Esempio n. 1
0
def gfl_cvx_2():
    '''   Graph Fused Lasso   '''
    numNode = 100
    data = np.zeros((numNode, 2))
    for k in range(11):
        data[k] = data[k] + np.array([1, 1])
        data[k + 11] = data[k + 11] + np.array([-1, 1])
        data[k + 22] = data[k + 22] + np.array([2, 2])
        data[k + 33] = data[k + 33] + np.array([-1, -1])

    np.random.seed(1)
    noise = 0.5 * np.random.normal(0, 1, (numNode, 2))
    data = data + noise

    edgelist = [[k, k + 1] for k in range(numNode - 1)]

    graph = nx.Graph()
    graph.add_edges_from(edgelist)
    coloredge(graph)
    graph0, graph1 = decompose_graph(graph)

    runtime = []
    objerror = []
    rho_set = [2**k for k in range(7)]
    # rho_set = [1]

    _, obj, _, _ = solve_GroupFL(y=data,
                                 G=graph,
                                 G0=graph0,
                                 G1=graph1,
                                 maxsteps=3000,
                                 rho=8,
                                 lam=0.1)
    optobj = obj[-1]

    for rho in rho_set:
        clock_time = time.time()
        _, tevolution, obj, _ = solve_GroupFL_pool(y=data,
                                                   G=graph,
                                                   G0=graph0,
                                                   G1=graph1,
                                                   maxsteps=250,
                                                   rho=rho,
                                                   lam=0.1)
        print('GraphLasso_pool:', time.time() - clock_time, ',rho:', rho)
        objerror.append([i - optobj for i in obj])
        runtime.append(tevolution)

    # plotting
    plt.figure()
    plt.plot(runtime[0], objerror[0], label=r'$\rho=2^0$')
    plt.plot(runtime[1], objerror[1], label=r'$\rho=2^1$')
    plt.plot(runtime[2], objerror[2], label=r'$\rho=2^2$')
    plt.plot(runtime[3], objerror[3], label=r'$\rho=2^3$')
    plt.plot(runtime[4], objerror[4], label=r'$\rho=2^4$')
    plt.plot(runtime[5], objerror[5], label=r'$\rho=2^5$')
    plt.plot(runtime[6], objerror[6], label=r'$\rho=2^6$')
    plt.legend(loc='upper right')
    plt.xlim((0, 50))
    plt.ylim((1e-5, 1e2))
    # plt.title('Graph Fused Lasso Algorithm-cvxpy (lambda=0.1)')
    plt.title(r'Graph Fused Lasso Algorithm '
              r'(cvx, $\lambda=0.1$)',
              fontsize=12)
    plt.xlabel('Running time (seconds)', fontsize=12)
    plt.ylabel('Objective Value Error', fontsize=12)
    plt.yscale('log')
    plt.savefig('Figures/ChainGraph_cvx/GraphFL_lamb=01_chain_graph_cvx',
                bbox_inches='tight')
Esempio n. 2
0
def gfl_explicit_1():
    ''' group fused lasso algorithm with explicit updates '''
    # img = mpimg.imread('Data/ABC.png')
    # img = np.ones((10,10,3))
    img = np.zeros((10, 10, 3))
    for i in range(10):
        for j in range(10):
            if (i - 4)**2 + (j - 4)**2 < 5:
                img[i][j] = np.array([0.5, 0.5, 0.5])

    imgsize1, imgsize2, _ = img.shape
    # img=img.astype(float)/255
    noise = 0.2 * np.random.normal(0, 1, (imgsize1, imgsize2, 3))
    imgnoise = img + noise
    data = np.reshape(imgnoise, (-1, 3))

    nodes = imgsize1 * imgsize2
    # graph settings

    graph = nx.grid_2d_graph(imgsize1, imgsize2)
    graph = nx.convert_node_labels_to_integers(graph)

    temp_edges = []
    for k in range(0, imgsize1 * imgsize2):
        for j in graph.neighbors(k):
            if j == k + 1:
                temp_edges.append([k, j])

    graph0_edges = []
    graph0_edges.append(temp_edges[0])
    for edge in temp_edges:
        if (edge[0] != graph0_edges[-1][1]) and (edge != graph0_edges[-1]):
            graph0_edges.append(edge)

    graph1_edges = [
        edge for edge in list(graph.edges()) if edge not in graph0_edges
    ]
    graph1_edges.sort()

    graph0_nodes = np.unique(np.array(graph0_edges))
    graph1_nodes = np.unique(np.array(graph1_edges))

    graph0 = nx.Graph()
    graph1 = nx.Graph()
    graph0.add_nodes_from(graph0_nodes)
    graph1.add_nodes_from(graph1_nodes)
    graph0.add_edges_from(graph0_edges)
    graph1.add_edges_from(graph1_edges)

    runtime = []
    objerror = []
    rho_set = [2**k for k in range(7)]
    # rho_set = [32]

    _, obj, _, _ = solve_GroupFL(y=data,
                                 G=graph,
                                 G0=graph0,
                                 G1=graph1,
                                 maxsteps=10000,
                                 rho=4,
                                 lam=0.5)
    optobj = obj[-1]

    for rho in rho_set:
        clock_time = time.time()
        _, obj, _, tevolution = solve_GroupFL(y=data,
                                              G=graph,
                                              G0=graph0,
                                              G1=graph1,
                                              maxsteps=3000,
                                              rho=rho,
                                              lam=0.5)
        print('running time GraphLasso_explicit:',
              time.time() - clock_time, ',rho:', rho)
        objerror.append([i - optobj for i in obj])
        runtime.append(tevolution)

    # plotting
    plt.figure()
    plt.plot(runtime[0], objerror[0], label=r'$\rho=2^0$')
    plt.plot(runtime[1], objerror[1], label=r'$\rho=2^1$')
    plt.plot(runtime[2], objerror[2], label=r'$\rho=2^2$')
    plt.plot(runtime[3], objerror[3], label=r'$\rho=2^3$')
    plt.plot(runtime[4], objerror[4], label=r'$\rho=2^4$')
    plt.plot(runtime[5], objerror[5], label=r'$\rho=2^5$')
    plt.plot(runtime[6], objerror[6], label=r'$\rho=2^6$')
    plt.legend(loc='upper right')
    plt.xlim((0, 20))
    plt.ylim((1e-12, 1e1))
    # plt.title('Graph Fused Lasso Algorithm (explict)')
    plt.title(r'Graph Fused Lasso Algorithm '
              r'(explicit, $\lambda=0.5$)',
              fontsize=12)
    plt.xlabel('Running time (seconds)', fontsize=12)
    # plt.xlabel('Number of Iterations')
    plt.ylabel('Objective Value Error', fontsize=12)
    plt.yscale('log')
    plt.savefig('Figures/GridGraph/GraphFL_lamb=05_grid_graph_explicit',
                bbox_inches='tight')
Esempio n. 3
0
def gfl_cvx_1():
    ''' group fused lasso algorithm with explicit updates '''
    # img = mpimg.imread('Data/ABC.png')
    img = np.zeros((10, 10, 3))
    for i in range(10):
        for j in range(10):
            if (i - 4)**2 + (j - 4)**2 < 5:
                img[i][j] = np.array([0.5, 0.5, 0.5])

    imgsize1, imgsize2, _ = img.shape
    # img=img.astype(float)/255
    np.random.seed(1)
    noise = 0.2 * np.random.normal(0, 1, (imgsize1, imgsize2, 3))
    imgnoise = img + noise
    data = np.reshape(imgnoise, (-1, 3))

    nodes = imgsize1 * imgsize2
    # graph settings

    graph = nx.grid_2d_graph(imgsize1, imgsize2)
    graph = nx.convert_node_labels_to_integers(graph)

    graph0_edges = []
    for k in range(0, nodes, 2):
        for j in graph.neighbors(k):
            if j == k + 1:
                graph0_edges.append([k, j])

    graph1_edges = [edge for edge in list(
        graph.edges()) if edge not in graph0_edges]

    graph1_edges.sort()

    # print(graph1_edges)

    # return

    graph0_nodes = np.unique(np.array(graph0_edges))
    graph1_nodes = np.unique(np.array(graph1_edges))

    graph0 = nx.Graph()
    graph1 = nx.Graph()
    graph0.add_nodes_from(graph0_nodes)
    graph1.add_nodes_from(graph1_nodes)
    graph0.add_edges_from(graph0_edges)
    graph1.add_edges_from(graph1_edges)

    runtime = []
    objerror = []
    rho_set = [2**k for k in range(7)]
    # rho_set = [2]

