=1e-5): 47 | return np.log(np.clip(x, a_min=clip_val, a_max=None) * C) 48 | 49 | def dynamic_range_decompression(x, C=1): 50 | return np.exp(x) / C 51 | 52 | def dynamic_range_compression_torch(x, C=1, clip_val=1e-5): 53 | return torch.log(torch.clamp(x, min=clip_val) * C) 54 | 55 | def dynamic_range_decompression_torch(x, C=1): 56 | return torch.exp(x) / C 57 | 58 | class STFT(): 59 | def __init__(self, sr=22050, n_mels=80, n_fft=1024, win_size=1024, hop_length=256, fmin=20, fmax=11025, clip_val=1e-5): 60 | self.target_sr = sr 61 | 62 | self.n_mels = n_mels 63 | self.n_fft = n_fft 64 | self.win_size = win_size 65 | self.hop_length = hop_length 66 | self.fmin = fmin 67 | self.fmax = fmax 68 | self.clip_val = clip_val 69 | self.mel_basis = {} 70 | self.hann_window = {} 71 | 72 | def get_mel(self, y, center=False): 73 | sampling_rate = self.target_sr 74 | n_mels = self.n_mels 75 | n_fft = self.n_fft 76 | win_size = self.win_size 77 | hop_length = self.hop_length 78 | fmin = self.fmin 79 | fmax = self.fmax 80 | clip_val = self.clip_val 81 | 82 | if torch.min(y) < -1.: 83 | print('min value is ', torch.min(y)) 84 | if torch.max(y) > 1.: 85 | print('max value is ', torch.max(y)) 86 | 87 | if fmax not in self.mel_basis: 88 | mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax) 89 | self.mel_basis[str(fmax)+'_'+str(y.device)] = torch.from_numpy(mel).float().to(y.device) 90 | self.hann_window[str(y.device)] = torch.hann_window(self.win_size).to(y.device) 91 | 92 | y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_length)/2), int((n_fft-hop_length)/2)), mode='reflect') 93 | y = y.squeeze(1) 94 | 95 | spec = torch.stft(y, n_fft, hop_length=hop_length, win_length=win_size, window=self.hann_window[str(y.device)], 96 | center=center, pad_mode='reflect', normalized=False, onesided=True) 97 | # print(111,spec) 98 | spec = torch.sqrt(spec.pow(2).sum(-1)+(1e-9)) 99 | # print(222,spec) 100 | spec = torch.matmul(self.mel_basis[str(fmax)+'_'+str(y.device)], spec) 101 | # print(333,spec) 102 | spec = dynamic_range_compression_torch(spec, clip_val=clip_val) 103 | # print(444,spec) 104 | return spec 105 | 106 | def __call__(self, audiopath): 107 | audio, sr = load_wav_to_torch(audiopath, target_sr=self.target_sr) 108 | spect = self.get_mel(audio.unsqueeze(0)).squeeze(0) 109 | return spect 110 | 111 | stft = STFT() 112 | -------------------------------------------------------------------------------- /vdecoder/hifigan/utils.py: -------------------------------------------------------------------------------- 1 | import glob 2 | import os 3 | import matplotlib 4 | import torch 5 | from torch.nn.utils import weight_norm 6 | matplotlib.use("Agg") 7 | import matplotlib.pylab as plt 8 | 9 | 10 | def plot_spectrogram(spectrogram): 11 | fig, ax = plt.subplots(figsize=(10, 2)) 12 | im = ax.imshow(spectrogram, aspect="auto", origin="lower", 13 | interpolation='none') 14 | plt.colorbar(im, ax=ax) 15 | 16 | fig.canvas.draw() 17 | plt.close() 18 | 19 | return fig 20 | 21 | 22 | def init_weights(m, mean=0.0, std=0.01): 23 | classname = m.__class__.__name__ 24 | if classname.find("Conv") != -1: 25 | m.weight.data.normal_(mean, std) 26 | 27 | 28 | def apply_weight_norm(m): 29 | classname = m.__class__.__name__ 30 | if classname.find("Conv") != -1: 31 | weight_norm(m) 32 | 33 | 34 | def get_padding(kernel_size, dilation=1): 35 | return int((kernel_size*dilation - dilation)/2) 36 | 37 | 38 | def load_checkpoint(filepath, device): 39 | assert os.path.isfile(filepath) 40 | print("Loading '{}'".format(filepath)) 41 | checkpoint_dict = torch.load(filepath, map_location=device) 42 | print("Complete.") 43 | return checkpoint_dict 44 | 45 | 46 | def save_checkpoint(filepath, obj): 47 | print("Saving checkpoint to {}".format(filepath)) 48 | torch.save(obj, filepath) 49 | print("Complete.") 50 | 51 | 52 | def del_old_checkpoints(cp_dir, prefix, n_models=2): 53 | pattern = os.path.join(cp_dir, prefix + '????????') 54 | cp_list = glob.glob(pattern) # get checkpoint paths 55 | cp_list = sorted(cp_list)# sort by iter 56 | if len(cp_list) > n_models: # if more than n_models models are found 57 | for cp in cp_list[:-n_models]:# delete the oldest models other than lastest n_models 58 | open(cp, 'w').close()# empty file contents 59 | os.unlink(cp)# delete file (move to trash when using Colab) 60 | 61 | 62 | def scan_checkpoint(cp_dir, prefix): 63 | pattern = os.path.join(cp_dir, prefix + '????????') 64 | cp_list = glob.glob(pattern) 65 | if len(cp_list) == 0: 66 | return None 67 | return sorted(cp_list)[-1] 68 | 69 | --------------------------------------------------------------------------------