Machine learning essentially involves a ton of trial and error. You're letting a program try millions of different settings to land on an algorithm that sort of does what you want it to do. This process is really really slow unless you have the hardware required to speed this up.
The type of computations that the process does are well suited for graphics cards, rather than regular processors. **It is pretty much required that you run the training process on a desktop or server capable GPU.** Running this on your CPU means it can take weeks to train your model, compared to several hours on a GPU.
## Hardware Requirements
**TL;DR: you need at least one of the following:**
- **A powerful CPU**
- Laptop CPUs can often run the software, but will not be fast enough to train at reasonable speeds
- **A powerful GPU**
- Currently only Nvidia GPUs are supported. AMD graphics cards are not supported.
This is not something that we have control over. It is a requirement of the Tensorflow library.
- The GPU needs to support at least CUDA Compute Capability 3.0 or higher.
To see which version your GPU supports, consult this list: https://developer.nvidia.com/cuda-gpus
Desktop cards later than the 7xx series are most likely supported.
**In its current iteration, the project relies heavily on the use of the command line, although a gui is available. if you are unfamiliar with command line tools, you may have difficulty setting up the environment and should perhaps not attempt any of the steps described in this guide.** This guide assumes you have intermediate knowledge of the command line.
- [virtualenv](https://github.com/pypa/virtualenv) and [virtualenvwrapper](https://virtualenvwrapper.readthedocs.io) may help when you are not using docker.
- If you are using an Nvidia graphics card You should install CUDA (https://developer.nvidia.com/cuda-zone) and CUDNN (https://developer.nvidia.com/cudnn). If you do not plan to build Tensorflow yourself, make sure you install no higher than version 9.0 of CUDA and 7.0.x of CUDNN
- dlib is required for face recognition and is compiled as part of the setup process. You will need the following applications for your os to successfully install dlib (nb: list may be incomplete. Please raise an issue if another prerequisite is required for your OS):
Simply download the code from http://github.com/deepfakes/faceswap - For development it is recommended to use git instead of downloading the code and extracting it.
For now, extract the code to a directory where you're comfortable working with it. Navigate to it with the command line. For our example we will use `~/faceswap/` as our project directory.
If you are experiencing issues, please raise them in the [faceswap-playground](https://github.com/deepfakes/faceswap-playground) repository instead of the main repo.
Setting up Faceswap can seem a little intimidating to new users, but it isn't that complicated, although a little time consuming. It is recommended to use Linux where possible as Windows will hog about 20% of your GPU Memory, making Faceswap run a little slower, however using Windows is perfectly fine and 100% supported.
**GPU Only** If you do not have an Nvidia GPU you can skip this step.
At the time of writing Tensorflow (version 1.12) only supports Cuda up to version 9.0, but check https://www.tensorflow.org/install/gpu for the latest supported version. It is crucial that you download the correct version of Cuda.
**GPU Only** If you do not have an Nvidia GPU you can skip this step.
As with Cuda you will need to install the correct version of cuDNN that the latest Tensorflow supports. At the time of writing this is Tensorflow v1.12 which supports cuDNN version 7.2, but check https://www.tensorflow.org/install/gpu for the latest supported version.
At the bottom of the list of latest cuDNN release will be a link to "Archived cuDNN Releases". Select this and choose the latest version of cuDNN that supports the version of Cuda you installed and has a minor version greater than or equal to the latest version that Tensorflow supports. (Eg Tensorflow 1.12 supports Cuda 9.0 and cuDNN 7.2. There is not an archived version of cuDNN 7.2 for Cuda 9.0, so select cuDNN version 7.3)
Install the latest stable release of CMake from https://cmake.org/download/. (Scroll down the page for Latest Releases and select the relevant Binary distribution installer for your OS).
When installing CMake make sure to enable the option to CMake to the system path:
### Anaconda
Download and install the latest Python 3 Anacconda from: https://www.anaconda.com/download/. Unless you know what you are doing, you can leave all the options at default.
### Git
Download and install Git for Windows: https://git-scm.com/download/win. Unless you know what you are doing, you can leave all the options at default.
## Setup
Reboot your PC, so that everything you have just installed gets registered.
You should only need to follow these steps if you want the latest Dlib code or the process was unable to install Dlib for you.
For reasons outside of our control, this is the trickiest part of the process, and most of the prerequisites you installed are to support just Dlib. It is recommended to build Dlib from source for 3 main reasons:
1) To get the latest version
2) Enable GPU Support in Dlib
3) To prevent yourself running into a whole host of issues later in the process.
If you are not bothered about having GPU support or the latest version, scroll to the end of this section for a simple one-liner to install the CPU version of Dlib.
### Build Latest Dlib with GPU Support
- If you are not already in your virtual environment follow [these steps](#entering-your-virtual-environment)
- In the terminal type: `git clone https://github.com/davisking/dlib.git`
- Enter the dlib folder: `cd dlib`
- Add Visual Studio to your path by typing: `SET PATH=%PATH%;C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin`
This will build and install dlib for you. It is worth backing up the generated .egg file somewhere so that you can re-install it if you ever need to rather than having to re-compile:
- From within the dlib folder copy the file named `dlib-xx.yy.zz-py3.5-win-amd64.egg` to somewhere safe
- If you ever need to install it again, then from within your virtual environment enter: `python -m easy_install <path to saved .egg>`
Once Dlib is built, you can remove Visual Studio and CMake from your PC.
### Easy install of Dlib without GPU Support
NB: Don't do this if you have already compiled Dlib with GPU support.
- If you are not already in your virtual environment follow [these steps](#entering-your-virtual-environment)
- In the terminal type: `conda install -c conda-forge dlib`
