def bench_eigvals():
numpy_eigvals = nl.eigvals
scipy_eigvals = sl.eigvals
print()
print(' Finding matrix eigenvalues')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size,repeat in [(20,150),(100,7),(200,2)]:
repeat *= 1
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size,size])
print('| %6.2f ' % measure('scipy_eigvals(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_eigvals(a)',repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1,-1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_eigvals(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_eigvals(a)',repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_load_trk():
rng = np.random.RandomState(42)
dtype = 'float32'
NB_STREAMLINES = 5000
NB_POINTS = 1000
points = [rng.rand(NB_POINTS, 3).astype(dtype)
for i in range(NB_STREAMLINES)]
scalars = [rng.rand(NB_POINTS, 10).astype(dtype)
for i in range(NB_STREAMLINES)]
repeat = 10
with InTemporaryDirectory():
trk_file = "tmp.trk"
tractogram = Tractogram(points, affine_to_rasmm=np.eye(4))
TrkFile(tractogram).save(trk_file)
streamlines_old = [d[0] - 0.5
for d in tv.read(trk_file, points_space="rasmm")[0]]
mtime_old = measure('tv.read(trk_file, points_space="rasmm")', repeat)
print("Old: Loaded {:,} streamlines in {:6.2f}".format(NB_STREAMLINES,
mtime_old))
trk = nib.streamlines.load(trk_file, lazy_load=False)
streamlines_new = trk.streamlines
mtime_new = measure('nib.streamlines.load(trk_file, lazy_load=False)',
repeat)
print("\nNew: Loaded {:,} streamlines in {:6.2}".format(NB_STREAMLINES,
mtime_new))
print("Speedup of {:.2f}".format(mtime_old / mtime_new))
for s1, s2 in zip(streamlines_new, streamlines_old):
assert_array_equal(s1, s2)
# Points and scalars
with InTemporaryDirectory():
trk_file = "tmp.trk"
tractogram = Tractogram(points,
data_per_point={'scalars': scalars},
affine_to_rasmm=np.eye(4))
TrkFile(tractogram).save(trk_file)
streamlines_old = [d[0] - 0.5
for d in tv.read(trk_file, points_space="rasmm")[0]]
scalars_old = [d[1]
for d in tv.read(trk_file, points_space="rasmm")[0]]
mtime_old = measure('tv.read(trk_file, points_space="rasmm")', repeat)
msg = "Old: Loaded {:,} streamlines with scalars in {:6.2f}"
print(msg.format(NB_STREAMLINES, mtime_old))
trk = nib.streamlines.load(trk_file, lazy_load=False)
scalars_new = trk.tractogram.data_per_point['scalars']
mtime_new = measure('nib.streamlines.load(trk_file, lazy_load=False)',
repeat)
msg = "New: Loaded {:,} streamlines with scalars in {:6.2f}"
print(msg.format(NB_STREAMLINES, mtime_new))
print("Speedup of {:2f}".format(mtime_old / mtime_new))
for s1, s2 in zip(scalars_new, scalars_old):
assert_array_equal(s1, s2)
def bench_svd():
numpy_svd = nl.svd
scipy_svd = sl.svd
print()
print(' Finding the SVD decomposition')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size,repeat in [(20,150),(100,7),(200,2)]:
repeat *= 1
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size,size])
print('| %6.2f ' % measure('scipy_svd(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_svd(a)',repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1,-1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_svd(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_svd(a)',repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_set_number_of_points():
repeat = 1
nb_points_per_streamline = 100
nb_points = 42
nb_streamlines = int(1e4)
streamlines = [
np.random.rand(nb_points_per_streamline, 3).astype("float32")
for i in range(nb_streamlines)
]
print("Timing set_number_of_points() in Cython"
"({0} streamlines)".format(nb_streamlines))
cython_time = measure("set_number_of_points(streamlines, nb_points)",
repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
streamlines = [
np.random.rand(nb_points_per_streamline, 3).astype("float32")
for i in range(nb_streamlines)
]
python_time = measure(
"[set_number_of_points_python(s, nb_points)"
" for s in streamlines]", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time / cython_time))
del streamlines
def bench_load_save():
rng = np.random.RandomState(20111001)
repeat = 4
img_shape = (128, 128, 64)
arr = rng.normal(size=img_shape)
img = Nifti1Image(arr, np.eye(4))
sio = BytesIO()
img.file_map['image'].fileobj = sio
hdr = img.get_header()
sys.stdout.flush()
print("\nImage load save")
print("----------------")
hdr.set_data_dtype(np.float32)
mtime = measure('img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save float64 to float32', mtime))
mtime = measure('img.from_file_map(img.file_map)', repeat)
print('%30s %6.2f' % ('Load from float32', mtime))
hdr.set_data_dtype(np.int16)
mtime = measure('img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save float64 to int16', mtime))
mtime = measure('img.from_file_map(img.file_map)', repeat)
print('%30s %6.2f' % ('Load from int16', mtime))
arr = np.random.random_integers(low=-1000,high=-1000, size=img_shape)
arr = arr.astype(np.int16)
img = Nifti1Image(arr, np.eye(4))
sio = BytesIO()
img.file_map['image'].fileobj = sio
hdr = img.get_header()
hdr.set_data_dtype(np.float32)
mtime = measure('img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save Int16 to float32', mtime))
sys.stdout.flush()
def bench_load_save():
rng = np.random.RandomState(20111001)
repeat = 4
img_shape = (128, 128, 64)
arr = rng.normal(size=img_shape)
img = Nifti1Image(arr, np.eye(4))
sio = BytesIO()
img.file_map['image'].fileobj = sio
hdr = img.get_header()
sys.stdout.flush()
print "\nImage load save"
print "----------------"
hdr.set_data_dtype(np.float32)
mtime = measure('img.to_file_map()', repeat)
print '%30s %6.2f' % ('Save float64 to float32', mtime)
mtime = measure('img.from_file_map(img.file_map)', repeat)
print '%30s %6.2f' % ('Load from float32', mtime)
hdr.set_data_dtype(np.int16)
mtime = measure('img.to_file_map()', repeat)
print '%30s %6.2f' % ('Save float64 to int16', mtime)
mtime = measure('img.from_file_map(img.file_map)', repeat)
print '%30s %6.2f' % ('Load from int16', mtime)
arr = np.random.random_integers(low=-1000,high=-1000, size=img_shape)
arr = arr.astype(np.int16)
img = Nifti1Image(arr, np.eye(4))
sio = BytesIO()
img.file_map['image'].fileobj = sio
hdr = img.get_header()
hdr.set_data_dtype(np.float32)
mtime = measure('img.to_file_map()', repeat)
print '%30s %6.2f' % ('Save Int16 to float32', mtime)
sys.stdout.flush()
def bench_vec_val_vect():
# nosetests -s --match '(?:^|[\\b_\\.//-])[Bb]ench'
repeat = 100
etime = measure("np.einsum('...ij,...j,...kj->...ik', evecs, evals, evecs)",
repeat)
vtime = measure("vec_val_vect(evecs, evals)", repeat)
print("einsum %4.2f; vec_val_vect %4.2f" % (etime, vtime))
def bench_eigvals():
numpy_eigvals = nl.eigvals
scipy_eigvals = sl.eigvals
print()
print(' Finding matrix eigenvalues')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size, repeat in [(20, 150), (100, 7), (200, 2)]:
repeat *= 1
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size, size])
print('| %6.2f ' % measure('scipy_eigvals(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_eigvals(a)', repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1, -1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_eigvals(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_eigvals(a)', repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_random(self):
numpy_det = nl.det
scipy_det = sl.det
print()
print(' Finding matrix determinant')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size, repeat in [(20, 1000), (100, 150), (500, 2), (1000, 1)][:-1]:
repeat *= 2
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size, size])
print('| %6.2f ' % measure('scipy_det(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_det(a)', repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1, -1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_det(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_det(a)', repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_set_number_of_points():
repeat = 5
nb_streamlines = DATA['nb_streamlines']
msg = "Timing set_number_of_points() with {0:,} streamlines."
