def test_xform_params_distribution():
pg = synthetic_positives.BasicPositivesGenerator(
dilate_time_stdev_factor=2.0,
dilate_flux_stdev_factor=3.0,
rng=random.Random(1234),
max_time_dilation = 6,
max_flux_dilation = 16
)
xforms = [pg.make_xform() for _ in range(1000)]
time_dilations = [xf._dilate_time for xf in xforms]
flux_dilations = [xf._dilate_flux for xf in xforms]
# Hard rules - should never be violated
assert 1/6. < min(time_dilations)
assert max(time_dilations) < 6
assert 1/16. < min(flux_dilations)
assert max(flux_dilations) < 16
# These *could* be violated by chance. If one fails, graph some distributions,
# manually check sanity, and adjust bounds if it looks good.
assert abs(np.median(time_dilations) - 1.0) < 0.06
assert abs(np.median(flux_dilations) - 1.0) < 0.06
def adjusted_percentile(max_dilation):
# need to rebalance the CDF since we are clipping the tails of the distribution
n_stds_clipped = -np.log(max_dilation)
return 100 * (norm.cdf(1) - norm.cdf(n_stds_clipped)) / (1 - 2 * norm.cdf(n_stds_clipped))
first_percentile_time_dilations = np.percentile(time_dilations, adjusted_percentile(6/2.0))
first_percentile_flux_dilations = np.percentile(flux_dilations, adjusted_percentile(16/3.0))
assert abs(first_percentile_time_dilations - 2) < 0.1
assert abs(first_percentile_flux_dilations - 3) < 0.14
def __init__(self):
super(FakePositivesBuilder, self).__init__(seed=65,
lc_lookup_chunk_size=4)
self.basic_synthetic_positives = synthetic_positives.BasicPositivesGenerator(
)
self.positives_fex = example_pair.PairFexFromPointwiseFex(
fex=metadata_features.MetadataValueExtractor(),
label=True,
)
def test_smoke_subsampler():
pg = synthetic_positives.BasicPositivesGenerator(
dilate_time_stdev_factor=2.0,
dilate_flux_stdev_factor=3.0,
rng=random.Random(1234)
)
pair_subsampler = synthetic_positives.PositivePairSubsampler(
rng=np.random.RandomState(123)
)
fpp = pg.make_positive_pair(simulate.TestLC.make_realistic_gauss())
fpp2 = pair_subsampler.apply(fpp)
assert _total_points(fpp) == (115, 115)
assert _total_points(fpp2) == (104, 78)
def __init__(
self, *, seed, bands, window_size, subsample_min_rate, subsample_max_rate
):
super(BasicPositivesBuilder, self).__init__(seed=seed)
self.basic_synthetic_positives = synthetic_positives.BasicPositivesGenerator()
self.subsampler = synthetic_positives.PositivePairSubsampler(
min_rate=subsample_min_rate,
max_rate=subsample_max_rate,
rng=np.random.RandomState(seed)
)
self.positives_fex = synthetic_positives.RawValuesFullPositives(
bands, window_size
)
def test_feature_extraction(tf_sess):
pg = synthetic_positives.BasicPositivesGenerator(
dilate_time_stdev_factor=2.0,
dilate_flux_stdev_factor=3.0,
rng=random.Random(1234)
)
params = {"batch_size": 20, "window_size": 10, "lc_bands": ["b"]}
rv_fex = synthetic_positives.RawValuesFullPositives.from_params(params)
lcs = [simulate.TestLC.make_realistic_gauss(scale) for scale in [10.0, 15.0, 30.0]]
fpp_gen = (pg.make_positive_pair(lc) for lc in lcs)
dataset = rv_fex.make_dataset(fpp_gen)
assert isinstance(dataset, tf.data.Dataset)
get_next_tensor = dataset.make_one_shot_iterator().get_next()
values = []
while True:
try:
values.append(tf_sess.run(get_next_tensor))
except tf.errors.OutOfRangeError:
break
assert len(values) == 3