Draw pictures on a canvas and have the AI guess what you are drawing after each stroke with high accuracy.
The model is a custom PyTorch model similar to ResNet 18 but adapted for our specific use case and trained on the sample data provided. The model converts the strokes into a 256x256 image before running classifications. The web portion was built on a Next.js frontend and django backend to allow easy interaction with our model.
Overfitting was a recurring problem during training. This was mitigated by lowering learning rate, increasing dropout rate, and adjusting batch size continuously until we could no longer decrease loss by any of these methods.
The model completed training with a 0.9 validation accuracy. Furthermore the classes had relatively consistent F1 scores with the lowest at 0.82.
We learned about all of the nuances it takes to build and train a successful image classification model.
We would like to experiment with making the model larger to increase data. We also would like to incorporate the user feedback that we collect to continuously train the model on new classes.
