Slice sampler shrank to zero

[JohnReid]

I'm trying a relatively vanilla optimisation over 9 dimensions, 5 of which are categorical (2 values), the other numeric. I'm getting the following exception:

Exception: Slice sampler shrank to zero! in File "/home/john/src/spearmint/spearmint/spearmint/util.py", line 69, in direction_slice

What does it mean and do you have any ideas what I can do to get around this? I've tried with and without noise. I'm using GPEIOptChooser with a local driver, max concurrent of 1

Thanks, John.

[JasperSnoek]

Aha, is it possible that your code is returning NaNs or infs? This error basically means that the sampling algorithm is not finding any reasonably valid setting for the hyperparameters (e.g. everything returns a NaN).

Would you mind trying to set the categorical parameters to ints (with two values) and seeing if that changes anything?

Jasper

[JohnReid]

I haven't had time to re-run this yet but I notice you put a fix in for cases where the first two results are the same. Was the fix for this issue? I suspect my values don't change for some large parts of parameter-space.

[JohnReid]

I fetch/merged your latest changes and switched the categorical parameters to ints. Now I get a different error after two evaluations (-0.775600 and -0.790467) have completed:

----------------------------------------
Current best: -0.775600 (job 0)
19998 candidates   1 pending   1 complete
Maximum number of jobs (1) pending.

----------------------------------------
Current best: -0.790467 (job 1)
19998 candidates   0 pending   2 complete
Choosing next candidate...
Traceback (most recent call last):
File "/home/john/src/spearmint/spearmint/bin/../spearmint/main.py", line 319, in <module>
    main()
File "/home/john/src/spearmint/spearmint/bin/../spearmint/main.py", line 165, in main
    while attempt_dispatch(experiment_config, expt_dir, chooser, driver, options):
File "/home/john/src/spearmint/spearmint/bin/../spearmint/main.py", line 244, in attempt_dispatch
    job_id = chooser.next(grid, values, durations, candidates, pending, complete)
File "/home/john/src/spearmint/spearmint/spearmint/chooser/GPEIOptChooser.py", line 217, in next
    self._real_init(grid.shape[1], values[complete])
File "/home/john/src/spearmint/spearmint/spearmint/chooser/GPEIOptChooser.py", line 173, in _real_init
    self.hyper_samples = state['hyper_samples']
KeyError: 'hyper_samples'
Exception TypeError: "'NoneType' object is not callable" in <bound method Locker.__del__ of <Locker.Locker instance at 0x45f20e0>> ignored

My config is:

language: PYTHON
name:     "MotEvo"

variable {
name: "EMPRIOR"
type: INT
size: 1
min: 1
max: 0
}

variable {
name: "UFEWMPRIOR"
type: FLOAT
size: 1
min: 0.
max: 800.
}

variable {
name: "BGPRIOR"
type: FLOAT
size: 1
min: 0.5
max: 0.9999999
}

#variable {
# name: "MARKOVORDERBG"
# type: INT
# size: 1
# min:  0
# max:  3
#}

and I'm running spearmint with these options:

    --max-concurrent=1 \
    --max-finished-jobs=10000 \
    --method-args=noiseless=0 \
    --driver=local \
    --method=GPEIOptChooser \

[JohnReid]

I've since changed the max value for BGPRIOR from 0.9999999 to 0.99 and spearmint seems happier now.

[Quanticles]

Is there a known solution to the "Slice sampler shrank to zero" error? Many of my variables are log-domain, like 0.0001-0.1. I'm considering making the variable an exponent (the spearmint parameter is -5 to -1, and I use it as 10^parameter in the script).

[JasperSnoek]

Ah, so there are solutions that will be rolled out hopefully very shortly, depending on the cause of this. One potential cause is that there are invalid values being thrown in as results (like NaNs, Infs, etc), in which case we have a solution but that will break the code currently. In other cases, in the meantime, usually deleting the *Chooser.pkl and rerunning spearmint usually clears the issue (this causes the sampler to reset and reburn).

It is definitely a good idea to log transform your parameters if you think that's a sensible transformation. We're also about to roll out code to learn the transformation automatically (i.e. according to this paper: http://people.seas.harvard.edu/~jsnoek/bayesopt-warping.pdf).

On Thu, Jan 9, 2014 at 5:13 PM, Quanticles [email protected] wrote:

Is there a known solution to the "Slice sampler shrank to zero" error? Many of my variables are log-domain, like 0.0001-0.1. I'm considering making the variable an exponent (the spearmint parameter is -5 to -1, and I use it as 10^parameter in the script).

— Reply to this email directly or view it on GitHubhttps://github.com/JasperSnoek/spearmint/issues/9#issuecomment-31982666 .

[Quanticles]

Deleting the *Chooser.pkl file worked for me. My cleaning script wasn't removing it properly due to a bug in VMWare Workstation. Thanks for continuing to maintain the code and whatnot