print(msg.format(nb_streamlines * repeat))
cython_time = measure("set_number_of_points(streamlines, nb_points)",
repeat)
print("Cython time: {0:.3f} sec".format(cython_time))
python_time = measure("[set_number_of_points_python(s, nb_points)"
" for s in streamlines]", repeat)
print("Python time: {0:.2f} sec".format(python_time))
print("Speed up of {0:.2f}x".format(python_time/cython_time))
# Make sure it produces the same results.
assert_array_almost_equal([set_number_of_points_python(s) for s in DATA["streamlines"]],
set_number_of_points(DATA["streamlines"]))
cython_time_arrseq = measure("set_number_of_points(streamlines, nb_points)", repeat)
print("Cython time (ArrSeq): {0:.3f} sec".format(cython_time_arrseq))
print("Speed up of {0:.2f}x".format(python_time/cython_time_arrseq))
# Make sure it produces the same results.
assert_array_equal(set_number_of_points(DATA["streamlines"]),
set_number_of_points(DATA["streamlines_arrseq"]))
def test_fprop_faster(self):
seed = 1234
repeat = 100
lstm = LSTM(input_size=DATA['features_size'],
hidden_sizes=[DATA['hidden_size']],
)
lstm.initialize(initer.UniformInitializer(seed))
lstm2 = LSTMFaster(input_size=DATA['features_size'],
hidden_sizes=[DATA['hidden_size']],
)
# Wi, Wo, Wf, Wm
# Make sure the weights are the same.
lstm2.layers_lstm[0].W.set_value(np.concatenate([lstm.layers_lstm[0].Wi.get_value(), lstm.layers_lstm[0].Wo.get_value(), lstm.layers_lstm[0].Wf.get_value(), lstm.layers_lstm[0].Wm.get_value()], axis=1))
lstm2.layers_lstm[0].U.set_value(np.concatenate([lstm.layers_lstm[0].Ui.get_value(), lstm.layers_lstm[0].Uo.get_value(), lstm.layers_lstm[0].Uf.get_value(), lstm.layers_lstm[0].Um.get_value()], axis=1))
input = T.tensor3('input')
input.tag.test_value = DATA['batch']
fprop = theano.function([input], lstm.get_output(input))
fprop2 = theano.function([input], lstm2.get_output(input))
fprop_time = measure("out = fprop(DATA['batch'])", repeat)
print("fprop time: {:.2f} sec.", fprop_time)
fprop2_time = measure("out = fprop2(DATA['batch'])", repeat)
print("fprop faster time: {:.2f} sec.", fprop2_time)
print("Speedup: {:.2f}x".format(fprop_time/fprop2_time))
out = fprop(DATA['batch'])
out2 = fprop2(DATA['batch'])
assert_array_equal(out, out2)
def bench_length():
repeat = 10
nb_streamlines = DATA['nb_streamlines']
streamlines = DATA["streamlines"] # Streamlines as a list of ndarrays.
print("Timing length() with {0:,} streamlines.".format(nb_streamlines))
python_time = measure("[length_python(s) for s in streamlines]", repeat)
print("Python time: {0:.2}sec".format(python_time))
cython_time = measure("length(streamlines)", repeat)
print("Cython time: {0:.3}sec".format(cython_time))
print("Speed up of {0:.2f}x".format(python_time / cython_time))
# Make sure it produces the same results.
assert_array_equal([length_python(s) for s in DATA["streamlines"]],
length(DATA["streamlines"]))
streamlines = DATA['streamlines_arrseq']
cython_time_arrseq = measure("length(streamlines)", repeat)
print("Cython time (ArrSeq): {0:.3}sec".format(cython_time_arrseq))
print("Speed up of {0:.2f}x".format(python_time / cython_time_arrseq))
# Make sure it produces the same results.
assert_array_equal(length(DATA["streamlines"]),
length(DATA["streamlines_arrseq"]))
def bench_set_number_of_points():
repeat = 5
nb_streamlines = DATA['nb_streamlines']
msg = "Timing set_number_of_points() with {0:,} streamlines."
print(msg.format(nb_streamlines * repeat))
cython_time = measure("set_number_of_points(streamlines, nb_points)",
repeat)
print("Cython time: {0:.3f} sec".format(cython_time))
python_time = measure("[set_number_of_points_python(s, nb_points)"
" for s in streamlines]", repeat)
print("Python time: {0:.2f} sec".format(python_time))
print("Speed up of {0:.2f}x".format(python_time/cython_time))
# Make sure it produces the same results.
assert_array_almost_equal([set_number_of_points_python(s) for s in DATA["streamlines"]],
set_number_of_points(DATA["streamlines"]))
cython_time_arrseq = measure("set_number_of_points(streamlines, nb_points)", repeat)
print("Cython time (ArrSeq): {0:.3f} sec".format(cython_time_arrseq))
print("Speed up of {0:.2f}x".format(python_time/cython_time_arrseq))
