enigmato | Devpost

Dataset examples we looked at while exploring the dataset.

Inspiration

We, as a team, have always loved puzzles. The pensiveness of it all almost makes those 100+ piece puzzles a bit meditative. However, we are all still Computer Science students, so as much as we live for puzzles, we also live for automating everything in our life. One thing led to another, and here we are, presenting a neural net that can solve puzzles!

What it does

it solves puzzles

How we built it

Because we wanted to optimize prototyping speed, we used fast.ai to make the model, specifically working with their vision library. To avoid having to retrain an entire model from scratch, we worked off of xresnet101.

Challenges we ran into

Yash used integer division to compute the model's accuracy and worked for half an hour to figure out why the accuracy was always 0%. For real, though, the most complex challenge we ran into was trying to hyperparameter tune and figured out how to do the necessary data augmentation for a more robust model.

Accomplishments that we're proud of

We're really proud not only of our model's performance but also of the speed at which we were able to prototype to achieve that level of performance. We were able to work separately on the challenge when needed yet used GitHub extensively to properly manage code.

What we learned

I feel like, as a team, we've learned about the importance of managing collaboration and individual work. Within the first hours, we quickly realized that 3 people watching one person live-stream their VS Code wasn't very productive. We needed the benefits of collaboration while maintaining the speed of multiple people working simultaneously on the same problem. For that, we had to learn to use Git and GitHub well.