def test_canberra_32_2():
for i in range(10):
X = random.randn(10, 2).astype(np.float32)
Y = X[[0, 1, 2], :]
X_indices = random.random_integers(low=0, high=9,
size=5).astype(np.intp)
assignments, inertia = assign_nearest(X,
Y,
'canberra',
X_indices=X_indices)
cdist = scipy.spatial.distance.cdist(X[X_indices],
Y,
metric='canberra')
ref = cdist.argmin(axis=1)
if not np.all(ref == assignments):
different = np.where(assignments != ref)[0]
row = cdist[different, :]
# if there are differences between assignments and the 'reference',
# make sure that there is actually some difference between the
# entries in that row of the distance matrix before throwing
# an error
if not np.all(row == row[0]):
assert False
def assign_(xtcf, cf, top, sel, dt=1, outfname=None):
if not outfname:
outfname = os.path.split(xtcf)[1].split(".")[0]
top = md.load_pdb(top)
ndx = top.top.select(sel)
centerpdbs = md.load(cf, top=top, atom_indices=ndx)
xtc = md.load_xtc(xtcf, top=top, atom_indices=ndx, stride=dt)
seq, iter0 = libdistance.assign_nearest(xtc, centerpdbs, metric="rmsd")
np.savez_compressed(outfname + '.npz', a=seq)
return seq
def test_kcenters_8():
X = np.random.RandomState(1).randn(100, 2)
X32 = X.astype(np.float32)
X64 = X.astype(np.float64)
m1 = KCenters(n_clusters=10, random_state=0).fit([X32])
m2 = KCenters(n_clusters=10, random_state=0).fit([X64])
eq(m1.cluster_centers_, m2.cluster_centers_)
eq(m1.distances_[0], m2.distances_[0])
eq(m1.labels_[0], m2.labels_[0])
assert np.all(np.logical_not(np.isnan(m1.distances_[0])))
eq(m1.predict([X32])[0], m2.predict([X64])[0])
eq(m1.predict([X32])[0], m1.labels_[0])
eq(float(m1.inertia_), libdistance.assign_nearest(X32, m1.cluster_centers_, "euclidean")[1])
def test_assign_nearest_rmsd_1():
# rmsd assign nearest without X_indices
assignments, inertia = assign_nearest(X_rmsd, Y_rmsd, "rmsd")
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist = np.array([md.rmsd(X_rmsd, Y_rmsd[i], precentered=True) for i in range(len(Y_rmsd))]).T
assert cdist.shape == (10, 3)
np.testing.assert_array_equal(
assignments,
cdist.argmin(axis=1))
np.testing.assert_almost_equal(
inertia,
cdist[np.arange(10), assignments].sum(),
decimal=6)
def test_assign_nearest_rmsd_1():
# rmsd assign nearest without X_indices
assignments, inertia = assign_nearest(X_rmsd, Y_rmsd, "rmsd")
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist_rmsd = cdist(X_rmsd, Y_rmsd, 'rmsd')
assert cdist_rmsd.shape == (10, 3)
np.testing.assert_array_equal(
assignments,
cdist_rmsd.argmin(axis=1))
np.testing.assert_almost_equal(
inertia,
cdist_rmsd[np.arange(10), assignments].sum(),
decimal=6)
def test_assign_nearest_rmsd_1():
# rmsd assign nearest without X_indices
assignments, inertia = assign_nearest(X_rmsd, Y_rmsd, "rmsd")
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist = np.array([
md.rmsd(X_rmsd, Y_rmsd[i], precentered=True)
for i in range(len(Y_rmsd))
]).T
assert cdist.shape == (10, 3)
np.testing.assert_array_equal(assignments, cdist.argmin(axis=1))
np.testing.assert_almost_equal(inertia,
cdist[np.arange(10), assignments].sum(),
decimal=6)
def test_assign_nearest_rmsd_1():
# rmsd assign nearest without X_indices
assignments, inertia = assign_nearest(X_rmsd, Y_rmsd, "rmsd")
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist_rmsd = cdist(X_rmsd, Y_rmsd, 'rmsd')
assert cdist_rmsd.shape == (10, 3)
np.testing.assert_array_equal(
assignments,
cdist_rmsd.argmin(axis=1))
np.testing.assert_almost_equal(
inertia,
cdist_rmsd[np.arange(10), assignments].sum(),
decimal=6)
def test_assign_nearest_rmsd_2():
# rmsd assign nearest with X_indices
assignments, inertia = assign_nearest(X_rmsd, Y_rmsd, "rmsd", X_indices)
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist_rmsd = cdist(X_rmsd, Y_rmsd, 'rmsd')
cdist_rmsd = cdist_rmsd[X_indices].astype(np.double)
assert cdist_rmsd.shape == (5, 3)
np.testing.assert_array_equal(
assignments,
cdist_rmsd.argmin(axis=1))
np.testing.assert_almost_equal(
inertia,
cdist_rmsd[np.arange(5), assignments].sum(),
decimal=5)
def test_assign_nearest_rmsd_2():
# rmsd assign nearest with X_indices
assignments, inertia = assign_nearest(X_rmsd, Y_rmsd, "rmsd", X_indices)
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist_rmsd = cdist(X_rmsd, Y_rmsd, 'rmsd')
cdist_rmsd = cdist_rmsd[X_indices].astype(np.double)
assert cdist_rmsd.shape == (5, 3)
np.testing.assert_array_equal(
assignments,
cdist_rmsd.argmin(axis=1))
np.testing.assert_almost_equal(
inertia,
cdist_rmsd[np.arange(5), assignments].sum(),
decimal=5)
def test_canberra_32_2():
for i in range(10):
X = random.randn(10,2).astype(np.float32)
Y = X[[0,1,2], :]
X_indices = random.random_integers(low=0, high=9, size=5)
assignments, inertia = assign_nearest(X, Y, 'canberra', X_indices=X_indices)