    _, obj, _, _ = solve_GroupFL(
        y=data, G=graph, G0=graph0, G1=graph1, maxsteps=4500, rho=8, lam=0.5)

    optobj = obj[-1]
    print(optobj)

    # with open('Figures/GridGraph_cvx/optobj.csv', newline='') as csvfile:
    #     spamreader = csv.reader(csvfile, delimiter=' ', quotechar='|')
    #     for value in spamreader:
    #         obj = value

    for rho in rho_set:
        clock_time = time.time()
        x_hat, tevolution, obj, _ = solve_GroupFL_pool(
            y=data, G=graph, G0=graph0, G1=graph1, maxsteps=800, rho=rho, lam=0.5)
        print('GraphLasso_cvx:',
              time.time() - clock_time, ',rho:', rho)
        objerror.append([i - optobj for i in obj])
        # objerror.append([i - obj[-1] for i in obj])
        runtime.append(tevolution)
        # print(obj)
        # print('obj:',obj)
        # print('time:',runtime)

    # print('obj_cvx:', obj)

    # img_recovered = np.reshape(x_hat, (imgsize1, imgsize2, 3))
    #
    # fig = plt.figure()
    # fig.add_subplot(1, 3, 1)
    # plt.imshow(img)
    # fig.add_subplot(1, 3, 2)
    # plt.imshow(imgnoise)
    # fig.add_subplot(1, 3, 3)
    # plt.imshow(img_recovered)
    # plt.savefig(
    #     'Figures/GridGraph_cvx/img', bbox_inches='tight')

    # plotting
    plt.figure()
    plt.plot(runtime[0], objerror[0], label=r'$\rho=2^0$')
    plt.plot(runtime[1], objerror[1], label=r'$\rho=2^1$')
    plt.plot(runtime[2], objerror[2], label=r'$\rho=2^2$')
    plt.plot(runtime[3], objerror[3], label=r'$\rho=2^3$')
    plt.plot(runtime[4], objerror[4], label=r'$\rho=2^4$')
    plt.plot(runtime[5], objerror[5], label=r'$\rho=2^5$')
    plt.plot(runtime[6], objerror[6], label=r'$\rho=2^6$')
    plt.legend(loc='upper right')
    plt.xlim((0, 120))
    plt.ylim((1e-6, 1e1))
    # plt.title('Graph Fused Lasso Algorithm (cvx)')
    plt.title(r'Graph Fused Lasso Algorithm ' r'(cvx, $\lambda=0.5$)', fontsize=12)
    plt.xlabel('Running time (seconds)', fontsize=12)
    # plt.xlabel('Number of Iterations')
    plt.ylabel('Objective Value Error', fontsize=12)
    plt.yscale('log')
    plt.savefig(
        'Figures/GridGraph_cvx/GraphFL_lamb=05_grid_graph_cvx', bbox_inches='tight')
Esempio n. 4
0
def obj_find():
    ''' find optimal objective vlaue '''

    # img = mpimg.imread('Data/ABC.png')
    # imgsize1, imgsize2, _ = img.shape
    img = np.ones(16, 16, 3)

    np.random.seed(1)
    noise = 0.2 * np.random.normal(0, 1, (imgsize1, imgsize2, 3))
    imgnoise = img + noise
    data = np.reshape(imgnoise, (-1, 3))

    nodes = imgsize1 * imgsize2
    # graph settings

    graph = nx.grid_2d_graph(imgsize1, imgsize2)
    graph = nx.convert_node_labels_to_integers(graph)
    # print(list(graph.edges()))

    temp_edges = []
    for k in range(0, imgsize1 * imgsize2):
        for j in graph.neighbors(k):
            if j == k + 1:
                temp_edges.append([k, j])

    graph0_edges = []
    graph0_edges.append(temp_edges[0])
    for edge in temp_edges:
        if (edge[0] != graph0_edges[-1][1]) and (edge != graph0_edges[-1]):
            graph0_edges.append(edge)

    graph_edges = list(graph.edges())
    graph_edges.sort()
    graph1_edges = [edge for edge in list(
        graph.edges()) if edge not in graph0_edges]

    graph1_edges.sort()

    graph0_nodes = np.unique(np.array(graph0_edges))
    graph1_nodes = np.unique(np.array(graph1_edges))

    graph = nx.Graph()
    graph0 = nx.Graph()
    graph1 = nx.Graph()
    graph0.add_nodes_from(graph0_nodes)
    graph1.add_nodes_from(graph1_nodes)
    graph0.add_edges_from(graph0_edges)
    graph1.add_edges_from(graph1_edges)
    graph.add_nodes_from(range(imgsize1 * imgsize2))
    graph.add_edges_from(graph_edges)

