def test_lomb_scargle_regular_single_freq():
"""
Test Lomb-Scargle model features on regularly-sampled periodic data with one
frequency/multiple harmonics. Estimated parameters should be very accurate in
this case.
"""
frequencies = np.hstack((WAVE_FREQS[0], np.zeros(len(WAVE_FREQS)-1)))
amplitudes = np.zeros((len(frequencies),4))
amplitudes[0,:] = [8,4,2,1]
phase = 0.1
times, values, errors = regular_periodic(frequencies, amplitudes, phase)
all_lomb = sft.generate_science_features(times, values, errors,
lomb_features)
# Only test the first (true) frequency; the rest correspond to noise
npt.assert_allclose(all_lomb['freq1_freq'], frequencies[0])
# Hard-coded value from previous solution
npt.assert_allclose(0.001996007984, all_lomb['freq1_lambda'], rtol=1e-7)
for (i,j), amplitude in np.ndenumerate(amplitudes):
npt.assert_allclose(amplitude,
all_lomb['freq{}_amplitude{}'.format(i+1,j+1)], rtol=1e-2,
atol=1e-2)
# Only test the first (true) frequency; the rest correspond to noise
for j in range(1, NUM_HARMONICS):
npt.assert_allclose(phase*j*(-1**j),
all_lomb['freq1_rel_phase{}'.format(j+1)], rtol=1e-2, atol=1e-2)
# Frequency ratio not relevant since there is only; only test amplitude/signif
for i in [2,3]:
npt.assert_allclose(0., all_lomb['freq_amplitude_ratio_{}1'.format(i)], atol=1e-3)
npt.assert_array_less(10., all_lomb['freq1_signif'])
# Only one frequency, so this should explain basically all the variance
npt.assert_allclose(0., all_lomb['freq_varrat'], atol=5e-3)
# Exactly periodic, so the same minima/maxima should reoccur
npt.assert_allclose(0., all_lomb['freq_model_max_delta_mags'], atol=1e-6)
npt.assert_allclose(0., all_lomb['freq_model_min_delta_mags'], atol=1e-6)
# Linear trend should be zero since the signal is exactly sinusoidal
npt.assert_allclose(0., all_lomb['linear_trend'], atol=1e-4)
folded_times = times % 1./(frequencies[0]/2.)
sort_indices = np.argsort(folded_times)
folded_times = folded_times[sort_indices]
folded_values = values[sort_indices]
# Residuals from doubling period should be much higher
npt.assert_array_less(10., all_lomb['medperc90_2p_p'])
# Slopes should be the same for {un,}folded data; use unfolded for stability
slopes = np.diff(values) / np.diff(times)
npt.assert_allclose(np.percentile(slopes,10),
all_lomb['fold2P_slope_10percentile'], rtol=1e-2)
npt.assert_allclose(np.percentile(slopes,90),
all_lomb['fold2P_slope_90percentile'], rtol=1e-2)
def test_lomb_scargle_irregular_multi_freq():
"""
Test Lomb-Scargle model features on irregularly-sampled periodic data with
multiple frequencies, each with a single harmonic. More difficult than
regularly-sampled case, so we allow parameter estimates to be slightly noisy.
"""
frequencies = WAVE_FREQS
amplitudes = np.zeros((len(frequencies),4))
amplitudes[:,0] = [4,2,1]
phase = 0.1
times, values, errors = irregular_periodic(frequencies, amplitudes, phase)
all_lomb = sft.generate_science_features(times, values, errors,
lomb_features)
for i, frequency in enumerate(frequencies):
npt.assert_allclose(frequency,
all_lomb['freq{}_freq'.format(i+1)], rtol=1e-2)
for (i,j), amplitude in np.ndenumerate(amplitudes):
npt.assert_allclose(amplitude,
all_lomb['freq{}_amplitude{}'.format(i+1,j+1)],
rtol=1e-1, atol=1e-1)
for i in [2,3]:
npt.assert_allclose(amplitudes[i-1,0] / amplitudes[0,0],
all_lomb['freq_amplitude_ratio_{}1'.format(i)], atol=2e-2)
npt.assert_allclose(frequencies[i-1] / frequencies[0],
all_lomb['freq_frequency_ratio_{}1'.format(i)], atol=5e-2)
npt.assert_array_less(10., all_lomb['freq1_signif'])
def test_lomb_scargle_regular_multi_freq():
"""
Test Lomb-Scargle model features on regularly-sampled periodic data with
multiple frequencies, each with a single harmonic. Estimated parameters should
be very accurate in this case.
