15_Aligatou AI Feb 23th

Inspiration

Meredith, Sofia & Feileacan has presented us a new concept which it make us think OUT-OF-THE BOX. Find an inspiration from Ai self- generator which it come up with new concept on the spot. Here we have the challenge to create something new on the spot. We’re quiet excited to use what we practice in the last few weeks on this fun challenge and how we can use it on our benefit to learn and continue on our project

An article provided by Ellen Nickles, our instructor for machine learning. On Wednesday evening, Ellen has presented the difference styles of GANS. one of which by Helena Sarin. A Cycle GAN artist which she create her own dataset. Here’s the link for this article of her work and why she decided to use Cycle GAN model https://thegradient.pub/playing-a-game-of-ganstruction/

We do like how she generate her own dataset to have the uniqueness of her artwork.

What it does

The CycleGAN is a technique that involves the automatic training of image-to-image translation models without paired examples. The models are trained in an unsupervised manner using a collection of images from the source and target domain that do not need to be related in any way.

How we built it

With the help of Meredith, Sofia & Feileacan, we used a pre-trained dataset (e.g. zebra and horse model) where we upload our dataset which is only 75 frames and modified to 400 frames using the side angle, flip and etc. We devided our data set to two elements; abstract graffiti and cars. We train the cycle GAN for 6 hours to get some promising results.

Challenges we ran into

Starting and figure out our path to make Cycle GAN since sometimes it get tricky since it need to take time to train the model

Accomplishments that we're proud of

The first time we saw our 1st results of Cycle GAN. It felt really great to realize how beneficial and powerful the cycle gan can be when you train it and how good it become as tool to help the artist to grow with his work.