In cd_fast.pyx, the function sparse_enet_coordinate_descent at lines 483 & 484 doesn't take the absolute value of the weight to check the max, it uses the signed value. This means it will ignore negative updates, and probably trigger the duality gap check more frequently than it should. Looking at lines 270 and 271 of enet_coordinate_descent the absolute value of the weight is used.
I think this is a replica of issue #283, which was fixed many years ago in enet_coordinate_descent, enet_coordinate_descent_gram and enet_coordinate_descent_multi_task.
The text was updated successfully, but these errors were encountered: