Lower traing speed between simp_iteration1 and simp_iteration2

[Embracely]

I have noticed that the training speed is dramatically decreased after the first simplification. Depth initialisation is included in that iteration, however it happens in each 5000 step as well. I am curious why the training speed decreases even if the gaussian number is fewer?

[Embracely]

Hello @fatPeter and everyone, that problem confused me for a while, really grateful for your help!