Pictionary Plunge in PyTorch

Inspiration

Having worked with PyTorch before with image data, this Pictionary challenge was one we were immediately gravitating towards upon the reveal of TD2023's challenges for its challenging aspect and opportunity to exercise some serious ML skills.

What it does

Our PyTorch model takes in inputs of .ndjson files of different drawing and outputs predictions of what the doodle is classified as.

How we built it

We utilized PyTorch's transfer learning capabilities to use a pre-trained VGG neural network and applied it to our data, along with some layer modifications to fit our needs.

Challenges we ran into

Our data was not easy to work with since it was in the form of JSON objects, so we worked hard to change each drawing to image data so that it was something our model could have fed in and learn from.

Accomplishments that we're proud of

We were able to achieve a decent accuracy in multiclass prediction, which is something we all haven't exercised vey often in our courses or other work.

What we learned

PyTorch image processing for model input and advanced multiclass prediction techniques were some skills we are proudly walking away with to use in future work and projects.