# Make sure it produces the same results.
assert_array_equal(set_number_of_points(DATA["streamlines"]),
set_number_of_points(DATA["streamlines_arrseq"]))
def bench_load_trk():
rng = np.random.RandomState(42)
dtype = 'float32'
NB_STREAMLINES = 5000
NB_POINTS = 1000
points = [rng.rand(NB_POINTS, 3).astype(dtype)
for i in range(NB_STREAMLINES)]
scalars = [rng.rand(NB_POINTS, 10).astype(dtype)
for i in range(NB_STREAMLINES)]
repeat = 10
with InTemporaryDirectory():
trk_file = "tmp.trk"
tractogram = Tractogram(points, affine_to_rasmm=np.eye(4))
TrkFile(tractogram).save(trk_file)
streamlines_old = [d[0] - 0.5
for d in tv.read(trk_file, points_space="rasmm")[0]]
mtime_old = measure('tv.read(trk_file, points_space="rasmm")', repeat)
print("Old: Loaded {:,} streamlines in {:6.2f}".format(NB_STREAMLINES,
mtime_old))
trk = nib.streamlines.load(trk_file, lazy_load=False)
streamlines_new = trk.streamlines
mtime_new = measure('nib.streamlines.load(trk_file, lazy_load=False)',
repeat)
print("\nNew: Loaded {:,} streamlines in {:6.2}".format(NB_STREAMLINES,
mtime_new))
print("Speedup of {:.2f}".format(mtime_old / mtime_new))
for s1, s2 in zip(streamlines_new, streamlines_old):
assert_array_equal(s1, s2)
# Points and scalars
with InTemporaryDirectory():
trk_file = "tmp.trk"
tractogram = Tractogram(points,
data_per_point={'scalars': scalars},
affine_to_rasmm=np.eye(4))
TrkFile(tractogram).save(trk_file)
streamlines_old = [d[0] - 0.5
for d in tv.read(trk_file, points_space="rasmm")[0]]
scalars_old = [d[1]
for d in tv.read(trk_file, points_space="rasmm")[0]]
mtime_old = measure('tv.read(trk_file, points_space="rasmm")', repeat)
msg = "Old: Loaded {:,} streamlines with scalars in {:6.2f}"
print(msg.format(NB_STREAMLINES, mtime_old))
trk = nib.streamlines.load(trk_file, lazy_load=False)
scalars_new = trk.tractogram.data_per_point['scalars']
mtime_new = measure('nib.streamlines.load(trk_file, lazy_load=False)',
repeat)
msg = "New: Loaded {:,} streamlines with scalars in {:6.2f}"
print(msg.format(NB_STREAMLINES, mtime_new))
print("Speedup of {:2f}".format(mtime_old / mtime_new))
for s1, s2 in zip(scalars_new, scalars_old):
assert_array_equal(s1, s2)
def bench_svd():
numpy_svd = nl.svd
scipy_svd = sl.svd
print()
print(' Finding the SVD decomposition')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size, repeat in [(20, 150), (100, 7), (200, 2)]:
repeat *= 1
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size, size])
print('| %6.2f ' % measure('scipy_svd(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_svd(a)', repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1, -1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_svd(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_svd(a)', repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_random(self):
numpy_det = nl.det
scipy_det = sl.det
print()
print(' Finding matrix determinant')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size,repeat in [(20,1000),(100,150),(500,2),(1000,1)][:-1]:
repeat *= 2
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size,size])
print('| %6.2f ' % measure('scipy_det(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_det(a)',repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1,-1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_det(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_det(a)',repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_vec_val_vect():
# nosetests -s --match '(?:^|[\\b_\\.//-])[Bb]ench'
repeat = 100
shape = (100, 100)
evecs, evals = randn(*(shape + (3, 3))), randn(*(shape + (3,)))
etime = measure("np.einsum('...ij,...j,...kj->...ik', evecs, evals, evecs)",
repeat)
vtime = measure("vec_val_vect(evecs, evals)", repeat)
print("einsum %4.2f; vec_val_vect %4.2f" % (etime, vtime))
def bench_vec_val_vect():
# nosetests -s --match '(?:^|[\\b_\\.//-])[Bb]ench'
repeat = 100
shape = (100, 100)
evecs, evals = randn(*(shape + (3, 3))), randn(*(shape + (3, )))
etime = measure(
"np.einsum('...ij,...j,...kj->...ik', evecs, evals, evecs)", repeat)
vtime = measure("vec_val_vect(evecs, evals)", repeat)
print("einsum %4.2f; vec_val_vect %4.2f" % (etime, vtime))
def bench_quickbundles():
dtype = "float32"
repeat = 10
nb_points = 18
streams, hdr = nib.trackvis.read(get_data('fornix'))
fornix = [s[0].astype(dtype) for s in streams]
fornix = streamline_utils.set_number_of_points(fornix, nb_points)
# Create eight copies of the fornix to be clustered (one in each octant).
streamlines = []
streamlines += [s + np.array([100, 100, 100], dtype) for s in fornix]
streamlines += [s + np.array([100, -100, 100], dtype) for s in fornix]
streamlines += [s + np.array([100, 100, -100], dtype) for s in fornix]
streamlines += [s + np.array([100, -100, -100], dtype) for s in fornix]
streamlines += [s + np.array([-100, 100, 100], dtype) for s in fornix]
streamlines += [s + np.array([-100, -100, 100], dtype) for s in fornix]
streamlines += [s + np.array([-100, 100, -100], dtype) for s in fornix]
streamlines += [s + np.array([-100, -100, -100], dtype) for s in fornix]
# The expected number of clusters of the fornix using threshold=10 is 4.
threshold = 10.
expected_nb_clusters = 4 * 8
print("Timing QuickBundles 1.0 vs. 2.0")
qb = QB_Old(streamlines, threshold, pts=None)
qb1_time = measure("QB_Old(streamlines, threshold, nb_points)", repeat)
print("QuickBundles time: {0:.4}sec".format(qb1_time))
assert_equal(qb.total_clusters, expected_nb_clusters)
sizes1 = [qb.partitions()[i]['N'] for i in range(qb.total_clusters)]
indices1 = [
qb.partitions()[i]['indices'] for i in range(qb.total_clusters)
]
qb2 = QB_New(threshold)
qb2_time = measure("clusters = qb2.cluster(streamlines)", repeat)
print("QuickBundles2 time: {0:.4}sec".format(qb2_time))
print("Speed up of {0}x".format(qb1_time / qb2_time))
clusters = qb2.cluster(streamlines)
sizes2 = map(len, clusters)
indices2 = map(lambda c: c.indices, clusters)
assert_equal(len(clusters), expected_nb_clusters)
assert_array_equal(sizes2, sizes1)
assert_arrays_equal(indices2, indices1)
qb = QB_New(threshold, metric=MDFpy())
qb3_time = measure("clusters = qb.cluster(streamlines)", repeat)
print("QuickBundles2_python time: {0:.4}sec".format(qb3_time))
print("Speed up of {0}x".format(qb1_time / qb3_time))
clusters = qb.cluster(streamlines)
sizes3 = map(len, clusters)
indices3 = map(lambda c: c.indices, clusters)
assert_equal(len(clusters), expected_nb_clusters)
assert_array_equal(sizes3, sizes1)
assert_arrays_equal(indices3, indices1)
def bench_quickbundles():
dtype = "float32"
repeat = 10
nb_points = 12
streams, hdr = nib.trackvis.read(get_fnames('fornix'))
fornix = [s[0].astype(dtype) for s in streams]
fornix = streamline_utils.set_number_of_points(fornix, nb_points)
# Create eight copies of the fornix to be clustered (one in each octant).
streamlines = []
streamlines += [s + np.array([100, 100, 100], dtype) for s in fornix]
streamlines += [s + np.array([100, -100, 100], dtype) for s in fornix]
streamlines += [s + np.array([100, 100, -100], dtype) for s in fornix]
streamlines += [s + np.array([100, -100, -100], dtype) for s in fornix]
streamlines += [s + np.array([-100, 100, 100], dtype) for s in fornix]
streamlines += [s + np.array([-100, -100, 100], dtype) for s in fornix]
streamlines += [s + np.array([-100, 100, -100], dtype) for s in fornix]
streamlines += [s + np.array([-100, -100, -100], dtype) for s in fornix]
# The expected number of clusters of the fornix using threshold=10 is 4.
threshold = 10.
