Hyper-scale images generated on the Cloud computing environment

Image by Oshidori LLC

Today I would like to talk about the beautiful illustrations crafted with Machine learning. Probably the closest date of Machine learning origin is 1950, but in the computer vision field, it was pioneered in the 1980s. Its main goal was to automatically identify objects in images. However, today’s computer vision field is more about understanding the underlying algorithms and using that to predict the state of a computer application.

Machine learning has shown itself to be an immensely powerful tool for almost any application, particularly when applied to image recognition. It does this using a set of different, specialized Neural Networks that can learn to recognize features from images. The algorithms used to train such models are more complex than the most simple Neural Networks that you can think of. They require a large collection of different features for each image — the features that can be used to differentiate the objects from their background. Each of these different features can make the final model extremely flexible.
The first Neural Networks to have this quality of being able to learn such complicated models were based on the Long Short-Term Memory (LSTM). In these type of models, a very simple set of feature detectors are used as a basis for the network. They allow the data to be fed back into the system, so those features can be added in and taken out to keep the model from becoming very specialized. However, this method is not very flexible, because what you have as a base is not all that flexible. It is extremely specific and does not allow for the possibility of the data to change along the way.

The Deep Neural Network (DNN) was developed to solve this problem. The concept itself is very simple: a network starts from some input and learns the model of the picture by iteratively adding more instances along the way till it gets enough of that data in the input. Then, the model is modified and passed along to the next iteration.

One of the particularly interesting uses of such DNN is the Neural Style Transfer (NST) algorithm. It can transform digital images in are such a way that simulates the appearance or visual style of another image. Popular applications for NST are the production of artificial artwork from photography images. Let’s look for the example provided below:

Photography image original
Photography image after NST

The first image is the original photography image made on regular DSLR Camera and the second image — is the image generated with the NST algorithm.

NST process is difficult because of the small size and low spectral representation of the image. It is a process of organizing information based on many patterns corresponding to the weakening: a large number of inputs, consistency in prioritizing the assets, certainty to calculate the output of the sampling, chance to subdivide the collection of pixels, the approximate size of an insight area, etc.

To optimize and make this process even possible for the hyper-scale images, this type of DNN model produced on the multi-GPU clustered computing environment, also known as the Cloud environment. The Cloud environment implements autoprocessing client-server architecture, which is capable of changing its characteristic attributes in an unannounced way and can be executed on an on-demand basis. Here is an example of the same photography image produced on Amazon SageMaker Cloud environment:

Cloud computed image produced with NST

Note: You can see more examples of the Cloud computed images in the magazine created with the advanced NST algorithms: https://www.amazon.com/dp/B08BJGJVM9

Summing up, we can say that a multi-GPU clustered computing environment is capable of producing highly scaled images and has a remarkable effect on the future of DNN training and applications.

P.S. This article was written with the help of GPT-2 model.

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store