"""
frequencies = WAVE_FREQS
amplitudes = np.zeros((len(frequencies),4))
amplitudes[:,0] = [4,2,1]
phase = 0.1
times, values, errors = regular_periodic(frequencies, amplitudes, phase)
all_lomb = sft.generate_science_features(times, values, errors,
lomb_features)
for i, frequency in enumerate(frequencies):
npt.assert_allclose(frequency,
all_lomb['freq{}_freq'.format(i+1)])
for (i,j), amplitude in np.ndenumerate(amplitudes):
npt.assert_allclose(amplitude,
all_lomb['freq{}_amplitude{}'.format(i+1,j+1)],
rtol=5e-2, atol=5e-2)
for i in [2,3]:
npt.assert_allclose(amplitudes[i-1,0] / amplitudes[0,0],
all_lomb['freq_amplitude_ratio_{}1'.format(i)], atol=2e-2)
npt.assert_array_less(10., all_lomb['freq1_signif'])
def test_lomb_scargle_irregular_single_freq():
"""
Test Lomb-Scargle model features on irregularly-sampled periodic data with one
frequency/multiple harmonics. More difficult than regularly-sampled case, so
we allow parameter estimates to be slightly noisy.
"""
frequencies = np.hstack((WAVE_FREQS[0], np.zeros(len(WAVE_FREQS)-1)))
amplitudes = np.zeros((len(WAVE_FREQS),4))
amplitudes[0,:] = [8,4,2,1]
phase = 0.1
times, values, errors = irregular_periodic(frequencies, amplitudes, phase)
all_lomb = sft.generate_science_features(times, values, errors,
lomb_features)
# Only test the first (true) frequency; the rest correspond to noise
npt.assert_allclose(all_lomb['freq1_freq'], frequencies[0], rtol=1e-2)
# Only test first frequency here; noise gives non-zero amplitudes for residuals
for j in range(NUM_HARMONICS):
npt.assert_allclose(amplitudes[0,j],
all_lomb['freq1_amplitude{}'.format(j+1)], rtol=5e-2, atol=5e-2)
if j >= 1:
npt.assert_allclose(phase*j*(-1**j),
all_lomb['freq1_rel_phase{}'.format(j+1)], rtol=1e-1, atol=1e-1)
npt.assert_array_less(10., all_lomb['freq1_signif'])
# Only one frequency, so this should explain basically all the variance
npt.assert_allclose(0., all_lomb['freq_varrat'], atol=5e-3)
npt.assert_allclose(-np.mean(values), all_lomb['freq_y_offset'], rtol=5e-2)
def test_feature_generation():
"""Compare generated features to reference values."""
