云端炼丹,算力白嫖,基于云端GPU(Colab)使用So-vits库制作AI特朗普演唱《国际歌》

人工智能AI技术早已深入到人们生活的每一个角落，君不见AI孙燕姿的歌声此起彼伏，不绝于耳，但并不是每个人都拥有一块N卡，没有GPU的日子总是不好过的，但是没关系，山人有妙计，本次我们基于Google的Colab免费云端服务器来搭建深度学习环境，制作AI特朗普，让他高唱《国际歌》。

Colab（全名Colaboratory ），它是Google公司的一款基于云端的基础免费服务器产品，可以在B端，也就是浏览器里面编写和执行Python代码，非常方便，贴心的是，Colab可以给用户分配免费的GPU进行使用，对于没有N卡的朋友来说，这已经远远超出了业界良心的范畴，简直就是在做慈善事业。

配置Colab

Colab是基于Google云盘的产品，我们可以将深度学习的Python脚本、训练好的模型、以及训练集等数据直接存放在云盘中，然后通过Colab执行即可。

首先访问Google云盘：drive.google.com

随后点击新建，选择关联更多应用：

接着安装Colab即可：

至此，云盘和Colab就关联好了，现在我们可以新建一个脚本文件my_sovits.ipynb文件，键入代码：

hello colab

随后，按快捷键 ctrl + 回车，即可运行代码：

这里需要注意的是，Colab使用的是基于Jupyter Notebook的ipynb格式的Python代码。

Jupyter Notebook是以网页的形式打开，可以在网页页面中直接编写代码和运行代码，代码的运行结果也会直接在代码块下显示。如在编程过程中需要编写说明文档，可在同一个页面中直接编写，便于作及时的说明和解释。

随后设置一下显卡类型：

接着运行命令，查看GPU版本：

!/usr/local/cuda/bin/nvcc --version  

!nvidia-smi

程序返回：

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:33:58_PDT_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0
Tue May 16 04:49:23 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.85.12    Driver Version: 525.85.12    CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            Off  | 00000000:00:04.0 Off |                    0 |
| N/A   65C    P8    13W /  70W |      0MiB / 15360MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+  

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

这里建议选择Tesla T4的显卡类型，性能更突出。

至此Colab就配置好了。

配置So-vits

下面我们配置so-vits环境，可以通过pip命令安装一些基础依赖：

!pip install pyworld==0.3.2
!pip install numpy==1.23.5

注意jupyter语言是通过叹号来运行命令。

注意，由于不是本地环境，有的时候colab会提醒：

Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/
Collecting numpy==1.23.5
  Downloading numpy-1.23.5-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (17.1 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 17.1/17.1 MB 80.1 MB/s eta 0:00:00
Installing collected packages: numpy
  Attempting uninstall: numpy
    Found existing installation: numpy 1.22.4
    Uninstalling numpy-1.22.4:
      Successfully uninstalled numpy-1.22.4
Successfully installed numpy-1.23.5
WARNING: The following packages were previously imported in this runtime:
  [numpy]
You must restart the runtime in order to use newly installed versions.

此时numpy库需要重启runtime才可以导入操作。

重启runtime后，需要再重新安装一次，直到系统提示依赖已经存在：

Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/
Requirement already satisfied: numpy==1.23.5 in /usr/local/lib/python3.10/dist-packages (1.23.5)

随后，克隆so-vits项目，并且安装项目的依赖：

import os
import glob
!git clone https://github.com/effusiveperiscope/so-vits-svc -b eff-4.0
os.chdir('/content/so-vits-svc')
# install requirements one-at-a-time to ignore exceptions
!cat requirements.txt | xargs -n 1 pip install --extra-index-url https://download.pytorch.org/whl/cu117
!pip install praat-parselmouth
!pip install ipywidgets
!pip install huggingface_hub
!pip install pip==23.0.1 # fix pip version for fairseq install
!pip install fairseq==0.12.2
!jupyter nbextension enable --py widgetsnbextension
existing_files = glob.glob('/content/**/*.*', recursive=True)
!pip install --upgrade protobuf==3.9.2
!pip uninstall -y tensorflow
!pip install tensorflow==2.11.0

