A nevergrad extension for hpman
After using hpman to define the hyperparameters, call hpnevergrad through a single line of code to define hyperparameters in nevergrad parametrization type, to specify what are the parameters that the optimization should be performed upon.
basic.py
from hpman.m import _
from hpnevergrad import hpng
import nevergrad as ng
def train(learning_rate: float, batch_size: int, architecture: str) -> float:
lr = _('learning_rate', 1e-3, range=[1e-3, 1.0], scale='log')
bs = _('batch_size', 1, range=[1, 12])
architecture = _('architecture', 'conv', choices=['conv', 'fc'])
accuracy = (learning_rate - 0.2)**2 + (batch_size - 4)**2 + (
0 if architecture == "conv" else 10)
return accuracy
if __name__ == "__main__":
_.parse_file(__file__)
# define hyperparameters in nevergrad parametrization type
parametrization = hpng.get_parametrization(_)
# load hyperparameter in hpman to objective_function: a warpper
objective_function = hpng.get_objective_function(train, _)
optimizer = ng.optimizers.NGO(parametrization=parametrization, budget=100)
recommendation = optimizer.minimize(objective_function)
print(recommendation.kwargs)
>>>{'architecture': 'conv', 'batch_size': 1.4889356356452565, 'learning_rate': 0.0016799143450021905}Besides using hpng.get_parametrization and hpng.get_objective_function in you code, we also come with a command line tool hpng to provide similar functions to any existing file using hpman.
src.py
def train() -> float:
lr = _('lr', 1e-3, range=[1e-3, 1.0], scale='log')
bs = _('bs', 1, range=[1, 12])
architecture = _('architecture', 'conv', choices=['conv', 'fc'])
accuracy = (lr - 0.2)**2 + (bs - 4)**2 + (0 if architecture == "conv" else 10)
return accuracyIn shell:
hpng src.py:train
{'architecture': 'fc', 'bs': 1.70302632074525, 'lr': 0.0012628999415326785}
hpng src.py:train --budget 10
{'architecture': 'fc', 'bs': 2.039529723147301, 'lr': 0.0016307524751455055}