def test_lasso():
np.random.seed(0)
print("test lasso")
##############################################
# Decomposition of a large number of signals
##############################################
# data generation
X = np.asfortranarray(np.random.normal(size=(100,100000)))
#* X=X./repmat(sqrt(sum(X.^2)),[size(X,1) 1]);
X = np.asfortranarray(X / np.tile(np.sqrt((X*X).sum(axis=0)),(X.shape[0],1)),dtype= myfloat)
D = np.asfortranarray(np.random.normal(size=(100,200)))
D = np.asfortranarray(D / np.tile(np.sqrt((D*D).sum(axis=0)),(D.shape[0],1)),dtype= myfloat)
# parameter of the optimization procedure are chosen
#param.L=20; # not more than 20 non-zeros coefficients (default: min(size(D,1),size(D,2)))
param = {
'lambda1' : 0.15, # not more than 20 non-zeros coefficients
'numThreads' : -1, # number of processors/cores to use; the default choice is -1
# and uses all the cores of the machine
'mode' : spams.PENALTY} # penalized formulation
tic = time.time()
alpha = spams.lasso(X,D = D,return_reg_path = False,**param)
tac = time.time()
t = tac - tic
print("%f signals processed per second\n" %(float(X.shape[1]) / t))
########################################
# Regularization path of a single signal
########################################
X = np.asfortranarray(np.random.normal(size=(64,1)),dtype= myfloat)
D = np.asfortranarray(np.random.normal(size=(64,10)))
D = np.asfortranarray(D / np.tile(np.sqrt((D*D).sum(axis=0)),(D.shape[0],1)),dtype= myfloat)
(alpha,path) = spams.lasso(X,D = D,return_reg_path = True,**param)
return None
def test_trainDL_Memory():
img_file = 'lena.png'
try:
img = Image.open(img_file)
except:
print("Cannot load image %s : skipping test" %img_file)
return None
I = np.array(img) / 255.
if I.ndim == 3:
A = np.asfortranarray(I.reshape((I.shape[0],I.shape[1] * I.shape[2])))
rgb = True
else:
A = np.asfortranarray(I)
rgb = False
m = 8
n = 8
X = spams.im2col_sliding(A,m,n,rgb)
X = X - np.tile(np.mean(X,0),(X.shape[0],1))
X = np.asfortranarray(X / np.tile(np.sqrt((X * X).sum(axis=0)),(X.shape[0],1)))
X = np.asfortranarray(X[:,np.arange(0,X.shape[1],10)],dtype = myfloat)
param = { 'K' : 200, # learns a dictionary with 100 elements
'lambda1' : 0.15, 'numThreads' : 4,
'iter' : 100}
############# FIRST EXPERIMENT ##################
tic = time.time()
D = spams.trainDL_Memory(X,**param)
tac = time.time()
t = tac - tic
print('time of computation for Dictionary Learning: %f' %t)
print('Evaluating cost function...')
lparam = _extract_lasso_param(param)
alpha = spams.lasso(X,D = D,**lparam)
xd = X - D * alpha
R = np.mean(0.5 * (xd * xd).sum(axis=0) + param['lambda1'] * np.abs(alpha).sum(axis=0))
print("objective function: %f" %R)
#* ? DISPLAY
############# SECOND EXPERIMENT ##################
tic = time.time()
D = spams.trainDL(X,**param)
tac = time.time()
t = tac - tic
print('time of computation for Dictionary Learning: %f' %t)
print('Evaluating cost function...')
alpha = spams.lasso(X,D = D,**lparam)
xd = X - D * alpha
R = np.mean(0.5 * (xd * xd).sum(axis=0) + param['lambda1'] * np.abs(alpha).sum(axis=0))
print("objective function: %f" %R)
#* ? DISPLAY
return None
def test_trainDL_Memory():
img_file = 'lena.png'
try:
img = Image.open(img_file)
except:
print("Cannot load image %s : skipping test" %img_file)
return None
I = np.array(img) / 255.
if I.ndim == 3:
A = np.asfortranarray(I.reshape((I.shape[0],I.shape[1] * I.shape[2])))
rgb = True
else:
A = np.asfortranarray(I)
rgb = False
m = 8;n = 8;
X = spams.im2col_sliding(A,m,n,rgb)
X = X - np.tile(np.mean(X,0),(X.shape[0],1))
X = np.asfortranarray(X / np.tile(np.sqrt((X * X).sum(axis=0)),(X.shape[0],1)))
X = np.asfortranarray(X[:,np.arange(0,X.shape[1],10)],dtype = myfloat)
param = { 'K' : 200, # learns a dictionary with 100 elements
'lambda1' : 0.15, 'numThreads' : 4,
'iter' : 100}
############# FIRST EXPERIMENT ##################
tic = time.time()
D = spams.trainDL_Memory(X,**param)
tac = time.time()
t = tac - tic
print('time of computation for Dictionary Learning: %f' %t)
print('Evaluating cost function...')
lparam = _extract_lasso_param(param)
alpha = spams.lasso(X,D = D,**lparam)
xd = X - D * alpha
R = np.mean(0.5 * (xd * xd).sum(axis=0) + param['lambda1'] * np.abs(alpha).sum(axis=0))
print("objective function: %f" %R)
#* ? DISPLAY
############# SECOND EXPERIMENT ##################
tic = time.time()
D = spams.trainDL(X,**param)
tac = time.time()
t = tac - tic
print('time of computation for Dictionary Learning: %f' %t)
print('Evaluating cost function...')
alpha = spams.lasso(X,D = D,**lparam)
xd = X - D * alpha
R = np.mean(0.5 * (xd * xd).sum(axis=0) + param['lambda1'] * np.abs(alpha).sum(axis=0))
print("objective function: %f" %R)
#* ? DISPLAY
return None
def smaf(X, d, lda1, lda2, maxItr=10, UW=None, posW=False, posU=True, use_chol=False, module_lower=500,
activity_lower=5, donorm=False, mode=1, mink=5, U0=[], U0_delta=0.1, doprint=False):
# use Cholesky when we expect a very sparse result
# this tends to happen more on the full vs subsampled matrices
if UW == None:
U, W = spams.nmf(np.asfortranarray(X), return_lasso=True, K=d, numThreads=THREADS)
W = np.asarray(W.todense())
else:
U, W = UW
Xhat = U.dot(W)
Xnorm = np.linalg.norm(X) ** 2 / X.shape[1]
for itr in range(maxItr):
if mode == 1:
# In this mode the ldas correspond to an approximate desired fit
# Higher lda will be a worse fit, but will result in a sparser sol'n
U = spams.lasso(np.asfortranarray(X.T), D=np.asfortranarray(W.T),
lambda1=lda2 * Xnorm, mode=1, numThreads=THREADS, cholesky=use_chol, pos=posU)
U = np.asarray(U.todense()).T
elif mode == 2:
if len(U0) > 0:
U = projected_grad_desc(W.T, X.T, U.T, U0.T, lda2, U0_delta, maxItr=400)
U = U.T
else:
U = spams.lasso(np.asfortranarray(X.T), D=np.asfortranarray(W.T),
lambda1=lda2, lambda2=0.0, mode=2, numThreads=THREADS, cholesky=use_chol, pos=posU)
U = np.asarray(U.todense()).T
if donorm:
U = U / np.linalg.norm(U, axis=0)
U[np.isnan(U)] = 0
if mode == 1:
wf = (1 - lda2)
W = sparse_decode(X, U, lda1, worstFit=wf, mink=mink)
elif mode == 2:
if len(U0) > 0:
W = projected_grad_desc(U, X, W, [], lda1, 0., nonneg=posW, maxItr=400)
else:
W = spams.lasso(np.asfortranarray(X), D=np.asfortranarray(U),
lambda1=lda1, lambda2=1.0, mode=2, numThreads=THREADS, cholesky=use_chol, pos=posW)
W = np.asarray(W.todense())
Xhat = U.dot(W)
module_size = np.average([np.exp(entropy(u)) for u in U.T if u.sum() > 0])
activity_size = np.average([np.exp(entropy(abs(w))) for w in W.T])
if doprint:
print distance.correlation(X.flatten(), Xhat.flatten()), module_size, activity_size, lda1, lda2
if module_size < module_lower:
lda2 /= 2.
if activity_size < activity_lower:
lda2 /= 2.
return U, W
def get_concentration(
Im,
stain_matrix,
lamda=0.01,
):
return sp.lasso(Im.T, D=stain_matrix.T, mode=2, lambda1=lamda,
pos=True).toarray().T
def fit_strf_lasso(input, output, lags, lambda1=1.0, lambda2=1.0, num_threads=10):
#convert the input into a toeplitz-like matrix
stime = time.time()
A = make_toeplitz(input, lags, include_bias=True, fortran_style=True)
etime = time.time() - stime
print '[fit_strf_lasso] Time to make Toeplitz matrix: %d seconds' % etime
fy = np.asfortranarray(output.reshape(len(output), 1))
#print 'fy.shape=',fy.shape
#print 'fA.shape=',fA.shape
#fit the STRF
stime = time.time()
fit_params = spams.lasso(fy, A, mode=2, lambda1=lambda1, lambda2=lambda2, numThreads=num_threads)
etime = time.time() - stime
print '[fit_strf_lasso] Time to fit STRF: %d seconds' % etime
#reshape the STRF so that it makes sense
nt = input.shape[0]
nf = input.shape[1]
d = len(lags)
strf = np.array(fit_params[:-1].todense()).reshape([nf, d])
bias = fit_params[-1].todense()[0, 0]
return strf,bias
def dictEval( X, D, param, lam=None, dsfactor=None, patchSize=None, patchFnGrp=None, kind='avg'):
if dsfactor is not None:
X_useme,dsz = downsamplePatchList( X, patchSize, dsfactor, kind=kind )
D_useme,Ddsz = downsamplePatchList( D, patchSize, dsfactor, kind=kind )
if patchFnGrp:
patchFnGrp.create_dataset('patchesDown', data=X_useme)
else:
X_useme = X
D_useme = D
if lam is None:
lam = param['lambda1']
alpha = spams.lasso( np.asfortranarray(X_useme), D = np.asfortranarray(D_useme), **param )
Xre = ( D * alpha )
if patchFnGrp:
patchFnGrp.create_dataset('patchesRecon', data=Xre)
xd = X - Xre
R = np.mean( (xd * xd).sum(axis=0))
if lam > 0:
print " dictEval - lambda: ", lam
R = R + lam * np.mean( np.abs(alpha).sum(axis=0))
return R
def transform(self, X):
'''
Compute the loadings of X using the learned components
Parameters
----------
X : array, shape (n_samples, n_features)
data matrix
Returns
-------
coefs : array, shape (n_samples, n_components)
transformed data
'''
coefs = spams.lasso(
# data
X=np.asfortranarray(X.T),
# dict
D=self.components_.T,
# pos
pos=True,
lambda1=0,
lambda2=0,
)
return coefs.toarray().T
def project_data(self, X, agreg=1):
"""
Projects data on the model dictionary
Parameters
----------
X : array, shape (n_samples, n_features)
Matrix to project on dictoonary
agreg : int
Specifies size of the groups to average after projection
Returns
-------
projections: array, shape(n_samples/agreg, n_components)
Projection matrix
"""
alpha_mat = spams.lasso(np.asfortranarray(np.transpose(X)),
D=np.asfortranarray(self.D),
lambda1=self.lambda1,
lambda2=self.lambda2,
mode=2,
pos=self.pos)
alpha_mat = alpha_mat.toarray()
if agreg > 1:
return np.mean(np.reshape(alpha_mat,
(alpha_mat.shape[0] / agreg, agreg, -1)),
axis=1)
else:
return alpha_mat
def fit(self, y, dirs, KERNELS, params):
nD = dirs.shape[0]
if nD > 1: # model works only with one direction
raise RuntimeError('"%s" model requires exactly 1 orientation' %
self.name)
n1 = len(self.d_perps)
n2 = len(self.d_isos)
nATOMS = n1 + n2
if nATOMS == 0: # empty dictionary
return [0, 0], None, None, None
# prepare DICTIONARY from dir and lookup tables
i1, i2 = amico.lut.dir_TO_lut_idx(dirs[0])
A = np.zeros((len(y), nATOMS), dtype=np.float64, order='F')
A[:, :(nD * n1)] = KERNELS['D'][:, i1, i2, :].T
A[:, (nD * n1):] = KERNELS['CSF'].T
# fit
x = spams.lasso(np.asfortranarray(y.reshape(-1, 1)), D=A,
**params).todense().A1
# return estimates
v = x[:n1].sum() / (x.sum() + 1e-16)
# checking that there is more than 1 isotropic compartment
if self.type == 'Mouse':
v_blood = x[n1] / (x.sum() + 1e-16)
v_csf = x[n1 + 1] / (x.sum() + 1e-16)
return [v, 1 - v, v_blood, v_csf], dirs, x, A
else:
return [v, 1 - v], dirs, x, A
def learn_sparse_coeffs(self, matrix, D, params, weighting=None):
if weighting is not None:
alphas = spams.lassoWeighted(matrix, D=D, W=weighting, **params)
else:
alphas = spams.lasso(matrix, D=D, **params)
#logging.info('Alphas shape and sparsity:\t{}\t{:.4f}'.format(alphas.shape, 100*alphas.nnz/np.prod(alphas.shape)))
return alphas
def cod2sparseLASSO(dictionary, feat, lambdaLasso=0.35, numThreads=-1):
init_time = time.time()
X_ = np.asfortranarray(feat)
D_ = np.asfortranarray(dictionary)
param = {
'lambda1': lambdaLasso, # not more than 20 non-zeros coefficients
'numThreads':
numThreads, # number of processors/cores to use; the default choice is -1
# and uses all the cores of the machine
'pos': False,
'mode': spams.PENALTY
} # penalized formulation
print "Using LASSO for sparse codification. Please wait..."
alpha = spams.lasso(X_, D_, return_reg_path=False, verbose=True, **param)
end_time = time.time()
t = end_time - init_time
print "%f signals processed per second\n" % (float(alpha.shape[1]) / t)
print "Total time: ", t, "seconds"
return alpha.todense()
def get_concentrations(self, image, stain_matrix):
"""Get concentration matrix.
Parameters
----------
image: array_like
rgb
Returns
-------
concentration: array_like
N x 2 matrix for an N x M case.
"""
OD = RGB2OD(image).reshape((-1, 3))
if self.maskout_white:
nonwhite_mask = get_nonwhite_mask(image,
self.nonwhite_threshold).reshape(
(-1, ))
OD = OD[nonwhite_mask]
coefs = (spams.lasso(OD.T,
D=stain_matrix.T,
mode=2,
lambda1=self.lambda1,
pos=True).toarray().T)
return coefs
def fit( self, y, dirs, KERNELS, params ) :
nD = dirs.shape[0]
if nD > 1 : # model works only with one direction
raise RuntimeError( '"%s" model requires exactly 1 orientation' % self.name )
n1 = len(self.d_perps)
n2 = len(self.d_isos)
nATOMS = n1+n2
if nATOMS == 0 : # empty dictionary
return [0, 0], None, None, None
# prepare DICTIONARY from dir and lookup tables
i1, i2 = amico.lut.dir_TO_lut_idx( dirs[0] )
A = np.zeros( (len(y), nATOMS), dtype=np.float64, order='F' )
A[:,:(nD*n1)] = KERNELS['D'][:,i1,i2,:].T
A[:,(nD*n1):] = KERNELS['CSF'].T
# fit
x = spams.lasso( np.asfortranarray( y.reshape(-1,1) ), D=A, **params ).todense().A1
# return estimates
v = x[ :n1 ].sum() / ( x.sum() + 1e-16 )
# checking that there is more than 1 isotropic compartment
if self.type == 'Mouse' :
v_blood = x[ n1 ] / ( x.sum() + 1e-16 )
v_csf = x[ n1+1 ] / ( x.sum() + 1e-16 )
return [ v, 1-v, v_blood, v_csf ], dirs, x, A
else :
return [ v, 1-v ], dirs, x, A
def get_concentrations(I, stain_matrix, **kwargs):
"""
Estimate concentration matrix given an image, stain matrix and relevant method parameters.
"""
OD = convert_RGB_to_OD(I).reshape((-1, 3))
lasso_regularizer = kwargs['lasso_regularizer'] if 'lasso_regularizer' in kwargs.keys() else 0.01
return spams.lasso(X=OD.T, D=stain_matrix.T, mode=2, lambda1=lasso_regularizer, pos=True).toarray().T
def get_concentrations(I, stain_matrix, lamda=0.01):
"""
Get the concentration matrix. Suppose the input image is H x W x 3 (uint8). Define Npix = H * W.
Then the concentration matrix is Npix x 2 (or we could reshape to H x W x 2).
The first element of each row is the Hematoxylin concentration.
The second element of each row is the Eosin concentration.
We do this by 'solving' OD = C*S (Matrix product) where OD is optical density (Npix x 3),\
C is concentration (Npix x 2) and S is stain matrix (2 x 3).
See docs for spams.lasso.
We restrict the concentrations to be positive and penalise very large concentration values,\
so that background pixels (which can not easily be expressed in the Hematoxylin-Eosin basis) have \
low concentration and thus appear white.
:param I: Image. A np array HxWx3 of type uint8.
:param stain_matrix: a 2x3 stain matrix. First row is Hematoxylin stain vector, second row is Eosin stain vector.
