Install VALL-E X on your own PC to imitate the voice of a voice actor.

2023-09-132024-11-21

Install VALL-E X for multilingual text-to-speech synthesis and speech cloning. Before we begin, we will describe the differences between the Python virtual environment and Docker.

You cannot install CUDA directly within a Python virtual environment (venv); CUDA is a system-level component and must be installed on the host system.

The Python virtual environment is used to isolate Python packages and their dependencies, but cannot be used to install or manage system-level software or drivers such as CUDA.

To use CUDA, the following steps must be taken

Install CUDA on the host system: Follow the CUDA Toolkit Installation Guide to install CUDA on the host system. After installation, you will be able to access the CUDA Toolkit and its dependencies by setting the environment variables appropriately.
Create a Python Virtual Environment: Create a Python virtual environment on the host system where CUDA is installed.
Install CUDA-supporting packages in the virtual environment: Install Python packages that utilize CUDA (e.g., TensorFlow, PyTorch, etc.) in the virtual environment.
Verify that CUDA is functioning properly: Use a Python script or notebook to verify that CUDA is functioning properly. This step can be performed using the test scripts provided in the official documentation of the specific deep learning framework.

Once this process is complete, CUDA can be used from within the virtual environment.

Docker uses WSL 2 as its backend, and containers running on a Linux distribution have access to the distribution’s resources and functionality. Therefore, by installing CUDA on a Linux distribution, applications in Docker containers will also be able to take advantage of CUDA’s functionality.

You can proceed with the following steps

Install a Linux distribution on WSL 2: First, install the appropriate Linux distribution on WSL 2.
Install CUDA on the Linux distribution: Next, install CUDA on the Linux distribution.
Set up Docker in WSL 2 mode: Install Docker Desktop for Windows and enable WSL 2 mode.
Create a Docker container: Create a Docker container based on the appropriate Docker image for your project.
Deploy a CUDA-enabled application to the container: Deploy a CUDA-enabled application to the container and run the application using CUDA functionality.

Note that in order to use CUDA, the Docker container must be able to access the GPU. This can be done using Docker’s --gpus option (available in Docker 19.03 and later).

Python Virtual Environment (`venv`)

Purpose: Isolate Python packages and their dependencies on a per-project basis.
Isolation level: Python packages and their versions.
Use Case: Manage different package versions in different Python projects.
Setup: Relatively easy to set up.

Docker Container

Purpose: Isolate an application and its entire dependencies (including OS, system libraries, environment variables, etc.) to create a portable container.
Isolation level: Encompass the entire environment within the container, down to the OS level.
Use Case: Complete control over the application’s environment, ensuring the same behavior on any system.
Setup: Building and managing the environment can be somewhat complex.

Using CUDA in the Python Virtual Environment:

The Python virtual environment can utilize host system resources (e.g., CUDA drivers), but cannot contain these system-level dependencies within itself.

Using CUDA with Docker:

Docker containers can utilize CUDA environments installed on the host system or on a WSL 2 Linux distribution. It is also possible to contain specific NVIDIA GPU drivers and CUDA versions within the container.

These are the differences between python virtual environments and Docker. Proceed with the installation of VALL-E-X using the following GitHub as a reference.

https://github.com/Plachtaa/VALL-E-X

This time, since we do not have the CUDA toolkit installed on Windows 11, we will use the python virtual environment to achieve this. First, we will build the conditions for VALL-E-X to work as described on GitHub.

Install with pip, recommended with Python 3.10, CUDA 11.7 ~ 12.0, PyTorch 2.0

I downloaded CUDA Toolkit 11.8 from the following page and installed it on Windows 11. I chose this version because I see 11.8 a lot.

https://developer.nvidia.com/cuda-11-8-0-download-archive

Next, create a clone in the appropriate folder.

git clone https://github.com/Plachtaa/VALL-E-X.git

‘git’ is an internal or external command,
It is not recognized as an operable program or batch file.

You did not install the Windows version of git. Download and install it at the following page

https://gitforwindows.org/

Launch Git CMD and type the command

cd youtube
git clone https://github.com/Plachtaa/VALL-E-X.git
cd VALL-E-X

Create a python virtual environment and move to the given directory.

python -m venv vall
cd vall
cd Scripts

Activate the virtual environment.

activate

Go back two levels up in the virtual environment.

(vall) C:\youtube\VALL-E-X\vall\Scripts>cd ..
(vall) C:\youtube\VALL-E-X\vall>cd ..

Install PyTorch, but this is important: I have installed CUDA Toolkit 11.8 on Windows, so the virtual environment should match it.

https://pytorch.org/get-started/locally/

The following two commands have different purposes. Understanding the purpose of each command will help you understand why you need both.

Install PyTorch and related packages:

(vall) C:\youtube\VALL-E-X>pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

This command installs a specific version of PyTorch and related packages. It specifies a specific PyTorch build that depends on a specific CUDA version (in this case CUDA 11.8).

Install project dependencies:

(vall) C:\youtube\VALL-E-X>pip install -r requirements.txt

This command installs all Python packages and their specific versions required for the project to work. This may include torch, torchvision, and torchaudio, but there are many other packages.

Therefore, after installing the PyTorch-related packages with the first command, the second command will install the other dependencies of the project.

However, if the requirements.txt file specifies the appropriate versions of torch, torchvision, and torchaudio, the first command may be unnecessary. That is, if the specification in requirements.txt specifies a version that is compatible with CUDA 11.8, the correct version will be installed without the first command.

In this case, it seems that no specific version was specified in requirements.txt, so a specific PyTorch build had to be installed first. After this, use requirements.txt to install the remaining dependencies.

To check if the GPU is available, I created the following script, named test.py and saved it

import torch

# Check if CUDA is available
if torch.cuda.is_available():
    # print GPU details
    print(torch.cuda.get_device_name(0))
else:
    print("CUDA is not available.")