def get_model(method, n=None, n_neighbors=None, max_iter=None, random_state=None, n_jobs=None, method_args=None):
model = None
kwargs = {}
if method_args is not None:
for arg_pair in method_args:
arg_par = arg_pair.split('=')
kwargs[arg_pair[0].trim()] = arg_pair[1].trim()
if method == 'ICA':
model = FastICA(whiten=True, **kwargs)
elif method == 'ICAexp':
model = FastICA(whiten=True, fun='exp', **kwargs)
elif method == 'ICAcube':
model = FastICA(whiten=True, fun='cube', **kwargs)
elif method == 'PCA':
model = PCA()
elif method == 'SPCA':
model = SparsePCA()
if n_jobs is not None:
model.n_jobs = n_jobs
elif method == 'NMF':
model = NMF(solver='cd')
elif method == 'ISO':
model = Isomap()
elif method == 'KPCA':
model = KernelPCA(kernel='rbf', fit_inverse_transform=False, gamma=1, alpha=0.0001)
elif method == 'FA':
#model = FactorAnalysis(svd_method='lapack') #(tol=0.0001, iterated_power=4)
model = FactorAnalysis(tol=0.0001, iterated_power=4)
elif method == 'DL':
model = DictionaryLearning(split_sign=True, fit_algorithm='cd', alpha=1)
if n_jobs is not None:
model.n_jobs = n_jobs
if n is not None:
model.n_components = n
if max_iter is not None:
model.max_iter = max_iter
if random_state is not None:
model.random_state = random_state
if n_neighbors is not None:
model.n_neighbors = n_neighbors
return model
def optimize_components(X, feature_names, label, abbrev):
# model selection (optimal number of components)
# choose number of components by highest average kurtosis
n_components = np.arange(1, len(feature_names) + 1)
ica = FastICA(random_state=SEED)
ica_scores = []
for n in n_components:
ica.n_components = n
result = pd.DataFrame(ica.fit_transform(X))
ica_scores.append(result.kurtosis().mean())
n_components_ica = n_components[np.argmax(ica_scores)]
print(label +
": best n_components by average kurtosis = %d" % n_components_ica)
# create plot
plt.figure()
plt.plot(n_components,
ica_scores,
'b',
label='average kurtosis by number of components')
plt.axvline(n_components_ica,
color='b',
label='chosen number of components: %d' % n_components_ica,
linestyle='--')
ax = plt.gca()
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
plt.xlabel('number of components')
plt.ylabel('average kurtosis')
plt.legend(loc='lower right')
plt.title(label + ": ICA model selection")
plt.savefig(path.join(PLOT_DIR, abbrev + "_ica_components.png"),
bbox_inches='tight')
plt.show()
plt.close()
return n_components_ica
def compute_scores(X):
pca = PCA(svd_solver='auto')
kpca = KernelPCA(fit_inverse_transform=True)
ica = FastICA()
nmf = NMF(init='nndsvda')
pca_scores, ica_scores, nmf_scores, kpca_scores = [], [], [], []
for n in n_components:
pca.n_components = n
ica.n_components = n
nmf.n_components = n
kpca.n_components = n
print(n)
Xpca = pca.inverse_transform(pca.fit_transform(Xs))
pca_scores.append(explained_variance_score(Xs, Xpca))
Xica = ica.inverse_transform(ica.fit_transform(Xs))
ica_scores.append(explained_variance_score(Xs, Xica))
Xkpca = kpca.inverse_transform(kpca.fit_transform(Xs))
kpca_scores.append(explained_variance_score(Xs, Xkpca))
Xnmf = nmf.inverse_transform(nmf.fit_transform(X))
nmf_scores.append(explained_variance_score(X, Xnmf))
return pca_scores, ica_scores, nmf_scores, kpca_scores
x = A @ sig
# x += 0.1*np.random.randn(*x.shape)
xm = x - x.mean(axis=1)[:, None]
# #### Perform PCA
u, s, v = np.linalg.svd(xm, full_matrices=False)
nm = np.sum(s > 1e-6)
s_pca = v[:nm, :].T * s[:nm]
# #### Apply ICA
xw = (u / s).T[:sig.shape[0]] @ xm
xw *= np.sqrt(xm.shape[1])
ica = FastICA(algorithm='deflation', tol=1e-12)
ica.whiten = True
ica.n_components = sig.shape[0]
s_ica = ica.fit_transform(xm.T)
# ica.whiten = False
# s_ica = ica.fit_transform(xw.T)
s_ica *= s[:nm].mean()
# #### Plot data
f = mplt.figure(figsize=(10, 10))
gs = mgs.GridSpec(4, 3)
gs.update(top=0.96,
bottom=0.05,
left=0.08,
right=0.98,
hspace=0.3,
wspace=0.35)