:return:
"""
# param = {
# 'lambda1': 0.15, # not more than 20 non-zeros coefficients
# 'numThreads': -1, # number of processors/cores to use, the default choice is -1
# # and uses all the cores of the machine
# 'mode': spams.PENALTY} # penalized formulation
# alpha = spams.lasso(X,D = D,return_reg_path = False,**param)
OD = mu.RGB_to_OD(I).reshape((-1, 3)) # convert to optical density and flatten to (H*W)x3.
return spams.lasso(OD.T, D=stain_matrix.T, mode=2, numThreads=6, lambda1=lamda, pos=True).toarray().T
def get_concentrations(I, stain_matrix, lamda=0.01):
"""
Get the concentration matrix. Suppose the input image is H x W x 3 (uint8). Define Npix = H * W.
Then the concentration matrix is Npix x 2 (or we could reshape to H x W x 2).
The first element of each row is the Hematoxylin concentration.
The second element of each row is the Eosin concentration.
We do this by 'solving' OD = C*S (Matrix product) where OD is optical density (Npix x 3),\
C is concentration (Npix x 2) and S is stain matrix (2 x 3).
See docs for spams.lasso.
We restrict the concentrations to be positive and penalise very large concentration values,\
so that background pixels (which can not easily be expressed in the Hematoxylin-Eosin basis) have \
low concentration and thus appear white.
:param I: Image. A np array HxWx3 of type uint8.
:param stain_matrix: a 2x3 stain matrix. First row is Hematoxylin stain vector, second row is Eosin stain vector.
:return:
"""
OD = mu.RGB_to_OD(I).reshape(
(-1, 3)) # convert to optical density and flatten to (H*W)x3.
return spams.lasso(OD.T,
D=stain_matrix.T,
mode=2,
lambda1=lamda,
pos=True).toarray().T
def coding_series(segment_list, D, lambda1):
"""Represent given series with learned dictionary
Args:
segment_list(list): each item contains m subsequences which is
sliced from original time series instance
D(numpy 2d-array): learning dictionary
lambda1(float): lambda conefficient in sparse coding,
|X-D*a|^2 + lambda*|a|^1
For more information, see spams packages:
http://spams-devel.gforge.inria.fr/doc-python/html/doc_spams005.html#sec15
Returns:
bow_data(numpy array): transformed data
"""
k = D.shape[1]
bow_data = numpy.zeros([len(segment_list), k])
# BoW for data
for index, item in enumerate(segment_list):
# Log lasso information
log_msg = "Solve lasso problem for %d series" % (index)
logger.info(log_msg)
code = spams.lasso(numpy.asfortranarray(item),
D,
lambda1=lambda1,
pos=True)
code = numpy.sum(code.todense(), axis=1)
bow_data[index:index + 1, :] += code.reshape([1, k])
div = numpy.linalg.norm(bow_data[index, :])
if div > 0:
bow_data[index, :] = bow_data[index, :] / div
# Log bow result for debug
log_msg = "%d series: " % (index)
logger.debug(log_msg + str(bow_data[index, :]))
return bow_data
def test_cd():
np.random.seed(0)
X = np.asfortranarray(np.random.normal(size = (64,100)))
X = np.asfortranarray(X / np.tile(np.sqrt((X*X).sum(axis=0)),(X.shape[0],1)),dtype=myfloat)
D = np.asfortranarray(np.random.normal(size = (64,100)))
D = np.asfortranarray(D / np.tile(np.sqrt((D*D).sum(axis=0)),(D.shape[0],1)),dtype=myfloat)
# parameter of the optimization procedure are chosen
lambda1 = 0.015
mode = spams.PENALTY
tic = time.time()
alpha = spams.lasso(X,D,lambda1 = lambda1,mode = mode,numThreads = 4)
tac = time.time()
t = tac - tic
xd = X - D * alpha
E = np.mean(0.5 * (xd * xd).sum(axis=0) + lambda1 * np.abs(alpha).sum(axis=0))
print("%f signals processed per second for LARS" %(X.shape[1] / t))
print('Objective function for LARS: %g' %E)
tol = 0.001
itermax = 1000
tic = time.time()
# A0 = ssp.csc_matrix(np.empty((alpha.shape[0],alpha.shape[1])))
A0 = ssp.csc_matrix((alpha.shape[0],alpha.shape[1]),dtype=myfloat)
alpha2 = spams.cd(X,D,A0,lambda1 = lambda1,mode = mode,tol = tol, itermax = itermax,numThreads = 4)
tac = time.time()
t = tac - tic
print("%f signals processed per second for CD" %(X.shape[1] / t))
xd = X - D * alpha2
E = np.mean(0.5 * (xd * xd).sum(axis=0) + lambda1 * np.abs(alpha).sum(axis=0))
print('Objective function for CD: %g' %E)
print('With Random Design, CD can be much faster than LARS')
return None
def _compute_codes(X, D, sc_mode, W, lambda1, lambda2,
n_jobs, pos_coef, **args):
""" Deprecated for very-large datasets! Use sparse_decode instead.
"""
X = np.asfortranarray(X)
D = np.asfortranarray(D)
gram = None
cov = None
if W is None:
gram = np.dot(D.T, D)
gram = np.asfortranarray(gram)
cov = np.dot(D.T, X)
cov = np.asfortranarray(cov)
if sc_mode in [0, 1, 2]:
if W is None:
A = spams.lasso(X , D, gram, cov, lambda1=lambda1, lambda2=lambda2,
numThreads=n_jobs, mode=sc_mode, pos=pos_coef)
else:
A = spams.lassoWeighted(X , D, W, lambda1=lambda1,mode=sc_mode,
pos=pos_coef, numThreads=n_jobs)
else:
L = lambda1 if sc_mode == 3 else None
eps = lambda1 if sc_mode == 4 else None
lambda_1 = lambda1 if sc_mode == 5 else None
A = spams.omp(X, D, L, eps, lambda_1, numThreads=n_jobs)
return A.toarray()
def fit( self, y, dirs, KERNELS, params ) :
nD = dirs.shape[0]
n1 = len(self.Rs)
n2 = len(self.ICVFs)
n3 = len(self.d_ISOs)
# prepare DICTIONARY from dirs and lookup tables
A = np.zeros( (len(y), nD*(n1+n2)+n3 ), dtype=np.float64, order='F' )
o = 0
for i in xrange(nD) :
i1, i2 = amico.lut.dir_TO_lut_idx( dirs[i] )
A[:,o:(o+n1)] = KERNELS['wmr'][:,i1,i2,:].T
o += n1
for i in xrange(nD) :
i1, i2 = amico.lut.dir_TO_lut_idx( dirs[i] )
A[:,o:(o+n2)] = KERNELS['wmh'][:,i1,i2,:].T
o += n2
A[:,o:] = KERNELS['iso'].T
# empty dictionary
if A.shape[1] == 0 :
return [0, 0, 0], None, None, None
# fit
x = spams.lasso( np.asfortranarray( y.reshape(-1,1) ), D=A, **params ).todense().A1
# return estimates
f1 = x[ :(nD*n1) ].sum()
f2 = x[ (nD*n1):(nD*(n1+n2)) ].sum()
v = f1 / ( f1 + f2 + 1e-16 )
xIC = x[:nD*n1].reshape(-1,n1).sum(axis=0)
a = 1E6 * 2.0 * np.dot(self.Rs,xIC) / ( f1 + 1e-16 )
d = (4.0*v) / ( np.pi*a**2 + 1e-16 )
return [v, a, d], dirs, x, A
def get_concentrations(I, stain_matrix, regularizer=0.01):
OD = convert_RGB_to_OD(I).reshape((-1, 3))
# 稀疏优化
return spams.lasso(X=OD.T,
D=stain_matrix.T,
mode=2,
lambda1=regularizer,
pos=True).toarray().T
def fit(self, y, dirs, KERNELS, params):
singleb0 = True if len(y) == (1 + self.scheme.dwi_count) else False
nD = dirs.shape[0]
if nD != 1:
raise RuntimeError('"%s" model requires exactly 1 orientation' %
self.name)
# prepare DICTIONARY from dir and lookup tables
nWM = len(self.IC_ODs) * len(self.IC_VFs)
nATOMS = nWM + 1
if self.isExvivo == True:
nATOMS += 1
i1, i2 = amico.lut.dir_TO_lut_idx(dirs[0])
A = np.ones((len(y), nATOMS), dtype=np.float64, order='F')
A[:, :nWM] = KERNELS['wm'][:, i1, i2, :].T
A[:, -1] = KERNELS['iso']
# estimate CSF partial volume (and isotropic restriction, if exvivo) and remove from signal
x, _ = scipy.optimize.nnls(A, y)
yy = y - x[-1] * A[:, -1]
if self.isExvivo == True:
yy = yy - x[-2] * A[:, -2]
yy[yy < 0] = 0
# estimate IC and EC compartments and promote sparsity
if singleb0:
An = A[1:, :nWM] * KERNELS['norms']
yy = yy[1:].reshape(-1, 1)
else:
An = A[self.scheme.dwi_idx, :nWM] * KERNELS['norms']
yy = yy[self.scheme.dwi_idx].reshape(-1, 1)
x = spams.lasso(np.asfortranarray(yy),
D=np.asfortranarray(An),
**params).todense().A1
# debias coefficients
x = np.append(x, 1)
if self.isExvivo == True:
x = np.append(x, 1)
idx = x > 0
x[idx], _ = scipy.optimize.nnls(A[:, idx], y)
# return estimates
xx = x / (x.sum() + 1e-16)
xWM = xx[:nWM]
fISO = xx[-1]
xWM = xWM / (xWM.sum() + 1e-16)
f1 = np.dot(KERNELS['icvf'], xWM)
f2 = np.dot((1.0 - KERNELS['icvf']), xWM)
v = f1 / (f1 + f2 + 1e-16)
k = np.dot(KERNELS['kappa'], xWM)
od = 2.0 / np.pi * np.arctan2(1.0, k)
if self.isExvivo:
return [v, od, fISO, xx[-2]], dirs, x, A
else:
