Does the model retraining work correctly?


I have submitted and executed a notebook with a training frequency of 25. When checking the execution log, I noticed that my model was first trained on moon 268 and then got retrained on 275 (with a difference of 8 moons).

However, the retraining moon (moon 275) looks incorrect. The starting moon used to generate scores on the LB is 268, and my model was correctly trained on that moon at the start. With a training frequency of 25, the retraining must be on moon 293 (268 + 25), not 275.

Can you check this?

My run ID is 4683.

weird, it’s kind of like 32-25=7… (and 275-268=7) somehow it’s counting from the end of OOS or something.

Actually @bigfish, I tried with retrain_period = 11, and it also started at 275…

This is because, from the documentation:

" Frequency of which your submission’s train function will be called. The frequency is a modulo on the moon, not loop"

And so 275 % 25 == 0, and 275 % 11 == 0… Imo it should be a modulo of the loop for sure. At least that’s the more natural thing

@vivacious-thomas is right and correctly cited the documentation.

Also, if you do not submit a resources folder, the first loop is always a train. (unless forced by force_first_train or the modulo is hit)

I see. Thanks for the clarification.