this_dir = os.path.join(os.path.dirname(__file__))
test_files = [
os.path.join(this_dir, 'data/257141.dat'),
os.path.join(this_dir, 'data/245486.dat'),
os.path.join(this_dir, 'data/247327.dat'),
]
features_extracted = None
values_computed = None
for i, ts_data_file_path in enumerate(test_files):
t, m, e = ctt.parse_ts_data(ts_data_file_path)
features = sft.generate_science_features(t, m, e, cfg.features_list_science)
sorted_features = sorted(features.items())
if features_extracted is None:
features_extracted = [f[0] for f in sorted_features]
values_computed = np.zeros((len(test_files),
len(features_extracted)))
values_computed[i,:] = [f[1] for f in sorted_features]
def features_from_csv(filename):
with open(filename) as f:
feature_names = f.readline().strip().split(",")
feature_values = np.loadtxt(f, delimiter=',')
return feature_names, feature_values
this_dir = os.path.join(os.path.dirname(__file__))
features_expected, values_expected = features_from_csv(
os.path.join(this_dir, "data/expected_features.csv"))
npt.assert_equal(features_extracted, features_expected)
npt.assert_array_almost_equal(values_computed, values_expected)
def test_weighted_average():
"""Test weighted average and distance from weighted average features."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['weighted_average'])
weighted_std_err = 1. / sum(errors**2)
error_weights = 1. / (errors)**2 / weighted_std_err
weighted_avg = np.average(values, weights=error_weights)
npt.assert_allclose(f.values()[0], weighted_avg)
dists_from_weighted_avg = values - weighted_avg
stds_from_weighted_avg = (dists_from_weighted_avg /
np.sqrt(weighted_std_err))
# TODO broken feature
f = sft.generate_science_features(times, values, errors,
['percent_beyond_1_std'])
npt.assert_equal(f.values()[0], np.mean(stds_from_weighted_avg > 1.))
def test_stetson():
"""Test Stetson variability features."""
times, values, errors = irregular_random(size=201)
f = sft.generate_science_features(times, values, errors, ['stetson_j'])
# Stetson mean approximately equal to standard mean for large inputs
dists = np.sqrt(float(len(values)) / (len(values) - 1.)) * (values - np.mean(values)) / 0.1
npt.assert_allclose(f.values()[0],
np.mean(np.sign(dists**2-1)*np.sqrt(np.abs(dists**2-1))),
rtol=1e-2)
# Stetson_j should be somewhat close to (scaled) variance for normal data
npt.assert_allclose(f.values()[0]*0.1, np.var(values), rtol=2e-1)
# Hard-coded original value
npt.assert_allclose(f.values()[0], 7.591347175195703)
f = sft.generate_science_features(times, values, errors, ['stetson_k'])
npt.assert_allclose(f.values()[0], 1./0.798 * np.mean(np.abs(dists)) / np.sqrt(np.mean(dists**2)), rtol=5e-4)
# Hard-coded original value
npt.assert_allclose(f.values()[0], 1.0087218792719013)
def test_scatter_res_raw():
"""Test feature that measures scatter of Lomb-Scargle residuals."""
times, values, errors = irregular_random()
lomb_model = sft.sf.lomb_scargle_model(times, values, errors)
residuals = values - lomb_model['freq_fits'][0]['model']
resid_mad = np.median(np.abs(residuals - np.median(residuals)))
value_mad = np.median(np.abs(values - np.median(values)))
f = sft.generate_science_features(times, values, errors, ['scatter_res_raw'])
npt.assert_allclose(f['scatter_res_raw'], resid_mad / value_mad, atol=3e-2)
def test_percent_close_to_median():
"""Test feature which finds the percentage of points near the median value."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors,
['percent_close_to_median'])
amplitude = (np.abs(max(values)) - np.abs(min(values))) / 2.
#amplitude = (max(values) - min(values)) / 2.
within_buffer = np.abs(values - np.median(values)) < 0.2*amplitude
npt.assert_allclose(f.values()[0], np.mean(within_buffer))
def test_qso_features():
"""
Test features which measure fit of QSO model.
Reference values are hard-coded values from previous implementation; not sure
of examples with a closed-form solution.
