Generating Pokemon using GANs, machine learning,deep learning

2019-2020 tavasz

Nincs megadva

Téma leírása

Generative Adversarial Networks consist of two Parts

●Generative Model: The aim is to learn to model images(or any kind of data) that  resemble the original data
●Discriminator: We don't actually generate any data here instead we try to classify our images
●The goal is for both the generator and discriminator to get better over time, the generator improves how it generates data and the discriminator improves its ability to classify real data from fake data
●Back Propagation is used to train the models
●For this project, the specific GAN implemented is WGAN
Problem:

These results are not bad but we want to see if we can improve the results and make even more refined and believable pokemon.

 We believe that this problem stems from our lack of data as that is a common problem for all generative networks

To achieve better results we can artificially expand upon our dataset by using augmentation techniques

Optimized distributed gradient boosting library

We change our source images by rotating, scaling, warping, etc. to artificially expand source image set(Augmenting the Data set)

We train the xgboost model to be able to differentiate between good and bad augmentations

We filter out the bad augmentation

The goal is to create a larger data set to have more material to train the GAN on

 

 

 

Feltételek

  • gépi tanulás, machine learning, deep learning, GANs

Maximális létszám: 3 fő