neural-filter

The work is still in progress but it's possible to run the script and obtain decent results...

However, you must fill a couple of conditions :

• Especially some python modules (listed below):
  • Tensorflow ( 1.0 or higher, for my part I use the 1.0 )
  • Scipy ( that's a must have ! You shall update it to the newest version or just install it with the command : ' sudo pip install scipy' ==> If you're on a Linux system. Though, you may have a Windows system ; So if pip doesn't work properly you can use easy_install provided with latest python version (For now, I think it's 3.5 ) . if you still encounter some problems, google it 'how to install Scipy on my laptop) ==> it's an easy one, everything is well documented, you should get rid of it ;)
  • os ( already installed )
  • argparse ( already installed )
  • numpy ( you must get the newest version )
  • math (already installed)
  • PIL ( Python Imaging Library )
  • functools

Install dependencies :

* install on your laptop python 3.5 ( or Anaconda 3.5 ) 
* then go to the terminal and enter the following commands :
    `pip3 install tensorflow`
    or 
    `pip install tensorflow`
    if it doesn't work try : 
    `pip install --upgrade https://storage.googleapis.com/tensorflow/windows/cpu/tensorflow-0.12.1-cp35-cp35m-win_amd64.whl`
    
     ( for more information you can go here https://www.tensorflow.org/versions/r0.12/get_started/os_setup )
     
* do the same thing for Scipy :
`pip install scipy` 
or 
`pip3 install scipy`

* (for those on Windows) for Numpy ( you'll have to install mkl with it), go to http://www.lfd.uci.edu/~gohlke/pythonlibs/ and download `numpy‑1.11.3+mkl‑cp35‑cp35m‑win_amd64.whl` ( for 64 bits)  or
`numpy‑1.11.3+mkl‑cp35‑cp35m‑win32.whl` ( for 32 bits )

Once downloaded, change your directory to the directory containing your .whl file and enter : 

`pip install numpy‑1.11.3+mkl‑cp35‑cp35m‑win32.whl` 
or 
`pip install numpy‑1.11.3+mkl‑cp35‑cp35m‑win_amd64.whl` 
( according to your system 32 or 64 bits)

* PIL is easily installed with pip or pip3 with the command `pip install PIL`

(N.B : if `pip` is not installed, install it or you can also use `easy_install` instead of `pip`, the command is similar)

• Moreover, you've to download ' imagenet-vgg-verydeep-19.mat ' which is the network you're going to use. That's a big one ( ~550Mo) but it's quite powerful... You can find it at http://www.vlfeat.org/matconvnet/pretrained/ You can put it in your working directory... It'll work for the future. Or you can use the argument --network <path/of/your/file>

HOW TO

• first you have to open a terminal and change your current directory to your working directory

• Then you can enter the following command ( it's and example ):

  py neural_filter.py --content C:\Users\gabri\Pictures\max.png --styles C:\Users\gabri\Pictures\gogh.png --output image_sortie.png --iterations 10

  (if your PATH isn't 'py', replace it with your PATH calling python)

the files max.png and gogh.png in the command, are respectively the base image and the style image ( see below...)

the resulting picture is quite disgusting ( Though, I precised only 10 iterations in argument so it's normal)

• And now with 100 iterations ( I have to say that the process is extremely loooong ! (if you are using your CPU instead of your GPU)

  • For my part, I had a new laptop ( running on windows 10 ) and the compatibility between TensorFlow en widows suck ! So I am still trying to install TensorFlow-GPU but you may encounter a bunch of problems !! ( Vive Linux ! )

• command :

  py neural_filter.py --content C:\Users\gabri\neural-filter\max.png --styles C:\Users\gabri\neural-filter\gogh.png --output output_image.png --iterations 100

• Result :

(It is important to notice that color-preservation has been avoided in this picture...) We can easily figure it out with the argument : --preserve-colors (see below )

Quick Conclusion :

Running the program on windows 10 (i.e with no TensorFlow-GPU ), I had to wait almost 2 hours ( 1h40min to be exact) to have the 100 iterations picture. The speed could be really increased ( 3-4 minutes for 1000 iterations) with a good GPU ( very good...) and a Linux system with TensorFlow-GPU installed and configured !