Hello!
I'm trying play around with the stochastic Frank-Wolfe benchmarks you have in this example and I'm having trouble reproducing the figure by running the provided example. The expected results are as follows:
However when I run the script out of the box, the first thing to note is that it takes a very long time; for 10 iters, these are the timing results:
CPU times:
user 16min 12s
sys: 7min 29s
total: 23min 42s
Wall time: 12min 25s
Perhaps this is fine seeing as ~10 steps should be sufficient (this would correspond to ~10^8 gradient evaluations which would match the figure in the example), but then the problem comes with the actual convergence results; if I set max_iter=500 (vs 1e4) and let it run for the full 2h x 3 runs, I get worse results:
In particular, after 10^8 gradient evaluations, the relative suboptimality
Some additional info that may be useful:
- Python 3.7.6
- SciPy 1.7.3
- NumPy 1.21.6
- Latest (development) version of COPT.
If useful, we can play around with the example in a colab notebook.
Thanks for any help!
Hello!
I'm trying play around with the stochastic Frank-Wolfe benchmarks you have in this example and I'm having trouble reproducing the figure by running the provided example. The expected results are as follows:
However when I run the script out of the box, the first thing to note is that it takes a very long time; for 10 iters, these are the timing results:
Perhaps this is fine seeing as ~10 steps should be sufficient (this would correspond to ~10^8 gradient evaluations which would match the figure in the example), but then the problem comes with the actual convergence results; if I set
max_iter=500(vs1e4) and let it run for the full 2h x 3 runs, I get worse results:In particular, after 10^8 gradient evaluations, the relative suboptimality
Some additional info that may be useful:
If useful, we can play around with the example in a colab notebook.
Thanks for any help!