"""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors,
['qso_log_chi2_qsonu', 'qso_log_chi2nuNULL_chi2nu'])
npt.assert_allclose(f['qso_log_chi2_qsonu'], 6.9844064754)
npt.assert_allclose(f['qso_log_chi2nuNULL_chi2nu'], -0.456526327522)
def test_amplitude():
"""Test features related to amplitude/magnitude percentiles."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['amplitude'])
npt.assert_allclose(f.values()[0], (max(values) - min(values)) /2.)
f = sft.generate_science_features(times, values, errors, ['percent_amplitude'])
max_scaled = 10**(-0.4 * max(values))
min_scaled = 10**(-0.4 * min(values))
med_scaled = 10**(-0.4 * np.median(values))
peak_from_median = max(abs((max_scaled - med_scaled) / med_scaled),
abs((min_scaled - med_scaled)) / med_scaled)
npt.assert_allclose(f.values()[0], peak_from_median, rtol=5e-4)
f = sft.generate_science_features(times, values, errors, ['percent_difference_flux_percentile'])
band_offset = 13.72
w_m2 = 10**(-0.4*(values+band_offset)-3) # 1 erg/s/cm^2 = 10^-3 w/m^2
npt.assert_allclose(f.values()[0], np.diff(
np.percentile(w_m2, [5, 95])) / np.median(w_m2))
f = sft.generate_science_features(times, values, errors, ['flux_percentile_ratio_mid20'])
npt.assert_allclose(f.values()[0],
np.diff(np.percentile(w_m2, [40, 60])) /
np.diff(np.percentile(w_m2, [5, 95])))
f = sft.generate_science_features(times, values, errors, ['flux_percentile_ratio_mid35'])
npt.assert_allclose(f.values()[0],
np.diff(np.percentile(w_m2, [32.5, 67.5])) /
np.diff(np.percentile(w_m2, [5, 95])))
f = sft.generate_science_features(times, values, errors, ['flux_percentile_ratio_mid50'])
npt.assert_allclose(f.values()[0],
np.diff(np.percentile(w_m2, [25, 75])) /
np.diff(np.percentile(w_m2, [5, 95])))
f = sft.generate_science_features(times, values, errors, ['flux_percentile_ratio_mid65'])
npt.assert_allclose(f.values()[0],
np.diff(np.percentile(w_m2, [17.5, 82.5])) /
np.diff(np.percentile(w_m2, [5, 95])))
f = sft.generate_science_features(times, values, errors, ['flux_percentile_ratio_mid80'])
npt.assert_allclose(f.values()[0],
np.diff(np.percentile(w_m2, [10, 90])) /
np.diff(np.percentile(w_m2, [5, 95])))
def test_lomb_scargle_period_folding():
"""
Tests for features derived from fitting a Lomb-Scargle periodic model and
period-folding the data by the estimated period.
"""
frequencies = np.hstack((WAVE_FREQS[0], np.zeros(len(WAVE_FREQS)-1)))
amplitudes = np.zeros((len(WAVE_FREQS),4))
amplitudes[0,:] = [8,4,2,1]
phase = 0.1
times, values, errors = irregular_periodic(frequencies, amplitudes, phase)
all_lomb = sft.generate_science_features(times, values, errors,
lomb_features)
# Folding is numerically unstable so we need to use the exact fitted frequency
freq_est = all_lomb['freq1_freq']
# Fold by 1*period
fold1ed_times = (times-times[0]) % (1./freq_est)
sort_indices = np.argsort(fold1ed_times)
fold1ed_times = fold1ed_times[sort_indices]
fold1ed_values = values[sort_indices]
# Fold by 2*period
fold2ed_times = (times-times[0]) % (2./freq_est)
sort_indices = np.argsort(fold2ed_times)
fold2ed_times = fold2ed_times[sort_indices]
fold2ed_values = values[sort_indices]
npt.assert_allclose(np.sum(np.diff(fold2ed_values)**2) /
np.sum(np.diff(values)**2), all_lomb['p2p_scatter_2praw'])
npt.assert_allclose(np.sum(np.diff(values)**2) / ((len(values) - 1) *
np.var(values)), all_lomb['p2p_ssqr_diff_over_var'])
npt.assert_allclose(np.median(np.abs(np.diff(values))) /
np.median(np.abs(values-np.median(values))),
all_lomb['p2p_scatter_over_mad'])
npt.assert_allclose(np.median(np.abs(np.diff(fold1ed_values))) /
np.median(np.abs(values-np.median(values))),
all_lomb['p2p_scatter_pfold_over_mad'])
def test_std():
"""Test standard deviation feature."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['std'])
npt.assert_allclose(f.values()[0], np.std(values))
def test_skew():
"""Test statistical skew feature."""
