UStar is a big approach in the field of generative adversarial networks with the goal of being a game changer to astrology. Our goal is to create simulations in astrology using generative networks.In this case, we aimed to provide images that resemble to realistic stars which is an Image-to-Image problem that we try to transform a basic painting to an image of a complex star.We used heavy image processing techniques to create our data using a single grayscale image of a star.Pix2Pix network is considered for training in this repo.We rank the model of each epoch using Mean-Squared-Error and then pick the top-5 best models with the lowest evaluation loss for our final insights
Every single human being has once tried to draw fancy and imaginary astronomical things in their life including planets, galaxies, stars and so on, but considering stars, have you ever wondered how would your paintings actually look like in real life? Would they look like white dwarfs? Or red giant? Or cosmic dusts? Well, by developing UStarV1 we have found a way to take these paintings into realistic images of stars.
The letter U in UStar is inspired by two reasons
- Letter U being a reference to the network used as our generator(U-Net)
- Letter U being a reference to word Unreal for having the capability to generate unreal stars.
CNNs(Convolutional Neural Networks) have been a game changer to the variety of image prediction problems such as semantic and instance segmentation, paired and unpaired image translation, image reconstruction and so on.In the first version of the UStar model, we use the UNet network as our cGAN(Conditional Generative Adversarial Network) where we wish to solve the paired Image-To-Image translation of basic paintings to realistic star images.This model is as well provided with a basic application to demonstrate the usage of UStarV1
https://github.com/Moeed1mdnzh/UStar-GUI
However, the translation of dark stars or pure white stars are not guaranteed to be perfectly realistic or clear as these colors are rarely seen in between stars.
The following URLs represent the explanation to each component of our approach
About how the the dataset was created
About the augmentation techniques used to increase data
Describes architectures of the following
Generator's architecture
Discriminator's architecture
Open up command prompt or terminal(for mac and linux users) and run the following commands in a directory you desire or if you're a google colab user, follow UStarV1 Colab
git clone https://github.com/Moeed1mdnzh/UStarV1
cd UStarV1
Install the required packages
python setup.py
Generate the dataset
python generate_data.py
Train using the following arguments
- --VISUALIZE: To wether save the visualizations or not
- --EVALUATE: To wether evaluate the model or not
- --SAVE: To wether save the best model or not
Just a reminder that you can modify the training configurations in UNet/configs.py
python train.py -v True -e True -s True
After training, you can find the result of your evaluation in the main directory in two .jpg and .json formats.Your trained models and visualizations can as well be
found in the UNet directory.