def test_add(self):
n1 = 30
n2 = 20
dx = 2
dy = 1
X1 = np.random.randn(n1, dx)
Y1 = np.random.rand(n1, dy)
X2 = np.random.rand(n2, dx)
Y2 = np.random.randn(n2, dy) + 1
pdata1 = data.PairedData(X1, Y1)
pdata2 = data.PairedData(X2, Y2)
# merge
pdata = pdata1 + pdata2
# check internals
X = pdata.X
Y = pdata.Y
np.testing.assert_array_almost_equal(X[:n1], X1)
np.testing.assert_array_almost_equal(X[n1:], X2)
np.testing.assert_array_almost_equal(Y[:n1], Y1)
np.testing.assert_array_almost_equal(Y[n1:], Y2)
self.assertTrue(pdata != pdata1)
self.assertTrue(pdata != pdata2)
# test size
self.assertEqual(pdata.sample_size(), n1 + n2)
self.assertEqual(pdata1.sample_size(), n1)
self.assertEqual(pdata2.sample_size(), n2)
def get_problem_pickle(folder_path, prob_label):
"""
- folder_path: path to a folder containing the data file relative to
fsic/data/ folder.
- prob_label: string of the form described in parse_prob_label() so that
fsic/data/(folder_path)/(name).p exists.
_n%d specifies the sample size to resample in each trial.
Return a (PairedSource object, n, is_h0).
"""
dataset_dir = glo.data_file(folder_path)
if not os.path.exists(dataset_dir):
raise ValueError('dataset directory does not exist: %s' % dataset_dir)
pl = parse_prob_label(prob_label)
data_path = os.path.join(dataset_dir, pl['name'] + '.p')
if not os.path.exists(data_path):
raise ValueError('dataset does not exist: %s' % data_path)
loaded = glo.pickle_load(data_path)
# Expected "loaded" to be a dictionary {'X': ..., 'Y': ..., ...}
X, Y = loaded['X'], loaded['Y']
is_h0 = pl['is_h0']
is_c = pl['is_classification']
if is_c:
assert Y.shape[1] == 1, 'Y should have one column. Shape = %s' % str(
Y.shape)
classes = Y[:, 0]
# If the data is a classification problem, make a 1-of-K coding
# of the label. We assume that Y has one column.
# modify Y
if len(np.unique(Y)) > 2:
# multiclass problem. Use 1-of-K coding.
# Only for #classes > 2
Y = util.one_of_K_code(classes)
is_std = pl['is_std']
n = pl['n']
ndx = pl['ndx']
ndy = pl['ndy']
# Standardization after resampling can cause a 0 standard deviation.
# We will do it as the first step.
#ps = data.PSStandardize(ps) if is_std else ps
if is_std:
X = util.standardize(X)
Y = util.standardize(Y)
pdata = data.PairedData(X, Y, label=prob_label)
ps = data.PSStraResample(pdata,
classes) if is_c else data.PSResample(pdata)
ps = data.PSNullShuffle(ps) if is_h0 else ps
if not (ndx == 0 and ndy == 0):
ps = data.PSGaussNoiseDims(ps, ndx, ndy)
return ps, n, is_h0
def job_nfsicJ10_cperm_stoopt(paired_source, tr, te, r):
"""
- Copula transform the data
- Use permutations to simulate from the null distribution.
"""
n_permute = 500
with util.ContextTimer() as t:
# copula transform to both X and Y
cop_map = fea.MarginalCDFMap()
xtr, ytr = tr.xy()
xte, yte = te.xy()
xtr = cop_map.gen_features(xtr)
ytr = cop_map.gen_features(ytr)
xte = cop_map.gen_features(xte)
yte = cop_map.gen_features(yte)
tr = data.PairedData(xtr, ytr)
te = data.PairedData(xte, yte)
to_return = job_nfsicJ10_stoopt(paired_source, tr, te, r, n_permute)
to_return['time_secs'] = t.secs
return to_return
def test_sample(self):
for s in [27, 91]:
n_ori = 200
p_fracs = [0.1, 0.5, 0.4]
X = np.random.randn(n_ori, 3)
Y = np.array([0] * int(p_fracs[0] * n_ori) +
[1] * int(p_fracs[1] * n_ori) +
[2] * int(p_fracs[2] * n_ori))[:, np.newaxis]
pdata_ori = data.PairedData(X, Y)
ps = data.PSStraResample(pdata_ori, Y[:, 0])
m = 79
pdata = ps.sample(m, seed=s)
self.assertEqual(pdata.sample_size(), m)
_, y = pdata.xy()
yu, counts = np.unique(y, return_counts=True)
for i, _ in enumerate(yu):
self.assertTrue(counts[i] - int(p_fracs[i] * m) <= 1)
def get_pdata_mean(n, dx=2):
X = np.random.randn(n, dx)
Y = np.mean(X, 1)[:, np.newaxis] + np.random.randn(n, 1) * 0.01
return data.PairedData(X, Y, label='mean')
