You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Expand the tutorial on batching comparing SGD and SVRG:
Show that, if you set the stepsize of the right size, the algorithm converge to the correct estimate
Compare with SGD, which won't converge without a decreasing stepsize schedule carefully calibrated.
Compare the iteration number, showing that SVRG requires less epochs to get to same log-likelihood level than SGD
Ideally, when we will have a run_iterator method we should showcase a simplified interface that will look something like:
# add a batcher parameter to GLM init, optional, if set it will try to run the algorithm using the iterator.model=GLM(solver_name="SVRG", batcher=my_iterator)
model.fit(X, y) # X could be lazy loaded
The text was updated successfully, but these errors were encountered:
As part of this, need to show what extra steps are required for SVRG beyond just calling update (and update the update docstring to point users here. @bagibence said this is already shown in a test -- which one was that?
The docstrings of the update method of SVRG should point to our tutorial on how to construct the external loop. Eventually, no external loop will be needed, since we will implement a run iterator method.
Expand the tutorial on batching comparing SGD and SVRG:
Ideally, when we will have a
run_iterator
method we should showcase a simplified interface that will look something like:The text was updated successfully, but these errors were encountered: