Exemplo n.º 1
0
def test_prediction():
    test_data = np.repeat(np.array(
        [[-1, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0], [-1, 1, 1, 1, 1, 1, 1, 1],
         [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]],
        dtype=int),
                          3,
                          axis=0)
    s = sieve.Sieve(max_layers=5, verbose=verbose, seed=seed).fit(test_data)
    assert len(s.layers) == 1, \
        'Only one layer is needed. TC and remainder at level 0: %f, %f' % (s.tcs[0], s.remainders[0])  # TODO: Could fix this with more versatile Pred(xi | f(y))
    #assert np.allclose(s.transform(z)[0][0, :, :-1], 0., atol=1e-4), \
    #    'Residual info should be small. Largest value was: %f' % np.max(np.absolute(s.transform(z)[0][0, :, :-1]))
    assert np.allclose(s.labels[:, 0], test_data[:, -1]) or np.allclose(s.labels[:,0], 1 - test_data[:, -1]), \
        'Check that labels are correct.'
    xbar, labels = s.transform(test_data)
    assert np.allclose(np.where(test_data >= 0,
                                s.invert(xbar) - test_data, 0),
                       0,
                       atol=0.01), "Invert should be near perfect"
    print s.predict2(labels)[:, 0]
    print s.predict2(labels)[:, 1]
    assert np.allclose(
        s.predict2(labels)[:, 0], test_data[:, -1], atol=0.15
    ), "Prediction should be close"  # Interesting reason for discrepancy here...
Exemplo n.º 2
0
def test_k_max():
    np.random.seed(seed)
    n, ns = 3, 300
    x_count_a = np.random.randint(0, 5, (ns, 1)) + np.random.randint(
        0, 1, (ns, n + 1))  # NOISELESS
    x_count_b = np.random.randint(0, 5,
                                  (ns, 1)) + np.random.randint(0, 1, (ns, n))
    x_count = np.hstack([x_count_a, x_count_b])

    out = sieve.Sieve(max_layers=3,
                      dim_hidden=5,
                      verbose=verbose,
                      seed=seed,
                      k_max=8,
                      n_repeat=10).fit(x_count)
    print len(out.layers)
    print x_count[:10]
    print out.layers[0].labels[:10]
    print out.layers[1].labels[:10]
    print 'r0 at l0', out.layers[0].remainders[0].pz_xy
    print 'r5 at l0', out.layers[0].remainders[5].pz_xy
    print[
        'tc: %0.3f (-) %0.3f (+) %0.3f' % (layer.corex.tc, layer.lb, layer.ub)
        for layer in out.layers
    ]
    assert len(
        out.layers) == 2, "2 latent factors of cardinality 5 should be enough."
    xbar, labels = out.transform(x_count)
    print 'pred', out.invert(xbar)[:10]
    print 'true', x_count[:10]
    assert np.allclose(out.invert(xbar), x_count)
Exemplo n.º 3
0
def test_structure():
    # Gets stuck for some seeds.
    test_data = np.repeat(np.array(
        [[0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 1, 1, 1],
         [1, 1, 1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1]],
        dtype=int),
                          3,
                          axis=0)
    s = sieve.Sieve(max_layers=5, verbose=verbose, seed=seed,
                    n_repeat=5).fit(test_data)
    assert len(s.layers) == 2, 'Only two latent factors required.'
    assert np.all(
        np.argmax(s.mis, axis=0)[:-1] == np.array(
            [0, 0, 0, 0, 0, 1, 1, 1])), 'Correct structure has two groups.'
    assert np.allclose(s.mis[:, -1],
                       0), 'Latent factors should not be correlated.'
Exemplo n.º 4
0
def test_sieve():
    test_data = np.repeat(np.array(
        [[0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1]], dtype=int),
                          10,
                          axis=0)
    s = sieve.Sieve(max_layers=5, verbose=verbose, seed=seed).fit(test_data)
    assert len(s.layers) == 1, \
        'Only one layer is needed. TC and remainder at level 0: %f, %f' % (s.tc, s.lb)
    print s.transform(test_data)
    assert np.allclose(s.transform(test_data)[0][:, :-1], 0., atol=1e-4), \
        'Residual info should be small. Largest value was: %f' % np.max(np.absolute(s.transform(test_data)[0][:, :-1]))
    xbar, labels = s.transform(test_data)
    print s.invert(xbar)
    print test_data
    assert np.allclose(s.invert(xbar) - test_data, 0,
                       atol=0.01), "Invert should be near perfect"
    assert np.allclose(s.labels, test_data[:, :1]) or np.allclose(s.labels, 1 - test_data[:, :1]), \
        'Check that labels are correct.'
Exemplo n.º 5
0
def test_invertibility():
    np.random.seed(seed)
    n, ns = 3, 7
    x_count_a = np.random.randint(0, 5,
                                  (ns, 1)) + np.random.randint(0, 1, (ns, n))
    x_count_b = np.random.randint(0, 5,
                                  (ns, 1)) + np.random.randint(0, 1, (ns, n))
    x_count = np.hstack([x_count_a, x_count_b])
    out = sieve.Sieve(max_layers=2, verbose=verbose, seed=seed,
                      k_max=7).fit(x_count)
    print 'stats for each layer'
    print[[r.h for r in layer.remainders] for layer in out.layers]
    xbar, labels = out.transform(x_count)
    x_predict = out.invert(xbar)
    print 'predicted'
    print x_predict
    print 'actual', x_count
    assert np.allclose(x_predict, x_count)
Exemplo n.º 6
0
def test_vis():
    ns = 100
    xa = np.random.randint(0, 3, (ns, 1))
    xb = np.random.randint(0, 3, (ns, 1))
    xc = np.random.randint(0, 3, (ns, 1))
    test_data = np.hstack([
        np.repeat(xa, 7, axis=1),
        np.repeat(xb, 5, axis=1),
        np.repeat(xc, 3, axis=1)
    ])
    s = sieve.Sieve(max_layers=4,
                    k_max=3,
                    dim_hidden=3,
                    verbose=verbose,
                    seed=seed,
                    n_repeat=20).fit(test_data)
    vis.output_dot(s, filename='test.dot')
    vis.output_plots(s, test_data)


#  TODO: Add this test
#  r_test = lambda n, m, q: np.array([np.random.random((n,m)) < q,np.ones((n,m))] ).astype(int)
#  I've been using this to generate random data and test whether we find nonzero TC. Challenging. Of course,
#  we could always just put in a bootstrap test, but that's inelegant.
Exemplo n.º 7
0
# Xint[:, 3] += (np.random.random(n_samples) < 0.05).astype(int)
X = Xint.astype(float)  # ICA likes floats.

# Sieve
# out = ce.Corex(dim_hidden=3, verbose=True, n_repeat=5, smooth_marginals=False).fit(Xint)
# print zip(s1, out.labels)[:20]
# print set(zip(s1, out.labels))
# print set(zip(s2, out.labels))
# print set(zip(s3, out.labels))
# print len(set(zip(s1, out.labels)))
# print out.alpha
# print out.mis
# sys.exit()
s = sieve.Sieve(max_layers=4,
                dim_hidden=2,
                k_max=0,
                verbose=1,
                n_repeat=20,
                smooth_marginals=False).fit(Xint)
xbar1 = s.layers[0].transform(Xint)
xbar2 = s.layers[1].transform(xbar1)
xbar3 = s.layers[2].transform(xbar2)
# xbar4 = s.layers[3].transform(xbar3)

sieve_labels, y = s.transform(Xint)
ybar = sieve_labels[:, 3:]
xbar = sieve_labels[:, :3]
print 'MIS', s.mis

# Compute ICA
ica = FastICA(n_components=3, max_iter=10000)
S_ = ica.fit_transform(X)  # Reconstruct signals
Exemplo n.º 8
0
import sieve as sv
import frontier as fr
import url_filter as uf
import web_graph as wg

if __name__ == '__main__':

    #Parameters
    sieve_limit = 10  # max number of urls in the sieve
    host_politeness = 60.0  # seconds to wait before visit the host again, float type
    requests_number = 5  # max number of urls, associated with an host, to visit
    number_of_threads = 4  # max number of active threads

    # Data structures instantiation
    seed = uf.fs_url_filter(
        sl.load('seed.txt'))  # Seed instantiation and loading
    web_graph = wg.Web_graph()  # WebGraph instantiation for Centralities

    frontier = fr.Frontier(
        sv.Sieve(seed,
                 sieve_limit), host_politeness, requests_number, web_graph
    )  # Frontier instantiation with data structures and Parameters

    # Execution
    print('Crawler Execution\n')
    frontier.execute(number_of_threads)

    #Centrality Measures
    print('\nCentralities:\n')
    web_graph.print_metrics()