test_darn.py 文件源码

def test_darn(dim_in=5, dim_h=3, dim_out=7, n_samples=13):
    darn = DARN(dim_in, dim_h, dim_out, 2, h_act='T.tanh', out_act='T.nnet.sigmoid')
    tparams = darn.set_tparams()

    X = T.matrix('X', dtype=floatX)
    H = T.matrix('H', dtype=floatX)
    C = darn(H)
    NLP = darn.neg_log_prob(X, C)

    f = theano.function([X, H], [C, NLP])

    x = np.random.randint(0, 2, size=(n_samples, dim_out)).astype(floatX)
    h = np.random.randint(0, 2, size=(n_samples, dim_in)).astype(floatX)

    c_t, nlp_t = f(x, h)
    print c_t.shape

    d_np = np.tanh(np.dot(h, darn.params['W0']) + darn.params['b0'])
    c_np = np.dot(d_np, darn.params['W1']) + darn.params['b1']

    assert np.allclose(c_t, c_np), (c_t, c_np)

    z_np = np.zeros((n_samples, dim_out)).astype(floatX) + darn.params['bar'][None, :] + c_np

    for i in xrange(dim_out):
        for j in xrange(i + 1, dim_out):
            z_np[:, i] += darn.params['War'][j, i] * x[:, j]

    p_np = sigmoid(z_np)

    p_np = np.clip(p_np, 1e-7, 1 - 1e-7)
    nlp_np = (- x * np.log(p_np) - (1 - x) * np.log(1 - p_np)).sum(axis=1)

    assert np.allclose(nlp_t, nlp_np), (nlp_t, nlp_np)

    samples, updates_s = darn.sample(C, n_samples=n_samples-1)
    f = theano.function([H], samples, updates=updates_s)
    print f(h)