return [v, od, fISO], dirs, x, A
def get_concentrations(I, stain_matrix, lamda=0.01):
"""
Get concentrations, a npix x 2 matrix
:param I:
:param stain_matrix: a 2x3 stain matrix
:return:
"""
OD = RGB_to_OD(I).reshape((-1, 3))
return spams.lasso(OD.T, D=stain_matrix.T, mode=2, lambda1=lamda, pos=True).toarray().T
def sparse_decode(Y, D, lda):
Ynorm = np.linalg.norm(Y)**2 / Y.shape[1]
W = spams.lasso(np.asfortranarray(Y),
np.asfortranarray(D),
lambda1=lda * Ynorm,
mode=1,
numThreads=THREADS,
pos=False)
W = np.asarray(W.todense())
return W
def sparse_coding_lasso(self, i, vectors):
A = np.zeros((self.K, self.K))
B = np.zeros((self.n, self.K))
inds = np.arange(i, i + self.chunk_size)
v = vectors[:, inds]
a = spams.lasso(v, D=self.D, **self.lasso_params).toarray()
for k in np.arange(len(inds)):
A += np.outer(a[:, k], a[:, k])
B += np.outer(v[:, k], a[:, k])
return A, B
def run(self, n_iter):
y = np.expand_dims(np.asfortranarray(self.y), axis=1)
if (scipy.sparse.issparse(self.X)):
W0 = np.zeros((self.X.shape[1], 1), dtype=y.dtype, order="F")
self.w = fistaFlat(y, self.X, W0, **self.solver_parameter,
regul='l1', it0=10000, loss='square', tol=1e-12,
max_it=n_iter).flatten()
else:
self.w = lasso(y, D=self.X, L=n_iter,
**self.solver_parameter).toarray().flatten()
def __call__(self, X, D):
import spams
lasso_params = {
'lambda1': self._lambda,
'lambda2': 0,
'numThreads': self.n_jobs,
'mode': 2}
return np.array(spams.lasso(np.asfortranarray(X, np.float64), D=np.asfortranarray(D, np.float64),
return_reg_path=False, **lasso_params).todense())
def predict(self, imgs, neuron_idx=None, penalty_lambda=None, algorithm=None):
""" get neuron response to images
Parameters
----------
imgs
Returns
-------
"""
imgs_array = make_2d_input_matrix(imgs)
if neuron_idx is None:
dict_to_use = self.w
else:
dict_to_use = self.w[neuron_idx:(neuron_idx + 1), :]
if penalty_lambda is None:
_lambda = self._lambda
else:
_lambda = penalty_lambda
assert np.isscalar(_lambda)
if algorithm is None:
_algorithm = self.algorithm
else:
_algorithm = algorithm
# let's call sparse encoder to do it!
# no scaling at all!
# having /nsample in objective function is exactly the same as sovling each problem separately.
# the underlying function called is elastic net, and that function fits each column of y separately.
# each column of y is each stimulus. This is because when passing imgs_array and dict_to_use to Elastic Net,
# they are transposed. That is, y = imgs_array.T
#
# in the code there's also a subtle detail, where alpha is divided by number of pixels in each stimulus.
# I haven't figured that out, but seems that's simply a detail for using ElasticNet to do this.
if _algorithm in ['lasso_lars', 'lasso_cd']:
response = sparse_encode(imgs_array, dict_to_use, alpha=_lambda, algorithm=_algorithm, max_iter=10000)
else:
assert _algorithm == 'spams'
#print(imgs_array.dtype, dict_to_use.dtype, _lambda.shape)
response = lasso(np.asfortranarray(imgs_array.T), D=np.asfortranarray(dict_to_use.T), lambda1=_lambda,
mode=2)
response = response.T.toarray() # because lasso returns sparse matrix...
# this can be used for debugging, for comparison with SPAMS.
# notice here I give per sample cost.
self.last_cost_recon = 0.5 * np.sum((imgs_array - np.dot(response, dict_to_use)) ** 2, axis=1)
self.last_cost_sparsity = _lambda * np.abs(response).sum(axis=1)
assert self.last_cost_sparsity.shape == (imgs_array.shape[0], )
assert self.last_cost_recon.shape == (imgs_array.shape[0],)
self.last_cost = np.mean(self.last_cost_recon + self.last_cost_sparsity)
return response
def forward(self, x: Tensor):
# x is B x input_size matrix to reconstruct
# init_guess is B x output_size matrix code guess
if self.solver_type == 'spams':
raise RuntimeError('spams deprecated')
from spams import lasso # intially I put this line later; and it failed.
x_cpu = x.data.cpu().numpy()
assert x_cpu.ndim == 2 and x_cpu.shape[1] == self.linear_module.out_features
# init_guess_shape = (x.shape[0], self.linear_module.in_features)
# if init_guess is not None:
# init_guess = init_guess.data.cpu().numpy()
# else:
# init_guess = np.zeros(init_guess_shape)
# assert init_guess.shape == init_guess_shape
# spams does not need any init guess.
response = lasso(np.asfortranarray(x_cpu.T),
D=np.asfortranarray(self.linear_module.weight.data.cpu().numpy()),
lambda1=self.lam / 2, # so to remove 1/2 factor for reconstruction loss.
mode=2)
response = response.T.toarray() # because lasso returns sparse matrix...
response_cost = abs(response) * self.lam
if self.size_average_l1:
response_cost = float(response_cost.mean())
else:
response_cost = float(response_cost.sum())
# then pass response into self.linear_module
response = tensor(response, dtype=torch.float32)
if x.is_cuda:
response = response.cuda()
recon = self.linear_module(response)
# return reconstruction loss, so that we can do gradient descent on weight later.
return self.cost(recon, x) + response_cost
elif self.solver_type == 'fista_custom':
# using my hand written fista code, which should reproduce the lua code exactly (up to machine precision).
xinit = torch.zeros(x.size()[0], self.linear_module.in_features)
if x.is_cuda:
xinit = xinit.cuda()
with torch.no_grad():
response, L_new = fista_ls(partial(self.f_fista, x.data), self.g_fista, self.pl, xinit,
verbose=False, L=self.L)
# i think this is kind of speed up trick.
# TODO: in FistaL1.lua (not LinearFistaL1.lua),
# there's another cap on L.
if L_new == self.L:
self.L = L_new / 2
else:
self.L = L_new
recon = self.linear_module(response)
return self.cost(recon, x) + self.g_fista(response), response
else:
raise NotImplementedError
def fit( self, y, dirs, KERNELS, params ) :
singleb0 = True if len(y) == (1+self.scheme.dwi_count) else False
nD = dirs.shape[0]
if nD != 1 :
raise RuntimeError( '"%s" model requires exactly 1 orientation' % self.name )
# prepare DICTIONARY from dir and lookup tables
nWM = len(self.IC_ODs)*len(self.IC_VFs)
nATOMS = nWM + 1
if self.isExvivo == True :
nATOMS += 1
i1, i2 = amico.lut.dir_TO_lut_idx( dirs[0] )
A = np.ones( (len(y), nATOMS), dtype=np.float64, order='F' )
A[:,:nWM] = KERNELS['wm'][:,i1,i2,:].T
A[:,-1] = KERNELS['iso']
# estimate CSF partial volume (and isotropic restriction, if exvivo) and remove from signal
x, _ = scipy.optimize.nnls( A, y )
yy = y - x[-1]*A[:,-1]
if self.isExvivo == True :
yy = yy - x[-2]*A[:,-2]
yy[ yy<0 ] = 0
# estimate IC and EC compartments and promote sparsity
if singleb0:
An = A[1:, :nWM] * KERNELS['norms']
yy = yy[1:].reshape(-1,1)
else:
An = A[ self.scheme.dwi_idx, :nWM ] * KERNELS['norms']
yy = yy[ self.scheme.dwi_idx ].reshape(-1,1)
x = spams.lasso( np.asfortranarray(yy), D=np.asfortranarray(An), **params ).todense().A1
# debias coefficients
x = np.append( x, 1 )
if self.isExvivo == True :
x = np.append( x, 1 )
idx = x>0
x[idx], _ = scipy.optimize.nnls( A[:,idx], y )
# return estimates
xx = x / ( x.sum() + 1e-16 )
xWM = xx[:nWM]
if self.isExvivo == True :
fISO = xx[-2]
else :
fISO = xx[-1]
xWM = xWM / ( xWM.sum() + 1e-16 )
f1 = np.dot( KERNELS['icvf'], xWM )
f2 = np.dot( (1.0-KERNELS['icvf']), xWM )
v = f1 / ( f1 + f2 + 1e-16 )
k = np.dot( KERNELS['kappa'], xWM )
od = 2.0/np.pi * np.arctan2(1.0,k)
return [v, od, fISO], dirs, x, A
def __call__(self, X, D):
import spams
lasso_params = {
'lambda1': self._lambda,
'lambda2': 0,
'numThreads': self.n_jobs,
'mode': 2
}
return np.array(spams.lasso(np.asfortranarray(X, np.float64), D=np.asfortranarray(D, np.float64),
return_reg_path=False, **lasso_params).todense())
def get_concentrations(I, stain_matrix, regularizer=0.01):
"""
Estimate concentration matrix given an image and stain matrix.
:param I:
:param stain_matrix:
:param regularizer:
:return:
"""
OD = convert_RGB_to_OD(I).reshape((-1, 3))
return spams.lasso(X=OD.T, D=stain_matrix.T, mode=2, lambda1=regularizer, pos=True).toarray().T
def test_lasso():
np.random.seed(0)
print("test lasso")
##############################################
# Decomposition of a large number of signals
##############################################
# data generation
X = np.asfortranarray(np.random.normal(size=(100, 100000)))
#* X=X./repmat(sqrt(sum(X.^2)),[size(X,1) 1])
X = np.asfortranarray(X / np.tile(np.sqrt((X * X).sum(axis=0)),
(X.shape[0], 1)),
dtype=myfloat)
D = np.asfortranarray(np.random.normal(size=(100, 200)))
D = np.asfortranarray(D / np.tile(np.sqrt((D * D).sum(axis=0)),
(D.shape[0], 1)),
dtype=myfloat)
# parameter of the optimization procedure are chosen
#param.L=20 # not more than 20 non-zeros coefficients (default: min(size(D,1),size(D,2)))
param = {
'lambda1': 0.15, # not more than 20 non-zeros coefficients
'numThreads':
-1, # number of processors/cores to use, the default choice is -1
# and uses all the cores of the machine
'mode': spams.PENALTY
} # penalized formulation
tic = time.time()
alpha = spams.lasso(X, D=D, return_reg_path=False, **param)
tac = time.time()
t = tac - tic
print("%f signals processed per second\n" % (float(X.shape[1]) / t))
########################################
# Regularization path of a single signal
########################################
X = np.asfortranarray(np.random.normal(size=(64, 1)), dtype=myfloat)
D = np.asfortranarray(np.random.normal(size=(64, 10)))
D = np.asfortranarray(D / np.tile(np.sqrt((D * D).sum(axis=0)),
(D.shape[0], 1)),
dtype=myfloat)
(alpha, path) = spams.lasso(X, D=D, return_reg_path=True, **param)
return None
def write_part(k):
global nz_code
print("start " + str(k))
sup = min(nz - 1, chunk_size * (k + 1))
print("compute matrix")
inds = global_inds[chunk_size * k:sup]
size_of_matrix = sup - chunk_size * k
print(size_of_matrix)
abundance_matrix = np.zeros((n, size_of_matrix), dtype=np.float32)
for i in range(n0, n + n0):
chunk = np.zeros(size_of_matrix)
inds_of_values = (nzis[i] >= global_inds[chunk_size * k]) * (
nzis[i] < global_inds[sup])
inds_absolute = nzis[i][inds_of_values]
chunk[inverted_index[inds_absolute] -
inverted_index[global_inds[chunk_size *
k]]] = values[i][inds_of_values]
abundance_matrix[i - n0, :] = chunk / norm[inds]
print("matrix computed")
print(abundance_matrix)
x = np.asfortranarray(abundance_matrix)
a = spams.lasso(x, D=D, **lparam)
nz_chunk = np.shape(a.data)[0]
indices[nz_code:(nz_chunk + nz_code)] = a.indices[:]
indptr[chunk_size * k:(sup + 1)] = a.indptr[:] + indptr[chunk_size * k]
data[nz_code:(nz_chunk + nz_code)] = a.data[:]
print(nz_code, nz_chunk)
nz_code = nz_code + nz_chunk
a = a.toarray()
print(np.sum(a > 0, 0))
clusters = np.argsort(a, axis=0)[-clusters_nb:][::-1]
mask = a[clusters, np.arange(size_of_matrix)]
mask = mask > thres
clusters[~mask] = -1
alpha[:clusters_nb, inds] = clusters + 1
alpha.flush()
print(str(i) + " ok !")
return 0
def active_support_elastic_net(X, y, alpha, tau=1.0, algorithm='spams', support_init='knn',
support_size=100, maxiter=40):
n_samples = X.shape[0]
if n_samples <= support_size: # skip active support search for small scale data
supp = np.arange(n_samples, dtype=int) # this results in the following iteration to converge in 1 iteration
else:
if support_init == 'L2':
L2sol = np.linalg.solve(np.identity(y.shape[1]) * alpha + np.dot(X.T, X), y.T)
c0 = np.dot(X, L2sol)[:, 0]
supp = np.argpartition(-np.abs(c0), support_size)[0:support_size]
elif support_init == 'knn':
supp = np.argpartition(-np.abs(np.dot(y, X.T)[0]), support_size)[0:support_size]
curr_obj = float("inf")
for _ in range(maxiter):
Xs = X[supp, :]
if algorithm == 'spams':
cs = spams.lasso(np.asfortranarray(y.T), D=np.asfortranarray(Xs.T),
lambda1=tau*alpha, lambda2=(1.0-tau)*alpha)
cs = np.asarray(cs.todense()).T
else:
cs = sparse_encode(y, Xs, algorithm=algorithm, alpha=alpha)
delta = (y - np.dot(cs, Xs)) / alpha
obj = tau * np.sum(np.abs(cs[0])) + (1.0 - tau)/2.0 * np.sum(np.power(cs[0], 2.0)) + alpha/2.0 * np.sum(np.power(delta, 2.0))
if curr_obj - obj < 1.0e-10 * curr_obj:
break
curr_obj = obj
coherence = np.abs(np.dot(delta, X.T))[0]
coherence[supp] = 0
addedsupp = np.nonzero(coherence > tau + 1.0e-10)[0]
