def test_split():
# test on DataArray with number of samples multiple of new length
X_da = xr.DataArray(
np.random.random((100, 10)),
coords={
"sample": range(100),
"feature": range(10),
"coord_1": (["sample", "feature"], np.tile("Test", (100, 10))),
},
dims=("sample", "feature"),
)
estimator = Splitter(
new_dim="split_sample",
new_len=5,
reduce_index="subsample",
axis=1,
keep_coords_as="sample_coord",
)
Xt_da = estimator.fit_transform(X_da)
assert Xt_da.shape == (20, 5, 10)
npt.assert_allclose(Xt_da[0, :, 0], X_da[:5, 0])
Xit_da = estimator.inverse_transform(Xt_da)
xrt.assert_allclose(X_da, Xit_da)
# test on Dataset with number of samples NOT multiple of new length
X_ds = xr.Dataset(
{"var_1": (["sample", "feature"], np.random.random((100, 10)))},
coords={
"sample": range(100),
"feature": range(10)
},
)
Xt_ds = split(
X_ds,
new_dim="split_sample",
new_len=7,
reduce_index="head",
axis=1,
new_index_func=None,
)
assert Xt_ds["var_1"].shape == (14, 7, 10)
npt.assert_allclose(Xt_ds.var_1[0, :, 0], X_ds.var_1[:7, 0])
def test_split():
# test on DataArray with number of samples multiple of new length
X_da = xr.DataArray(np.random.random((100, 10)),
coords={
'sample':
range(100),
'feature':
range(10),
'coord_1':
(['sample', 'feature'], np.tile('Test', (100, 10)))
},
dims=('sample', 'feature'))
estimator = Splitter(new_dim='split_sample',
new_len=5,
reduce_index='subsample',
axis=1,
keep_coords_as='sample_coord')
Xt_da = estimator.fit_transform(X_da)
assert Xt_da.shape == (20, 5, 10)
npt.assert_allclose(Xt_da[0, :, 0], X_da[:5, 0])
Xit_da = estimator.inverse_transform(Xt_da)
xrt.assert_allclose(X_da, Xit_da)
# test on Dataset with number of samples NOT multiple of new length
X_ds = xr.Dataset(
{'var_1': (['sample', 'feature'], np.random.random((100, 10)))},
coords={
'sample': range(100),
'feature': range(10)
})
Xt_ds = split(X_ds,
new_dim='split_sample',
new_len=7,
reduce_index='head',
axis=1,
new_index_func=None)
assert Xt_ds['var_1'].shape == (14, 7, 10)
npt.assert_allclose(Xt_ds.var_1[0, :, 0], X_ds.var_1[:7, 0])
axarr[2].set_xlabel('Time [s]')
axarr[2].set_title('Acceleration along z-axis')
##############################################################################
# Then we define a pipeline with various preprocessing steps and a classifier.
#
# The preprocessing consists of splitting the data into segments, removing
# segments with `nan` values and standardizing. Since the accelerometer data is
# three-dimensional but the standardizer and classifier expect a one-dimensional
# feature vector, we have to vectorize the samples.
#
# Finally, we use PCA and a naive Bayes classifier for classification.
pl = Pipeline([
('splitter', Splitter(
groupby=['subject', 'activity'], new_dim='timepoint', new_len=30)),
('sanitizer', Sanitizer()),
('featurizer', Featurizer()),
('scaler', wrap(StandardScaler)),
('pca', wrap(PCA, reshapes='feature')),
('cls', wrap(GaussianNB, reshapes='feature'))
])
##############################################################################
# Since we want to use cross-validated grid search to find the best model
# parameters, we define a cross-validator. In order to make sure the model
# performs subject-independent recognition, we use a `GroupShuffleSplit`
# cross-validator that ensures that the same subject will not appear in both
# training and validation set.
cv = CrossValidatorWrapper(
axarr[2].plot(X_plot.sample, X_plot.sel(axis='z'), color='#2ca02c')
axarr[2].set_xlabel('Time [s]')
axarr[2].set_title('Acceleration along z-axis')
##############################################################################
# Then we define a pipeline with various preprocessing steps and a classifier.
#
# The preprocessing consists of splitting the data into segments, removing
# segments with `nan` values and standardizing. Since the accelerometer data is
# three-dimensional but the standardizer and classifier expect a one-dimensional
# feature vector, we have to vectorize the samples.
#
# Finally, we use PCA and logistic regression to perform the classification.
pl = Pipeline([('splitter',
Splitter(groupby=['subject', 'activity'],
new_dim='timepoint')), ('sanitizer', Sanitizer()),
('featurizer', Featurizer()), ('scaler', wrap(StandardScaler)),
('pca', wrap(PCA, reshapes='feature')),
('lr', wrap(LogisticRegression, reshapes='feature'))])
##############################################################################
# Since we want to use cross-validated grid search to find the best model
# parameters, we define a cross-validator. In order to make sure the model
# performs subject-independent recognition, we use a `GroupShuffleSplit`
# cross-validator that ensures that the same subject will not appear in both
# training and validation set.
cv = CrossValidatorWrapper(GroupShuffleSplit(n_splits=3, test_size=0.3),
groupby=['subject'])
##############################################################################
##############################################################################
# Then we define a pipeline with various preprocessing steps and a classifier.
#
# The preprocessing consists of splitting the data into segments, removing
# segments with `nan` values and standardizing. Since the accelerometer data is
# three-dimensional but the standardizer and classifier expect a
# one-dimensional feature vector, we have to vectorize the samples.
#
# Finally, we use PCA and a naive Bayes classifier for classification.
pl = Pipeline([
(
"splitter",
Splitter(
groupby=["subject", "activity"],
new_dim="timepoint",
new_len=30,
),
),
("sanitizer", Sanitizer()),
("featurizer", Featurizer()),
("scaler", wrap(StandardScaler)),
("pca", wrap(PCA, reshapes="feature")),
("cls", wrap(GaussianNB, reshapes="feature")),
])
##############################################################################
# Since we want to use cross-validated grid search to find the best model
# parameters, we define a cross-validator. In order to make sure the model
# performs subject-independent recognition, we use a `GroupShuffleSplit`
# cross-validator that ensures that the same subject will not appear in both