def _coroutine(self):
# type: (...) -> TunerCoroutine
"""Function that returns generator coroutine that optimizes hyperparams."""
def key_for_hparams(hparams):
return tuple(sorted(six.iteritems(hparams)))
hparams_to_point = {}
for hparams in self.initial_hparams:
point = yield hparams
key = key_for_hparams(hparams)
hparams_to_point[key] = point
def dist(hparams0, hparams1):
assert set(six.iterkeys(hparams0)) == set(six.iterkeys(hparams1))
squared_dist = sum((hparams0[k] - hparams1[k])**2 for k in hparams0)
return squared_dist**0.5
while True:
points = list(six.itervalues(hparams_to_point))
_, coeffs = learn_lin_reg.learn_linear_regularizer(points)
hparams = dict(zip(self.hparam_names, coeffs))
# If the hparams we get from solving the LP are very close to ones
# we've already tried, revert to random sampling instead.
min_dist = min(dist(hparams, dict(prev_hparams))
for prev_hparams in hparams_to_point)
if min_dist < self.eps:
hparams = self.sample_hparams()
point = yield hparams
key = key_for_hparams(hparams)
hparams_to_point[key] = point
def test_learn_linear_regularizer(self, data_points,
expected_normalized_coefficients):
alpha, coefficients = learn_lin_reg.learn_linear_regularizer(
data_points)
normalized_coefficients = coefficients / alpha
np.testing.assert_almost_equal(expected_normalized_coefficients,
normalized_coefficients)