Closed Expertium closed 2 months ago
While it's not a catastrophically large difference
It's rather small as far as I can tell, and I'm not sure it makes sense to throw a bunch of extra computing resources at it to squeeze out a tiny improvement.
IMO, the current result is good enough. I have tried different combinations of hype-parameters, but none of them improved the metrics significantly.
According to the benchmark results, FSRS-rs underperforms the Python version. While it's not a catastrophically large difference, it would be nice to try and bridge the gap. The only simple way of doing it that I can think of is training FSRS-rs for more epochs.