Python models.LatentGridCRF类代码示例

def test_blocks_crf_directional():
    # test latent directional CRF on blocks
    # test that all results are the same as equivalent LatentGridCRF
    X, Y = toy.generate_blocks(n_samples=1)
    x, y = X[0], Y[0]
    pairwise_weights = np.array([0, 0, 0, -4, -4, 0, -4, -4, 0, 0])
    unary_weights = np.repeat(np.eye(2), 2, axis=0)
    w = np.hstack([unary_weights.ravel(), pairwise_weights])
    pw_directional = np.array(
        [0, 0, -4, -4, 0, 0, -4, -4, -4, -4, 0, 0, -4, -4, 0, 0, 0, 0, -4, -4, 0, 0, -4, -4, -4, -4, 0, 0, -4, -4, 0, 0]
    )
    w_directional = np.hstack([unary_weights.ravel(), pw_directional])
    crf = LatentGridCRF(n_labels=2, n_states_per_label=2)
    directional_crf = LatentDirectionalGridCRF(n_labels=2, n_states_per_label=2)
    h_hat = crf.inference(x, w)
    h_hat_d = directional_crf.inference(x, w_directional)
    assert_array_equal(h_hat, h_hat_d)

    h = crf.latent(x, y, w)
    h_d = directional_crf.latent(x, y, w_directional)
    assert_array_equal(h, h_d)

    h_hat = crf.loss_augmented_inference(x, y, w)
    h_hat_d = directional_crf.loss_augmented_inference(x, y, w_directional)
    assert_array_equal(h_hat, h_hat_d)

    psi = crf.psi(x, h_hat)
    psi_d = directional_crf.psi(x, h_hat)
    assert_array_equal(np.dot(psi, w), np.dot(psi_d, w_directional))

def test_continuous_y():
    for inference_method in ["lp", "ad3"]:
        X, Y = toy.generate_blocks(n_samples=1)
        x, y = X[0], Y[0]
        w = np.array([1, 0, 0, 1, 0, -4, 0])  # unary  # pairwise

        crf = LatentGridCRF(n_labels=2, n_states_per_label=1, inference_method=inference_method)
        psi = crf.psi(x, y)
        y_cont = np.zeros_like(x)
        gx, gy = np.indices(x.shape[:-1])
        y_cont[gx, gy, y] = 1
        # need to generate edge marginals
        vert = np.dot(y_cont[1:, :, :].reshape(-1, 2).T, y_cont[:-1, :, :].reshape(-1, 2))
        # horizontal edges
        horz = np.dot(y_cont[:, 1:, :].reshape(-1, 2).T, y_cont[:, :-1, :].reshape(-1, 2))
        pw = vert + horz

        psi_cont = crf.psi(x, (y_cont, pw))
        assert_array_almost_equal(psi, psi_cont)

        const = find_constraint(crf, x, y, w, relaxed=False)
        const_cont = find_constraint(crf, x, y, w, relaxed=True)

        # dpsi and loss are equal:
        assert_array_almost_equal(const[1], const_cont[1])
        assert_almost_equal(const[2], const_cont[2])

        # returned y_hat is one-hot version of other
        assert_array_equal(const[0], np.argmax(const_cont[0][0], axis=-1))

        # test loss:
        assert_equal(crf.loss(y, const[0]), crf.continuous_loss(y, const_cont[0][0]))

def test_with_crosses_bad_init():
    # use less perfect initialization
    rnd = np.random.RandomState(0)
    X, Y = toy.generate_crosses(n_samples=20, noise=5, n_crosses=1,
                                total_size=8)
    X_test, Y_test = X[10:], Y[10:]
    X, Y = X[:10], Y[:10]
    n_labels = 2
    crf = LatentGridCRF(n_labels=n_labels, n_states_per_label=2)
    H_init = crf.init_latent(X, Y)

    mask = rnd.uniform(size=H_init.shape) > .7
    H_init[mask] = 2 * (H_init[mask] / 2)

    one_slack = OneSlackSSVM(crf, inactive_threshold=1e-8, cache_tol=.0001,
                             inference_cache=50, max_iter=10000)
    n_slack = NSlackSSVM(crf)
    subgradient = SubgradientSSVM(crf, max_iter=150, learning_rate=.01,
                                  momentum=0)

    for base_ssvm in [one_slack, n_slack, subgradient]:
        base_ssvm.C = 10. ** 2
        clf = LatentSSVM(base_ssvm)

        clf.fit(X, Y, H_init=H_init)
        Y_pred = clf.predict(X)

