Style Transfer Photography: Using your Nvidia graphics card to make artwork

A few years ago there were some breakthroughs in machine learning. Some examples would be the Deepdream project by Google. Which was in a way a CAPTCHA algorithm ran in reverse. Other algorithms that identify objects by picture were developed, interestingly these can be ran in reverse as well. And by doing that, the algorithm is able to impose what it sees between two images and does its best to merge them. Since the algorithm is trained on everyday objects, The lower layers close to the input image can only recognize different lines and edges in them while the upper layers have more advanced shapes such as circles, triangles.

The algorithm extracts the style from the lower layers (edges, curves) and the content from the upper layers (shapes). Then it compares them to the input style image and content image and minimizes the differences until it is close enough.

The algorithms above are also known as convolutional neural networks. And have such names as VGG16, or VGG19, Alexnet, NIN-Imagenet and many more.

Many CNN's simply function as a feature extractor, it extracts the most important features from the images, and it can be used for different tasks. VGG16 and VGG19 which were commonly used in 2015. These networks actually really large receptive fields early on which causes them to use far more parameters than more modern networks like ResNet and InceptionV3, which are being used now.

https://arxiv.org/abs/1512.03385.

A common trend is to use the weights trained on ImageNet as a starting point, and then to "fine-tune" your network for a new task (object detection, generative modeling (style transfer)). To learn more about CNNs check out the cs231n course

http://cs231n.stanford.edu/

Using a Python script, and installing such dependencies like Pytorch, PIP, and a few others. You can use a content image as your base image. And then add Style images that are used by the CNN to "Stylize" the content photo.

Besides the Python scripts, commands and the dependencies you will need a CUDA compatible video card. Which newer Nvidia cards are, I have a 1060 and a 1080 which are compatible and work great for this rendering small to medium size images.

A list of CUDA enabled cards:

https://developer.nvidia.com/cuda-gpus

Link to Python scripts:

https://github.com/ProGamerGov/neural-style-pt

#nvidia #machinelearning #tensorflow #pytorch #python #art #digitalart #technology #learning #education #cuda #rendering

---

Platform	URL/Username
Steemit	https://steemit.com/@solominer
D.Tube
Weku	https://deals.weku.io/@solominer
Bit.Tube	https://bit.tube/solominer
Mithril	Solominer
Bitcointalk	https://bitcointalk.org/index.php?action=profile;u=83228
CryptoPanic	https://cryptopanic.com/solominer
Whaleshares	https://whaleshares.io/@solominer
Bearshares	https://bearshares.com/@solominer

thanks @nevlu123 for the animation