expected_nb_clusters = 4 * 8
print("Timing QuickBundles 1.0 vs. 2.0")
qb = QB_Old(streamlines, threshold, pts=None)
qb1_time = measure("QB_Old(streamlines, threshold, nb_points)", repeat)
print("QuickBundles time: {0:.4}sec".format(qb1_time))
assert_equal(qb.total_clusters, expected_nb_clusters)
sizes1 = [qb.partitions()[i]['N'] for i in range(qb.total_clusters)]
indices1 = [qb.partitions()[i]['indices']
for i in range(qb.total_clusters)]
qb2 = QB_New(threshold)
qb2_time = measure("clusters = qb2.cluster(streamlines)", repeat)
print("QuickBundles2 time: {0:.4}sec".format(qb2_time))
print("Speed up of {0}x".format(qb1_time / qb2_time))
clusters = qb2.cluster(streamlines)
sizes2 = map(len, clusters)
indices2 = map(lambda c: c.indices, clusters)
assert_equal(len(clusters), expected_nb_clusters)
assert_array_equal(list(sizes2), sizes1)
assert_arrays_equal(indices2, indices1)
qb = QB_New(threshold, metric=MDFpy())
qb3_time = measure("clusters = qb.cluster(streamlines)", repeat)
print("QuickBundles2_python time: {0:.4}sec".format(qb3_time))
print("Speed up of {0}x".format(qb1_time / qb3_time))
clusters = qb.cluster(streamlines)
sizes3 = map(len, clusters)
indices3 = map(lambda c: c.indices, clusters)
assert_equal(len(clusters), expected_nb_clusters)
assert_array_equal(list(sizes3), sizes1)
assert_arrays_equal(indices3, indices1)
def test_fprop_faster(self):
activation = "tanh"
seed = 1234
repeat = 1000
layer = LayerLSTM(input_size=DATA['features_size'],
hidden_size=DATA['hidden_size'],
activation=activation)
layer.initialize(initer.UniformInitializer(seed))
layer_fast = LayerLSTMFast(input_size=DATA['features_size'],
hidden_size=DATA['hidden_size'],
activation=activation)
# Wi, Wo, Wf, Wm
layer_fast.W.set_value(
np.concatenate([
layer.Wi.get_value(),
layer.Wo.get_value(),
layer.Wf.get_value(),
layer.Wm.get_value()
],
axis=1))
layer_fast.U.set_value(
np.concatenate([
layer.Ui.get_value(),
layer.Uo.get_value(),
layer.Uf.get_value(),
layer.Um.get_value()
],
axis=1))
input = T.matrix('input')
input.tag.test_value = DATA['batch_one_step']
last_h = sharedX(DATA['state_h'])
last_m = sharedX(DATA['state_m'])
fprop = theano.function([input], layer.fprop(input, last_h, last_m))
fprop_faster = theano.function([input],
layer_fast.fprop(input, last_h, last_m))
fprop_time = measure("h, m = fprop(DATA['batch_one_step'])", repeat)
fprop_faster_time = measure(
"h, m = fprop_faster(DATA['batch_one_step'])", repeat)
print("fprop time: {:.2f} sec.", fprop_time)
print("fprop faster time: {:.2f} sec.", fprop_faster_time)
print("Speedup: {:.2f}x".format(fprop_time / fprop_faster_time))
for i in range(DATA['seq_len']):
h1, m1 = fprop(DATA['batch'][:, i, :])
h2, m2 = fprop_faster(DATA['batch'][:, i, :])
assert_array_equal(h1, h2)
assert_array_equal(m1, m2)
def bench_length():
repeat = 1000
streamline = np.random.rand(1000, 3)
print("Timing length() in Cython")
cython_time = measure("length(streamline)", repeat)
print("Cython time: {0:.2}sec".format(cython_time))
python_time = measure("length_python(streamline)", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time/cython_time))
def bench_resample():
repeat = 1000
nb_points = 42
streamline = np.random.rand(1000, 3)
print("Timing set_number_of_points() in Cython")
cython_time = measure("set_number_of_points(streamline, nb_points)", repeat)
print("Cython time: {0:.2}sec".format(cython_time))
python_time = measure("set_number_of_points_python(streamline, nb_points)", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time/cython_time))
def bench_bounding_box():
vol = np.zeros((100, 100, 100))
vol[0, 0, 0] = 1
times = 100
time = measure("bounding_box(vol)", times) / times
print("Bounding_box on a sparse volume: {}".format(time))
vol[:] = 10
times = 1
time = measure("bounding_box(vol)", times) / times
print("Bounding_box on a dense volume: {}".format(time))
def bench_bounding_box():
vol = np.zeros((100, 100, 100))
vol[0, 0, 0] = 1
times = 100
time = measure("bounding_box(vol)", times) / times
print("Bounding_box on a sparse volume: {}".format(time))
vol[:] = 10
times = 1
time = measure("bounding_box(vol)", times) / times
print("Bounding_box on a dense volume: {}".format(time))
def bench_local_maxima():
repeat = 10000
vertices, faces = default_sphere.vertices, default_sphere.faces
print('Timing peak finding')
timed0 = measure("local_maxima(odf, edges)", repeat)
print('Actual sphere: %0.2f' % timed0)
# Create an artificial odf with a few peaks
odf = np.zeros(len(vertices))
odf[1] = 1.
odf[143] = 143.
odf[505] = 505.