安装好依赖之后，定义一些前置工具方法：

os.chdir('/content/so-vits-svc') # force working-directory to so-vits-svc - this line is just for safety and is probably not required  

import tarfile
import os
from zipfile import ZipFile
# taken from https://github.com/CookiePPP/cookietts/blob/master/CookieTTS/utils/dataset/extract_unknown.py
def extract(path):
    if path.endswith(".zip"):
        with ZipFile(path, 'r') as zipObj:
           zipObj.extractall(os.path.split(path)[0])
    elif path.endswith(".tar.bz2"):
        tar = tarfile.open(path, "r:bz2")
        tar.extractall(os.path.split(path)[0])
        tar.close()
    elif path.endswith(".tar.gz"):
        tar = tarfile.open(path, "r:gz")
        tar.extractall(os.path.split(path)[0])
        tar.close()
    elif path.endswith(".tar"):
        tar = tarfile.open(path, "r:")
        tar.extractall(os.path.split(path)[0])
        tar.close()
    elif path.endswith(".7z"):
        import py7zr
        archive = py7zr.SevenZipFile(path, mode='r')
        archive.extractall(path=os.path.split(path)[0])
        archive.close()
    else:
        raise NotImplementedError(f"{path} extension not implemented.")  

# taken from https://github.com/CookiePPP/cookietts/tree/master/CookieTTS/_0_download/scripts  

# megatools download urls
win64_url = "https://megatools.megous.com/builds/builds/megatools-1.11.1.20230212-win64.zip"
win32_url = "https://megatools.megous.com/builds/builds/megatools-1.11.1.20230212-win32.zip"
linux_url = "https://megatools.megous.com/builds/builds/megatools-1.11.1.20230212-linux-x86_64.tar.gz"
# download megatools
from sys import platform
import os
import urllib.request
import subprocess
from time import sleep  

if platform == "linux" or platform == "linux2":
        dl_url = linux_url
elif platform == "darwin":
    raise NotImplementedError('MacOS not supported.')
elif platform == "win32":
        dl_url = win64_url
else:
    raise NotImplementedError ('Unknown Operating System.')  

dlname = dl_url.split("/")[-1]
if dlname.endswith(".zip"):
    binary_folder = dlname[:-4] # remove .zip
elif dlname.endswith(".tar.gz"):
    binary_folder = dlname[:-7] # remove .tar.gz
else:
    raise NameError('downloaded megatools has unknown archive file extension!')  

if not os.path.exists(binary_folder):
    print('"megatools" not found. Downloading...')
    if not os.path.exists(dlname):
        urllib.request.urlretrieve(dl_url, dlname)
    assert os.path.exists(dlname), 'failed to download.'
    extract(dlname)
    sleep(0.10)
    os.unlink(dlname)
    print("Done!")  

binary_folder = os.path.abspath(binary_folder)  

def megadown(download_link, filename='.', verbose=False):
    """Use megatools binary executable to download files and folders from MEGA.nz ."""
    filename = ' --path "'+os.path.abspath(filename)+'"' if filename else ""
    wd_old = os.getcwd()
    os.chdir(binary_folder)
    try:
        if platform == "linux" or platform == "linux2":
            subprocess.call(f'./megatools dl{filename}{" --debug http" if verbose else ""} {download_link}', shell=True)
        elif platform == "win32":
            subprocess.call(f'megatools.exe dl{filename}{" --debug http" if verbose else ""} {download_link}', shell=True)
    except:
        os.chdir(wd_old) # don't let user stop download without going back to correct directory first
        raise
    os.chdir(wd_old)
    return filename  

import urllib.request
from tqdm import tqdm
import gdown
from os.path import exists  

def request_url_with_progress_bar(url, filename):
    class DownloadProgressBar(tqdm):
        def update_to(self, b=1, bsize=1, tsize=None):
            if tsize is not None:
                self.total = tsize
            self.update(b * bsize - self.n)  

    def download_url(url, filename):
        with DownloadProgressBar(unit='B', unit_scale=True,
                                 miniters=1, desc=url.split('/')[-1]) as t:
            filename, headers = urllib.request.urlretrieve(url, filename=filename, reporthook=t.update_to)
            print("Downloaded to "+filename)
    download_url(url, filename)  

def download(urls, dataset='', filenames=None, force_dl=False, username='', password='', auth_needed=False):
    assert filenames is None or len(urls) == len(filenames), f"number of urls does not match filenames. Expected {len(filenames)} urls, containing the files listed below.\n{filenames}"
    assert not auth_needed or (len(username) and len(password)), f"username and password needed for {dataset} Dataset"
    if filenames is None:
        filenames = [None,]*len(urls)
    for i, (url, filename) in enumerate(zip(urls, filenames)):
        print(f"Downloading File from {url}")
        #if filename is None:
        #    filename = url.split("/")[-1]
        if filename and (not force_dl) and exists(filename):
            print(f"{filename} Already Exists, Skipping.")
            continue
        if 'drive.google.com' in url:
            assert 'https://drive.google.com/uc?id=' in url, 'Google Drive links should follow the format "https://drive.google.com/uc?id=1eQAnaoDBGQZldPVk-nzgYzRbcPSmnpv6".\nWhere id=XXXXXXXXXXXXXXXXX is the Google Drive Share ID.'
            gdown.download(url, filename, quiet=False)
        elif 'mega.nz' in url:
            megadown(url, filename)
        else:
            #urllib.request.urlretrieve(url, filename=filename) # no progress bar
            request_url_with_progress_bar(url, filename) # with progress bar  

import huggingface_hub
import os
import shutil  

class HFModels:
    def __init__(self, repo = "therealvul/so-vits-svc-4.0",
            model_dir = "hf_vul_models"):
        self.model_repo = huggingface_hub.Repository(local_dir=model_dir,
            clone_from=repo, skip_lfs_files=True)
        self.repo = repo
        self.model_dir = model_dir  

        self.model_folders = os.listdir(model_dir)
        self.model_folders.remove('.git')
        self.model_folders.remove('.gitattributes')  

    def list_models(self):
        return self.model_folders  

    # Downloads model;
    # copies config to target_dir and moves model to target_dir
    def download_model(self, model_name, target_dir):
        if not model_name in self.model_folders:
            raise Exception(model_name + " not found")
        model_dir = self.model_dir
        charpath = os.path.join(model_dir,model_name)  

        gen_pt = next(x for x in os.listdir(charpath) if x.startswith("G_"))
        cfg = next(x for x in os.listdir(charpath) if x.endswith("json"))
        try:
          clust = next(x for x in os.listdir(charpath) if x.endswith("pt"))
        except StopIteration as e:
          print("Note - no cluster model for "+model_name)
          clust = None  

        if not os.path.exists(target_dir):
            os.makedirs(target_dir, exist_ok=True)  

        gen_dir = huggingface_hub.hf_hub_download(repo_id = self.repo,
            filename = model_name + "/" + gen_pt) # this is a symlink  

        if clust is not None:
          clust_dir = huggingface_hub.hf_hub_download(repo_id = self.repo,
              filename = model_name + "/" + clust) # this is a symlink
          shutil.move(os.path.realpath(clust_dir), os.path.join(target_dir, clust))
          clust_out = os.path.join(target_dir, clust)
        else:
          clust_out = None  

        shutil.copy(os.path.join(charpath,cfg),os.path.join(target_dir, cfg))
        shutil.move(os.path.realpath(gen_dir), os.path.join(target_dir, gen_pt))  

        return {"config_path": os.path.join(target_dir,cfg),
            "generator_path": os.path.join(target_dir,gen_pt),
            "cluster_path": clust_out}  

# Example usage
# vul_models = HFModels()
# print(vul_models.list_models())
# print("Applejack (singing)" in vul_models.list_models())
# vul_models.download_model("Applejack (singing)","models/Applejack (singing)")  

    print("Finished!")

这些方法可以帮助我们下载、解压和加载模型。

音色模型下载和线上推理

接着将特朗普的音色模型和配置文件进行下载，下载地址是：

https://huggingface.co/Nardicality/so-vits-svc-4.0-models/tree/main/Trump18.5k

随后模型文件放到项目的models文件夹，配置文件则放入config文件夹。

接着将需要转换的歌曲上传到和项目平行的目录中。

运行代码：

import os
import glob
import json
import copy
import logging
import io
from ipywidgets import widgets
from pathlib import Path
from IPython.display import Audio, display  

os.chdir('/content/so-vits-svc')  

import torch
from inference import infer_tool
from inference import slicer
from inference.infer_tool import Svc
import soundfile
import numpy as np  

MODELS_DIR = "models"  

def get_speakers():
  speakers = []
  for _,dirs,_ in os.walk(MODELS_DIR):
    for folder in dirs:
      cur_speaker = {}
      # Look for G_****.pth
      g = glob.glob(os.path.join(MODELS_DIR,folder,'G_*.pth'))
      if not len(g):
        print("Skipping "+folder+", no G_*.pth")
        continue
      cur_speaker["model_path"] = g[0]
      cur_speaker["model_folder"] = folder  

      # Look for *.pt (clustering model)
      clst = glob.glob(os.path.join(MODELS_DIR,folder,'*.pt'))
      if not len(clst):
        print("Note: No clustering model found for "+folder)
        cur_speaker["cluster_path"] = ""
      else:
        cur_speaker["cluster_path"] = clst[0]  

      # Look for config.json
      cfg = glob.glob(os.path.join(MODELS_DIR,folder,'*.json'))
      if not len(cfg):
        print("Skipping "+folder+", no config json")
        continue
      cur_speaker["cfg_path"] = cfg[0]
      with open(cur_speaker["cfg_path"]) as f:
        try:
          cfg_json = json.loads(f.read())
        except Exception as e:
          print("Malformed config json in "+folder)
        for name, i in cfg_json["spk"].items():
          cur_speaker["name"] = name
          cur_speaker["id"] = i
          if not name.startswith('.'):
            speakers.append(copy.copy(cur_speaker))  

    return sorted(speakers, key=lambda x:x["name"].lower())  

logging.getLogger('numba').setLevel(logging.WARNING)
chunks_dict = infer_tool.read_temp("inference/chunks_temp.json")
existing_files = []
slice_db = -40
wav_format = 'wav'  

class InferenceGui():
  def __init__(self):
    self.speakers = get_speakers()
    self.speaker_list = [x["name"] for x in self.speakers]
    self.speaker_box = widgets.Dropdown(
        options = self.speaker_list
    )
    display(self.speaker_box)  

    def convert_cb(btn):
      self.convert()
    def clean_cb(btn):
      self.clean()  

    self.convert_btn = widgets.Button(description="Convert")
    self.convert_btn.on_click(convert_cb)
    self.clean_btn = widgets.Button(description="Delete all audio files")
    self.clean_btn.on_click(clean_cb)  

    self.trans_tx = widgets.IntText(value=0, description='Transpose')
    self.cluster_ratio_tx = widgets.FloatText(value=0.0,
      description='Clustering Ratio')
    self.noise_scale_tx = widgets.FloatText(value=0.4,
      description='Noise Scale')
    self.auto_pitch_ck = widgets.Checkbox(value=False, description=
      'Auto pitch f0 (do not use for singing)')  

