I have a fairly large dataset for which I am calculating the SVD on then I want to get back the matrix. My matrix shape is: (33388, 104) which is a lot of columns and rows. I want an energy of 80%, which is k=51.
When I go to return my matrix in the return, I get the following error:
operands could not be broadcast together with shapes (33388,51) (51,51)
img is a numpy array of an image file
k is the number of singular vectors to use
How can I correct my function to fix this error?
def rank_k_approx(img, k):
"""Return a rank-k approximation
img: an image (as a 2D grayscale array)
k: number of singular vectors used"""
u, sigma, vt = np.linalg.svd(img)
energy = np.linalg.norm(sigma)**2
approx_energy = np.linalg.norm(sigma[:k])**2
percentage = approx_energy*100/energy
print ("Energy retained = %4.2f"%percentage)
return u[:,:k]*np.diag(sigma[:k])*vt[:k,:]
The outputs of np.linalg.svd are ndarray objects, for which the operator * is element-wise multiplication. You want to compute the matrix product, for which you need to use np.dot() or the @ operator.
The multiplication should be:
u[:, :k].dot(np.diag(sigma[:k])).dot(vt[:k])