occ suspiciouslogin:optimize --stats causes an exception to be thrown about suspicious_login HOT 5 CLOSED

Please test #189 https://patch-diff.githubusercontent.com/raw/nextcloud/suspicious_login/pull/189.patch

from suspicious_login.

@ChristophWurst Incredible, you're faster than light 😉 I think it looks much better now, although I don't know how many coffees I have to drink until it finishes:

> ./occ suspiciouslogin:optimize --stats
Optimizing a MLP trainer in 100 steps. Enjoy your coffee!

Epoch 0: epochs= 443 layers=12 shuffledRate=1.500 randomRate=1.081, learningRate=0.057
  Step width for next config neighbor: 0.8
Epoch 1: epochs= 390 layers=13 shuffledRate=1.500 randomRate=0.639, learningRate=0.076
  Step width for next config neighbor: 0.8
  Result got worse: 0.95188714303814->0.94655344382768
Epoch 2: epochs= 279 layers= 9 shuffledRate=1.176 randomRate=0.900, learningRate=0.070
  Step width for next config neighbor: 0.788
  Result got worse: 0.95188714303814->0.94614394707304
Epoch 3: epochs= 548 layers=12 shuffledRate=0.990 randomRate=1.377, learningRate=0.076
  Step width for next config neighbor: 0.77618
  Result got worse: 0.95188714303814->0.93428117339547
Epoch 4: epochs= 291 layers=10 shuffledRate=1.500 randomRate=0.606, learningRate=0.065
  Step width for next config neighbor: 0.7645373
  Result got worse: 0.95188714303814->0.94304861644041
Epoch 5: epochs= 264 layers=14 shuffledRate=1.245 randomRate=1.212, learningRate=0.034
  Step width for next config neighbor: 0.7530692405
...

Just one small cosmetic thing I realized, there is a supernumerary white-space between the epochs= and the following figure e.g. 291, which is different to the following values being displayed.

from suspicious_login.

lol, yeah, that message is from a time where I had very little training data. I think you shouldn't wait for the completion. Haven't done such a optimization run in a few months (default params seem to do well on most systems) but please let me know the results!

It might take a day to finish. It totally depends on which area of the parameter space the algorithm settles 🙈

from suspicious_login.

Also please note that I later realized that this optimization run has a little flaw in that it doesn't ensure to always test on the exact same data. The results aren't scientifically correct.

from suspicious_login.

Just one small cosmetic thing I realized, there is a supernumerary white-space between the epochs= and the following figure e.g. 291, which is different to the following values being displayed.

If you shoot a PR I'll integrate it!

from suspicious_login.