timed1 = measure("local_maxima(odf, edges)", repeat)
print('Few-peak sphere: %0.2f' % timed1)
def bench_quick_squash():
# nosetests -s --match '(?:^|[\\b_\\.//-])[Bb]ench'
repeat = 10
shape = (300, 200)
arrs = np.zeros(shape, dtype=object)
scalars = np.zeros(shape, dtype=object)
for ijk in ndindex(arrs.shape):
arrs[ijk] = np.ones((3, 5))
scalars[ijk] = np.float32(0)
print('\nSquashing benchmarks')
for name, objs in (
('floats', np.zeros(shape, float).astype(object)),
('ints', np.zeros(shape, int).astype(object)),
('arrays', arrs),
('scalars', scalars),
):
print(name)
timed0 = measure("quick_squash(objs)", repeat)
timed1 = measure("old_squash(objs)", repeat)
print("fast %4.2f; slow %4.2f" % (timed0, timed1))
objs[50, 50] = None
timed0 = measure("quick_squash(objs)", repeat)
timed1 = measure("old_squash(objs)", repeat)
print("With None: fast %4.2f; slow %4.2f" % (timed0, timed1))
msk = objs != np.array(None)
timed0 = measure("quick_squash(objs, msk)", repeat)
timed1 = measure("old_squash(objs, msk)", repeat)
print("With mask: fast %4.2f; slow %4.2f" % (timed0, timed1))
objs[50, 50] = np.float32(0)
timed0 = measure("quick_squash(objs, msk)", repeat)
timed1 = measure("old_squash(objs, msk)", repeat)
print("Other dtype: fast %4.2f; slow %4.2f" % (timed0, timed1))
def bench_quick_squash():
# nosetests -s --match '(?:^|[\\b_\\.//-])[Bb]ench'
repeat = 10
shape = (300, 200)
arrs = np.zeros(shape, dtype=object)
scalars = np.zeros(shape, dtype=object)
for ijk in ndindex(arrs.shape):
arrs[ijk] = np.ones((3, 5))
scalars[ijk] = np.float32(0)
print('\nSquashing benchmarks')
for name, objs in (
('floats', np.zeros(shape, float).astype(object)),
('ints', np.zeros(shape, int).astype(object)),
('arrays', arrs),
('scalars', scalars),
):
print(name)
timed0 = measure("quick_squash(objs)", repeat)
timed1 = measure("old_squash(objs)", repeat)
print("fast %4.2f; slow %4.2f" % (timed0, timed1))
objs[50, 50] = None
timed0 = measure("quick_squash(objs)", repeat)
timed1 = measure("old_squash(objs)", repeat)
print("With None: fast %4.2f; slow %4.2f" % (timed0, timed1))
msk = objs != np.array(None)
timed0 = measure("quick_squash(objs, msk)", repeat)
timed1 = measure("old_squash(objs, msk)", repeat)
print("With mask: fast %4.2f; slow %4.2f" % (timed0, timed1))
objs[50, 50] = np.float32(0)
timed0 = measure("quick_squash(objs, msk)", repeat)
timed1 = measure("old_squash(objs, msk)", repeat)
print("Other dtype: fast %4.2f; slow %4.2f" % (timed0, timed1))
def bench_local_maxima():
repeat = 10000
sphere = get_sphere('symmetric724')
vertices, faces = sphere.vertices, sphere.faces
print('Timing peak finding')
timed0 = measure("local_maxima(odf, edges)", repeat)
print('Actual sphere: %0.2f' % timed0)
# Create an artificial odf with a few peaks
odf = np.zeros(len(vertices))
odf[1] = 1.
odf[143] = 143.
odf[505] = 505.
timed1 = measure("local_maxima(odf, edges)", repeat)
print('Few-peak sphere: %0.2f' % timed1)
def test_fprop_faster(self):
seed = 1234
repeat = 100
lstm = LSTM(
input_size=DATA['features_size'],
hidden_sizes=[DATA['hidden_size']],
)
lstm.initialize(initer.UniformInitializer(seed))
lstm2 = LSTMFaster(
input_size=DATA['features_size'],
hidden_sizes=[DATA['hidden_size']],
)
# Wi, Wo, Wf, Wm
# Make sure the weights are the same.
lstm2.layers_lstm[0].W.set_value(
np.concatenate([
lstm.layers_lstm[0].Wi.get_value(),
lstm.layers_lstm[0].Wo.get_value(),
lstm.layers_lstm[0].Wf.get_value(),
lstm.layers_lstm[0].Wm.get_value()
],
axis=1))
lstm2.layers_lstm[0].U.set_value(
np.concatenate([
lstm.layers_lstm[0].Ui.get_value(),
lstm.layers_lstm[0].Uo.get_value(),
lstm.layers_lstm[0].Uf.get_value(),
lstm.layers_lstm[0].Um.get_value()
],
axis=1))
input = T.tensor3('input')
input.tag.test_value = DATA['batch']
fprop = theano.function([input], lstm.get_output(input))
fprop2 = theano.function([input], lstm2.get_output(input))
fprop_time = measure("out = fprop(DATA['batch'])", repeat)
print("fprop time: {:.2f} sec.", fprop_time)
fprop2_time = measure("out = fprop2(DATA['batch'])", repeat)
print("fprop faster time: {:.2f} sec.", fprop2_time)
print("Speedup: {:.2f}x".format(fprop_time / fprop2_time))
out = fprop(DATA['batch'])
out2 = fprop2(DATA['batch'])
assert_array_equal(out, out2)
def bench_local_maxima():
repeat = 10000
sphere = get_sphere('symmetric724')
vertices, faces = sphere.vertices, sphere.faces
odf = abs(vertices.sum(-1))
edges = unique_edges(faces)
print('Timing peak finding')
timed0 = measure("local_maxima(odf, edges)", repeat)
print('Actual sphere: %0.2f' % timed0)
# Create an artificial odf with a few peaks
odf = np.zeros(len(vertices))
odf[1] = 1.
odf[143] = 143.
odf[505] = 505.
timed1 = measure("local_maxima(odf, edges)", repeat)
print('Few-peak sphere: %0.2f' % timed1)
def bench_length():
repeat = 1
nb_points_per_streamline = 100
nb_streamlines = int(1e5)
streamlines = [np.random.rand(nb_points_per_streamline, 3).astype("float32") for i in range(nb_streamlines)]
print("Timing length() in Cython ({0} streamlines)".format(nb_streamlines))
cython_time = measure("length(streamlines)", repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
streamlines = [np.random.rand(nb_points_per_streamline, 3).astype("float32") for i in range(nb_streamlines)]
python_time = measure("[length_python(s) for s in streamlines]", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time/cython_time))
del streamlines
def bench_local_maxima():
repeat = 10000
sphere = get_sphere("symmetric724")
vertices, faces = sphere.vertices, sphere.faces
odf = abs(vertices.sum(-1))
edges = unique_edges(faces)
print("Timing peak finding")
timed0 = measure("local_maxima(odf, edges)", repeat)
print("Actual sphere: %0.2f" % timed0)
# Create an artificial odf with a few peaks
odf = np.zeros(len(vertices))
odf[1] = 1.0
odf[143] = 143.0
odf[505] = 505.0
timed1 = measure("local_maxima(odf, edges)", repeat)
print("Few-peak sphere: %0.2f" % timed1)
def bench_zhang():
print 'Zhang min'
print '=' * 10
#ref_time = measure('tm.most_similar_track_zhang(tracks300)')
#print 'reference time: %f' % ref_time
opt_time = measure('pf.most_similar_track_zhang(tracks300)')
print 'optimized time: %f' % opt_time
print
def bench_zhang():
print 'Zhang min'
print '=' * 10
#ref_time = measure('tm.most_similar_track_zhang(tracks300)')
#print 'reference time: %f' % ref_time
opt_time = measure('pf.most_similar_track_zhang(tracks300)')
print 'optimized time: %f' % opt_time
print
def bench_array_to_file():
rng = np.random.RandomState(20111001)
repeat = 10
