def rebuild(patch, dic, lambd):
return skl
D = np.asarray([x.getdata() for x in dic]).transpose()
X = img.getdata()
alpha = optim.lasso_seq(D, X, lambd)
X_ = D.dot(alpha)
nimg = Image.new(img.mode, img.size)
nimg.putdata(X_)
nimg.save("restored.png")
def learn_dic(patches, dic_size, T=1000, lambd=0.1, batch_size=100):
"""
Learns sparse coding dictionary using the online algorithm exposed in
"J. Mairal, F. Bach, J. Ponce, and G. Sapiro. Online dictionary learning for sparse coding.
In Proceedings of the International Conference on Machine Learning (ICML), 2009."
:param patches: array of patches patches[i, j] represents the jth pixel of the ith patch
:param dic_size: size of dictionary
:param T: number of iterations
:param lambd: Lasso regularization parameter
:param batch_size: size of batches processed during each iteration
:return: learned dictionary
"""
# Extract patch size
patch_size = patches.shape[1]
# Initialize A and B at 0
A = np.zeros((dic_size, dic_size))
B = np.zeros((patch_size, dic_size))
# Initialize D at random
D = np.random.rand(patch_size, dic_size)
for t in range(T):
# Extract batch as array of column-vectors
batch_indices = np.random.choice(range(len(patches)), size=batch_size)
x = patches[batch_indices].T
# Compute sparse coding of x using lars lasso
alpha = optim.lasso_seq(D, x, lambd)
# Update A and B
A += alpha.dot(alpha.T) / batch_size
B += x.dot(alpha.T) / batch_size
# Update D
D = update_dic(D, A, B)
# Print percentage
print("\rDictionary learning: {0}".format(int((float(t + 1) / T) * 100)), "%", end="", flush=True)
return D
