SpaceOpt: hyperparameter optimization via gradient boosting regression
SpaceOpt is a hyperparameter optimization algorithm that uses gradient boosting regression to find the most promising candidates for the next trial by predicting their evaluation score.
Installation
$ pip install spaceoptUsage
- Define a discrete hyperparameter search space, for example:
search_space = {
'a': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], # list of ordered numbers: ints
'b': [-3.5, -0.1, 0.0, 2.5, 10.0], # list of ordered numbers: floats
'c': [256, 512, 1024, 2048], # another list of ordered numbers
'd': ['ABC', 'IJK', 'XYZ'], # categorical variable
'e': [True, False], # boolean variable
# ... (add as many as you need)
}- Define your evaluation function:
def evaluation_function(spoint: dict) -> float:
# your code (e.g. model fit)
return y # score (e.g. model accuracy)
spoint = {'a': 4, 'b': 0.0, 'c': 512, 'd': 'XYZ', 'e': False}
y = evaluation_function(spoint)- Use SpaceOpt for a hyperparameter optimization:
from spaceopt import SpaceOpt
spaceopt = SpaceOpt(search_space=search_space,
target_name='y',
objective='maximize') # or 'minimize'
for iteration in range(200):
if iteration < 20:
spoint = spaceopt.get_random() # exploration
else:
spoint = spaceopt.fit_predict() # exploitation
spoint['y'] = evaluation_function(spoint)
spaceopt.append_evaluated_spoint(spoint)More examples here.
Advanced
- get multiple points by setting
num_spoints:
spoints = spaceopt.get_random(num_spoints=2)
# or
spoints = spaceopt.fit_predict(num_spoints=5)- control exploration vs. exploitation behaviour by adjusting
sample_size(default=10000), which is the number of candidates sampled for ranking:
spoint = spaceopt.fit_predict(sample_size=1000) # decreasing `sample_size` increses exploration
spoint = spaceopt.fit_predict(sample_size=100000) # increasing `sample_size` increses exploitation- add manually selected evaluation points to SpaceOpt:
my_spoint = {'a': 8, 'b': -3.5, 'c': 256, 'd': 'IJK', 'e': False}
my_spoint['y'] = evaluation_function(my_spoint)
spaceopt.append_evaluated_spoint(my_spoint)License
MIT License (see LICENSE).