img_shape = (128, 128, 64, 10)
arr = rng.normal(size=img_shape)
sys.stdout.flush()
print_git_title("\nArray to file")
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save float64 to float32', mtime))
mtime = measure('array_to_file(arr, BytesIO(), np.int16)', repeat)
print('%30s %6.2f' % ('Save float64 to int16', mtime))
# Set a lot of NaNs to check timing
arr[:, :, :, 1] = np.nan
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save float64 to float32, NaNs', mtime))
mtime = measure('array_to_file(arr, BytesIO(), np.int16)', repeat)
print('%30s %6.2f' % ('Save float64 to int16, NaNs', mtime))
# Set a lot of infs to check timing
arr[:, :, :, 1] = np.inf
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save float64 to float32, infs', mtime))
mtime = measure('array_to_file(arr, BytesIO(), np.int16)', repeat)
print('%30s %6.2f' % ('Save float64 to int16, infs', mtime))
# Int16 input, float output
arr = np.random.random_integers(low=-1000, high=1000, size=img_shape)
arr = arr.astype(np.int16)
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save Int16 to float32', mtime))
sys.stdout.flush()
def bench_array_to_file():
rng = np.random.RandomState(20111001)
repeat = 10
img_shape = (128, 128, 64, 10)
arr = rng.normal(size=img_shape)
sys.stdout.flush()
print_git_title("\nArray to file")
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save float64 to float32', mtime))
mtime = measure('array_to_file(arr, BytesIO(), np.int16)', repeat)
print('%30s %6.2f' % ('Save float64 to int16', mtime))
# Set a lot of NaNs to check timing
arr[:, :, :, 1] = np.nan
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save float64 to float32, NaNs', mtime))
mtime = measure('array_to_file(arr, BytesIO(), np.int16)', repeat)
print('%30s %6.2f' % ('Save float64 to int16, NaNs', mtime))
# Set a lot of infs to check timing
arr[:, :, :, 1] = np.inf
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save float64 to float32, infs', mtime))
mtime = measure('array_to_file(arr, BytesIO(), np.int16)', repeat)
print('%30s %6.2f' % ('Save float64 to int16, infs', mtime))
# Int16 input, float output
arr = np.random.random_integers(low=-1000, high=1000, size=img_shape)
arr = arr.astype(np.int16)
mtime = measure('array_to_file(arr, BytesIO(), np.float32)', repeat)
print('%30s %6.2f' % ('Save Int16 to float32', mtime))
sys.stdout.flush()
def bench_compress_streamlines():
repeat = 10
fname = get_data('fornix')
streams, hdr = tv.read(fname)
streamlines = [i[0] for i in streams]
print("Timing compress_streamlines() in Cython ({0} streamlines)".format(len(streamlines)))
cython_time = measure("compress_streamlines(streamlines)", repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
fname = get_data('fornix')
streams, hdr = tv.read(fname)
streamlines = [i[0] for i in streams]
python_time = measure("map(compress_streamlines_python, streamlines)", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time/cython_time))
del streamlines
def bench_compress_streamlines():
repeat = 10
fname = get_fnames('fornix')
fornix = load_tractogram(fname, 'same', bbox_valid_check=False).streamlines
streamlines = Streamlines(fornix)
print("Timing compress_streamlines() in Cython"
" ({0} streamlines)".format(len(streamlines)))
cython_time = measure("compress_streamlines(streamlines)", repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
streamlines = Streamlines(fornix)
python_time = measure("map(compress_streamlines_python, streamlines)",
repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time / cython_time))
del streamlines
def bench_mdl_traj():
t=np.concatenate(tracks300)
#t=tracks300[0]
print 'MDL traj'
print '=' * 10
opt_time = measure('pf.approximate_mdl_trajectory(t)')
#opt_time = measure('tm.approximate_trajectory_partitioning(t)')
#opt_time= measure('tm.minimum_description_length_unpartitoned(t)')
print 'optimized time: %f' % opt_time
print
def bench_mdl_traj():
t = np.concatenate(tracks300)
#t=tracks300[0]
print 'MDL traj'
print '=' * 10
opt_time = measure('pf.approximate_mdl_trajectory(t)')
#opt_time = measure('tm.approximate_trajectory_partitioning(t)')
#opt_time= measure('tm.minimum_description_length_unpartitoned(t)')
print 'optimized time: %f' % opt_time
print
def bench_compress_streamlines():
repeat = 10
fname = get_fnames('fornix')
streams, hdr = tv.read(fname)
streamlines = [i[0] for i in streams]
print("Timing compress_streamlines() in Cython"
" ({0} streamlines)".format(len(streamlines)))
cython_time = measure("compress_streamlines(streamlines)", repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
fname = get_fnames('fornix')
streams, hdr = tv.read(fname)
streamlines = [i[0] for i in streams]
python_time = measure("map(compress_streamlines_python, streamlines)",
repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time / cython_time))
del streamlines
def bench_finite_range():
rng = np.random.RandomState(20111001)
repeat = 10
img_shape = (128, 128, 64, 10)
arr = rng.normal(size=img_shape)
sys.stdout.flush()
print_git_title("\nFinite range")
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('float64 all finite', mtime))
arr[:, :, :, 1] = np.nan
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('float64 many NaNs', mtime))
arr[:, :, :, 1] = np.inf
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('float64 many infs', mtime))
# Int16 input, float output
arr = np.random.random_integers(low=-1000, high=-1000, size=img_shape)
arr = arr.astype(np.int16)
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('int16', mtime))
sys.stdout.flush()
def bench_finite_range():
rng = np.random.RandomState(20111001)
repeat = 10
img_shape = (128, 128, 64, 10)
arr = rng.normal(size=img_shape)
sys.stdout.flush()
print_git_title("\nFinite range")
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('float64 all finite', mtime))
arr[:, :, :, 1] = np.nan
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('float64 many NaNs', mtime))
arr[:, :, :, 1] = np.inf
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('float64 many infs', mtime))
# Int16 input, float output
arr = np.random.random_integers(low=-1000, high=1000, size=img_shape)
arr = arr.astype(np.int16)
mtime = measure('finite_range(arr)', repeat)
print('%30s %6.2f' % ('int16', mtime))
sys.stdout.flush()
def test_fprop_faster(self):
activation = "tanh"
seed = 1234
repeat = 1000
layer = LayerLSTM(input_size=DATA['features_size'],
hidden_size=DATA['hidden_size'],
activation=activation)
layer.initialize(initer.UniformInitializer(seed))