if addedsupp.size == 0: # converged
break
# Find the set of nonzero entries of cs.
activesupp = supp[np.abs(cs[0]) > 1.0e-10]
if activesupp.size > 0.8 * support_size: # this suggests that support_size is too small and needs to be increased
support_size = min([round(max([activesupp.size, support_size]) * 1.1), n_samples])
if addedsupp.size + activesupp.size > support_size:
ord = np.argpartition(-coherence[addedsupp], support_size - activesupp.size)[0:support_size - activesupp.size]
addedsupp = addedsupp[ord]
supp = np.concatenate([activesupp, addedsupp])
c = np.zeros(n_samples)
c[supp] = cs
return c
def lasso_nn(Y, D):
# lambda1 controls sparsity, numThreads=-1 if you want to use all cores
params = {
'lambda1': 0.007,
'numThreads': -1,
'mode': spams.PENALTY,
'pos': True
}
alpha = spams.lasso(np.asfortranarray(Y, dtype=np.float32),
D=np.asfortranarray(D, dtype=np.float32),
**params)
return alpha.todense()
def elastic_net_subspace_clustering(X, gamma=50.0, gamma_nz=True, tau=1.0, algorithm='lasso_lars',
active_support=True, active_support_params=None, n_nonzero=50):
if algorithm in ('lasso_lars', 'lasso_cd') and tau < 1.0 - 1.0e-10:
warnings.warn('algorithm {} cannot handle tau smaller than 1. Using tau = 1'.format(algorithm))
tau = 1.0
if active_support == True and active_support_params == None:
active_support_params = {}
n_samples = X.shape[0]
rows = np.zeros(n_samples * n_nonzero)
cols = np.zeros(n_samples * n_nonzero)
vals = np.zeros(n_samples * n_nonzero)
curr_pos = 0
for i in progressbar.progressbar(range(n_samples)):
y = X[i, :].copy().reshape(1, -1)
X[i, :] = 0
if algorithm in ('lasso_lars', 'lasso_cd', 'spams'):
if gamma_nz == True:
coh = np.delete(np.absolute(np.dot(X, y.T)), i)
alpha0 = np.amax(coh) / tau # value for which the solution is zero
alpha = alpha0 / gamma
else:
alpha = 1.0 / gamma
if active_support == True:
c = active_support_elastic_net(X, y, alpha, tau, algorithm, **active_support_params)
else:
if algorithm == 'spams':
c = spams.lasso(np.asfortranarray(y.T), D=np.asfortranarray(X.T),
lambda1=tau * alpha, lambda2=(1.0-tau) * alpha)
c = np.asarray(c.todense()).T[0]
else:
c = sparse_encode(y, X, algorithm=algorithm, alpha=alpha)[0]
else:
warnings.warn("algorithm {} not found".format(algorithm))
index = np.flatnonzero(c)
if index.size > n_nonzero:
# warnings.warn("The number of nonzero entries in sparse subspace clustering exceeds n_nonzero")
index = index[np.argsort(-np.absolute(c[index]))[0:n_nonzero]]
rows[curr_pos:curr_pos + len(index)] = i
cols[curr_pos:curr_pos + len(index)] = index
vals[curr_pos:curr_pos + len(index)] = c[index]
curr_pos += len(index)
X[i, :] = y
# affinity = sparse.csr_matrix((vals, (rows, cols)), shape=(n_samples, n_samples)) + sparse.csr_matrix((vals, (cols, rows)), shape=(n_samples, n_samples))
return sparse.csr_matrix((vals, (rows, cols)), shape=(n_samples, n_samples))
def coding_series(self, segment_list, D, a=None, batch=False, iter=-5):
k = D.shape[1]
bow_data = numpy.zeros([len(segment_list), k])
# BoW for data
for index, item in enumerate(segment_list):
code = spams.lasso(numpy.asfortranarray(item), D, lambda1=a, pos=True)
code = numpy.sum(code.todense(), axis=1)
bow_data[index : index + 1, :] += code.reshape([1, k])
div = numpy.linalg.norm(bow_data[index, :])
if div > 0:
bow_data[index, :] = bow_data[index, :] / div
return bow_data
def fit(self, x, y):
self.coef_ = (
spams.lasso(
np.asfortranarray(y, dtype=np.float),
D=np.asfortranarray(x, dtype=np.float),
pos=self.positive,
lambda1=self.lambda_,
**self.params
)
.toarray()
.flatten()
)
return self
def sparseCoding(X):
X = np.asfortranarray(X)
param = { 'K' : NCLUSTER, # size of the dictionary
'lambda1' : 0.15,
#'posD' : True, # dictionary positive constrain
#'modeD' : 1, # L1 regulization regularization on D
'iter' : ITER} # runtime limit 15mins
D = spams.trainDL_Memory(X,**param)
lparam = _extract_lasso_param(param)
print 'genrating codes...'
alpha = spams.lasso(X,D = D,**lparam)
return D, alpha
def findSparseRep(A, y, epsilon):
# A: data in columns
# y: image to match
# epsilon: allowed error
yCol = np.asfortranarray(np.reshape(y, (len(y), 1)), dtype=float)
A = np.asfortranarray(np.array(A), dtype=float)
param = {'lambda1' : epsilon,
'numThreads' : 4,
'pos' : True,
'mode' : 1}
x = sp.lasso(yCol, D=A, **param).toarray()
residuals = [np.linalg.norm(A[:,i]*x[i] - y) for i in range(0,np.shape(A)[1])]
return (x, residuals)
def sparse_decode(analysis_coefficients,
analysis_operator,
lasso_scaling_factor=0.1,
ridge_sparsity=None,
minimum_loss=1.,
minimum_ridge_sparsity=0,
method='omp'):
"""
"""
print(analysis_coefficients.shape)
print(analysis_operator.shape)
analysis_coefficients_fortran = np.asfortranarray(analysis_coefficients)
analysis_operator_fortran = np.asfortranarray(analysis_operator)
coefficient_norm = np.linalg.norm(analysis_coefficients)
num_bases, dimensionality = analysis_coefficients.shape
_, latent_dimensionality = analysis_operator.shape
if method == 'omp':
if not ridge_sparsity:
ridge_sparsity = min(num_bases, latent_dimensionality)
while ridge_sparsity < minimum_ridge_sparsity:
sparse_synthesis_coefficients = spams.omp(
analysis_coefficients_fortran,
analysis_operator_fortran,
L=ridge_sparsity,
numThreads=4)
synthesis_coefficients = np.asarray(
sparse_synthesis_coefficients.todense())
loss = 1 - np.linalg.norm(analysis_coefficients - analysis_operator
.dot(W))**2 / coefficient_norm**2
if loss < minimium_loss:
break
k -= 1
elif method == 'lasso':
print("Using lasso")
lasso_penalty = lasso_scaling_factor * coefficient_norm**2 / dimensionality
print("Lasso penalty is ", lasso_penalty)
# TODO: set numThreads to some reasonable amount based on mp.cpu_count() and parallelization level
sparse_synthesis_coefficients = spams.lasso(
analysis_coefficients_fortran,
analysis_operator_fortran,
lambda1=lasso_penalty,
mode=1,
numThreads=4,
pos=False)
synthesis_coefficients = np.asarray(
sparse_synthesis_coefficients.todense())
print(synthesis_coefficients.sum())
return synthesis_coefficients
def solve(self, A, y, x0, as_signs):
#print 'dense_solver: y.shape=',y.shape
#print 'dense_solver: A.shape=',A.shape
fy = np.asfortranarray(y.reshape(len(y), 1))
fA = np.asfortranarray(A)
#print 'fy.shape=',fy.shape
#print 'fA.shape=',fA.shape
xnew = lasso(fy, fA, mode=2, lambda1=self.lambda1, lambda2=self.lambda2)
xnew = np.array(xnew.todense()).reshape(x0.shape)
#print 'dense_solver: xnew.shape=',xnew.shape
return xnew
def generateCode(X,D):
X = np.asfortranarray(X)
print X.shape
D = np.asfortranarray(D)
print D.shape
param = { 'K' : NCLUSTER, # size of the dictionary
'lambda1' : 0.05,
#'posD' : True, # dictionary positive constrain
#'modeD' : 1, # L1 regulization regularization on D
'iter' : ITER} # runtime limit 15mins
lparam = _extract_lasso_param(param)
print 'genrating codes...'
alpha = spams.lasso(X,D = D,**lparam)
return alpha
def _objective(X,D,param,imgname = None):
print('Evaluating cost function...')
lparam = _extract_lasso_param(param)
alpha = spams.lasso(X,D = D,**lparam)
# NB : as alpha is sparse, D*alpha is the dot product
xd = X - D * alpha
R = np.mean(0.5 * (xd * xd).sum(axis=0) + param['lambda1'] * np.abs(alpha).sum(axis=0))
print("objective function: %f" %R)
#* display ?
if imgname != None:
img = spams.displayPatches(D)
print("IMG %s" %str(img.shape))
x = np.uint8(img[:,:,0] * 255.)
