search

jolars / slopecd

4 stars 2 forks source link

Oracle divergence on Scheetz #48

Closed mathurinm closed 2 years ago

mathurinm commented 2 years ago

investigates #44

mathurinm commented 2 years ago

If we run the hybrid solver for a long time, it has a sharp drop where it finally converges (note that this is the gap, so it's sure that there is convergence):

Epoch: 11281, loss: 0.705251785459538, gap: 1.09e-05
Epoch: 11291, loss: 0.7052517853959062, gap: 1.09e-05
Epoch: 11301, loss: 0.7052517853324648, gap: 1.09e-05
Epoch: 11311, loss: 0.7052517852958691, gap: 4.50e-14

Then, with the correct clusters, the oracle solver converges in 1 epoch:

Epoch: 1, loss: 0.7052517852958691, gap: 6.99e-15

image

jolars commented 2 years ago

Ah, thanks for looking into this. Then there is no problem! Sorry for the noise!

mathurinm commented 2 years ago

It's not noise IMO, it is very surprising that the gap drops suddenly like that. I wonder what happens at the last iteration