# The root of the problem may be that the dictionary is biased towards high expression
# (because the loss function in SMAF wasn't told to specifically care about lowly expressed genes)
# So we might change the loss function to be sensitive to the average loss for each gene
# ...possibly normalized somehow so that we care the same about each gene individually
# This may also help with the poor performance for each gene across tissues
# In addition, we might weight Phi by a prior on (inverse) expression
# Note that applying this weight without making the dictionary sensitive to low expression did not improve the results
k = min(int(xa.shape[1] * 1.5), 150)
UW = (np.random.random((xa.shape[0], k)), np.random.random(
(k, xa.shape[1])))
ua, va = smaf(xa,
k,
5,
0.0005,
maxItr=10,
use_chol=True,
activity_lower=0.,
module_lower=xa.shape[0] / 10,
UW=UW,
donorm=True,
mode=1,
mink=3.)
x2a, phi, y, w, d, psi = recover_system_knownBasis(
xa,
measurements,
sparsity,
Psi=ua,
snr=SNR,
use_ridge=False,
f="smaf_train_measurement")
Results['SMAF (training)'] = compare_results(xa, x2a)
np.save("SMAF_gene_actual_train_100", xa)
module_size_nmf = np.array([np.exp(entropy(abs(x))) for x in ua.T])
usage_nmf = np.array([np.exp(entropy(abs(x))) for x in va.T])
ua_nmf = ua
w_nmf = va
k = min(int(xa.shape[1] * 1.5), ds * 4)
k = min(k, MAX_BASIS)
UW = (np.random.random((xa.shape[0], k)), np.random.random(
(k, xa.shape[1])))
lda2 = ERROR_THRESH
while True:
U, W = smaf(xa,
k,
10,
lda2,
maxItr=10,
use_chol=True,
activity_lower=4.,
module_lower=400,
UW=UW,
donorm=True,
mode=1,
mink=5)
nz = np.nonzero(U.sum(0))[0]
U = U[:, nz]
W = W[nz]
xh_smaf = U.dot(W)
fit_smaf = 1 - np.linalg.norm(xa - xh_smaf)**2 / np.linalg.norm(xa)**2
if (len(nz) > MIN_BASIS) and (fit_smaf > MIN_FIT):
break
elif lda2 < ERROR_THRESH / 16:
break
else: