def spec_iterative_pca(outfile, n_ev=10, n_iter=20, norm='L2'):
"""
This function takes the file outputted above, performs an iterative
PCA to fill in the gaps, and appends the results to the same file.
"""
data_in = np.load(outfile)
spectra = data_in['spectra']
mask = data_in['mask']
res = iterative_pca(spectra,
mask,
n_ev=n_ev,
n_iter=n_iter,
norm=norm,
full_output=True)
input_dict = {key: data_in[key] for key in data_in.files}
# don't save the reconstructed spectrum: this can easily
# be recomputed from the other parameters.
input_dict['mu'] = res[1]
input_dict['evecs'] = res[2]
input_dict['evals'] = res[3]
input_dict['norms'] = res[4]
input_dict['coeffs'] = res[5]
np.savez(outfile, **input_dict)
def spec_iterative_pca(outfile, n_ev=10, n_iter=20, norm='L2'):
"""
This function takes the file outputted above, performs an iterative
PCA to fill in the gaps, and appends the results to the same file.
"""
data_in = np.load(outfile)
spectra = data_in['spectra']
mask = data_in['mask']
res = iterative_pca(spectra, mask,
n_ev=n_ev, n_iter=n_iter, norm=norm,
full_output=True)
input_dict = dict([(key, data_in[key]) for key in data_in.files])
# don't save the reconstructed spectrum: this can easily
# be recomputed from the other parameters.
input_dict['mu'] = res[1]
input_dict['evecs'] = res[2]
input_dict['evals'] = res[3]
input_dict['norms'] = res[4]
input_dict['coeffs'] = res[5]
np.savez(outfile,
**input_dict)
def test_iterative_PCA(n_samples=50, n_features=40):
np.random.seed(0)
# construct some data that is well-approximated
# by two principal components
x = np.linspace(0, np.pi, n_features)
x0 = np.linspace(0, np.pi, n_samples)
X = np.sin(x) * np.cos(0.5 * (x - x0[:, None]))
# mask 10% of the pixels
M = (np.random.random(X.shape) > 0.9)
# reconstruct and check accuracy
for norm in (None, 'L1', 'L2'):
X_recons = iterative_pca(X, M, n_ev=2, n_iter=10, norm=norm)
assert_array_almost_equal(X, X_recons, decimal=2)
def eigenspectra(data, output='spectra', wvrange=(6600,9000)):
# data is a fits table (or similar) 2d array
notnan = (data.LBIN[0] > wvrange[0]) & (data.LBIN[0] < wvrange[1])
N = len(data)
n = len(data.LBIN[0][notnan])
smooth=True
spectra = np.empty((N,n))
mask = np.empty((N,n))
for i in range(N):
flx = data.SPECCALIB[i][notnan]
wv = data.LBIN[i][notnan]
if smooth:
spec = Spectrum(wv,flx)
spec.smooth(width='5', filtertype='gaussian', replace=True)
flx = spec.flux
wv = spec.x
spectra[i,:] = flx
mask[i,:] = (~np.isfinite(flx)) | (flx == 0)
x_recon, mu, evecs, evals, norms, coeffs = iterative_pca(spectra, mask, n_ev=8, n_iter=15, norm='L2', full_output=True)
evals_cs = (evals**2).cumsum()
evals_cs /= evals_cs[-1]
spec_mean = x_recon.mean(0)
coeffs = np.empty((N,N))
for i, spec in enumerate(x_recon):
cfs = np.dot(evecs, spec/norms[i] - mu)
coeffs[:,i] = cfs
if output == 'spectra':
return x_recon, norms, notnan, mask
elif output == 'components':
return evecs, evals, coeffs
