Function test_WGAN_GP_loss

numpy_ml/tests/test_nn.py:189–238

(N=5)

def test_WGAN_GP_loss(N=5):
    from numpy_ml.neural_nets.losses import WGAN_GPLoss

    np.random.seed(12345)

    N = np.inf if N is None else N

    i = 1
    while i < N:
        lambda_ = np.random.randint(0, 10)
        n_ex = np.random.randint(1, 10)
        n_feats = np.random.randint(2, 10)
        Y_real = random_tensor([n_ex], standardize=True)
        Y_fake = random_tensor([n_ex], standardize=True)
        gradInterp = random_tensor([n_ex, n_feats], standardize=True)

        mine = WGAN_GPLoss(lambda_=lambda_)
        C_loss = mine(Y_fake, "C", Y_real, gradInterp)
        G_loss = mine(Y_fake, "G")

        C_dY_fake, dY_real, dGradInterp = mine.grad(Y_fake, "C", Y_real, gradInterp)
        G_dY_fake = mine.grad(Y_fake, "G")

        golds = TorchWGANGPLoss(lambda_).extract_grads(Y_real, Y_fake, gradInterp)
        if np.isnan(golds["C_dGradInterp"]).any():
            continue

        params = [
            (Y_real, "Y_real"),
            (Y_fake, "Y_fake"),
            (gradInterp, "gradInterp"),
            (C_loss, "C_loss"),
            (G_loss, "G_loss"),
            (-dY_real, "C_dY_real"),
            (-C_dY_fake, "C_dY_fake"),
            (dGradInterp, "C_dGradInterp"),
            (G_dY_fake, "G_dY_fake"),
        ]

        print("\nTrial {}".format(i))
        for ix, (mine, label) in enumerate(params):
            np.testing.assert_allclose(
                mine,
                golds[label],
                err_msg=err_fmt(params, golds, ix),
                rtol=0.1,
                atol=1e-2,
            )
            print("\tPASSED {}".format(label))
        i += 1


def test_NCELoss(N=1):