A dockerized version of neural style transfer algorithms. nvidia-docker is used to make use of GPU hardware.
- docker
- nvidia-docker
- Appropriate nvidia drivers for your GPU
You can either pull the Docker image from Docker Hub with
docker pull albarji/neural-style
or build the image locally with
make
This docker container operates by receiving images through a volume to be mounted at the /images directory. For instance, to apply a style image somestyle.png onto a content image somecontent.png located at the current directory, run:
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content somecontent.png --style somestyle.png
All paths referenced in the arguments are regarded as relative to the /images folder within the container. So in case of having a local structure such as
contents/
docker.png
whatever.jpg
styles/
picasso.png
vangogh.png
results/
something.md
applying the vangogh.png style to the docker.png image amounts to
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content contents/docker.png --style styles/vangogh.png
You can provide several content and style images, in which case all cross-combinations will be generated.
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content contents/docker.png contents/whatever.jpg --style styles/vangogh.png styles/picasso.png
By default all generated images are saved into the container /images. When running the commands above the results will get saved into the host local folder (through the mounted volume). If you want to locate the generated images at a different folder you can use the --output parameters. Note that, similarly to the --content and --style arguments, the path provided in --output will be regarded as relative to the /images folder. For instance, using the local structure from the example above, saving the generated images in the "results" folder will require the following:
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content contents/docker.png --style styles/vangogh.png --output results
Better results can be attained by modifying some of the transfer parameters.
The --alg parameter allows changing the neural style transfer algorithm to use.
- gatys: highly detailed transfer, slow processing times (default)
- gatys-multiresolution: multipass version of Gatys method, provides even better quality, but is also much slower
- chen-schmidt: fast patch-based style transfer
- chen-schmidt-inverse: even faster aproximation to chen-schmidt through the use of an inverse network
The following example illustrates kind of results to be expected by these different algorithms
Content image | Algorithm | Style image |
---|---|---|
Gatys | ||
Gatys Multiresolution | ||
Chen-Schmidt | ||
Chen-Schmidt Inverse |
By default the output image will have the same size as the input content image, but a different target size can be specified through the --size parameter. For example, to produce a 512 image
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content contents/docker.png --style styles/vangogh.png --size 512
Note the proportions of the image are maintained, therefore the value of the size parameter is understood as the width of the target image, the height being scaled accordingly to keep proportion.
If the image to be generated is large, a tiling strategy will be used, applying the neural style transfer method to small tiles of the image and stitching them together. Tiles overlap to provide some guarantees on overall consistency, though results might vary depending on the algorithm used.
The size of these tiles is defined through the configuration file gpuconfig.json inside the container. This file contains dictionary keys for different GPU models and each neural style algorithm. Your GPU will be automatically checked against the registered configurations and the appropriate tile size will be selected. These values have been chosen to maximize the use of the available GPU memory, asumming the whole GPU is available for the style transfer task.
If your GPU is not included in the configuration file, the default values will we used instead, though to obtain better performance you might want to edit this file and rebuild the docker images.
Note also that since the full style image is applied to each tile separately, as a result the style features will appear as smaller in the rendered image.
Gatys and Gatys Multiresolution algorithms allow to adjust the amount of style imposed over the content image, by means of the --sw parameter. By default a value of 5 is used, meaning the importance of the style is 5 times the importance of the content. Smaller weight values result in the transfer of colors, while higher values transfer textures and even objects of the style.
If several weight values are provided, all combinations will be generated. For instance, to generate the same style transfer with three different weights, use
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content contents/docker.png --style styles/vangogh.png --sw 5 10 20
Note also that they Gatys Multiresolution algorithm tends to produce a stronger style imprint, and this you might want to use weight values smaller than the default (e.g. 3).
If the transferred style results in too large or too small features, the scaling can be modified through the --ss parameter. A value of 1 keeps the style at its original scale. Smaller values reduce the scale of the style, resulting in smaller style features in the output image. Conversely, larger values produce larger features. Similarly to the style weight, several values can be provided
nvidia-docker run --rm -v $(pwd):/images albarji/neural-style --content contents/docker.png --style styles/vangogh.png --ss 0.75 1 1.25
Warning: using a value larger than 1 will increasy the memory consumption.
Transparency values (alpha channels) are preserved by the neural style transfer. Note for instance how in the Wikipedia logo example above the transparent background is not transformed.