Fix dtype handing in `SklearnSGDModel`
With the release of scikit-learn 1.3.0, the sklearn.linear_model.SGDClassifier
and SGDRegressor
classes now adjust the dtype of their coefficients to the input data (rather than forcing the use of the numpy default float64 dtype).
As a result, our current integration tests fail, due to the use of float32 input data but default-dtype (hence, float64) initialized coefficients in declearn.model.sklearn.SklearnSGDModel
.
This issue needs fixing, for example using either the data-info field on data dtype, or casting the initial coefficients when processing the first batch of data.