This is my first post in the forum and I hope this is the correct place to ask this.
I’m following the structured data predictors as shown as lesson 4 to do the kaggle competition of housing prices(link).
When trying to find the learning rate, the lr_find() function is stuck on 0% and the learn.sched() function is plotting an empty plot.
If I may hazard a guess: Your batch size is large compared to the overall size. If there’s not that many minibatches being passed to the learner, the lr finder doesn’t have enough data to do a proper plot.
You can either use a lower batch size (in my personal experimentation I found much better performance for a SMALL batchsize, in terms of optimizing) or use the lr_find2 method instead, which works with a fixed number of iterations.
On a more general note, I’m wondering about feature engineering such as your step where you compute the total square footage by adding up basement, 1st floor and 2nd floor.
I feel like if total footage is important, then the NN should recognize that as a feature, so one of the first layer steps should have a “filter”, or row, that sums up those relevant variables…