def fit(self, data):
max_sir = None
self.ica = None
for _ in range(self.n_iter):
fca = FastICA(n_components=self.n_components,
max_iter=self.max_iter,
algorithm='deflation')
fca.fit(data)
# try:
new_components = sorted(zip(fca.components_, np.argmax(np.abs(fca.components_), axis=1)), key= lambda x: x[1])
new_components = np.stack(new_components)
new_components = np.array([x[0] for x in new_components])
fca.components_ = new_components
m = new_components/(abs(norm(new_components, axis=1)).reshape(-1,1)) - np.identity(data.shape[1])
sir = norm(m, axis=1)
# except:
# continue
sir = -10*np.log10(sir).sum()
# in the paper its >
if max_sir is None or sir > max_sir:
self.max_sir = sir
max_sir = sir
self.sir_log.append(sir)
self.ica = fca
return self.ica
def PerformIca(X,Y,num_components,random_state):
result = {}
algo = FastICA(random_state=random_state,max_iter=800)
algo.fit(X)
full_mixing_matrix = algo.mixing_
full_unmixing_matrix = algo.components_
_x = algo.transform(X)
kt_value = np.abs(kt(_x))
largest_kt_values_idx = np.argsort(kt_value)[::-1]
result["ica_kt_all"] = kt_value
for n in num_components:
prefix = "ica_" + str(n) + "_"
component_idx_to_select = largest_kt_values_idx[0:n]
mixing_matrix = full_mixing_matrix.T[component_idx_to_select,:].T
unmixing_matrix = full_unmixing_matrix[component_idx_to_select,:]
algo.components_ = unmixing_matrix
algo.mixing_ = mixing_matrix
result[prefix+"mm"] = mixing_matrix
result[prefix+"umm"] = unmixing_matrix
_x = algo.transform(X)
result[prefix+"data"] = _x
X_recons = algo.inverse_transform(_x)
result[prefix+"reconstruction_error"] = ComputeReconstructionSSE(X,X_recons)
n_kt_value = kt_value[component_idx_to_select]
avg_kt = n_kt_value.mean()
#print("ICA num dim {0} : reconstruction error {1} avg kt {2}".format(str(n),str(result[prefix+"reconstruction_error"]),str(avg_kt)))
#print(np.sort(n_kt_value))
return result
def generate_components(
images,
hemi,
term_scores=None,
n_components=20,
random_state=42,
out_dir=None,
memory=Memory(cachedir="nilearn_cache"),
):
"""
images: list
Can be nibabel images, can be file paths.
"""
# Create grey matter mask from mni template
target_img = datasets.load_mni152_template()
# Reshape & mask images
print("%s: Reshaping and masking images; may take time." % hemi)
if hemi == "wb":
masker = GreyMatterNiftiMasker(target_affine=target_img.affine, target_shape=target_img.shape, memory=memory)
else: # R and L maskers
masker = HemisphereMasker(
target_affine=target_img.affine, target_shape=target_img.shape, memory=memory, hemisphere=hemi
)
masker = masker.fit()
# Images may fail to be transformed, and are of different shapes,
# so we need to trasnform one-by-one and keep track of failures.
X = [] # noqa
xformable_idx = np.ones((len(images),), dtype=bool)
for ii, im in enumerate(images):
img = cast_img(im, dtype=np.float32)
img = clean_img(img)
try:
X.append(masker.transform(img))
except Exception as e:
print("Failed to mask/reshape image %d/%s: %s" % (im.get("collection_id", 0), op.basename(im), e))
xformable_idx[ii] = False
# Now reshape list into 2D matrix
X = np.vstack(X) # noqa
# Run ICA and map components to terms
print("%s: Running ICA; may take time..." % hemi)
fast_ica = FastICA(n_components=n_components, random_state=random_state)
fast_ica = memory.cache(fast_ica.fit)(X.T)
ica_maps = memory.cache(fast_ica.transform)(X.T).T
# Tomoki's suggestion to normalize components_
# X ~ ica_maps * fast_ica.components_
# = (ica_maps * f) * (fast_ica.components_ / f)
# = new_ica_map * new_components_
C = fast_ica.components_
factor = np.sqrt(np.multiply(C, C).sum(axis=1, keepdims=True)) # (n_components x 1)
ica_maps = np.multiply(ica_maps, factor)
fast_ica.components_ = np.multiply(C, 1.0 / (factor + 1e-12))
if term_scores is not None:
terms = term_scores.keys()
term_matrix = np.asarray(term_scores.values())
term_matrix[term_matrix < 0] = 0
term_matrix = term_matrix[:, xformable_idx] # terms x images
# Don't use the transform method as it centers the data
ica_terms = np.dot(term_matrix, fast_ica.components_.T).T
# 2015/12/26 - sign matters for comparison, so don't do this!
# 2016/02/01 - sign flipping is ok for R-L comparison, but RL concat
# may break this.
# Pretty up the results
for idx, ic in enumerate(ica_maps):
if -ic.min() > ic.max():
# Flip the map's sign for prettiness
ica_maps[idx] = -ic
if term_scores:
ica_terms[idx] = -ica_terms[idx]
# Create image from maps, save terms to the image directly
ica_image = NiftiImageWithTerms.from_image(masker.inverse_transform(ica_maps))
if term_scores:
ica_image.terms = dict(zip(terms, ica_terms.T))
# Write to disk
if out_dir is not None:
out_path = op.join(out_dir, "%s_ica_components.nii.gz" % hemi)
if not op.exists(op.dirname(out_path)):
os.makedirs(op.dirname(out_path))
ica_image.to_filename(out_path)
return ica_image