Function test_RNNCell

numpy_ml/tests/test_nn.py:1067–1127

(N=15)

def test_RNNCell(N=15):
    from numpy_ml.neural_nets.layers import RNNCell

    N = np.inf if N is None else N

    np.random.seed(12345)

    i = 1
    while i < N + 1:
        n_ex = np.random.randint(1, 10)
        n_in = np.random.randint(1, 10)
        n_out = np.random.randint(1, 10)
        n_t = np.random.randint(1, 10)
        X = random_tensor((n_ex, n_in, n_t), standardize=True)

        # initialize RNN layer
        L1 = RNNCell(n_out=n_out)

        # forward prop
        y_preds = []
        for t in range(n_t):
            y_pred = L1.forward(X[:, :, t])
            y_preds += [y_pred]

        # backprop
        dLdX = []
        dLdAt = np.ones_like(y_preds[t])
        for t in reversed(range(n_t)):
            dLdXt = L1.backward(dLdAt)
            dLdX.insert(0, dLdXt)
        dLdX = np.dstack(dLdX)

        # get gold standard gradients
        gold_mod = TorchRNNCell(n_in, n_out, L1.parameters)
        golds = gold_mod.extract_grads(X)

        params = [
            (X, "X"),
            (np.array(y_preds), "y"),
            (L1.parameters["ba"].T, "ba"),
            (L1.parameters["bx"].T, "bx"),
            (L1.parameters["Wax"].T, "Wax"),
            (L1.parameters["Waa"].T, "Waa"),
            (L1.gradients["ba"].T, "dLdBa"),
            (L1.gradients["bx"].T, "dLdBx"),
            (L1.gradients["Wax"].T, "dLdWax"),
            (L1.gradients["Waa"].T, "dLdWaa"),
            (dLdX, "dLdX"),
        ]

        print("Trial {}".format(i))
        for ix, (mine, label) in enumerate(params):
            np.testing.assert_allclose(
                mine,
                golds[label],
                err_msg=err_fmt(params, golds, ix),
                atol=1e-3,
                rtol=1e-3,