 So, let's briefly talk about the pipeline of de-blowning. What you really want to do is look at the data, calculate relevant summary statistics of the data, and ask yourself how would you know which thing is which labels. And then you build your de-blowning system. Central to the process are data loaders. And we generally will do data augmentation when we load the data. What kind of augmentations make sense? Look at the data here and understand it, and what kind of data augmentation is happening in the code. And why does that make any sense?