Two Questions about Regression

  1. In notebook 06_multicat, we use y_range (which uses sigmoid range) to rescale our output between -1 and +1. Jeremy says coordinates in fastai and PyTorch are always rescaled between -1 and +1). Why? Is this just a convention or there is actual advantages to do so?

  2. I am working on an Image Regression Problem. My independent variable is an image that contains people and my dependent variable is a scalar number, which is the number of people in the image. I am building a model to predict number of people in an image. I chose to use RegressionBlock in my blocks (I assume that is the correct block and not PointBlock). My question is, should I still use y_range? Is y_range required for each Regression Problem? And if yes, how do I know the suitable range to rescale (like how above we know that coordinates are scaled to -1 and +1) or is this just a choice the practitioner makes like choosing an optimizer?

1 Like