from scipy import stats
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['skew'])
npt.assert_allclose(f.values()[0], stats.skew(values))
def test_median():
"""Test median value feature."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['median'])
npt.assert_allclose(f.values()[0], np.median(values))
def test_median_absolute_deviation():
"""Test median absolute deviation (from the median) feature."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['median_absolute_deviation'])
npt.assert_allclose(f.values()[0], np.median(np.abs(values -
np.median(values))))
def featurize_single_ts(ts, features_to_use, custom_script_path=None,
custom_functions=None, use_docker=True):
"""Compute feature values for a given single time-series. Data is
returned as dictionaries/lists of lists (as opposed to a more
convenient DataFrame/DataSet) since it will be serialized as part of
`celery_tasks.featurize_ts_file`.
Parameters
----------
ts : TimeSeries object
Single time series to be featurized.
features_to_use : list of str
List of feature names to be generated.
custom_script_path : str, optional
Path to custom features script .py file to be run in Docker container.
custom_functions : dict, optional
Dictionary of custom feature functions to be evaluated for the given
time series, or a dictionary representing a dask graph of function
evaluations. Dictionaries of functions should have keys `feature_name`
and values functions that take arguments (t, m, e); in the case of a
dask graph, these arrays should be referenced as 't', 'm', 'e',
respectively, and any values with keys present in `features_to_use`
will be computed.
use_docker : bool, optional
Bool specifying whether to generate custom features inside a Docker
container. Defaults to True.
Returns
-------
dict
Dictionary with feature names as keys, lists of feature values (one per
channel) as values.
"""
all_feature_lists = {feature: [0.] * ts.n_channels
for feature in features_to_use}
for (t_i, m_i, e_i), i in zip(ts.channels(), range(ts.n_channels)):
obs_features = oft.generate_obs_features(t_i, m_i, e_i,
features_to_use)
science_features = sft.generate_science_features(t_i, m_i, e_i,
features_to_use)
if custom_script_path:
custom_features = cft.generate_custom_features(
custom_script_path, t_i, m_i, e_i,
features_already_known=dict(list(obs_features.items()) +
list(science_features.items()) +
list(ts.meta_features.items())),
use_docker=use_docker)
custom_features = {key: custom_features[key]
for key in custom_features.keys()
if key in features_to_use}
elif custom_functions:
# If all values in custom_functions are functions, evaluate each
if all(hasattr(v, '__call__') for v in custom_functions.values()):
custom_features = {feature: f(t_i, m_i, e_i)
for feature, f in custom_functions.items()
if feature in features_to_use}
# Otherwise, custom_functions is a dask graph
else:
dask_graph = {key: value
for key, value in custom_functions.items()
if key in features_to_use}
dask_keys = list(dask_graph.keys())
dask_graph['t'] = t_i
dask_graph['m'] = m_i
dask_graph['e'] = e_i
dask_graph.update(dict(list(obs_features.items()) +
list(science_features.items()) +
list(ts.meta_features.items())))
custom_features = dict(zip(dask_keys,
dask.async.get_sync(dask_graph,
dask_keys)))
else:
custom_features = {}
# We set values in this order so that custom features take priority
# over MLTSP features in the case of name conflicts
for feature, value in (list(obs_features.items()) +
list(science_features.items()) +
list(custom_features.items())):
all_feature_lists[feature][i] = value
return all_feature_lists
def test_min():
"""Test minimum value feature."""
times, values, errors = irregular_random()
f = sft.generate_science_features(times, values, errors, ['minimum'])
npt.assert_equal(f.values()[0], min(values))
