Closed BuenoRuben closed 4 months ago
Merging #108 (2a68953) into main (cd7647e) will decrease coverage by
0.02%
. The diff coverage is96.92%
.
@antoinedemathelin I added KMM to the reweighting methods' example, and noticed that when fitting with: n_training, n_source = 20, 20; it's almost instantly fitted n_training, n_source = 30, 30; it's taking some time n_training, n_source = 50, 50; it's taking about 2 minutes, while others are taking less than 5 sec
I don't know if your method is specifically having hight complexity, so I just prefer to tell you
@antoinedemathelin I added KMM to the reweighting methods' example, and noticed that when fitting with: n_training, n_source = 20, 20; it's almost instantly fitted n_training, n_source = 30, 30; it's taking some time n_training, n_source = 50, 50; it's taking about 2 minutes, while others are taking less than 5 sec
I don't know if your method is specifically having hight complexity, so I just prefer to tell you
Hi @BuenoRuben,
Yes, that's right, the qp solver used in KMM is pretty slow when n_source increase (note that n_source = 50 corresponds to 400 samples I think). This can be fasten by using cvxopt
, which we will add as an option solver. I also plan to implement the Frank-Wolfe algorithm for KMM, which really speed up the optim.
As a first workaround, maybe we can reduce the default number of max_iter
to 100 instead of 1000. The algorithm will not fully converge, but it will be faster...
Hello @BuenoRuben sory for this infinitre PR ;) could you also rename KMMAdapter
to KMMReweightAdapter
and same fro KLIEP ?
After that it seems everything will be OK
should I then rename KMM
into KMMReweight
then?
based on this paper: https://arxiv.org/pdf/2102.02291.pdf