    # print(list(graph.edges()))

    _, obj, _, _ = solve_GroupFL(
        y=data, G=graph, G0=graph0, G1=graph1, maxsteps=1000, rho=2, lam=10)

    optobj = obj[-1]
    print(obj)

    _, obj, _, _ = solve_NetworkLasso(
        y=data, G=graph, maxsteps=1000, rho=2, lam=10)
    optobj = obj[-1]
    print(obj)

    return
    plt.switch_backend('TkAgg')
    plt.figure()
    plt.plot(obj - obj[-1])
    # plt.ylim((1e-4, 1e3))
    plt.title('Graph Fused Lasso Algorithm (cvx)')
    plt.ylabel('error')
    plt.yscale('log')
    plt.show()

    with open('Figures/GridGraph_cvx/optobj.csv', 'w', newline='') as csvfile:
        objwriter = csv.writer(csvfile, delimiter=' ',
                               quotechar='|', quoting=csv.QUOTE_MINIMAL)
        objwriter.writerow([optobj] + ['lambda:'] + [10])
Esempio n. 5
0
def gfl_cvx_1():
    ''' group fused lasso algorithm with explicit updates '''
    img = np.random.normal(0, 1, (10, 10, 3))

    imgsize1, imgsize2, _ = img.shape

    # img=img.astype(float)/255
    np.random.seed(1)
    noise = 0.2 * np.random.normal(0, 1, (imgsize1, imgsize2, 3))
    imgnoise = img + noise
    data = np.reshape(imgnoise, (-1, 3))

    nodes = imgsize1 * imgsize2
    # graph settings

    graph = nx.grid_2d_graph(imgsize1, imgsize2)
    graph = nx.convert_node_labels_to_integers(graph)

    graph0_edges = []
    for k in range(0, nodes, 2):
        for j in graph.neighbors(k):
            if j == k + 1:
                graph0_edges.append([k, j])

    graph1_edges = [
        edge for edge in list(graph.edges()) if edge not in graph0_edges
    ]

    graph1_edges.sort()

    graph0_nodes = np.unique(np.array(graph0_edges))
    graph1_nodes = np.unique(np.array(graph1_edges))

    graph0 = nx.Graph()
    graph1 = nx.Graph()
    graph0.add_nodes_from(graph0_nodes)
    graph1.add_nodes_from(graph1_nodes)
    graph0.add_edges_from(graph0_edges)
    graph1.add_edges_from(graph1_edges)

    runtime = []
    objerror = []
    # rho_set = [2**k for k in range(7)]
    rho_set = [1]

    x_hat1, obj, _, _ = solve_GroupFL(y=data,
                                      G=graph,
                                      G0=graph0,
                                      G1=graph1,
                                      maxsteps=50,
                                      rho=1,
                                      lam=10)
    optobj = obj[-1]
    print("obj_exp:", obj)

    # lam = 10
    # objtemp = (np.linalg.norm(x_hat1 - data))**2
    #
    # for node1, node2 in graph.edges():
    #     objtemp = objtemp + lam * np.linalg.norm(x_hat1[node1] - x_hat1[node2])
    #
    # print(objtemp)

    obj = []

    for rho in rho_set:
        clock_time = time.time()
        x_hat2, tevolution, obj, _ = solve_GroupFL_pool(y=data,
                                                        G=graph,
                                                        G0=graph0,
                                                        G1=graph1,
                                                        maxsteps=50,
                                                        rho=rho,
                                                        lam=10)
        print('GraphLasso_cvx:', time.time() - clock_time, ',rho:', rho)
        objerror.append([i - optobj for i in obj])
        runtime.append(tevolution)
        # print('x_cvx:',x_hat)

