def generate_gaussians(name, mean_std_nums, dim, min_pts, max_pts):
bags = []
categories = []
for mean, std, num in mean_std_nums:
cat_name = 'mean{}-std{}'.format(mean, std)
for x in range(num):
n_pts = np.random.randint(min_pts, max_pts+1)
feats = np.random.normal(mean, std, size=(n_pts, dim))
bags.append(feats)
categories.append(cat_name)
features = Features(bags, categories=categories)
features.save_as_hdf5('data/{}.h5'.format(name))
def generate_gaussians(name, mean_std_nums, dim, min_pts, max_pts):
bags = []
categories = []
for mean, std, num in mean_std_nums:
cat_name = 'mean{}-std{}'.format(mean, std)
for x in range(num):
n_pts = np.random.randint(min_pts, max_pts + 1)
feats = np.random.normal(mean, std, size=(n_pts, dim))
bags.append(feats)
categories.append(cat_name)
features = Features(bags, categories=categories)
features.save_as_hdf5('data/{}.h5'.format(name))
def test_with_and_without_js():
dir = os.path.join(os.path.dirname(__file__), 'data')
name = 'gaussian-2d-mean0-std1,2'
feats = Features.load_from_hdf5(os.path.join(dir, name + '.h5'))
specs = ['hellinger']
Ks = [3, 5]
with h5py.File(os.path.join(dir, name + '.divs.h5'), 'r') as f:
expected = load_divs(f, specs, Ks)
min_dist = f.attrs['min_dist']
est = partial(estimate_divs, feats, Ks=Ks, min_dist=min_dist,
status_fn=None)
with capture_output(True, True, merge=False):
ds = est(specs=specs + ['js'])
oth_with = ds[:, :, :-1, :]
js_with = ds[:, :, -1, :]
js_without = est(specs=['js'])[:, :, 0, :]
assert_close(oth_with, expected, atol=5e-5,
msg="others ({}) broke with JS".format(', '.join(specs)))
assert_close(js_with, js_without, atol=5e-5,
msg="JS different with/without others")
def test_divs():
dir = os.path.join(os.path.dirname(__file__), 'data')
argses = [{'cores': cores, 'status_fn': status_fn}
for cores in [1, None]
for status_fn in [None]] # , True]]
# TODO: test a custom status_fn also
specs = ['hellinger', 'kl', 'l2', 'linear',
'renyi:0.5', 'renyi:0.7', 'renyi:0.9', 'renyi:0.99']
Ks = [1, 3, 5, 10]
for name in ['gaussian-2d-mean0-std1,2', 'gaussian-20d-mean0-std1,2']:
for dtype in [np.float64, np.float32]:
feats = Features.load_from_hdf5(
os.path.join(dir, name + '.h5'),
features_dtype=dtype)
with h5py.File(os.path.join(dir, name + '.divs.h5'), 'r') as f:
expected = load_divs(f, specs, Ks)
min_dist = f.attrs['min_dist']
tests = []
for args in argses:
tests.extend(check_div(feats, expected, specs, Ks, name,
min_dist=min_dist, **args))
for test in sorted(tests, key=lambda t: t[0].description):
yield test
def test_kl_simple():
# verified by hand
# Dhat(P||Q) = \log m/(n-1) + d / n \sum_{i=1}^n \log \nu_k(i)/rho_k(i)
x = np.reshape([0., 1, 3], (3, 1))
y = np.reshape([.2, 1.2, 3.2, 7.2], (4, 1))
n = x.shape[0]
m = y.shape[0]
x_to_y = np.log(
m /
(n - 1)) + 1 / n * (np.log(1.2 / 3) + np.log(.8 / 2) + np.log(1.8 / 3))
y_to_x = np.log(n / (m - 1)) + 1 / m * (np.log(.8 / 3) + np.log(1.2 / 2) +
np.log(2.2 / 3) + np.log(6.2 / 6))
# NOTE: clamping makes this test useless.
x_to_y = max(x_to_y, 0)
y_to_x = max(y_to_x, 0)
res = estimate_divs(Features([x, y]), specs=['kl'], Ks=[2]).squeeze()
assert res[0, 0] == 0
assert res[1, 1] == 0
assert np.allclose(res[1, 0], y_to_x), "{} vs {}".format(res[1, 0], y_to_x)
assert np.allclose(res[0, 1], x_to_y), "{} vs {}".format(res[0, 1], x_to_y)
def test_with_and_without_js():
dir = os.path.join(os.path.dirname(__file__), 'data')
name = 'gaussian-2d-mean0-std1,2'
feats = Features.load_from_hdf5(os.path.join(dir, name + '.h5'))
specs = ['hellinger']
Ks = [3, 5]
with h5py.File(os.path.join(dir, name + '.divs.h5'), 'r') as f:
expected = load_divs(f, specs, Ks)
min_dist = f.attrs['min_dist']
est = partial(estimate_divs,
feats,
Ks=Ks,
min_dist=min_dist,
status_fn=None)
with capture_output(True, True, merge=False):
ds = est(specs=specs + ['js'])
oth_with = ds[:, :, :-1, :]
js_with = ds[:, :, -1, :]
js_without = est(specs=['js'])[:, :, 0, :]
assert_close(oth_with,
expected,
atol=5e-5,
msg="others ({}) broke with JS".format(', '.join(specs)))
assert_close(js_with,
js_without,
atol=5e-5,
msg="JS different with/without others")
def test_divs():
dir = os.path.join(os.path.dirname(__file__), 'data')
argses = [{
'cores': cores,
'status_fn': status_fn
} for cores in [1, None] for status_fn in [None]] # , True]]
# TODO: test a custom status_fn also
specs = [
'hellinger', 'kl', 'l2', 'linear', 'renyi:0.5', 'renyi:0.7',
'renyi:0.9', 'renyi:0.99'
]
Ks = [1, 3, 5, 10]
for name in ['gaussian-2d-mean0-std1,2', 'gaussian-20d-mean0-std1,2']:
for dtype in [np.float64, np.float32]:
feats = Features.load_from_hdf5(os.path.join(dir, name + '.h5'),
features_dtype=dtype)
with h5py.File(os.path.join(dir, name + '.divs.h5'), 'r') as f:
expected = load_divs(f, specs, Ks)
min_dist = f.attrs['min_dist']
tests = []
for args in argses:
tests.extend(
check_div(feats,
expected,
specs,
Ks,
name,
min_dist=min_dist,
**args))
for test in sorted(tests, key=lambda t: t[0].description):
yield test
def test_js_simple():
# verified by hand
x = np.reshape([0, 1, 3], (3, 1))
y = np.reshape([.2, 1.2, 3.2, 6.2], (4, 1))
mix_ent = np.log(2) + np.log(3) + psi(2) \
+ (np.log(.2) + np.log(.8) + np.log(1.8) - psi(1) - 2*psi(2)) / 6 \
+ (np.log(.2) + np.log(2) + np.log(3.2) - psi(1) - 3*psi(2)) / 8
x_ent = np.log(2) + (np.log(3) + np.log(2) + np.log(3)) / 3
y_ent = np.log(3) + (np.log(3) + np.log(2) + np.log(3) + np.log(5)) / 4
right_js = mix_ent - (x_ent + y_ent) / 2
expected = np.array([[0, right_js], [right_js, 0]])
# TODO: what about clamping???
est = estimate_divs(Features([x, y]), specs=['js'], Ks=[2],
status_fn=None).squeeze()
assert_close(est, expected, atol=5e-5, msg="JS estimate not as expected")
