def test_numbaprog(self):
coeffs = np.zeros(shape=(self.num,self.num))
#Test for correct outputs for simple data
dictionary = np.diag(np.ones(self.num))
stimuli = np.diag(np.ones(self.num))
#Change dtype and enforce Fortran ordering
dictionary = np.array(dictionary,dtype=np.float32,order='F')
stimuli = np.array(stimuli,dtype=np.float32,order='F')
s,u,thresh = lcag.infer(dictionary,
stimuli,
self.eta,
self.lamb,
self.nIter,
self.adapt)
assert np.allclose(s,np.diag(np.ones(self.num)))
assert np.allclose(u,np.diag(np.ones(self.num)))
#Test on random data
dictionary = self.rng.randn(self.numDict,self.dataSize)
dictionary = np.sqrt(np.diag(1/np.diag(dictionary.dot(dictionary.T)))).dot(dictionary)
stimuli = self.rng.randn(self.numStim,self.dataSize)
#Change dtype and enforce Fortran ordering
dictionary = np.array(dictionary,dtype=np.float32,order='F')
stimuli = np.array(stimuli,dtype=np.float32,order='F')
s,u,thresh = lcag.infer(dictionary,
stimuli,
self.eta,
self.lamb,
self.nIter,
self.adapt)
assert np.allclose(stimuli,s.dot(dictionary),atol=1e-5)
def sparse_lca(basis=None,coeff=None,batch=50,LR=0.05,lamb=0.05,eta=.01,iter=10,InferIter=300,adapt=0.99,patchdim=32):
indoor_file_list = load_list('indoor_file_list')
outdoor_file_list = load_list('outdoor_file_list')
if basis is None and coeff is None:
basis = np.random.randn(patchdim**2,(patchdim**2)*4)
coeff = np.random.randn(2*batch,(patchdim**2)*4)
elif coeff is None:
coeff = np.random.randn(2*batch,(patchdim**2)*4)
elif basis is None:
basis = np.random.randn(patchdim**2,(patchdim**2)*4)
# to get the image-data from the hdf5 file:
#Let's create a file handle
try:
h = tables.open_file('/media/mudigonda/Gondor/Data/scene-sparse/places32.h5','r')
except:
print('Could not open file handle')
return 1
#Modify this to get a batch of images
#Generate two random numbers corresponding to the len(list)-batchsize
for ii in range(iter):
indoor_idx=random.randint(0,len(indoor_file_list)-batch-1)
outdoor_idx=random.randint(0,len(outdoor_file_list)-batch-1)
print(indoor_idx,outdoor_idx)
print('The range of the batch is then')
print(indoor_idx+batch,outdoor_idx+batch)
data = np.zeros([patchdim**2,2*batch])
counter=-1
for idx in range(batch):
image_indoor = h.root._f_get_child(indoor_file_list[indoor_idx+idx])[:]
image_outdoor = h.root._f_get_child(outdoor_file_list[outdoor_idx+idx])[:]
counter +=1
data[:,counter] = image_indoor.flatten()
counter +=1
data[:,counter] = image_outdoor.flatten()
#Infer Coefficients (on updated dictionary)
[coeff,u,thresh] = lcanumb.infer(basis.T,data.T,eta,lamb,InferIter,adapt)
if np.any(np.isnan(coeff)):
print('coeff has nans')
pdb.set_trace()
if np.linalg.norm(coeff)==0:
print('No active coeff, also a doo doo')
pdb.set_trace()
#Update basis (dictionary)
[basis,residual]=update_basis(basis,coeff,LR,data)
if np.any(np.isnan(basis)):
print('basis has nans')
pdb.set_trace()
if np.linalg.norm(basis)==0:
print('all basis be zero. that is a doo doo')
pdb.set_trace()
#For every x 1000 iterations, save the basis
print('Residual Error, Norm is ', np.linalg.norm(residual))
#Close the file handle
h.close()
return 1
def test_numbaprog(self):
coeffs = np.zeros(shape=(self.num, self.num))
#Test for correct outputs for simple data
dictionary = np.diag(np.ones(self.num))
stimuli = np.diag(np.ones(self.num))
#Change dtype and enforce Fortran ordering
dictionary = np.array(dictionary, dtype=np.float32, order='F')
stimuli = np.array(stimuli, dtype=np.float32, order='F')
s, u, thresh = lcag.infer(dictionary, stimuli, self.eta, self.lamb,
self.nIter, self.adapt)
assert np.allclose(s, np.diag(np.ones(self.num)))
assert np.allclose(u, np.diag(np.ones(self.num)))
#Test on random data
dictionary = self.rng.randn(self.numDict, self.dataSize)
dictionary = np.sqrt(np.diag(
1 / np.diag(dictionary.dot(dictionary.T)))).dot(dictionary)
stimuli = self.rng.randn(self.numStim, self.dataSize)
#Change dtype and enforce Fortran ordering
dictionary = np.array(dictionary, dtype=np.float32, order='F')
stimuli = np.array(stimuli, dtype=np.float32, order='F')
s, u, thresh = lcag.infer(dictionary, stimuli, self.eta, self.lamb,
self.nIter, self.adapt)
assert np.allclose(stimuli, s.dot(dictionary), atol=1e-5)