Feature request: sample weights

Feature Requests
#1

Please would you consider enabling the data wizard to accept a column of sample weights to be applied.

Comment: I don’t know whether there are any constraints on weights; I’m sure I have seen examples of weight >1, but obviously if they are too big that could be a problem. Normalise them automatically?

NB in future, if there are “serious” imbalances in #items in training categories, it might be worth at least alerting the user. Some “diagnostic” feedback after datawizard setup (this, presence of NaNs, etc.) could also prove helpful.

In pure TF2 I’m trying weights in the interval (0, 1] - that means some very small weights per sample when there are 100k+ in one weighting band, but we’ll see how it goes.

#2

Hey @JulianSMoore,
When you say weight, you mean a constant to multiply the column data with?
Or do you mean weighted categorical cross-entropy?
Or to load existing model weights on a specific columns recommended model?

NB in future, if there are “serious” imbalances in #items in training categories, it might be worth at least alerting the user. Some “diagnostic” feedback after datawizard setup (this, presence of NaNs, etc.) could also prove helpful.

We have been talking about this a lot internally, expect a “data inspection” view to appear in some months which will visualize the dataset similar to how Kaggle does it.

#3

Hi @robertl , I was thinking of sample_weight as the parameter to model.fit described here.

I hadn’t thought about weightings elsewhere.

+1 for the data inspection view :slight_smile: