Deep Dream: how to add photos and videos of robots dream

How to quickly make a cool psychedelic video or draw several unique patterns using a computer? With Deep Dream!

image processing technology using neural networks Deep Dream for a short time, the Internet exploded and disappeared from sight. But this does not mean that it ceased to exist, and it ceased to use.

Thus, a group Years & Years introduced the first music video (the first for the famous teams and major record labels), fully processed by machine intelligence - neural network from Google. Before the video was assembled for Desire lyrics, it took four days of video recording just deserts of the United States and a few hundred hours of operation of the computer.

Agree, the effect is stunning. You want to create something similar from your own records? This will require the material (photo or video), computer software for video editing, internet access and a while.

In short, working drawing using neural networks as follows:

1. Built a neural network.
2. Training is carried out by using certain images-standards with the required content.
instagram viewer
3. The picture is processed by the neural network. The network searches the photo familiar combinations of pixels and converts them in accordance with the algorithm internalized. If the network was trained to see the eyes, then the picture will be your eyes or the like.

The easiest way is to use a ready and trained neural networks. At the moment they are presented in the form of several online services. One of them - Deep DreamerWith whom we already work described. Today, it is primarily an online service, although the application for the Mac still exists. Deep Dreamer suitable for processing individual frames or pictures, but it is necessary to be careful: the neural network of the service is already trained. This means that they will be looking at the familiar images of a material and render it exactly.

Similarly, the service works with the title Deep Dream Generator. Like Deep Dreamer, he prefers to draw the eye and dogs.

There are also more hardcore way to use Deep Dream at home to draw anything. Build its own neural network is not required, and the glory of God: this requires hundreds of hours. But sweat when you set up the network have finished.

Use should prepare a project called GoogleNet. It is not difficult, the knowledge of programming languages ​​is not required. But reading this instructions required.

Then you have to find 200-1 000 different images on the same topic, and save them with a resolution of 256 × 256 with the regime True Color JPG (not black-and-white even if the original is as follows). Figures should contain exactly the information you want to display the processed image.

Description of further tweaking process, debugging and training network can be found in the Russian-speaking article. All this may seem a bit complicated, but in fact the only obstacle - prolonged high CPU usage.