image = Image.fromarray(x,mode = 'L')
image.save("%s.png" %imgname)
def bow(trajectory, D, a, w_len, interval=1):
k = D.shape[1]
histogram = numpy.zeros([len(trajectory), k])
for index, item in enumerate(trajectory):
temp = list()
stamp_index = range(0, item.shape[0]-w_len+1, interval)
for i in stamp_index:
temp.append(item[i:i+w_len, :1])
temp = numpy.hstack(temp)
code = spams.lasso(numpy.asfortranarray(temp), D, lambda1=a, pos=True)
code = numpy.sum(code.todense(), axis=1)
histogram[index:index+1, :] += code.reshape([1, k])
div = numpy.linalg.norm(histogram[index, :])
if div > 0:
histogram[index, :] = histogram[index, :] / div
return histogram
def doFeatureEncoding(self, features):
""" do feature encoding to original features"""
encodedFeatures = None
whitenedFeatures = whiten(features)
if self._featureEncodingMethod == 'vector-quantization':
# Vector quantization
# each row is a feature vector
index, _ = vq(whitenedFeatures, self._codebook)
row, _ = features.shape
col = config.codebookSize
encodedFeatures = np.zeros((row, col))
for i in xrange(len(index)):
encodedFeatures[i, index[i]] = 1
elif self._featureEncodingMethod == 'sparse-coding':
# Sparse coding
# each column is a feature vector
X = np.asfortranarray(whitenedFeatures.transpose())
X = np.asfortranarray(X / np.tile(np.sqrt((X*X).sum(axis=0)),
(X.shape[0],1)),
dtype= X.dtype)
D = np.asfortranarray(self._codebook.transpose())
D = np.asfortranarray(D / np.tile(np.sqrt((D*D).sum(axis=0)),
(D.shape[0],1)),
dtype = D.dtype)
# Parameters of the optimization are chosen
param = {
'lambda1': 0.15,
'numThreads': -1,
'mode': 0
}
alpha = spams.lasso(X, D, **param) # alpha is sparse matrix
alphaShape = (D.shape[1], X.shape[1])
denseMatrix = csc_matrix(alpha, shape = alphaShape).todense()
encodedFeatures = np.asarray(denseMatrix).transpose()
return encodedFeatures
def solve(self, A, y, x0, as_signs):
#print 'dense_solver: y.shape=',y.shape
#print 'dense_solver: A.shape=',A.shape
fy = np.asfortranarray(y.reshape(len(y), 1))
fA = np.asfortranarray(A)
#print 'fy.shape=',fy.shape
#print 'fA.shape=',fA.shape
if not self.positive:
xnew = spams.lasso(fy, fA, mode=2, lambda1=self.lambda1, lambda2=self.lambda2)
else:
W = np.ones(len(x0))
params = {'lambda1':self.lambda1, 'pos':True}
xnew = spams.lassoWeighted(fy, fA, W, **params)
xnew = np.array(xnew.todense()).reshape(x0.shape)
#print 'dense_solver: xnew.shape=',xnew.shape
return xnew
outputPredictOtherAlgo = open(sys.argv[len(sys.argv)-2],'w')
#Compare Output
outputCompare = open(sys.argv[len(sys.argv)-1],'w')
#Caculate
#args for Lasso
alpha1Lambda = 1
compareLambda = 0
alpha_lasso_m1_Ds = []
#a = spams.lasso(X,Ds[i],return_reg_path = False,lambda1 = alpha1Lambda,pos=True,mode=0)
for i in range(numOfDicts):
tic = time.time()
alpha_lasso_m1_Ds.append(spams.lasso(X,Ds[i],return_reg_path = False,lambda1 = alpha1Lambda,pos=True,mode=0))
# alpha_lasso_D = spams.lasso(X,Ds[i],return_reg_path = False,lambda1 = alpha1Lambda,pos=True,mode=0)
# alpha_lasso_D = spams.lasso(X,Ds[i],return_reg_path = False,lambda1 = compareLambda,pos=True,mode=1)
# alpha_lasso_m1_Ds.append(spams.lasso(X,Ds[i],return_reg_path = False,lambda1 = compareLambda,pos=True,mode=1))
tac = time.time()
t = tac - tic
# print 'alpha_lasoo_m1_D'+str(i)+':'
# print 'time:'+str(t)
#print alpha_lasso_m1_D1.getcol(0)
#print alpha_lasso_m1_Ds[0].getcol(0)
#save each Dict's Info
allInfos = []
#weight2Page in Dict
def reChooseDict(instNo,currDict,currMean,weightWindowDsCurDict,currStdCompare,currMeanCompare,dicts,testDWindow,numAttrs,numInsts,outputCompare,Lambda):
outputCompare.write('Re-Choose Dictionary!\n')
outputPredictSparse.write('Re-Choose Dictionary!\n')
weightWindowDsRe = []
# tmpsRe = []
tmpsRe = 0
# splitByEnterDsRe = []
splitByEnterDsRe = 0
numDicts = len(dicts)
for i in range(numDicts):
weightWindowDsRe.append(deque())
dictWeightMeanRe = {}
testWindow = aL.arffLoader()
testD = testWindow.fortranArrayPara(testDWindow,numAttrs,numInsts)
testD = np.asfortranarray(testD / np.tile(np.sqrt((testD*testD).sum(axis=0)),(testD.shape[0],1)))
for dictNo in range(numDicts):
# tmpsRe[:] = []
# splitByEnterDsRe[:] = []
if(dictNo==currDict):
weightWindowDsRe[dictNo] = weightWindowDsCurDict
continue
alpha_lasso_m1_Ds_batch = spams.lasso(testD,dicts[dictNo],return_reg_path = False,lambda1 = 1,pos=True,mode=0)
for j in range(alpha_lasso_m1_Ds_batch.shape[1]):
# tmpsRe.append(str(alpha_lasso_m1_Ds_batch.getcol(j)))
tmpsRe = str(alpha_lasso_m1_Ds_batch.getcol(j))
#instNo+j才是正確的instance Number
outputCompare.write(str(instNo-len(testDWindow)+j)+'-D'+str(dictNo)+':'+tmpsRe+'\n\n')
#split
#print tmpsRe[i].split('\n')
# splitByEnterDsRe.append(tmpsRe[j].split('\n'))
splitByEnterDsRe = tmpsRe.split('\n')
# splitByEnterD1 = tmp1.split('\n')
weightTmp = []
# for line in splitByEnterDsRe[j]:
for line in splitByEnterDsRe:
line = line.strip()
#mapping page
pageNo = int(line.split(',')[0].split('(')[1].strip())
#weight of mapping page
weight = float(line.split(',')[1].split(')')[1].strip())
if weight < Lambda + 0.02:
weightTmp.append(weight)
# if weight >= 1:
# print 'InstNo:'+str(instNo+j)+', DictNo:'+str(dictNo)+', page:'+str(pageNo)+', weight:'+str(weight)
maxWeight = max(weightTmp)
weightTmp[:] = []
#one page bug if weight = 0.0 transform to 1
if maxWeight==0:
maxWeight = 1
if maxWeight > 1:
maxWeight = 1
weightWindowDsRe[dictNo].append(maxWeight)
#choose dict
for dictNo in range(numOfDicts):
#若是目前的Dict則直接給值
if(dictNo==currDict):
dictWeightMeanRe[dictNo] = currMean
#若不是則要重新計算一次
else:
dictWeightMeanRe[dictNo] = np.mean(weightWindowDsRe[dictNo])
#找出值最大的Dict，以及平均值
maxWeightDict,meanCompareRe = max(dictWeightMeanRe.iteritems(), key=lambda x:x[1])
if(maxWeightDict==currDict):
#更新currMean
meanCompareRe = currMean
#保留舊的currMeanCompare
# meanCompareRe = currMeanCompare
#更新stdCompare
#保留舊的stdCompare
# stdCompareRe = currStdCompare
#更新stdCompare
stdCompareRe = np.std(weightWindowDsRe[maxWeightDict])
# weightWindowDsRe[maxWeightDict] = weightWindowDsCurDict
outputCompare.write('Keep same model'+str(maxWeightDict)+'!\n')
outputPredictSparse.write('Keep same model'+str(maxWeightDict)+'!\n')
else:
# changeModelRe = 1
stdCompareRe = np.std(weightWindowDsRe[maxWeightDict])
outputCompare.write('Change model to model-' + str(maxWeightDict) +',meanCompare:'+str(meanCompareRe)+',stdCompare:'+str(stdCompareRe)+'!\n')
outputPredictSparse.write('Change model to model-' + str(maxWeightDict) +',meanCompare:'+str(meanCompareRe)+',stdCompare:'+str(stdCompareRe)+'!\n')
return (maxWeightDict,meanCompareRe,stdCompareRe,weightWindowDsRe[maxWeightDict])
os.listdir(path)
except Exception as e:
os.mkdir(path)
test_cases = 5
DwR = []
for i in range(test_cases):
ratings = sc.parallelize(XOrged)
model = pmr.ALS.train(ratings, K, nonnegative=True)
D = np.array(model.userFeatures().sortByKey().map(lambda (w,fs): fs).collect())
lparam = {'lambda1': Lambda,
'pos': True,
'mode': 2,
'numThreads': -1}
alpha = spams.lasso(X, np.asfortranarray(D), **lparam)
R = np.mean(0.5 * sum((X - D * alpha) ** 2) + Lambda * sum(abs(alpha)))
DwR += [(D, R)]
dwr = sc.parallelize(DwR)
def findBest(E1, E2):
if E1[1] > E2[1]:
return E2
return E1
(D, R) = dwr.reduce(findBest)
sio.savemat(path + "bestMat", {'Dbest': D, "R": R})
toc = time.time()
print "time passed "+str(toc-tic)+" s."
