Bug in signal decomposition concerning navigation_mask

[w-markus]

Please run:

%matplotlib qt
import hyperspy.api as hs
import hyperspy
import numpy as np
import matplotlib
from matplotlib import pyplot as plt

# some fake data
e = np.arange(0,2048)
spectra = [np.exp(-((e-p)/100) ** 2) for p in [200,600,1400]]

data = np.zeros((192,256,2048))
comp = [(0,90,4,1,1), (90,100,1,4,0.8), (100,110,1,4,1.0),(110,120,1,4,1.2),(120,192,1,1,4)]
for b,t,c0,c1,c2 in comp:
    data[b:t] = spectra[0]*c0+spectra[1]*c1+spectra[2]*c2
data = np.random.poisson(lam=100*data)*1.0
signal = hs.signals.Signal1D(data, signal_type='EDS_TEM')
signal.axes_manager[-1].name = 'E'
signal.axes_manager['E'].units = 'eV'
signal.axes_manager['E'].scale = 1.0e1
signal.axes_manager['E'].offset = 0

signal.plot()

#--------------------------------------------------
# do NMF decomposition
# define a mask
mask = np.ones((192,256)).astype(bool)
mask[90:120] = False

# one would expect to use navigation_mask=mask
# but this leads to "The shape of `mask` must match the shape of the `navigation_shape`."
# using navigation_mask=mask.T does not give that error, but the result makes no sense
# see the resulting loadings plot

signal.decomposition(normalize_poissonian_noise=True, 
                     navigation_mask=mask.T, 
                     algorithm='NMF', 
                     output_dimension=3, 
                     centre=None, 
                     max_iter=20,
                     return_info=False)    

loadings = signal.get_decomposition_loadings()
loadings.plot()

[ericpre]

Thanks for the example, I can reproduce.

Until this is fixed, you can use a signal as a mask instead of numpy array:

s2 = signal.sum(-1)
nav_mask = (s2 < 1.08E5)

signal.decomposition(normalize_poissonian_noise=True,
                     navigation_mask=nav_mask,
                     algorithm='NMF',
                     output_dimension=3,
                     centre=None,
                     max_iter=20,
                     return_info=False)

[ericpre]

Should be fixed in #2679.