Function test_WGAN_GP

numpy_ml/tests/test_nn.py:2326–2446

(N=1)

def test_WGAN_GP(N=1):
    from numpy_ml.neural_nets.models.wgan_gp import WGAN_GP

    np.random.seed(12345)

    ss = np.random.randint(0, 1000)
    np.random.seed(ss)

    N = np.inf if N is None else N

    i = 1
    while i < N + 1:
        c_updates_per_epoch, n_steps = 1, 1
        n_ex = np.random.randint(1, 500)
        n_in = np.random.randint(1, 100)
        lambda_ = np.random.randint(0, 20)
        g_hidden = np.random.randint(2, 500)
        X = random_tensor((n_ex, n_in), standardize=True)

        # initialize WGAN_GP model
        L1 = WGAN_GP(g_hidden=g_hidden, debug=True)

        # forward prop
        batchsize = n_ex
        L1.fit(
            X,
            lambda_=lambda_,
            c_updates_per_epoch=c_updates_per_epoch,
            n_steps=n_steps,
            batchsize=batchsize,
        )

        # backprop
        dv = L1.derived_variables
        params = L1.parameters["components"]
        grads = L1.gradients["components"]
        params["noise"] = dv["noise"]
        params["alpha"] = dv["alpha"]
        params["n_in"] = n_in
        params["g_hidden"] = g_hidden
        params["c_updates_per_epoch"] = c_updates_per_epoch
        params["n_steps"] = n_steps

        # get gold standard gradients
        golds = WGAN_GP_tf(X, lambda_=lambda_, batch_size=batchsize, params=params)

        params = [
            (dv["X_real"], "X_real"),
            (params["generator"]["FC1"]["W"], "G_weights_FC1"),
            (params["generator"]["FC2"]["W"], "G_weights_FC2"),
            (params["generator"]["FC3"]["W"], "G_weights_FC3"),
            (params["generator"]["FC4"]["W"], "G_weights_FC4"),
            (dv["G_fwd_X_fake"]["FC1"], "G_fwd_X_fake_FC1"),
            (dv["G_fwd_X_fake"]["FC2"], "G_fwd_X_fake_FC2"),
            (dv["G_fwd_X_fake"]["FC3"], "G_fwd_X_fake_FC3"),
            (dv["G_fwd_X_fake"]["FC4"], "G_fwd_X_fake_FC4"),
            (dv["X_fake"], "X_fake"),
            (dv["X_interp"], "X_interp"),