Fresh-Diversion/DeepFaceLab_NVIDIA_Linux

Setting up environments for creating deepfake videos (DeepFaceLab)

Install DeepFaceLab

Note: This installation method uses Anaconda to install DFL on Centos 7/8 or Ubuntu 18.04
To install CentOS, refer here

Note: The directory structure is important.
This is how the project tree will look like:

DeepFaceLab_NVIDIA_Linux
|-- _internal               
|   |-- DeepFaceLab         # Created via git clone command (Step #2)
|   |-- pretrain_CelebA     # Manually copied after extracting .EXE (Step #3)
|   `-- pretrain_Quick96    # Manually copied after extracting .EXE (Step #3)
|-- install_script          # Created via git clone command (Step #0)
|-- scripts                 # Created via git clone command (Step #0)
`-- workspace               # Created through one of the scripts in /scripts

Step #0: Preparation

Clone this repository

git clone https://github.com/jwchoi09/DeepFaceLab_NVIDIA_Linux.git
cd DeepFaceLab_NVIDIA_Linux

Step #0-1: Install NVIDIA driver and Dependencies

Navigate to DeepFaceLab_NVIDIA_Linux/install_script directory and run:

1. CentOS:

   bash 1_CentOS_install_dependencies.sh

2. Ubuntu:

   bash 1_Ubuntu_install_dependencies.sh

Step #1: Install Anaconda

URL #4: https://www.anaconda.com/distribution/#linux
URL #5: https://docs.anaconda.com/anaconda/install/linux/

Note: Download Python 3.7 version (at the time of writing: 03/16/2020)
Through the terminal, locate to Anaconda download destination and run the following:

bash Anaconda3-2020.02-Linux-x86_64.sh

The installer prompts "In order to continue the installation process, please review the license agreement."
Click Enter to view license terms.
Scroll to the bottom of the license terms and enter "Yes" to agree.
Accept the default install location.
Type "yes" to the prompt, "Do you wish the installer to initialize Anaconda3 by running conda init?"
Close and open your terminal window for the installation to take effect, or you can enter the command:

source ~/.bashrc

Step #2: Install DeepFaceLab_NVIDIA

Create the DeepFaceLab environment with the following libraries.
Enter the following in the terminal:

conda create -n deepfacelab -c main python=3.6.8 cudnn=7.6.5 cudatoolkit=10.0.130

Activate the environment

conda activate deepfacelab

Navigate to _internal directory in DeepFaceLab_NVIDIA_Linux

cd DeepFaceLab_NVIDIA_Linux/_internal

Install Python Dependencies

python3 -m pip install -r ./DeepFaceLab/requirements-cuda.txt  

Delete the already existing DeepFaceLab directory, and
Clone DeepFaceLab from the main repository to the _internal directory

sudo rm -rf DeepFaceLab
git clone https://github.com/iperov/DeepFaceLab.git

Step #3: Download CelebA Dataset and Quick96 Model

Download the latest NVIDIA build(.EXE) from the main repo for Windows
To download, refer to the Google Drive link provided in here: https://github.com/iperov/DeepFaceLab/
Extract the build.

7z x DeepFaceLab_NVIDIA_build_XX_XX_XXX.exe

Once the build has been extracted, there will be _internal directory.
From it, copy both pretrain_CelebA and pretrain_Quick96 to the directory DeepFaceLab_NVIDIA_Linux/_internal/

Step #4: Navigate to the scripts directory and begin using DeepFaceLab

(Further guide coming soon!)
https://mrdeepfakes.com/forums/thread-guide-deepfacelab-2-0-explained-and-tutorials-recommended

Confirm the build directory is set up properly

There are many scripts. We will choose from each category to execute the extraction/trainging/merging process

The Scripts: SETUP
1_clear_workspace.sh

EXTRACT:
2_extract_images_from_video_data_src.sh
3_cut_video_drop_video_on_me.sh
3_extract_images_from_video_data_dst_FULL_FPS.sh
4.2_data_src_sort.sh
4.2_data_src_util_add_landmarks_debug_images.sh
4.2_data_src_util_faceset_enhance.sh
4.2_data_src_util_faceset_metadata_restore.sh
4.2_data_src_util_faceset_metadata_save.sh
4.2_data_src_util_faceset_pack.sh
4.2_data_src_util_faceset_unpack.sh
4.2_data_src_util_recover_original_filename.sh
4_data_src_extract_full_face_MANUAL.sh
4_data_src_extract_full_face_S3FD.sh
4_data_src_extract_whole_face_MANUAL.sh
4_data_src_extract_whole_face_S3FD.sh
5.2_data_dst_sort.sh
5.2_data_dst_util_faceset_pack.sh
5.2_data_dst_util_faceset_unpack.sh
5.2_data_dst_util_recover_original_filename.sh
5.3_data_dst_mask_editor.sh
5_data_dst_extract_full_face_MANUAL_RE-EXTRACT_DELETED_ALIGNED_DEBUG.sh
5_data_dst_extract_full_face_MANUAL.sh
5_data_dst_extract_full_face_S3FD_+manual_fix.sh
5_data_dst_extract_full_face_S3FD.sh
5_data_dst_extract_whole_face_MANUAL_RE+EXTRACT_DELETED_ALIGNED_DEBUG.sh
5_data_dst_extract_whole_face_MANUAL.sh
5_data_dst_extract_whole_face_S3FD+_manual_fix.sh
5_data_dst_extract_whole_face_S3FD.sh
5.XSeg_data_dst_edit.sh
5.XSeg_data_dst_merge.sh
5.XSeg_data_dst_split.sh
5.XSeg_data_src_edit.sh
5.XSeg_data_src_merge.sh
5.XSeg_data_src_split.sh
5.XSeg_train.sh

