Scikit-learn
human-activity-recognition
wearable
wrist
time-series
cpu
scikit-learn
WISP / tests /test_features_fast.py
Zipeng365's picture
wisp-har 0.2.0: accelerated implementation with speed (exact/fast) and HMM sequence (subject/recording) options
ac7f7cb verified
Raw History Blame Contribute Delete
3.81 kB
"""Feature-level release conformance, including discrete and singleton windows."""
import warnings
import numpy as np
import unittest
from unittest.mock import patch
from wisp.cpu import features as ref
from wisp_fast import features as fx
IMPLEMENTATIONS = [
('WearableFeatureExtractor', {}),
('IntervalDistributionSketch', {'n_random_intervals': 7}),
('RandomConvSketch', {'n_kernels': 36}),
('InvariantPhysicsFeatureExtractor', {'random_intervals': 7}),
('RandomShapeletSketch', {'n_shapelets': 12}),
]
def compare_features(name, kw, case):
rng = np.random.default_rng(781)
x = rng.standard_normal((1 if case == 'one_window' else 5, 37,
4 if case == 'four_channels' else 3))
if case == 'ties':
x = np.round(x)
x[:, :, 1] = 2.
if case == 'fortran':
x = np.asfortranarray(x)
if case == 'univariate':
x = x[:, :, 1].copy()
x[0, 5] = np.nan
original = getattr(ref, name)(**kw)
replacement = getattr(fx, name)(**kw)
with warnings.catch_warnings():
warnings.simplefilter('ignore', RuntimeWarning)
expected = original.fit_transform(x)
result = replacement.fit_transform(x)
np.testing.assert_array_equal(result, expected)
if hasattr(original, 'feature_names_'):
assert replacement.feature_names_ == original.feature_names_
def compare_turbo(name, kw):
x = np.random.default_rng(1024).normal(size=(7, 65, 3)).astype(np.float32)
model = getattr(fx, name)(**kw, fast_profile=True)
result = model.fit_transform(x)
assert len(result) == 7 and np.isfinite(result).all()
np.testing.assert_allclose(result[4:5], model.transform(x[4:5]), rtol=1e-14, atol=1e-14)
np.testing.assert_array_equal(result, model.fit_transform(x))
class TestFeatures(unittest.TestCase):
def test_release_conformance(self):
cases = ['triaxial', 'ties', 'one_window', 'univariate', 'four_channels', 'fortran']
for name, kw in IMPLEMENTATIONS:
for case in cases:
with self.subTest(name=name, case=case):
compare_features(name, kw, case)
def test_turbo_singleton_batch_and_refit(self):
for name, kw in IMPLEMENTATIONS:
with self.subTest(name=name):
compare_turbo(name, kw)
@patch.object(fx, '_kernels', None)
def test_conv_numpy_fallback(self):
x = np.random.default_rng(781).normal(size=(7, 40, 3))
kw = dict(n_kernels=64, dilations=(1, 2, 5))
a = ref.RandomConvSketch(**kw).fit_transform(x)
b = fx.RandomConvSketch(**kw).fit_transform(x)
np.testing.assert_array_equal(b, a)
@unittest.skipIf(fx._kernels is None, 'optional extension unavailable')
def test_native_dimension_guards(self):
native = fx._kernels
x = np.zeros((3, 5, 37))
weights = np.zeros((2, 4, 3))
biases = np.zeros(2)
out = np.zeros((2, 5, 34))
for args in [(x, weights, biases, 0, 1, out),
(x, weights, biases[:1], 1, 1, out),
(x.astype(np.float32), weights, biases, 1, 1, out),
(x[0:2], weights, biases, 1, 1, out),
(x, weights, biases, 2, 1, out),
(x, weights, biases, 1, 100, out)]:
with self.assertRaises((BufferError, ValueError)):
native.project(*args)
def test_identical_shapelet_matches_choose_first_window(self):
x = np.tile(np.arange(8), 7).reshape(1, -1, 1).astype(float)
kw = dict(n_shapelets=12, lengths=(8,), dilations=(1,))
expected = ref.RandomShapeletSketch(**kw).fit_transform(x)
actual = fx.RandomShapeletSketch(**kw).fit_transform(x)
np.testing.assert_array_equal(actual, expected)