def run(idx, reduction, alpha, mask, raw, n_components, init, func_filenames):
output_dir = join(trace_folder, 'experiment_%i' % idx)
try:
os.makedirs(output_dir)
except OSError:
pass
dict_fact = SpcaFmri(mask=mask,
smoothing_fwhm=3,
batch_size=40,
shelve=not raw,
n_components=n_components,
replacement=False,
dict_init=fetch_atlas_smith_2009().rsn70 if
init else None,
reduction=reduction,
alpha=alpha,
random_state=0,
n_epochs=2,
l1_ratio=0.5,
backend='c',
memory=expanduser("~/nilearn_cache"), memory_level=2,
verbose=5,
n_jobs=1,
trace_folder=output_dir
)
print('[Example] Learning maps')
t0 = time.time()
dict_fact.fit(func_filenames, raw=raw)
t1 = time.time() - t0
print('[Example] Dumping results')
# Decomposition estimator embeds their own masker
masker = dict_fact.masker_
components_img = masker.inverse_transform(dict_fact.components_)
components_img.to_filename(join(output_dir, 'components_final.nii.gz'))
print('[Example] Run in %.2f s' % t1)
# Show components from both methods using 4D plotting tools
import matplotlib.pyplot as plt
from nilearn.plotting import plot_prob_atlas, show
print('[Example] Displaying')
fig, axes = plt.subplots(2, 1)
plot_prob_atlas(components_img, view_type="filled_contours",
axes=axes[0])
plot_stat_map(index_img(components_img, 0),
axes=axes[1],
colorbar=False,
threshold=0)
plt.savefig(join(output_dir, 'components.pdf'))
show()
def run(idx, reduction, alpha, mask, raw, n_components, init, func_filenames):
output_dir = join(trace_folder, 'experiment_%i' % idx)
try:
os.makedirs(output_dir)
except OSError:
pass
dict_fact = SpcaFmri(
mask=mask,
smoothing_fwhm=3,
batch_size=40,
shelve=not raw,
n_components=n_components,
replacement=False,
dict_init=fetch_atlas_smith_2009().rsn70 if init else None,
reduction=reduction,
alpha=alpha,
random_state=0,
n_epochs=2,
l1_ratio=0.5,
backend='c',
memory=expanduser("~/nilearn_cache"),
memory_level=2,
verbose=5,
n_jobs=1,
trace_folder=output_dir)
print('[Example] Learning maps')
t0 = time.time()
dict_fact.fit(func_filenames, raw=raw)
t1 = time.time() - t0
print('[Example] Dumping results')
# Decomposition estimator embeds their own masker
masker = dict_fact.masker_
components_img = masker.inverse_transform(dict_fact.components_)
components_img.to_filename(join(output_dir, 'components_final.nii.gz'))
print('[Example] Run in %.2f s' % t1)
# Show components from both methods using 4D plotting tools
import matplotlib.pyplot as plt
from nilearn.plotting import plot_prob_atlas, show
print('[Example] Displaying')
fig, axes = plt.subplots(2, 1)
plot_prob_atlas(components_img, view_type="filled_contours", axes=axes[0])
plot_stat_map(index_img(components_img, 0),
axes=axes[1],
colorbar=False,
threshold=0)
plt.savefig(join(output_dir, 'components.pdf'))
show()
def test_component_sign():
# Regression test
# We should have a heuristic that flips the sign of components in
# DictLearning to have more positive values than negative values, for
# instance by making sure that the largest value is positive.
data, mask_img, components, rng = _make_test_data(n_subjects=2, noisy=True)
for mp in components:
assert_less_equal(-mp.min(), mp.max())
sparse_pca = SpcaFmri(n_components=4, random_state=rng, mask=mask_img, smoothing_fwhm=0.0)
sparse_pca.fit(data)
for mp in iter_img(sparse_pca.masker_.inverse_transform(sparse_pca.components_)):
mp = mp.get_data()
assert_less_equal(np.sum(mp[mp <= 0]), np.sum(mp[mp > 0]))
def test_component_sign():
# Regression test
# We should have a heuristic that flips the sign of components in
# DictLearning to have more positive values than negative values, for
# instance by making sure that the largest value is positive.
data, mask_img, components, rng = _make_test_data(n_subjects=2, noisy=True)
for mp in components:
assert_less_equal(-mp.min(), mp.max())
sparse_pca = SpcaFmri(n_components=4,
random_state=rng,
mask=mask_img,
smoothing_fwhm=0.)
