You can reach the code below;
Final Assignment Proposal
IDEA
To create awareness on misinformation by making reference to fake news. An interactive web application which gets data from a news API, and by using text-based generate model I can create fake news.
Lots of things you read online especially in your social media feeds may appear to be true, often is not. Fake news is news, stories or hoaxes created to deliberately misinform or deceive readers.
Traditionally we got our news from trusted sources, journalists and media outlets that are required to follow strict codes of practice. However, the internet has enabled a whole new way to publish, share and consume information and news with very little regulation or editorial standards.
Usually, these stories are created to either influence people’s views, push a political agenda or cause confusion and can often be a profitable business for online publishers.
My hometown Turkey is a country where fact and fiction are increasingly hard to distinguish. It is little wonder that Turkey ranks first in a list of countries where people complain about completely made-up stories. according to this year's Reuters Digital News Report.
So, I want to create awareness on unreal misinformation by making reference to fake news. An interactive web application which gets data from a news API, and by using text-based generate model I can create fake news. That might be a good idea for this project, so that I can also combine this assignment with my Dynamic Web Development class.
I’m afraid if I create a website with generated images and texts, I could be one of the fake news websites. So, I believe I should pay attention to that.
TRAINING A StyleGAN Model
I’ve trained a styleGAN model by using inputs from golden retriever photos, which I had one named “Mio”, he died from heart attack. I dowloaded images from Stanford dog datasheet. There are bunch of dogs from different breeds.
Here is the half golden retriever half portrait model.
800 / 3000 steps
1770/3000 steps
I guess my styleGAN image generation training wasn’t successful is because of my image sizes might be too small. But I’ve learned I better have images that are bigger than 512x512px.
PROCESS
StyleGAN
I’ve trained Kids-Self-Portrait-GAN model with your own data using Runway. My inspiration for this project was my nephew who draws beautiful pictures. So, I’ve chosen a styleGAN which reminds me of his works.
I’ve generated kids pictures with runway ML and by passing those to P5 JS, displayed them in the browser. This model generate self-portrait images of kids images. After adding it to my workspace I run it remotely, which runs remotely in cloud. After the images populated, I’ve choosen the generated image I want to work with and went to Network option, which shows me a variety of options of communication. I worked with HTTP. I sent the my request to Runway by httpPost() function;
httpPost(path, [datatype], [data], [callback], [errorCallback])DataType: “json”
Path: "http://localhost:8000/query"
In my sketch every time I click on kids button, I get a new style of StyleGAN geretated kids image in P5 web editor. I’m slightly happy with the results, and i’ve attempted to animate the images but couldn’t success. I want to work on it.
CODE
RunwayML for Adobe Photoshop
For this assignment I’ve accessed RunwayML directly from Adobe Photoshop and generated images by applying style transfers to one of my engaving composition that I’ve done in 2018. I’ve applied various models including; BDCN, DynamicStyleTransfer, BigBiGAN, DeepLab, EnlightenGAN, ImageInpailing to the same image.
BDCN
Dynamic style transfer
bigiban
deeplab
IMAGEINPAILING
PROCESS
Arbitrary Image Stylization with Runway
This project is from open-sources a machine learning system for performing fast artistic style transfer that may work on arbitrary painting styles. Because this system provides a learned representation, one may arbitrarily combine painting styles as well as dial in the strength of a painting style. I’ve combined the two well known paintings; The monalisa by the Italian artist Leonardo da Vinci and the Persistence of Memory by Salvador Dali. The result was pretty interesting.
Image Classification with ml5 and MobileNet
I attempted to classify images coming straight in a real time from the web camera by using ml5.js with pre-trained model called MobileNet. It doesn’t recognize most of the items but It was pretty amazing to experiment with! Machine learning is not magic. The model only knows about what it’s been trained with. In this case, the MobileNet model was trained on a somewhat arbitrary set of objects. Pre-trained models like this provide friendly starting points for beginners. With just a line or two of code, a variety of models can be run. Best thing about this project was the experiment itself, I was very curious each time if the model could predict the object, the worst part is using the browser, the speed slows down.