layer_fast = LayerLSTMFast(input_size=DATA['features_size'],
hidden_size=DATA['hidden_size'],
activation=activation)
# Wi, Wo, Wf, Wm
layer_fast.W.set_value(np.concatenate([layer.Wi.get_value(), layer.Wo.get_value(), layer.Wf.get_value(), layer.Wm.get_value()], axis=1))
layer_fast.U.set_value(np.concatenate([layer.Ui.get_value(), layer.Uo.get_value(), layer.Uf.get_value(), layer.Um.get_value()], axis=1))
input = T.matrix('input')
input.tag.test_value = DATA['batch_one_step']
last_h = sharedX(DATA['state_h'])
last_m = sharedX(DATA['state_m'])
fprop = theano.function([input], layer.fprop(input, last_h, last_m))
fprop_faster = theano.function([input], layer_fast.fprop(input, last_h, last_m))
fprop_time = measure("h, m = fprop(DATA['batch_one_step'])", repeat)
fprop_faster_time = measure("h, m = fprop_faster(DATA['batch_one_step'])", repeat)
print("fprop time: {:.2f} sec.", fprop_time)
print("fprop faster time: {:.2f} sec.", fprop_faster_time)
print("Speedup: {:.2f}x".format(fprop_time/fprop_faster_time))
for i in range(DATA['seq_len']):
h1, m1 = fprop(DATA['batch'][:, i, :])
h2, m2 = fprop_faster(DATA['batch'][:, i, :])
assert_array_equal(h1, h2)
assert_array_equal(m1, m2)
def bench_csdeconv(center=(50, 40, 40), width=12):
img, gtab, labels_img = read_stanford_labels()
data = img.get_data()
labels = labels_img.get_data()
shape = labels.shape
mask = np.in1d(labels, [1, 2])
mask.shape = shape
a, b, c = center
hw = width // 2
idx = (slice(a - hw, a + hw), slice(b - hw, b + hw), slice(c - hw, c + hw))
data_small = data[idx].copy()
mask_small = mask[idx].copy()
voxels = mask_small.sum()
cmd = "model.fit(data_small, mask_small)"
print("== Benchmarking CSD fit on %d voxels ==" % voxels)
msg = "SH order - %d, gradient directons - %d :: %g sec"
# Basic case
sh_order = 8
ConstrainedSphericalDeconvModel(gtab, None, sh_order=sh_order)
time = npt.measure(cmd)
print(msg % (sh_order, num_grad(gtab), time))
# Smaller data set
# data_small = data_small[..., :75].copy()
gtab = GradientTable(gtab.gradients[:75])
ConstrainedSphericalDeconvModel(gtab, None, sh_order=sh_order)
time = npt.measure(cmd)
print(msg % (sh_order, num_grad(gtab), time))
# Super resolution
sh_order = 12
ConstrainedSphericalDeconvModel(gtab, None, sh_order=sh_order)
time = npt.measure(cmd)
print(msg % (sh_order, num_grad(gtab), time))
def bench_csdeconv(center=(50, 40, 40), width=12):
img, gtab, labels_img = read_stanford_labels()
data = img.get_data()
labels = labels_img.get_data()
shape = labels.shape
mask = np.in1d(labels, [1, 2])
mask.shape = shape
a, b, c = center
hw = width // 2
idx = (slice(a - hw, a + hw), slice(b - hw, b + hw), slice(c - hw, c + hw))
data_small = data[idx].copy()
mask_small = mask[idx].copy()
voxels = mask_small.sum()
cmd = "model.fit(data_small, mask_small)"
print("== Benchmarking CSD fit on %d voxels ==" % voxels)
msg = "SH order - %d, gradient directons - %d :: %g sec"
# Basic case
sh_order = 8
model = ConstrainedSphericalDeconvModel(gtab, None, sh_order=sh_order)
time = npt.measure(cmd)
print(msg % (sh_order, num_grad(gtab), time))
# Smaller data set
data_small = data_small[..., :75].copy()
gtab = GradientTable(gtab.gradients[:75])
model = ConstrainedSphericalDeconvModel(gtab, None, sh_order=sh_order)
time = npt.measure(cmd)
print(msg % (sh_order, num_grad(gtab), time))
# Super resolution
sh_order = 12
model = ConstrainedSphericalDeconvModel(gtab, None, sh_order=sh_order)
time = npt.measure(cmd)
print(msg % (sh_order, num_grad(gtab), time))
def bench_length():
repeat = 1
nb_points_per_streamline = 100
nb_streamlines = int(1e5)
streamlines = [
np.random.rand(nb_points_per_streamline, 3).astype("float32")
for i in range(nb_streamlines)
]
print("Timing length() in Cython ({0} streamlines)".format(nb_streamlines))
cython_time = measure("length(streamlines)", repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
streamlines = [
np.random.rand(nb_points_per_streamline, 3).astype("float32")
for i in range(nb_streamlines)
]
python_time = measure("[length_python(s) for s in streamlines]", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time / cython_time))
del streamlines
def bench_random(self):
numpy_solve = nl.solve
scipy_solve = sl.solve
print()
print(' Solving system of linear equations')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size,repeat in [(20,1000),(100,150),(500,2),(1000,1)][:-1]:
repeat *= 2
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size,size])
# larger diagonal ensures non-singularity:
for i in range(size):
a[i,i] = 10*(.1+a[i,i])
b = random([size])
print('| %6.2f ' % measure('scipy_solve(a,b)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_solve(a,b)',repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1,-1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_solve(a,b)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_solve(a,b)',repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_random(self):
numpy_solve = nl.solve
scipy_solve = sl.solve
print()
print(' Solving system of linear equations')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy ')
for size, repeat in [(20, 1000), (100, 150), (500, 2), (1000, 1)][:-1]:
repeat *= 2
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size, size])
# larger diagonal ensures non-singularity:
for i in range(size):
a[i, i] = 10 * (.1 + a[i, i])
b = random([size])
print('| %6.2f ' % measure('scipy_solve(a,b)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_solve(a,b)', repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1, -1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_solve(a,b)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_solve(a,b)', repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def run_print(test_list):
print()
print(" Integrating sum(x**2) -- MISER Monte Carlo")
print(" ==========================================")
print()
print(" ndims | npoints | nprocs | time ")
print(" ------------------------------- ")
for ndims,npoints,nprocs,repeat in test_list:
print(" {ndims:5} | {npoints:7} | {nprocs:6} |".format(ndims=ndims,npoints=npoints,nprocs=nprocs),end="")
xl = [0.]*ndims
xu = [1.]*ndims
time = measure("mcmiser(lambda x: sum(x**2),{npoints},{xl},{xu},nprocs={nprocs})".format(npoints=npoints,xl=str(xl),xu=str(xu),nprocs=str(nprocs)),repeat)
print(" {time:.2f} (seconds for {ncalls} calls)".format(time=time,ncalls=repeat))
def bench_length():
repeat = 10
nb_streamlines = DATA['nb_streamlines']
msg = "Timing length() with {0:,} streamlines."
print(msg.format(nb_streamlines * repeat))
python_time = measure("[length_python(s) for s in streamlines]", repeat)
print("Python time: {0:.2f} sec".format(python_time))
cython_time = measure("length(streamlines)", repeat)
print("Cython time: {0:.3f} sec".format(cython_time))
print("Speed up of {0:.2f}x".format(python_time/cython_time))
# Make sure it produces the same results.
assert_array_almost_equal([length_python(s) for s in DATA["streamlines"]],
length(DATA["streamlines"]))
cython_time_arrseq = measure("length(streamlines)", repeat)
print("Cython time (ArrSeq): {0:.3f} sec".format(cython_time_arrseq))
print("Speed up of {0:.2f}x".format(python_time/cython_time_arrseq))
# Make sure it produces the same results.
assert_array_equal(length(DATA["streamlines"]),
length(DATA["streamlines_arrseq"]))
def bench_set_number_of_points():
repeat = 1
nb_points_per_streamline = 100
nb_points = 42
nb_streamlines = int(1e4)
streamlines = [np.random.rand(nb_points_per_streamline,
3).astype("float32")
for i in range(nb_streamlines)]
print("Timing set_number_of_points() in Cython"
"({0} streamlines)".format(nb_streamlines))
cython_time = measure("set_number_of_points(streamlines, nb_points)",
repeat)
print("Cython time: {0:.3}sec".format(cython_time))
del streamlines
streamlines = [np.random.rand(nb_points_per_streamline,
3).astype("float32")
for i in range(nb_streamlines)]
python_time = measure("[set_number_of_points_python(s, nb_points)"
" for s in streamlines]", repeat)
print("Python time: {0:.2}sec".format(python_time))
print("Speed up of {0}x".format(python_time/cython_time))
del streamlines
def bench_random(self):
numpy_inv = nl.inv
scipy_inv = sl.inv
print()
print(' Finding matrix inverse')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy')
for size,repeat in [(20,1000),(100,150),(500,2),(1000,1)][:-1]:
repeat *= 2
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size,size])
# large diagonal ensures non-singularity:
for i in range(size):
a[i,i] = 10*(.1+a[i,i])
print('| %6.2f ' % measure('scipy_inv(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_inv(a)',repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1,-1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_inv(a)',repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_inv(a)',repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def bench_random(self):
numpy_inv = nl.inv
scipy_inv = sl.inv
print()
print(' Finding matrix inverse')
print(' ==================================')
print(' | contiguous | non-contiguous ')
print('----------------------------------------------')
print(' size | scipy | numpy | scipy | numpy')
for size, repeat in [(20, 1000), (100, 150), (500, 2), (1000, 1)][:-1]:
repeat *= 2
print('%5s' % size, end=' ')
sys.stdout.flush()
a = random([size, size])
# large diagonal ensures non-singularity:
for i in range(size):
a[i, i] = 10 * (.1 + a[i, i])
print('| %6.2f ' % measure('scipy_inv(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_inv(a)', repeat), end=' ')
sys.stdout.flush()
a = a[-1::-1, -1::-1] # turn into a non-contiguous array
assert_(not a.flags['CONTIGUOUS'])
print('| %6.2f ' % measure('scipy_inv(a)', repeat), end=' ')
sys.stdout.flush()
print('| %6.2f ' % measure('numpy_inv(a)', repeat), end=' ')
sys.stdout.flush()
print(' (secs for %s calls)' % (repeat))
def run_print(test_list):
print()
print(" Integrating exp(-sum(x**2))*sum(x**2) w. importance sampling")
print(" ============================================================")
print()
print(" ndims | npoints | nprocs | time ")
print(" ------------------------------- ")
for ndims,npoints,nprocs,repeat in test_list:
print(" {ndims:5} | {npoints:7} | {nprocs:6} |".format(ndims=ndims,npoints=npoints,nprocs=nprocs),end="")
mean = "np.zeros(({ndims},))".format(ndims=ndims)
cov = "np.eye({ndims}) / np.sqrt(2.)".format(ndims=ndims)
time = measure(
"mcimport(lambda x: sum(x**2),{npoints},\
lambda size: multivariate_normal({mean},{cov},size),\
nprocs={nprocs})".format(
npoints=npoints,mean=mean,cov=cov,nprocs=str(nprocs)),repeat)
print(" {time:.2f} (seconds for {ncalls} calls)".format(time=time,ncalls=repeat))
def run_print(test_list):
print()
print(" Integrating sum(x**2) -- MISER Monte Carlo")
print(" ==========================================")
print()
print(" ndims | npoints | nprocs | time ")
print(" ------------------------------- ")
for ndims, npoints, nprocs, repeat in test_list:
print(" {ndims:5} | {npoints:7} | {nprocs:6} |".format(ndims=ndims,
npoints=npoints,
nprocs=nprocs),
end="")
xl = [0.] * ndims
xu = [1.] * ndims
time = measure(
"mcmiser(lambda x: sum(x**2),{npoints},{xl},{xu},nprocs={nprocs})".
format(npoints=npoints, xl=str(xl), xu=str(xu),
nprocs=str(nprocs)), repeat)
print(" {time:.2f} (seconds for {ncalls} calls)".format(
time=time, ncalls=repeat))
def test_fprop(self):
activation = "tanh"
seed = 1234
repeat = 1000
layer = LayerLSTM(input_size=DATA['features_size'],
hidden_size=DATA['hidden_size'],
activation=activation)
layer.initialize(initer.UniformInitializer(seed))
# input = T.tensor3('input')
input = T.matrix('input')
input.tag.test_value = DATA['batch_one_step']
last_h = sharedX(DATA['state_h'])
last_m = sharedX(DATA['state_m'])
fprop = theano.function([input],
layer.fprop_faster(input, last_h, last_m))
fprop_time = measure("h, m = fprop(DATA['batch_one_step'])", repeat)
print("fprop time: {:.2f} sec.", fprop_time)
h, m = fprop(DATA['batch_one_step'])
def bench_load_save():
rng = np.random.RandomState(20111001)
repeat = 10
img_shape = (128, 128, 64, 10)
arr = rng.normal(size=img_shape)
img = Nifti1Image(arr, np.eye(4))
sio = BytesIO()
img.file_map['image'].fileobj = sio
hdr = img.header
sys.stdout.flush()
print()
print_git_title("Image load save")
hdr.set_data_dtype(np.float32)
mtime = measure('sio.truncate(0); img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save float64 to float32', mtime))
mtime = measure('img.from_file_map(img.file_map)', repeat)
print('%30s %6.2f' % ('Load from float32', mtime))
hdr.set_data_dtype(np.int16)
mtime = measure('sio.truncate(0); img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save float64 to int16', mtime))
mtime = measure('img.from_file_map(img.file_map)', repeat)
print('%30s %6.2f' % ('Load from int16', mtime))
# Set a lot of NaNs to check timing
arr[:, :, :20] = np.nan
mtime = measure('sio.truncate(0); img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save float64 to int16, NaNs', mtime))
mtime = measure('img.from_file_map(img.file_map)', repeat)
print('%30s %6.2f' % ('Load from int16, NaNs', mtime))
# Int16 input, float output
arr = np.random.random_integers(low=-1000, high=1000, size=img_shape)
arr = arr.astype(np.int16)
img = Nifti1Image(arr, np.eye(4))
sio = BytesIO()
img.file_map['image'].fileobj = sio
hdr = img.header
hdr.set_data_dtype(np.float32)
mtime = measure('sio.truncate(0); img.to_file_map()', repeat)
print('%30s %6.2f' % ('Save Int16 to float32', mtime))
sys.stdout.flush()
