Improve usage for non technical people about faceswap HOT 13 CLOSED

In theory an installation like that works, but for development, you ideally you want it in some kind of separate environment. Virtualenv or something, so it does not conflict with any local packages you may already have installed.

Packaging is not necessary immediatelly, but I can imagine having prebuilt executables of a command line version is already a great improvement. It allows other devs to setup simple frontends that interface with the command-line tool. Having a pre-built command-line tool is a great step towards coupling it with both web and native frontends.

from faceswap.

When I tried running the convert.py script I had the following error :

Failed to extract from image: /home/edouard/Téléchargements/faceswap-data/source-images/lawrence/107.jpg. Reason: Unable to open contrib/shape_predictor_68_face_landmarks.dat

I had to look into the code to know what to do.

I guess we'll need a way to get this file without having to download it separatly if we want to make it user friendly. Or say explictily how and where to download it if it is missing when executing the script.

from faceswap.

I would see separate GPU and CPU versions for each target OS.

What I wonder, is how the process would look for the end user. Seeing as the training portion takes the most time, you could initially start by providing the basic "image processor" to the user, and have more advanced users share their models in ready made archives. I'm unfamiliar with the format, so I'm not sure if that opens it up to some kind of security issue. However, I feel like that could be an easy way of simplifying the process for non-technical people. Your video/picture + a pre-made swappable model = easy result.

from faceswap.

It's quite easy making the installation on windows x64, for CPU based processing:

• Install python 3.6 x64 from: https://www.python.org/ftp/python/3.6.4/python-3.6.4-amd64.exe

• And then, from a command line with supervisor permisions, the following commands:

• pip3 install --upgrade tensorflow

• pip3 install h5py

• pip3 install keras

• pip3 install opencv-python

Until some GUI is added, I don't think it's worth packaging it ...

from faceswap.

Basic command line usage is almost ready. Can probably look into cooking up pre-built packages once that's done. I might look into some CI stuff to make semi-automated builds, but that might be goals to shoot for after christmas or start of 2018.

from faceswap.

@Yutsa I merged your PR with the error message. We can most likely also improve the behavior for handling non-existent input/output directories with some "folder does not exist, create it? y/n" dialogues.

from faceswap.

I am working on using one main script that calls the other.

The idea is to get commands like this ./faceswap.py train, ./faceswap.py extract, ./faceswap.py convert.

I don't know if this is what you want though.

from faceswap.

Yes, you can achieve this easily using the subparser feature that comes as part of argparse. I was going to do this myself, but I may not have time today.

from faceswap.

I'm currently looking into creating an initial self-contained build of the command-line tool in its current state. PyInstaller seems like a good option, although I fear it might result in very large binaries.

Would be nice to release it on January 1st if possible!

from faceswap.

Shouldn't we provide to Dockerfiles ? One for GPU and the other for CPU ?

Ideally we should upload both images on Docker Hub, that way a user would only have to install docker and run the docker run -it --rm -v [sourceFolder]:/srv -it user/faceswap:tag bash

I also thought maybe we could add a function to create the video from the swapped pictures after convert maybe ?

Ideally if we could from the same tool cut the video into frames, cut faces, train, replace faces and assemble the video back it would be the best for a user.

from faceswap.

In order to use Docker with a GPU, you'd need to split it up further between OpenCL and CUDA. To use Nvidia CUDA with docker, the only possible way is nvidia-docker, which is only supported on Linux. In addition, the same is true for Docker with OpenCL. As far as I know, OpenCL support in Docker involves manually giving your docker container access to your host's GPU driver.

So, assuming that Linux users are per definition more tech-savvy than Windows or macOS users and considering that Dockerized GPU support is limited to Linux for the time being, I strongly believe that the usage of Docker does not help non-technical users at all. Yes, it's nice if you're savvy with Docker and containers, but is that really even a desirable method of distribution for people who have very limited command line knowledge?

I do not think Docker GPU support is going to help in getting this project usable for non-technical users. A Docker Hub image would be nice, but it would be limited to a CPU version, because getting it working on a GPU (CUDA or OpenCL) would still be a pain.

Still, I would encourage you to research and play around with this, and provide pull requests if you manage to get it working.

from faceswap.

Yeah I was actually looking at nvidia-docker, it looks complicated to have a Docker container with GPU.

I guess we'll have to wait for this to be better supported.

from faceswap.

for dependencies a useful command is:
pip3 install -r requirements{-gpu}.txt

we can also create 2x pypi packages.

from faceswap.