cdist = scipy.spatial.distance.cdist(X[X_indices], Y, metric='canberra')
ref = cdist.argmin(axis=1)
if not np.all(ref == assignments):
different = np.where(assignments != ref)[0]
row = cdist[different, :]
# if there are differences between assignments and the 'reference',
# make sure that there is actually some difference between the
# entries in that row of the distance matrix before throwing
# an error
if not np.all(row==row[0]):
assert False
def test_assign_nearest_float_double_2():
# test with X_indices
for metric in VECTOR_METRICS:
for X, Y in ((X_double, Y_double), (X_float, Y_float)):
if metric == 'canberra' and X.dtype == np.float32:
# this is tested separately
continue
assignments, inertia = assign_nearest(X, Y, metric, X_indices)
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist = scipy.spatial.distance.cdist(X[X_indices], Y, metric=metric)
yield lambda: np.testing.assert_array_equal(
assignments,
cdist.argmin(axis=1))
yield lambda: np.testing.assert_almost_equal(
inertia,
cdist[np.arange(5), assignments].sum(),
decimal=5 if X.dtype == np.float32 else 10)
def test_assign_nearest_float_double_2():
# test with X_indices
for metric in VECTOR_METRICS:
for X, Y in ((X_double, Y_double), (X_float, Y_float)):
if metric == 'canberra' and X.dtype == np.float32:
# this is tested separately
continue
assignments, inertia = assign_nearest(X, Y, metric, X_indices)
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist_1 = cdist(X[X_indices], Y, metric=metric)
yield lambda: np.testing.assert_array_equal(
assignments,
cdist_1.argmin(axis=1))
yield lambda: np.testing.assert_almost_equal(
inertia,
cdist_1[np.arange(5), assignments].sum(),
decimal=5 if X.dtype == np.float32 else 10)
def test_canberra_32_1():
# with canberra in float32, there is a rounding issue where many of
# the distances come out exactly the same, but due to finite floating
# point resolution, a different one gets picked than by argmin()
# on the cdist
for i in range(10):
X = random.randn(10,2).astype(np.float32)
Y = X[[0,1,2], :]
assignments, inertia = assign_nearest(X, Y, 'canberra')
cdist = scipy.spatial.distance.cdist(X, Y, metric='canberra')
ref = cdist.argmin(axis=1)
if not np.all(ref == assignments):
different = np.where(assignments != ref)[0]
row = cdist[different, :]
# if there are differences between assignments and the 'reference',
# make sure that there is actually some difference between the
# entries in that row of the distance matrix before throwing
# an error
if not np.all(row==row[0]):
assert False
def test_assign_nearest_double_float_1():
# test without X_indices
for metric in VECTOR_METRICS:
for X, Y in ((X_double, Y_double), (X_float, Y_float)):
if metric == 'canberra' and X.dtype == np.float32:
# this is tested separately
continue
assignments, inertia = assign_nearest(X, Y, metric)
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist = scipy.spatial.distance.cdist(X, Y, metric=metric)
assert cdist.shape == (10, 3)
f = lambda: np.testing.assert_array_equal(assignments,
cdist.argmin(axis=1))
f.description = 'assign_nearest: %s %s' % (metric, X.dtype)
yield lambda: np.testing.assert_almost_equal(
inertia,
cdist[np.arange(10), assignments].sum(),
decimal=5 if X.dtype == np.float32 else 10)
def test_canberra_32_1():
# with canberra in float32, there is a rounding issue where many of
# the distances come out exactly the same, but due to finite floating
# point resolution, a different one gets picked than by argmin()
# on the cdist
for i in range(10):
X = random.randn(10, 2).astype(np.float32)
Y = X[[0, 1, 2], :]
assignments, inertia = assign_nearest(X, Y, 'canberra')
cdist = scipy.spatial.distance.cdist(X, Y, metric='canberra')
ref = cdist.argmin(axis=1)
if not np.all(ref == assignments):
different = np.where(assignments != ref)[0]
row = cdist[different, :]
# if there are differences between assignments and the 'reference',
# make sure that there is actually some difference between the
# entries in that row of the distance matrix before throwing
# an error
if not np.all(row == row[0]):
assert False
def test_assign_nearest_double_float_1():
# test without X_indices
for metric in VECTOR_METRICS:
for X, Y in ((X_double, Y_double), (X_float, Y_float)):
if metric == 'canberra' and X.dtype == np.float32:
# this is tested separately
continue
assignments, inertia = assign_nearest(X, Y, metric)
assert isinstance(assignments, np.ndarray)
assert isinstance(inertia, float)
cdist_1 = cdist(X, Y, metric=metric)
assert cdist_1.shape == (10, 3)
f = lambda: np.testing.assert_array_equal(
assignments,
cdist_1.argmin(axis=1))
f.description = 'assign_nearest: %s %s' % (metric, X.dtype)
yield lambda: np.testing.assert_almost_equal(
inertia,
cdist_1[np.arange(10), assignments].sum(),
decimal=5 if X.dtype == np.float32 else 10)
