def test_FBPD_Norm1_cvx(self): if not cvx_not_installable: opt = {'memopt': True} # Problem data. m = 30 n = 20 np.random.seed(1) Amat = np.random.randn(m, n) A = LinearOperatorMatrix(Amat) bmat = np.random.randn(m) bmat.shape = (bmat.shape[0], 1) # A = Identity() # Change n to equal to m. b = DataContainer(bmat) # Regularization parameter lam = 10 opt = {'memopt': True} # Create object instances with the test data A and b. f = Norm2sq(A, b, c=0.5, memopt=True) g0 = ZeroFun() # Initial guess x_init = DataContainer(np.zeros((n, 1))) # Create 1-norm object instance g1 = Norm1(lam) # Compare to CVXPY # Construct the problem. x1 = Variable(n) objective1 = Minimize(0.5 * sum_squares(Amat * x1 - bmat.T[0]) + lam * norm(x1, 1)) prob1 = Problem(objective1) # The optimal objective is returned by prob.solve(). result1 = prob1.solve(verbose=False, solver=SCS, eps=1e-9) # The optimal solution for x is stored in x.value and optimal objective value # is in result as well as in objective.value print( "CVXPY least squares plus 1-norm solution and objective value:" ) print(x1.value) print(objective1.value) # Now try another algorithm FBPD for same problem: x_fbpd1, itfbpd1, timingfbpd1, criterfbpd1 = FBPD( x_init, Identity(), None, f, g1) print(x_fbpd1) print(criterfbpd1[-1]) self.assertNumpyArrayAlmostEqual(numpy.squeeze(x_fbpd1.array), x1.value, 6) else: self.assertTrue(cvx_not_installable)
# Construct the problem. x1 = Variable(n) objective1 = Minimize(0.5*sum_squares(Amat*x1 - bmat.T[0]) + lam*norm(x1,1) ) prob1 = Problem(objective1) # The optimal objective is returned by prob.solve(). result1 = prob1.solve(verbose=False,solver=SCS,eps=1e-9) # The optimal solution for x is stored in x.value and optimal objective value # is in result as well as in objective.value print("CVXPY least squares plus 1-norm solution and objective value:") print(x1.value) print(objective1.value) # Now try another algorithm FBPD for same problem: x_fbpd1, itfbpd1, timingfbpd1, criterfbpd1 = FBPD(x_init,Identity(), None, f, g1) print(x_fbpd1) print(criterfbpd1[-1]) # Plot criterion curve to see both FISTA and FBPD converge to same value. # Note that FISTA is very efficient for 1-norm minimization so it beats # FBPD in this test by a lot. But FBPD can handle a larger class of problems # than FISTA can. plt.figure() plt.loglog(iternum[[0,-1]],[objective1.value, objective1.value], label='CVX LS+1') plt.loglog(iternum,criter1,label='FISTA LS+1') plt.legend() plt.show() plt.figure() plt.loglog(iternum[[0,-1]],[objective1.value, objective1.value], label='CVX LS+1')
plt.imshow(x_fista1.array) plt.title('FISTA Least squares plus 1-norm regularisation') plt.show() plt.semilogy(criter1) plt.title('FISTA Least squares plus 1-norm regularisation criterion') plt.show() # The least squares plus 1-norm regularisation problem can also be solved by # other algorithms such as the Forward Backward Primal Dual algorithm. This # algorithm minimises the sum of three functions and the least squares and # 1-norm functions should be given as the second and third function inputs. # In this test case, this algorithm requires more iterations to converge, so # new options are specified. opt_FBPD = {'tol': 1e-4, 'iter': 20000} x_fbpd1, it_fbpd1, timing_fbpd1, criter_fbpd1 = FBPD(x_init, None, f, g0, opt_FBPD) plt.imshow(x_fbpd1.array) plt.title('FBPD for least squares plus 1-norm regularisation') plt.show() plt.semilogy(criter_fbpd1) plt.title('FBPD for least squares plus 1-norm regularisation criterion') plt.show() # The FBPD algorithm can also be used conveniently for TV regularisation: # Specify TV function object lamtv = 10 gtv = TV2D(lamtv)
lam = 0.1 g0 = Norm1(lam) # Run FISTA for least squares plus 1-norm function. x_fista1, it1, timing1, criter1 = FISTA(x_init, f, g0, opt=opt) plt.imshow(x_fista0.subset(horizontal_x=80).array) plt.title('FISTA LS+1') plt.colorbar() plt.show() # Run FBPD=Forward Backward Primal Dual method on least squares plus 1-norm print("Run FBPD for least squares plus 1-norm regularisation") x_fbpd1, it_fbpd1, timing_fbpd1, criter_fbpd1 = FBPD(x_init, None, f, g0, opt=opt) plt.imshow(x_fbpd1.subset(horizontal_x=80).array) plt.title('FBPD LS+1') plt.colorbar() plt.show() # Run CGLS, which should agree with the FISTA least squares print("Run CGLS for least squares") x_CGLS, it_CGLS, timing_CGLS, criter_CGLS = CGLS(x_init, Cop, padded_data2, opt=opt) plt.imshow(x_CGLS.subset(horizontal_x=80).array)
def test_FISTA_denoise_cvx(self): if not cvx_not_installable: opt = {'memopt': True} N = 64 ig = ImageGeometry(voxel_num_x=N, voxel_num_y=N) Phantom = ImageData(geometry=ig) x = Phantom.as_array() x[int(round(N / 4)):int(round(3 * N / 4)), int(round(N / 4)):int(round(3 * N / 4))] = 0.5 x[int(round(N / 8)):int(round(7 * N / 8)), int(round(3 * N / 8)):int(round(5 * N / 8))] = 1 # Identity operator for denoising I = TomoIdentity(ig) # Data and add noise y = I.direct(Phantom) y.array = y.array + 0.1 * np.random.randn(N, N) # Data fidelity term f_denoise = Norm2sq(I, y, c=0.5, memopt=True) # 1-norm regulariser lam1_denoise = 1.0 g1_denoise = Norm1(lam1_denoise) # Initial guess x_init_denoise = ImageData(np.zeros((N, N))) # Combine with least squares and solve using generic FISTA implementation x_fista1_denoise, it1_denoise, timing1_denoise, \ criter1_denoise = \ FISTA(x_init_denoise, f_denoise, g1_denoise, opt=opt) print(x_fista1_denoise) print(criter1_denoise[-1]) # Now denoise LS + 1-norm with FBPD x_fbpd1_denoise, itfbpd1_denoise, timingfbpd1_denoise,\ criterfbpd1_denoise = \ FBPD(x_init_denoise, I, None, f_denoise, g1_denoise) print(x_fbpd1_denoise) print(criterfbpd1_denoise[-1]) # Compare to CVXPY # Construct the problem. x1_denoise = Variable(N**2, 1) objective1_denoise = Minimize( 0.5 * sum_squares(x1_denoise - y.array.flatten()) + lam1_denoise * norm(x1_denoise, 1)) prob1_denoise = Problem(objective1_denoise) # The optimal objective is returned by prob.solve(). result1_denoise = prob1_denoise.solve(verbose=False, solver=SCS, eps=1e-12) # The optimal solution for x is stored in x.value and optimal objective value # is in result as well as in objective.value print( "CVXPY least squares plus 1-norm solution and objective value:" ) print(x1_denoise.value) print(objective1_denoise.value) self.assertNumpyArrayAlmostEqual(x_fista1_denoise.array.flatten(), x1_denoise.value, 5) self.assertNumpyArrayAlmostEqual(x_fbpd1_denoise.array.flatten(), x1_denoise.value, 5) x1_cvx = x1_denoise.value x1_cvx.shape = (N, N) # Now TV with FBPD lam_tv = 0.1 gtv = TV2D(lam_tv) gtv(gtv.op.direct(x_init_denoise)) opt_tv = {'tol': 1e-4, 'iter': 10000} x_fbpdtv_denoise, itfbpdtv_denoise, timingfbpdtv_denoise,\ criterfbpdtv_denoise = \ FBPD(x_init_denoise, gtv.op, None, f_denoise, gtv, opt=opt_tv) print(x_fbpdtv_denoise) print(criterfbpdtv_denoise[-1]) # Compare to CVXPY # Construct the problem. xtv_denoise = Variable((N, N)) objectivetv_denoise = Minimize(0.5 * sum_squares(xtv_denoise - y.array) + lam_tv * tv(xtv_denoise)) probtv_denoise = Problem(objectivetv_denoise) # The optimal objective is returned by prob.solve(). resulttv_denoise = probtv_denoise.solve(verbose=False, solver=SCS, eps=1e-12) # The optimal solution for x is stored in x.value and optimal objective value # is in result as well as in objective.value print( "CVXPY least squares plus 1-norm solution and objective value:" ) print(xtv_denoise.value) print(objectivetv_denoise.value) self.assertNumpyArrayAlmostEqual(x_fbpdtv_denoise.as_array(), xtv_denoise.value, 1) else: self.assertTrue(cvx_not_installable)
g1 = Norm1(lam) g1(x_init) g1.prox(x_init, 0.02) # Combine with least squares and solve using generic FISTA implementation x_fista1, it1, timing1, criter1 = FISTA(x_init, f, g1) x_fista1_m, it1_m, timing1_m, criter1_m = FISTA(x_init, f, g1, opt=opt) iternum = [i for i in range(len(criter1))] # Print for comparison print("FISTA least squares plus 1-norm solution and objective value:") print(x_fista1) print(criter1[-1]) # Now try another algorithm FBPD for same problem: x_fbpd1, itfbpd1, timingfbpd1, criterfbpd1 = FBPD(x_init, None, f, g1) iternum = [i for i in range(len(criterfbpd1))] print(x_fbpd1) print(criterfbpd1[-1]) # Plot criterion curve to see both FISTA and FBPD converge to same value. # Note that FISTA is very efficient for 1-norm minimization so it beats # FBPD in this test by a lot. But FBPD can handle a larger class of problems # than FISTA can. plt.figure() plt.loglog(iternum, criter1, label='FISTA LS+1') plt.loglog(iternum, criter1_m, label='FISTA LS+1 memopt') plt.legend() plt.show() plt.figure()
#%% #%% Then run FBPD algorithm for TV denoising # Initial guess x_init_denoise = ImageData(np.zeros((N, N))) # Set up TV function gtv = TV2D(lam_tv) # Evalutate TV of noisy image. gtv(gtv.op.direct(y)) # Specify FBPD options and run FBPD. opt_tv = {'tol': 1e-4, 'iter': 10000} x_fbpdtv_denoise, itfbpdtv_denoise, timingfbpdtv_denoise, criterfbpdtv_denoise = FBPD( x_init_denoise, None, f_denoise, gtv, opt=opt_tv) print("FBPD least squares plus TV solution and objective value:") plt.figure() plt.imshow(x_fbpdtv_denoise.as_array()) plt.title('FBPD TV with objective equal to {:.2f}'.format( criterfbpdtv_denoise[-1])) plt.show() print(criterfbpdtv_denoise[-1]) # Also plot history of criterion vs. CVX if use_cvxpy: plt.loglog([0, opt_tv['iter']], [objectivetv_denoise.value, objectivetv_denoise.value], label='CVX TV')
plt.title('FISTA Least squares plus 1-norm regularisation') plt.show() plt.semilogy(criter1) plt.title('FISTA Least squares plus 1-norm regularisation criterion') plt.show() # The least squares plus 1-norm regularisation problem can also be solved by # other algorithms such as the Forward Backward Primal Dual algorithm. This # algorithm minimises the sum of three functions and the least squares and # 1-norm functions should be given as the second and third function inputs. # In this test case, this algorithm requires more iterations to converge, so # new options are specified. x_fbpd1, it_fbpd1, timing_fbpd1, criter_fbpd1 = FBPD(x_init, None, f, g0, opt=opt) plt.imshow(x_fbpd1.subset(vertical=0).array) plt.title('FBPD for least squares plus 1-norm regularisation') plt.show() plt.semilogy(criter_fbpd1) plt.title('FBPD for least squares plus 1-norm regularisation criterion') plt.show() # Compare all reconstruction and criteria clims = (0, 1) cols = 3