Today, researchers from Google AI and Georgia Institute of Technology released an interactive website for learning GAN: GAN Lab! Driven by TensorFlow.js, GAN can be run in the browser.
Researchers from Google AI and Georgia Institute of Technology have released an interactive website for learning GAN: GAN Lab! Driven by TensorFlow.js, you can run GAN in the browser, and understand the mechanism of various GAN models very intuitively, which is a great artifact. After the release, it was quickly praised.
GAN Lab address:
https://poloclub.github.io/ganlab/
GitHub:
https://github.com/poloclub/ganlab
paper:
http://minsuk.com/research/papers/kahng-ganlab-vast2018.pdf
The website provides four types of data distribution, you can also customize your own data distribution, and you can use a pre-trained model.
After clicking the run button, you can view the visual presentation of the model:
Many machine learning systems need to look at some kind of complex input (such as an image) and then produce a simple output (such as a label like "cat"). In contrast, the goal of a generative model is just the opposite: take a small part of the input (maybe some random numbers), and then generate a complex output, such as a realistic-looking face image. Generative Adversarial Networks (GAN) is a particularly effective generative model. Since its introduction a few years ago, it has been a hot research topic in machine learning.
The idea of ​​a machine "creating" a realistic image from scratch may seem magical, but GAN uses two key techniques to turn the seemingly impossible goal into reality, namely, the establishment of two competing networks-generators. ) And discriminator. The generator tries to create random synthetic outputs (for example, facial images), while the discriminator tries to distinguish these outputs from actual outputs (for example, a database of celebrities). When two networks fight each other, they will get better and better-the end result is a generative network that can produce realistic output.
GAN is very complicated, and its visualization also has many problems. The basic idea of ​​the researcher is:
First, it didn't visualize anything as complicated as generating real images. Instead, it shows how a GAN learns the distribution of points in a two-dimensional space. Such a simple thing has no real application, but it helps to show the mechanism of the system. Because the probability distribution in the ordinary two-dimensional (x, y) space is easier to visualize than the distribution in the high-resolution image space.
Show data distribution in two locations
Click the play button on the toolbar to run the model. In addition to the actual samples in the selected distribution, you will also see fake samples generated by the model. As the training progresses, the positions of the fake samples are constantly updated. The distribution of fake samples created by a perfect GAN is almost indistinguishable from the distribution of actual samples. When this happens, in the hierarchical distribution view, you will see that the two distributions overlap very well.
As the training progresses, the positions of the fake samples are constantly updated. Finally, the distribution of real samples and fake samples overlap.
The generator's data transformation is visualized as a manifold, and the input noise (far left) is converted into fake samples (far right).
The performance of the discriminator can be explained by 2Dheatmap.
LED Tile ligh,Waterproof led tile lamp
Kindwin Technology (H.K.) Limited , https://www.ktlleds.com