ffx
ffx copied to clipboard
natekupp
How to increase iterations?
I'm trying to train the models but it is not converging. I tried to set max_iter in regressor, but it is not recognizing it. This is the warning I'm getting.
ConvergenceWarning: Objective did not converge. You might want to increase the number of iterations. Duality gap: 0.17428639502685428, tolerance: 0.0699999999999999
Should I just wait for it to finish, or, if not, how can I increase iterations?
Hey @RicardoFCJ - I've seen this a bunch in the past but have never had a chance to debug what's going on. If you have time would love to hear what you learn.
If you grep the codebase for max_iter you can see it's hard-coded to 5000 right now; it may be worth experimenting with that. Also it should certainly not be hard-coded :)
I'm sorry for the delayed response. (I'm not sure if that's what you're asking with "would love to hear what you learn", but ...) What I'm trying to learn is a dataset precision agriculture with 50 features and ~700 cases.
I am reviving this, because I have problems with it aswell. If the duality gap is too large, results are really not so good because the optimization did not converge.
A manual fix is to increase max_iter in ffx/core/model_factories.py inside the _pathwiseLearn() function (l. 414).
However, I was thinking of a more permanent solution. Since **fit_params get's passed to _pathwiseLearn() anyway, we could use that dict to drag along max_iter from run() and from the constructor of the FFXRegressor. That would be a bit ugly and difficult to test so I am open to better suggestions.
