Ganilla

• 

  dataset example

Inspiration

We wanted to implement a neural network that will be some artistic project, that is also a GAN. We found a great paper about a model named GANILLA. GANILLA is a type of GAN (Generative Adversarial Network) that solves the image-to-image translation problem, in the domain of children’s book illustrations. GANILLA excels in preserving both content from the original input photograph and the style of the desired illustrator. This was our challenge.

What it does

When you provide a photograph to a trained GANILLA model, it will output an illustration of the input image, based on the style of a specific illustrator.

How we built it

GANILLA is built with pytorch

Challenges we ran into

One big challenge was to find a free dataset. We overcame that with Kaggle and with the generosity of a good friend :)

You can see from the github timestamp that we didn't change the code at all after the deadline. We just had only some of the description for our project, and we missed it. Another big challenge was to adopt our code to the Gaudi HPU.

Since it wasn't a simple network, bug a GAN, i.e. a few networks that interact in the training process, it was a bit complex. We found it difficult to display the images, probably due to something related to converting KCRS to RSCK and back. We saw on the forums that we weren't the only one how faced this issue, but we didn't managed to solve it in the time we had.

Accomplishments that we're proud of

We are very proud that we got it to work, but we run our of time and couldn't actually run the training on the full dataset (we used "mim-dataset" of a few pictures, just to get it running). We feel that had we had more time - we could get impressive results, but we are still happy with what we achieved. As a side-project, we rather aim high and take the risk

What we learned

It's much more fun to work in a group and on a shared project, but it needs more time to coordinate between us, and squeezing it with a full-time job was less simple than we estimated :)

What's next for Ganilla

We are still aiming to train GANILLA with a new dataset, and see what result it can yield.

More information

[Medium]GANILLA — Fantasy Enhanced

Please note: from the github timestamp it can be seen that we didn't change the code at all after the deadline. We just had only some of the description for our project, and we missed it.

Built With

• pytorch