Some possible invariance tests:

1. Check that an error is raised if the number of features changes from one partial fit to another
2. Check that doing fit -> partial fit == partial_fit, partial_fit
3. Check that doing fit after a set of fit/partial_fit restart the estimator
4. Check that classifier handles correctly the classes argument in the partial_fit.

I didn't see this before commening on #3907. Please note that there I
express doubts about the fit() -> partial_fit() semantics, but agree that
whatever we choose, it should be consistent across estimators.

On 28 November 2014 at 18:25, Arnaud Joly notifications@github.com wrote:

Some possible invariance tests:

1. Check that an error is raised if the number of features changes
   from one partial fit to another
2. Check that doing fit -> partial fit == partial_fit, partial_fit
3. Check that doing fit after a set of fit/partial_fit restart the
   estimator
4. Check that classifier handles correctly the classes.

—
Reply to this email directly or view it on GitHub
#3896 (comment)
.

Check that doing fit -> partial fit == partial_fit, partial_fit

This is not true: a call to fit can go do several passes over the provided until it reaches convergence on that specific piece of training data, whereas a call to partial_fit will in general to an small incremental update assuming that there is more data to come and that there is no use trying to overfit this specific chunk.

@ogrisel extensive discussion here: #3907

Which is waiting for #4841 to get (completed :P) and merged.

We have quite a few tests in common tests now for partial_fit. Closing this.