Doing some data science tonight. When it came time to tune my hyperparameters, I remembered I still had an account at Weights & Biases and decided to give their “sweeps” feature a spin. Hyperparameter tuning is usually something I build into my notebooks/early scripts on a project and do it manually/simply. I have to say though, W&B made it pretty easy, their api is very easy to implement, is highly configurable, and has some pretty nice looking graphs to visualize what’s going on.
I especially apreciate the ability to filter the bad sweeps (as I did below) and calculate the importance/corelation of each parameter.
I think I’ll start using this, it’s pretty slick. I also notice they’ve added MANY new features since I last looked at it.
Note: this datset was very small, so, that’s why accuracy is still quite low on this one (only 50 samples)