TRAIN:
6_train_Quick96.sh
6_train_SAEHD.sh

MERGE:
7_merge_Quick96.sh
7_merge_SAEHD.sh

CONVERT TO VIDEO: 8_merged_to_avi.sh
8_merged_to_mov_lossless.sh
8_merged_to_mp4_lossless.sh
8_merged_to_mp4.sh
env.sh

Navigate to the DeepFaceLab_NVIDIA_Linux directory

cd ~/DeepFaceLab_NVIDIA_Linux 

give permissions on the scripts folder(check this)

chmod +x scripts -R  

Navigate to the scripts directory

cd ~/DeepFaceLab_NVIDIA_Linux/scripts 

The 3 main steps of creating a deepfake are Extract, Train, and Merge
Run the scripts in this order:

SET UP
This will create/clear the workspace folder - only use this at the beginning of the process.

./1_clear_workspace.sh  

Move the data_src.mp4 and data_dst.mp4 files into the workspace folder

the folder should look like this:

|-- workspace        
|   |-- data_dst 
|       |--aligned  
|   |-- data_src  
|       |--aligned  
|   |-- model  
|   -- data_dst.mp4  
|   -- data_src.mp4  

EXTRACT Extract frames from src file
The next script has 2 options:
A) FPS -> you can choose the FPS. The greater the FPS the more material will be available to work with = longer time
B) Output Image Format png/jpg ->

./2_extract_images_from_video_data_src.sh  

When Finished, the data_src folder should have folder named aligned and the extracted frames from the data_src movie together

Extract frames from dst file
The next script has 1 option:
A) Output Image Format png/jpg ->

./3_extract_images_from_video_data_dst_FULL_FPS.sh 

When finished, the data_dst folder should have folder named aligned and the extracted frames from the data_dst movie together

Extract Faces,align the landmarks from the src The next script has 2 options:
A)gpu or cpu
B)choosing whether to generate "aligned_debug" folder or not. -> I think doing this is a good idea

./4_data_src_extract_full_face_S3FD.sh

When finished, the data_src folder should have folder named aligned_debug with the extracted frames from the data_src frames inside. The frames in aligned_debug should have the alignments

Clean the source faceset/dataset of false positives/incorrectly aligned faces
The next script has 13 options:
Choose sorting method:
[0] blur
[1] face yaw direction
[2] face pitch direction
[3] face rect size in source image
[4] histogram similarity
[5] histogram dissimilarity
[6] brightness
[7] hue
[8] amount of black pixels
[9] original filename
[10] one face in image
[11] absolute pixel difference
[12] best faces
[13] best faces faster

./4.2_data_src_sort.sh

Choose 4 if you don't know what to choose

Add land marks to debug images if the option was not chosen before

4.2_data_src_util_add_landmarks_debug_images.sh  

Extract Faces,align the landmarks from the dst
The next script has 1 option:
A)gpu or cpu

./5_data_dst_extract_whole_face_S3FD.sh

When finished, the data_dst folder should have a folder named aligned_debug. The frames in aligned_debug should have the alignments.

Clean the dst faceset/dataset of false positives/incorrectly aligned faces
The next script has 13 options:
Choose sorting method:
[0] blur
[1] face yaw direction
[2] face pitch direction
[3] face rect size in source image
[4] histogram similarity
[5] histogram dissimilarity
[6] brightness
[7] hue
[8] amount of black pixels
[9] original filename
[10] one face in image
[11] absolute pixel difference
[12] best faces
[13] best faces faster

./5.2_data_dst_sort.sh  

Choose 4 if you don't know what to choose

When finished, the data_dst folder should have a folder named aligned with the extracted faces from the data_dst frames inside

TRAINING

There are 2 methods for training:
6_train_Quick96.sh - one option -> cpu/gpu
6_train_SAEHD.sh - many options

./6_train_Quick96.sh

The longer the training, the better when finished - around 150000 iterations press ctrl + c in the terminal to quit

MERGING

There are 2 methods for merging:
7_merge_Quick96.sh
7_merge_SAEHD.sh

./7_merge_Quick96.sh

CONVERSION There are 4 methods for converging: 8_merged_to_avi.sh
8_merged_to_mov_lossless.sh
8_merged_to_mp4_lossless.sh
8_merged_to_mp4.sh

The next script has 1 option:
bitrate - 16 is suggested

./8_merged_to_mp4.sh 

not sure about this: 5.2_data_dst_util_recover_original_filename.sh

batch rename for linux
Preview the change
rename 's/_0//g' *.jpg -n

Execute the command
rename 's/_0//g' *.jpg