lparam = dict( tmp )
toc = time.time()
t = toc - tic
print 'time to load dictionary: %f' % t
print "D shape",D.shape
print "X shape",X.shape
#if args.objective_function:
#if True:
if False:
print "computing objective function"
tic = time.time()
alpha = spams.lasso( X, D = D, **lparam )
xd = X - D * alpha
R = np.mean(0.5 * (xd * xd).sum(axis=0) + params['lambda1'] * np.abs(alpha).sum(axis=0))
toc = time.time()
print " objective function value: %f" % R
t = toc - tic
print 'time of computation for objective function: %f' % t
if doUpsamp:
print "doing upsampling evaluation"
tic = time.time()
# dsz - size of downsampled image patch
X_ds,dsz = evaluation.downsamplePatchList( X, patchSize, dsfactor )
X_ds = np.asfortranarray( X_ds )
# print X
if(len(testDataWindow)==numAlgoWindow):
testDataWindow.popleft()
testDataWindow.append(currentTestData)
instanceNo = instNo
tmps[:] = []
splitByEnterDs[:] = []
#Initial in slidingWindow decide the initial Dictionary
#Caculate
if(instNo < numAlgoWindow):
for dictNo in range(numOfDicts):
# alpha_lasso_m1_Ds = spams.lasso(X,Ds[dictNo],return_reg_path = False,lambda1 = alpha1Lambda,pos=True,mode=0)
# alpha_lasso_m1_Ds = spams.lasso(X,Ds[dictNo],return_reg_path = False,lambda1 = compareLambda,pos=True,mode=1)
alpha_lasso_m1_Ds = spams.lasso(X,Ds[dictNo],return_reg_path = False,lambda1 = 1,pos=True,mode=0)
# alpha_lasso_m2_Ds = spams.lasso(X,Ds[dictNo],return_reg_path = False,lambda1 = 0,pos=True,mode=1)
# alpha_lasso_m3_Ds = spams.lasso(X,Ds[dictNo],return_reg_path = False,lambda1 = 0.5,lambda2 = 1,pos=True,mode=2)
#spams.omp
# alpha_omp_m1_Ds = spams.omp(X,Ds[dictNo],L=2,return_reg_path = False,numThreads = -1)
# alpha_omp_m2_Ds = spams.omp(X,Ds[dictNo],eps= 0.9,return_reg_path = False,numThreads = -1)
# alpha_omp_m3_Ds = spams.omp(X,Ds[dictNo],lambda1=0.4,return_reg_path = False,numThreads = -1)
# alpha_lasso_m1_Ds.max()
# alpha_lasso_m1_Ds.append(spams.lasso(X,Ds[dictNo],return_reg_path = False,lambda1 = compareLambda,pos=True,mode=1))
tmps.append(str(alpha_lasso_m1_Ds.getcol(0)))
# tmps.append(str(alpha_omp_m1_Ds.getcol(0)))
outputCompare.write(str(instNo)+'-D'+str(dictNo)+':'+tmps[dictNo]+'\n\n')
def _encode(self, X, A):
S = lasso(X, A, return_reg_path = False, **self.spams_param).todense()
#print "LASSO nnz: %d" % count_nonzero(S)
return S
X = np.asfortranarray(X / np.tile(np.sqrt((X * X).sum(axis=0)),(X.shape[0],1)),dtype = myfloat)
param = { 'K' : 100, # learns a dictionary with 100 elements
'lambda1' : 0.15, 'numThreads' : 4, 'batchsize' : 400,
'iter' : 10}
paramL = {'lambda1' : 0.15, 'numThreads' : 4}
########## FIRST EXPERIMENT ###########
tic = time.time()
D = spams.trainDL(X,**param)
tac = time.time()
t = tac - tic
print 'time of computation for Dictionary Learning: %f' %t
print "DTYPE %s" %str(D.dtype)
#param['approx'] = 0
print 'Evaluating cost function...'
alpha = spams.lasso(X,D,**paramL)
print "XX X %s, D %s, alpha %s" %(str(X.shape),str(D.shape),str(alpha.shape))
y = X
if(alpha.shape[1] > 1000):
alpha = alpha[:,0:1000]
y = X[:,0:1000]
#Da = spams.calcXAt(D,ssp.csc_matrix(alpha.T))
a = alpha.todense()
print "XXa %s" %str(a.shape)
Da = np.dot(D,a)
#Da = D * alpha
xd = y - Da
print "YY D %s Da %s y %s %s alpah %s xd %s %s" %(str(D.shape),str(Da.shape),str(y.shape),type(y),str(alpha.shape),str(xd.shape),type(xd))
#R = np.mean(0.5 * (xd * xd).sum(axis=0) + param['lambda1'] * np.abs(alpha).sum(axis=0))
#R = 0.5 * (xd * xd).sum(axis=0)
R = np.multiply(xd,xd)
# for each fixed dictionary K, we will repeat dictionary
# learning for 100 times, each with a different initial value
test_cases = 10
R = np.zeros((test_cases, ))
for i in xrange(0, test_cases):
if randomStart == 1:
D0 = util.dictLearnInit(X, K, 'random', 0)
param['D'] = np.asfortranarray(D0)
lparam = {'lambda1': Lambda,
'pos': True,
'mode': 2,
'numThreads': -1
}
D = spams.trainDL(X, **param)
alpha = spams.lasso(X, D, **lparam)
R[i] = np.mean(0.5 * sum((X - D * alpha) ** 2) + param['lambda1'] * sum(abs(alpha)))
print R[i]
#(permInd, cost) = util.dissimilarityDict(Dtemplate, D, 'euclidean')
#D = D[:, permInd] # permute the columns of D to match the template Dtemplate
if i >= 1:
if R[i] < Rbest:
Dbest = D
Rbest = R[i]
else:
Dbest = D
Rbest = R[0]
# print path
def _processer(data, mask, variance, block_size, overlap, param_alpha, param_D, dtype=np.float64, n_iter=10, gamma=3., tau=1.):
# data, mask, variance, block_size, overlap, param_alpha, param_D, dtype, n_iter = arglist
# gamma = 3.
# tau = 1.
orig_shape = data.shape
mask_array = im2col_nd(mask, block_size[:3], overlap[:3])
train_idx = np.sum(mask_array, axis=0) > mask_array.shape[0]/2
# If mask is empty, return a bunch of zeros as blocks
if not np.any(train_idx):
return np.zeros_like(data)
X = im2col_nd(data, block_size, overlap)
var_mat = np.median(im2col_nd(variance[..., 0:orig_shape[-1]], block_size, overlap)[:, train_idx], axis=0).astype(dtype)
X_full_shape = X.shape
X = X[:, train_idx]
param_alpha['L'] = int(0.5 * X.shape[0])
D = param_alpha['D']
alpha = lil_matrix((D.shape[1], X.shape[1]))
W = np.ones(alpha.shape, dtype=dtype, order='F')
DtD = np.dot(D.T, D)
DtX = np.dot(D.T, X)
DtXW = np.empty_like(DtX, order='F')
alpha_old = np.ones(alpha.shape, dtype=dtype)
has_converged = np.zeros(alpha.shape[1], dtype=np.bool)
xi = np.random.randn(X.shape[0], X.shape[1]) * var_mat
eps = np.max(np.abs(np.dot(D.T, xi)), axis=0)
param_alpha['mode'] = 1
param_alpha['pos'] = True
for _ in range(n_iter):
not_converged = np.equal(has_converged, False)
DtXW[:, not_converged] = DtX[:, not_converged] / W[:, not_converged]
for i in range(alpha.shape[1]):
if not has_converged[i]:
param_alpha['lambda1'] = var_mat[i] * (X.shape[0] + gamma * np.sqrt(2 * X.shape[0]))
DtDW = (1. / W[..., None, i]) * DtD * (1. / W[:, i])
alpha[:, i:i+1] = spams.lasso(X[:, i:i+1], Q=np.asfortranarray(DtDW), q=DtXW[:, i:i+1], **param_alpha)
arr = alpha.toarray()
nonzero_ind = arr != 0
arr[nonzero_ind] /= W[nonzero_ind]
has_converged = np.max(np.abs(alpha_old - arr), axis=0) < 1e-5
if np.all(has_converged):
break
alpha_old = arr
W[:] = 1. / (np.abs(alpha_old**tau) + eps)
# compute_weights(alpha_old, alpha, W, tau, eps)
# alpha = arr
# X = D.dot(alpha)
# X = sparse_dot(D,alpha)
X = np.dot(D, arr)
weigths = np.ones(X_full_shape[1], dtype=dtype, order='F')
weigths[train_idx] = 1. / (alpha.getnnz(axis=0) + 1.)
X2 = np.zeros(X_full_shape, dtype=dtype, order='F')
X2[:, train_idx] = X
return col2im_nd(X2, block_size, orig_shape, overlap, weigths)
print np.tile(np.sqrt((D1*D1).sum(axis=0)),(D1.shape[0],1)) D1 = np.asfortranarray(D1 / np.tile(np.sqrt((D1*D1).sum(axis=0)),(D1.shape[0],1))) print 'newD:' print D1 X=np.array([1.0,4,7])#,1,2,3])#,2.0,5.0,8.0]) X = X.reshape(3,1, order='F') #X = X.reshape(3,2) print 'X:' print X #X = np.asfortranarray(X / np.tile(np.sqrt((X*X).sum(axis=0)),(X.shape[0],1))) print 'newX:' print X #for i in range(3): # X[i][0]=i*3+1 #for i in range(3): # X[i][1]=i*3+2 tic = time.time() #alpha = spams.lasso(X,D1,return_reg_path = False,lambda1 = 2,pos=True,mode=0) alpha = spams.lasso(X,D1,return_reg_path = False,lambda1 = 0.1,pos=True,mode=1,verbose=True) #alpha = spams.omp(X,D1,L=2,lambda1 = None,return_reg_path = False,numThreads = -1) #alpha = spams.omp(X,D1,L=None,eps= 0,lambda1 = None,return_reg_path = False,numThreads = -1) tac = time.time() t = tac - tic print 'time:'+str(t) print 'alpha:' print alpha