sparse_pca.fit(data)
for mp in iter_img(
sparse_pca.masker_.inverse_transform(sparse_pca.components_)):
mp = mp.get_data()
assert_less_equal(np.sum(mp[mp <= 0]), np.sum(mp[mp > 0]))
def test_sparse_pca(backend, var_red):
data, mask_img, components, rng = _make_test_data(n_subjects=10)
sparse_pca = SpcaFmri(n_components=4,
random_state=0,
mask=mask_img,
backend=backend,
var_red=var_red,
reduction=2 if var_red else 1,
smoothing_fwhm=0.,
n_epochs=3,
alpha=0.01)
sparse_pca.fit(data)
maps = sparse_pca.masker_. \
inverse_transform(sparse_pca.components_).get_data()
maps = np.reshape(np.rollaxis(maps, 3, 0), (4, 400))
S = np.sqrt(np.sum(components**2, axis=1))
S[S == 0] = 1
components /= S[:, np.newaxis]
S = np.sqrt(np.sum(maps**2, axis=1))
S[S == 0] = 1
maps /= S[:, np.newaxis]
G = np.abs(components.dot(maps.T))
# Hard
# if var_red:
# recovered_maps = min(np.sum(np.any(G > 0.5, axis=1)),
# np.sum(np.any(G > 0.5, axis=0)))
# else:
# recovered_maps = min(np.sum(np.any(G > 0.95, axis=1)),
# np.sum(np.any(G > 0.95, axis=0)))
if var_red:
recovered_maps = np.sum(G > 0.7)
else:
recovered_maps = np.sum(G > 0.95)
assert (recovered_maps >= 4)
# Smoke test n_epochs > 1
sparse_pca = SpcaFmri(n_components=4,
random_state=0,
mask=mask_img,
var_red=var_red,
smoothing_fwhm=0.,
n_epochs=2,
alpha=1)
sparse_pca.fit(data)
# Smoke test reduction_ratio < 1
sparse_pca = SpcaFmri(n_components=4,
random_state=0,
reduction=2,
mask=mask_img,
var_red=var_red,
smoothing_fwhm=0.,
n_epochs=1,
alpha=1)
sparse_pca.fit(data)
def test_sparse_pca(backend):
data, mask_img, components, rng = _make_test_data(n_subjects=10)
sparse_pca = SpcaFmri(
n_components=4,
random_state=0,
mask=mask_img,
backend=backend,
reduction=2,
smoothing_fwhm=0.0,
n_epochs=3,
alpha=0.01,
)
sparse_pca.fit(data)
maps = sparse_pca.masker_.inverse_transform(sparse_pca.components_).get_data()
maps = np.reshape(np.rollaxis(maps, 3, 0), (4, 400))
S = np.sqrt(np.sum(components ** 2, axis=1))
S[S == 0] = 1
components /= S[:, np.newaxis]
S = np.sqrt(np.sum(maps ** 2, axis=1))
S[S == 0] = 1
maps /= S[:, np.newaxis]
G = np.abs(components.dot(maps.T))
recovered_maps = np.sum(G > 0.95)
assert recovered_maps >= 4
# Smoke test n_epochs > 1
sparse_pca = SpcaFmri(n_components=4, random_state=0, mask=mask_img, smoothing_fwhm=0.0, n_epochs=2, alpha=1)
sparse_pca.fit(data)
# Smoke test reduction_ratio < 1
sparse_pca = SpcaFmri(
n_components=4, random_state=0, reduction=2, mask=mask_img, smoothing_fwhm=0.0, n_epochs=1, alpha=1
)
sparse_pca.fit(data)
dict_fact = SpcaFmri(
n_components=n_components,
smoothing_fwhm=6.,
memory="nilearn_cache",
memory_level=2,
reduction=3,
verbose=4,
alpha=0.001,
random_state=0,
n_epochs=1,
n_jobs=1,
)
print('[Example] Learning maps')
t0 = time.time()
dict_fact.fit(func_filenames)
print('[Example] Dumping results')
# Decomposition estimator embeds their own masker
masker = dict_fact.masker_
components_img = masker.inverse_transform(dict_fact.components_)
components_img.to_filename('components.nii.gz')
time = time.time() - t0
print('[Example] Run in %.2f s' % time)
# Show components from both methods using 4D plotting tools
from nilearn.plotting import plot_prob_atlas, show
print('[Example] Displaying')
plot_prob_atlas(components_img,
view_type="filled_contours",
title="Reduced sparse PCA",
verbose=10,
alpha=0.001,
random_state=0,
learning_rate=.8,
replacement=True,
batch_size=50,
offset=0,
n_epochs=1,
backend='c',
# trace_folder=trace_folder,
n_jobs=n_jobs,
)
print('[Example] Learning maps')
t0 = time.time()
dict_fact.fit(func_filenames)
print('[Example] Dumping results')
# Decomposition estimator embeds their own masker
masker = dict_fact.masker_
components_img = masker.inverse_transform(dict_fact.components_)
components_img.to_filename(join(trace_folder, 'components.nii.gz'))
time = time.time() - t0
print('[Example] Run in %.2f s' % time)
# Show components from both methods using 4D plotting tools
import matplotlib.pyplot as plt
from nilearn.plotting import plot_prob_atlas, plot_stat_map, show
from nilearn.image import index_img
print('[Example] Displaying')
fig, axes = plt.subplots(2, 1)
plot_prob_atlas(components_img, view_type="filled_contours",
init = True
dict_fact = SpcaFmri(mask=mask,
smoothing_fwhm=3,
shelve=not raw,
n_components=n_components,
dict_init=fetch_atlas_smith_2009().rsn70 if init else None,
reduction=12,
alpha=0.001,
random_state=0,
n_epochs=1,
memory=expanduser("~/nilearn_cache"), memory_level=2,
verbose=4,
n_jobs=1,
)
print('[Example] Learning maps')
timings = np.zeros(20)
for n_jobs in range(1, 21):
with num_threads(n_jobs):
t0 = time.time()
dict_fact.fit(func_filenames, raw=raw)
timings[n_jobs - 1] = time.time() - t0
print(timings)
import matplotlib
matplotlib.use('pdf')
import matplotlib.pyplot as plt
plt.plot(np.arange(1, 21), timings)
plt.savefig('bench_hcp_blas.pdf')
