gistfile1.py

k = 256  #number of clusters
c = 4   #number of subspaces
R = np.eye(m, m)
opq_codes = np.zeros((n, c))
opq_centroids = np.zeros((k, m/c, c))
iterations = 30
for it in xrange(iterations):
    print "Num iter", it
    mod_X = np.dot(X, R)
    opq_Y = np.zeros((n, m))
    for i in xrange(c):
        centroid, label = vq.kmeans2(mod_X[:, dim[i]:dim[i+1]], k, iter=1)
        opq_codes[:, i] = label
        opq_centroids[:, :, i] = centroid
        opq_Y[:, dim[i]:dim[i+1]] = centroid[label, :]
    mat_to_svd = np.dot(X.T, opq_Y)
    U, S, V = scipy.linalg.svd(mat_to_svd)
    R = np.dot(U, V)
    
    
def opq_retrieval(query):
    q = X[query,:].reshape((1, X.shape[1])).dot(R)
    distance = np.zeros((k, c))
    dist = np.zeros((n, ))
    for i in xrange(c):
        part_q = q[:, dim[i]:dim[i+1]]
        distance[:, i] = cdist(part_q, opq_centroids[:, :, i])
    for j in xrange(n):
        dist[j] = np.sum(distance[opq_codes[j, :].astype(int), :])
    indices = np.argsort(dist)
    return indices