    display(self.trans_tx)
    display(self.cluster_ratio_tx)
    display(self.noise_scale_tx)
    display(self.auto_pitch_ck)
    display(self.convert_btn)
    display(self.clean_btn)  

  def convert(self):
    trans = int(self.trans_tx.value)
    speaker = next(x for x in self.speakers if x["name"] ==
          self.speaker_box.value)
    spkpth2 = os.path.join(os.getcwd(),speaker["model_path"])
    print(spkpth2)
    print(os.path.exists(spkpth2))  

    svc_model = Svc(speaker["model_path"], speaker["cfg_path"],
      cluster_model_path=speaker["cluster_path"])  

    input_filepaths = [f for f in glob.glob('/content/**/*.*', recursive=True)
     if f not in existing_files and
     any(f.endswith(ex) for ex in ['.wav','.flac','.mp3','.ogg','.opus'])]
    for name in input_filepaths:
      print("Converting "+os.path.split(name)[-1])
      infer_tool.format_wav(name)  

      wav_path = str(Path(name).with_suffix('.wav'))
      wav_name = Path(name).stem
      chunks = slicer.cut(wav_path, db_thresh=slice_db)
      audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)  

      audio = []
      for (slice_tag, data) in audio_data:
          print(f'#=====segment start, '
              f'{round(len(data)/audio_sr, 3)}s======')  

          length = int(np.ceil(len(data) / audio_sr *
              svc_model.target_sample))  

          if slice_tag:
              print('jump empty segment')
              _audio = np.zeros(length)
          else:
              # Padding "fix" for noise
              pad_len = int(audio_sr * 0.5)
              data = np.concatenate([np.zeros([pad_len]),
                  data, np.zeros([pad_len])])
              raw_path = io.BytesIO()
              soundfile.write(raw_path, data, audio_sr, format="wav")
              raw_path.seek(0)
              _cluster_ratio = 0.0
              if speaker["cluster_path"] != "":
                _cluster_ratio = float(self.cluster_ratio_tx.value)
              out_audio, out_sr = svc_model.infer(
                  speaker["name"], trans, raw_path,
                  cluster_infer_ratio = _cluster_ratio,
                  auto_predict_f0 = bool(self.auto_pitch_ck.value),
                  noice_scale = float(self.noise_scale_tx.value))
              _audio = out_audio.cpu().numpy()
              pad_len = int(svc_model.target_sample * 0.5)
              _audio = _audio[pad_len:-pad_len]
          audio.extend(list(infer_tool.pad_array(_audio, length)))  

      res_path = os.path.join('/content/',
          f'{wav_name}_{trans}_key_'
          f'{speaker["name"]}.{wav_format}')
      soundfile.write(res_path, audio, svc_model.target_sample,
          format=wav_format)
      display(Audio(res_path, autoplay=True)) # display audio file
    pass  

  def clean(self):
     input_filepaths = [f for f in glob.glob('/content/**/*.*', recursive=True)
     if f not in existing_files and
     any(f.endswith(ex) for ex in ['.wav','.flac','.mp3','.ogg','.opus'])]
     for f in input_filepaths:
       os.remove(f)  

inference_gui = InferenceGui()

此时系统会自动在根目录，也就是content下寻找音乐文件，包含但不限于wav、flac、mp3等等，随后根据下载的模型进行推理，推理之前会自动对文件进行背景音分离以及降噪和切片等操作。

推理结束之后，会自动播放转换后的歌曲。

结语

如果是刚开始使用Colab，默认分配的显存是15G左右，完全可以胜任大多数训练和推理任务，但是如果经常用它挂机运算，能分配到的显卡配置就会渐进式地降低，如果需要长时间并且相对稳定的GPU资源，还是需要付费订阅Colab pro服务，另外Google云盘的免费使用空间也是15G，如果模型下多了，导致云盘空间不足，运行代码也会报错，所以最好定期清理Google云盘，以此保证深度学习任务的正常运行。