Closed hugoabonizio closed 3 years ago
I'm not sure if implementing group normalization as a gate in this scenario would work the same way. As I know, it is a common practice to scale/normalize the features before feeding it to a neural network (or SVMs, LRs, and other kinds of linear models), specially for tabular data like Iris.
I was think more of a sklearn's StandardScaler
, that has the whole dataset to compute the mean and standard deviation, instead of a batch normalization approach. I agree that a BN layer would be a great feature, specially for deep learning models, but I'm proposing a simpler method for using in MLP models.
What do you think @christopherzimmerman?
@hugoabonizio I think that would be a great addition. I know that StandardScaler
treats NaN
as missing values and doesn't include them in normalization, so I might work on some iterators that skip missing values, so that the mean
, std
methods can have an option to ignore them.
Maybe feature scalers should be on an utils module of this library, I can send a PR. WDYT?