Neural-Stylr

Neural-style transfer (NST) algorithm implementation using Tensorflow.

Using Neural-Stylr

Neural-stylr can be used by cloning this project, or by running a demo Docker image.

Setup

With Cloned Project

Clone the project

git clone https://github.com/IMBurbank/neural-stylr.git
cd neural-stylr

Download pretrained model

./model/scripts/download_vgg19.sh

Download any desired images to model/images/input/. There is a starter-pack of images already in the input folder that can be used for exploration.

The vgg19 model takes input images sized 300x400. Input images of any other size will be automatically resized and center-cropped to the default aspect-ratio. Output images will be 300x400.

Next, build a Docker image with the necessary dependencies.

# Image is built for gpu-support by default
./model/scripts/docker_build_dev.sh

# Add flag -c for to use CPU instead
./model/scripts/docker_build_dev.sh -c

Run Docker image.

# Image is built for gpu-support by default
./model/scripts/docker_run_dev.sh

# Add flag -c to use CPU instead
./model/scripts/docker_run_dev.sh -c

Change to model directory.

cd model

With Demo Docker Image

Run demo image hosted on Docker Hub

# GPU demo image
docker run --rm --runtime nvidia -u $(id -u):$(id -g) -it dget/dock-stylr:gpu-demo

# CPU demo image
docker run --rm -u $(id -u):$(id -g) -it dget/dock-stylr:demo

Usage

Once set up, run stylr.py and enjoy.

# Run stylr algorith using default images and settings
python stylr.py

# See parameter options
python stylr.py -h

# Make a beautiful kandinsky-hummingbird-sphinx-moth
python stylr.py \
    -c images/input/hummingbird-sphinx-moth-small.jpg \
    -g images/input/wassily-kandinsky-romantic-landscape-1911-small.jpg \
    -s 0.25 -r 2.5 -a 30 -w 0.1 0.1 0.1 0.1 0.1

Examples

Here are a few of my favorite creations. More can be found in assets/examples - along with the configurations used to create them.

Style	Content	NST