Not sure if it's relevant to the motivation behind the implementation as discussed in #4324 , but a two-class y array with two classes present in the classes param proceeds with the sum of the weights being equal to the number of classes:

compute_class_weight('auto', [0, 1], iris.target[0:100])
array([ 1.,  1.])

While a three-class y array with only two of the classes present in the classes param does something different altogether:

compute_class_weight('auto', [0, 1], iris.target[0:120])
array([ 0.66666667,  0.66666667])

I had sidestepped this in compute_sample_weight in #4190 by determining the present classes from y itself. I'm happy to open a PR to remove the param, and was going to, but while the function is somewhat private, it is exposed in partial_fit in BaseSGDClassifier:

"In order to use 'auto' weights, use compute_class_weight('auto', classes, y)."

So does this need a deprecation warning? Some more discussion?