        # lam = 10
        # objtemp = (np.linalg.norm(x_hat2 - data))**2
        #
        # for node1, node2 in graph.edges():
        #     objtemp = objtemp + lam * \
        #         np.linalg.norm(x_hat2[node1] - x_hat2[node2])
        #
        # print(objtemp)
        # print("x1-x2", x_hat1 - x_hat2)

    print('obj_cvx:', obj)
Esempio n. 6
0
def gfl_cvx_2():
    '''   Graph Fused Lasso   '''
    numNode = 100
    data = np.zeros((numNode, 2))
    for k in range(11):
        data[k] = data[k] + np.array([1, 1])
        data[k + 11] = data[k + 11] + np.array([-1, 1])
        data[k + 22] = data[k + 22] + np.array([2, 2])
        data[k + 33] = data[k + 33] + np.array([-1, -1])

    np.random.seed(1)
    noise = 0.5 * np.random.normal(0, 1, (numNode, 2))
    data = data + noise

    graph = nx.grid_2d_graph(10, 10)
    graph = nx.convert_node_labels_to_integers(graph)

    print('graph:', list(graph.edges()))

    temp_edges = []
    for k in range(0, numNode):
        for j in graph.neighbors(k):
            if j == k + 1:
                temp_edges.append([k, j])

    graph0_edges = []
    graph0_edges.append(temp_edges[0])
    for edge in temp_edges:
        if (edge[0] != graph0_edges[-1][1]) and (edge != graph0_edges[-1]):
            graph0_edges.append(edge)

    graph1_edges = [
        edge for edge in list(graph.edges()) if edge not in graph0_edges
    ]

    graph1_edges.sort()

    graph0_nodes = np.unique(np.array(graph0_edges))
    graph1_nodes = np.unique(np.array(graph1_edges))

    graph0 = nx.Graph()
    graph1 = nx.Graph()
    graph0.add_nodes_from(graph0_nodes)
    graph1.add_nodes_from(graph1_nodes)
    graph0.add_edges_from(graph0_edges)
    graph1.add_edges_from(graph1_edges)

    print("graph0:", graph0_edges, "graph1:", graph1_edges)

    # edgelist = [[k, k + 1] for k in range(numNode - 1)]

    # graph = nx.Graph()
    # graph.add_edges_from(edgelist)
    # coloredge(graph)
    # graph0, graph1 = decompose_graph(graph)

    runtime = []
    objerror = []
    # rho_set = [2**k for k in range(7)]
    rho_set = [1]

    x_hat1, obj1, _, _ = solve_GroupFL(y=data,
                                       G=graph,
                                       G0=graph0,
                                       G1=graph1,
                                       maxsteps=20,
                                       rho=1,
                                       lam=1)
    optobj = obj1[-1]
    print(obj1[:20])

    for rho in rho_set:
        clock_time = time.time()
        x_hat2, tevolution, obj, _ = solve_GroupFL_pool(y=data,
                                                        G=graph,
                                                        G0=graph0,
                                                        G1=graph1,
                                                        maxsteps=20,
                                                        rho=rho,
                                                        lam=1)
        print('GraphLasso_pool:', time.time() - clock_time, ',rho:', rho)
        objerror.append([i - optobj for i in obj])
        runtime.append(tevolution)
        print(obj[:20])
        print((x_hat1 - x_hat2)[:10])
        print([obj1[i] - obj[i] for i in range(len(obj))])

    # plotting
    plt.switch_backend('TkAgg')
    plt.figure()
    plt.plot(runtime[0], objerror[0], label=r'$\rho=2^0$')
    # plt.plot(runtime[1], objerror[1], label=r'$\rho=2^1$')
    # plt.plot(runtime[2], objerror[2], label=r'$\rho=2^2$')
    # plt.plot(runtime[3], objerror[3], label=r'$\rho=2^3$')
    # plt.plot(runtime[4], objerror[4], label=r'$\rho=2^4$')
    # plt.plot(runtime[5], objerror[5], label=r'$\rho=2^5$')
    # plt.plot(runtime[6], objerror[6], label=r'$\rho=2^6$')
    plt.legend()
    plt.xlim((0, 50))
    plt.ylim((1e-5, 1e2))
    plt.title('Graph Fused Lasso Algorithm (cvxpy)')
    plt.xlabel('Running time (seconds)', fontsize=12)
    plt.ylabel('Objective Value Error', fontsize=12)
    plt.yscale('log')