        assert_array_equal(np.array(Y_pred), Y)
        # test that score is not always 1
        assert_true(.98 < clf.score(X_test, Y_test) < 1)

def test_switch_to_ad3():
    # smoketest only
    # test if switching between qpbo and ad3 works inside latent svm
    # use less perfect initialization

    if not get_installed(['qpbo']) or not get_installed(['ad3']):
        return
    X, Y = generate_crosses(n_samples=20, noise=5, n_crosses=1, total_size=8)
    X_test, Y_test = X[10:], Y[10:]
    X, Y = X[:10], Y[:10]

    crf = LatentGridCRF(n_states_per_label=2,
                        inference_method='qpbo')
    crf.initialize(X, Y)
    H_init = crf.init_latent(X, Y)

    np.random.seed(0)
    mask = np.random.uniform(size=H_init.shape) > .7
    H_init[mask] = 2 * (H_init[mask] / 2)

    base_ssvm = OneSlackSSVM(crf, inactive_threshold=1e-8, cache_tol=.0001,
                             inference_cache=50, max_iter=10000,
                             switch_to=('ad3', {'branch_and_bound': True}),
                             C=10. ** 3)
    clf = LatentSSVM(base_ssvm)

    clf.fit(X, Y, H_init=H_init)
    assert_equal(clf.model.inference_method[0], 'ad3')

    Y_pred = clf.predict(X)

    assert_array_equal(np.array(Y_pred), Y)
    # test that score is not always 1
    assert_true(.98 < clf.score(X_test, Y_test) < 1)

def test_with_crosses_bad_init():
    # use less perfect initialization
    X, Y = toy.generate_crosses(n_samples=10, noise=5, n_crosses=1,
                                total_size=8)
    n_labels = 2
    crf = LatentGridCRF(n_labels=n_labels, n_states_per_label=2,
                        inference_method='lp')
    H_init = crf.init_latent(X, Y)

    mask = np.random.uniform(size=H_init.shape) > .7
    H_init[mask] = 2 * (H_init[mask] / 2)

    one_slack = OneSlackSSVM(crf, inactive_threshold=1e-8, cache_tol=.0001,
                             inference_cache=50, max_iter=10000)
    n_slack = StructuredSVM(crf)
    subgradient = SubgradientSSVM(crf, max_iter=150, learning_rate=5)

    for base_ssvm in [one_slack, n_slack, subgradient]:
        base_ssvm.C = 10. ** 3
        base_ssvm.n_jobs = -1
        clf = LatentSSVM(base_ssvm)

        clf.fit(X, Y, H_init=H_init)
        Y_pred = clf.predict(X)

        assert_array_equal(np.array(Y_pred), Y)

def test_latent_consistency_grid():
    crf = LatentGridCRF(n_labels=2, n_features=2, n_states_per_label=2)
    for i in range(10):
        w = np.random.normal(size=18)
        y = np.random.randint(2, size=(4, 4))
        x = np.random.normal(size=(4, 4, 2))
        h = crf.latent(x, y, w)
        assert_array_equal(h // 2, y)

def test_blocks_crf_unaries():
    X, Y = toy.generate_blocks(n_samples=1)
    x, y = X[0], Y[0]
    unary_weights = np.repeat(np.eye(2), 2, axis=0)
    pairwise_weights = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
    w = np.hstack([unary_weights.ravel(), pairwise_weights])
    crf = LatentGridCRF(n_labels=2, n_states_per_label=2)
    h_hat = crf.inference(x, w)
    assert_array_equal(h_hat / 2, np.argmax(x, axis=-1))

def test_loss_augmented_inference_exhaustive_grid():
    crf = LatentGridCRF(n_labels=2, n_features=2, n_states_per_label=2)
    for i in range(10):
        w = np.random.normal(size=18)
        y = np.random.randint(2, size=(2, 2))
        x = np.random.normal(size=(2, 2, 2))
        h_hat = crf.loss_augmented_inference(x, y * 2, w)
        h = exhaustive_loss_augmented_inference(crf, x, y * 2, w)
        assert_array_equal(h, h_hat)

def test_latent_consistency_zero_pw_grid():
    crf = LatentGridCRF(n_labels=2, n_features=2, n_states_per_label=2)
    for i in xrange(10):
        w = np.zeros(18)
        w[:8] = np.random.normal(size=8)
        y = np.random.randint(2, size=(5, 5))
        x = np.random.normal(size=(5, 5, 2))
        h = crf.latent(x, y, w)
        assert_array_equal(h / 2, y)

def test_blocks_crf():
    X, Y = toy.generate_blocks(n_samples=1)
    x, y = X[0], Y[0]
    pairwise_weights = np.array([0, 0, 0, -4, -4, 0, -4, -4, 0, 0])
    unary_weights = np.repeat(np.eye(2), 2, axis=0)
    w = np.hstack([unary_weights.ravel(), pairwise_weights])
    crf = LatentGridCRF(n_labels=2, n_states_per_label=2)
    h_hat = crf.inference(x, w)
    assert_array_equal(y, h_hat / 2)

    h = crf.latent(x, y, w)
    assert_equal(crf.loss(h, h_hat), 0)

def main():
    X, Y = toy.generate_crosses(n_samples=20, noise=5, n_crosses=1,
                                total_size=8)
    X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=.5)
    n_labels = len(np.unique(Y_train))
    crf = LatentGridCRF(n_labels=n_labels, n_states_per_label=[1, 2],
                        inference_method='lp')
    #clf = LatentSSVM(model=crf, max_iter=500, C=1000., verbose=2,
                     #check_constraints=True, n_jobs=-1, break_on_bad=True,
                     #base_svm='1-slack', inference_cache=20, tol=.1)
    clf = LatentSubgradientSSVM(
        model=crf, max_iter=500, C=1000., verbose=2,
        n_jobs=-1, learning_rate=0.1, show_loss_every=10)
    clf.fit(X_train, Y_train)

    #for X_, Y_, H, name in [[X_train, Y_train, clf.H_init_, "train"],
                            #[X_test, Y_test, [None] * len(X_test), "test"]]:
    for X_, Y_, H, name in [[X_train, Y_train, [None] * len(X_test), "train"],
                            [X_test, Y_test, [None] * len(X_test), "test"]]:
        Y_pred = clf.predict(X_)
        i = 0
        loss = 0
        for x, y, h_init, y_pred in zip(X_, Y_, H, Y_pred):
            loss += np.sum(y != y_pred)
            fig, ax = plt.subplots(3, 2)
            ax[0, 0].matshow(y, vmin=0, vmax=crf.n_labels - 1)
            ax[0, 0].set_title("ground truth")
            ax[0, 1].matshow(np.argmax(x, axis=-1),
                             vmin=0, vmax=crf.n_labels - 1)
            ax[0, 1].set_title("unaries only")
            if h_init is None:
                ax[1, 0].set_visible(False)
            else:
                ax[1, 0].matshow(h_init, vmin=0, vmax=crf.n_states - 1)
                ax[1, 0].set_title("latent initial")
            ax[1, 1].matshow(crf.latent(x, y, clf.w),
                             vmin=0, vmax=crf.n_states - 1)
            ax[1, 1].set_title("latent final")
            ax[2, 0].matshow(crf.inference(x, clf.w),
                             vmin=0, vmax=crf.n_states - 1)
            ax[2, 0].set_title("prediction latent")
            ax[2, 1].matshow(y_pred,
                             vmin=0, vmax=crf.n_labels - 1)
            ax[2, 1].set_title("prediction")
            for a in ax.ravel():
                a.set_xticks(())
                a.set_yticks(())
            fig.savefig("data_%s_%03d.png" % (name, i), bbox_inches="tight")
            i += 1
        print("loss %s set: %f" % (name, loss))
    print(clf.w)

def test_switch_to_ad3():
    # smoketest only
    # test if switching between qpbo and ad3 works inside latent svm
    # use less perfect initialization

    if not get_installed(["qpbo"]) or not get_installed(["ad3"]):
        return
    X, Y = generate_crosses(n_samples=20, noise=5, n_crosses=1, total_size=8)
    X_test, Y_test = X[10:], Y[10:]
    X, Y = X[:10], Y[:10]

    crf = LatentGridCRF(n_states_per_label=2, inference_method="qpbo")
    crf.initialize(X, Y)
    H_init = crf.init_latent(X, Y)

    np.random.seed(0)
    mask = np.random.uniform(size=H_init.shape) > 0.7
    H_init[mask] = 2 * (H_init[mask] / 2)

    base_ssvm = OneSlackSSVM(
        crf,
        inactive_threshold=1e-8,
        cache_tol=0.0001,
        inference_cache=50,
        max_iter=10000,
        switch_to=("ad3", {"branch_and_bound": True}),
        C=10.0 ** 3,
    )
    clf = LatentSSVM(base_ssvm)

    # evil hackery to get rid of ad3 output
    try:
        devnull = open("/dev/null", "w")
        oldstdout_fno = os.dup(sys.stdout.fileno())
        os.dup2(devnull.fileno(), 1)
        replaced_stdout = True
    except:
        replaced_stdout = False

    clf.fit(X, Y, H_init=H_init)

    if replaced_stdout:
        os.dup2(oldstdout_fno, 1)
    assert_equal(clf.model.inference_method[0], "ad3")

    Y_pred = clf.predict(X)

    assert_array_equal(np.array(Y_pred), Y)
    # test that score is not always 1
    assert_true(0.98 < clf.score(X_test, Y_test) < 1)

def test_continuous_y():
    for inference_method in get_installed(["lp", "ad3"]):
        X, Y = generate_blocks(n_samples=1)
        x, y = X[0], Y[0]
        w = np.array([1, 0,  # unary
                      0, 1,
                      0,     # pairwise
                      -4, 0])

        crf = LatentGridCRF(n_labels=2, n_features=2, n_states_per_label=1,
                            inference_method=inference_method)
        joint_feature = crf.joint_feature(x, y)
        y_cont = np.zeros_like(x)
        gx, gy = np.indices(x.shape[:-1])
        y_cont[gx, gy, y] = 1
        # need to generate edge marginals
        vert = np.dot(y_cont[1:, :, :].reshape(-1, 2).T,
                      y_cont[:-1, :, :].reshape(-1, 2))
        # horizontal edges
        horz = np.dot(y_cont[:, 1:, :].reshape(-1, 2).T,
                      y_cont[:, :-1, :].reshape(-1, 2))
        pw = vert + horz

        joint_feature_cont = crf.joint_feature(x, (y_cont, pw))
        assert_array_almost_equal(joint_feature, joint_feature_cont, 4)

        const = find_constraint(crf, x, y, w, relaxed=False)
        const_cont = find_constraint(crf, x, y, w, relaxed=True)

        # djoint_feature and loss are equal:
        assert_array_almost_equal(const[1], const_cont[1], 4)
        assert_almost_equal(const[2], const_cont[2], 4)

        if isinstance(const_cont[0], tuple):
            # returned y_hat is one-hot version of other
            assert_array_equal(const[0], np.argmax(const_cont[0][0], axis=-1))

            # test loss:
            assert_almost_equal(crf.loss(y, const[0]),
                                crf.continuous_loss(y, const_cont[0][0]), 4)

def test_with_crosses_bad_init():
    # use less perfect initialization
    rnd = np.random.RandomState(0)
    X, Y = generate_crosses(n_samples=20, noise=5, n_crosses=1, total_size=8)
    X_test, Y_test = X[10:], Y[10:]
    X, Y = X[:10], Y[:10]
    crf = LatentGridCRF(n_states_per_label=2)
    crf.initialize(X, Y)
    H_init = crf.init_latent(X, Y)

    mask = rnd.uniform(size=H_init.shape) > 0.7
    H_init[mask] = 2 * (H_init[mask] / 2)

    one_slack_ssvm = OneSlackSSVM(crf, inactive_threshold=1e-8, cache_tol=0.0001, inference_cache=50, C=100)
    clf = LatentSSVM(one_slack_ssvm)

    clf.fit(X, Y, H_init=H_init)
    Y_pred = clf.predict(X)

    assert_array_equal(np.array(Y_pred), Y)
    # test that score is not always 1
    assert_true(0.98 < clf.score(X_test, Y_test) < 1)

def test_objective():
    # test that SubgradientLatentSSVM does the same as SubgradientSVM,
    # in particular that it has the same loss, if there are no latent states.
    X, Y = generate_blocks_multinomial(n_samples=10, noise=.3, seed=1)
    inference_method = get_installed(["qpbo", "ad3", "lp"])[0]
    n_labels = 3
    crfl = LatentGridCRF(n_labels=n_labels, n_states_per_label=1,
                         inference_method=inference_method)
    clfl = SubgradientLatentSSVM(model=crfl, max_iter=20, C=10.,
                                 learning_rate=0.001, momentum=0.98)
    crfl.initialize(X, Y)
    clfl.w = np.zeros(crfl.size_joint_feature)  # this disables random init
    clfl.fit(X, Y)

    crf = GridCRF(n_states=n_labels, inference_method=inference_method)
    clf = SubgradientSSVM(model=crf, max_iter=20, C=10., learning_rate=0.001,
                          momentum=0.98)
    clf.fit(X, Y)
    assert_array_almost_equal(clf.w, clfl.w)
    assert_almost_equal(clf.objective_curve_[-1], clfl.objective_curve_[-1])
    assert_array_equal(clf.predict(X), clfl.predict(X))
    assert_array_equal(clf.predict(X), Y)