Tasting Success

Last week, I went on a mentor visit and made large strides with my toxicity model. During the visit, I got a better idea of how to structure my schedule for my final product so that I would be able to have a working product in the end and how to explain my methodology logically in a written format. For my model, I spent a lot of time over the weekend running benchmarks and trying to make improvements. I realized that my laptop, even though it doesn’t have a GPU, can train some of the models I’ve been using faster than Google Colab. I thought about buying Colab Premium, but that still doesn’t guarantee faster times. Thankfully, none of the benchmarks take ridiculously long—the latest models I’ve worked with take anywhere from 10 minutes to 75 minutes to be trained and evaluated. I was very, very happy to see that, after several trials and tweaks, my GCN model was able to beat first place on the leaderboard, and an ensemble combination of GCN and XGBoost performed even better. Along the way, I decided to log my progress and results in a Google Doc, which is far better than confusing comments in code. There’s still room for growth as I find better hyperparameter combinations, but I’m really excited to continue. I should probably also work on updating my ISM website this week, since I can finally find the time.