Tuner re-proposes already-tried parameters

[StaffanAldenfalk]

I use the one two three steps on the site and have noticed that if i specify my tunables as:

tunables = [
    ('dropout_precentage1', HyperParameter(ParamTypes.INT, [40, 50])),
    ('dropout_precentage2', HyperParameter(ParamTypes.INT, [40, 50]))
]

Where the range on each parameter is only 10 values, it will do two guesses, for example [44, 46] and [48,42] and then continue to run the one of them which gave the best result. So if [44,46] gave the best results it will continue to propose the variables [44,46] for all of my iterations.

[micahjsmith]

Which tuner class are you using? And are you able to provide the scores that you passed to the tuner that led to this behavior?

[micahjsmith]

Here is a reproducible example:

import random
import sys
from collections import Counter

import numpy as np

from btb import HyperParameter
from btb.tuning import GP

if len(sys.argv) == 2:
    n = int(sys.argv[1])
else:
    n = 10

np.random.seed(1)
random.seed(1)

tunables = [
    ('a', HyperParameter('int', [0, 5])),
    ('b', HyperParameter('int', [0, 5])),
]

tuner = GP(tunables)

counter = Counter()
for _ in range(n):
    params = tuner.propose()
    tuner.add(params, 1.0)
    counter.update([tuple(sorted(params.items()))])

print(counter)

Produces:

Counter({(('a', 2), ('b', 3)): 2, (('a', 2), ('b', 5)): 1, (('a', 4), ('b', 5)): 1, (('a', 2), ('b', 1)):
1, (('a', 5), ('b', 1)): 1, (('a', 1), ('b', 1)): 1, (('a', 0), ('b', 1)): 1, (('a', 4), ('b', 2)): 1, (('a', 3), ('b', 4)): 1})

With 10 iterations and 36 combinations of a/b, there should be no reason to try {'a': 2, 'b': 3} twice.

[pvk-developer]

With the new re-implementation, issues #131 #132 #133 #134 , this problem is fixed. The Tuner will give you new combinations only, unless passing the argument allow_duplicates=True when calling tuner.propose. However, the tuner can't propose more than the tunable cardinality.
A short example is shown below:

from btb.tuning import GPTuner, IntHyperParam, Tunable

tunable = Tunable({
    'a': IntHyperParam(40, 50),
    'b': IntHyperParam(40, 50)
})

tuner = GPTuner(tunable)

tuner.propose(tunable.cardinality)  # This will work and will generate 121 combinations
tuner.propose(tunable.cardinality + 1)  # This will raise a value error
tuner.propose(tunable.cardinality + 1, allow_duplicates=True)  # This will work and will generate 122 combinations with repeated values

Bear in mind that the tuner will generate a random combination until we don't use record, also after recording n amounts of combinations, the tuner will be able to propose tunable.cardinality - n combinations.