Function test_FullyConnected

numpy_ml/tests/test_nn.py:640–696

(N=15)

def test_FullyConnected(N=15):
    from numpy_ml.neural_nets.layers import FullyConnected
    from numpy_ml.neural_nets.activations import Tanh, ReLU, Sigmoid, Affine

    np.random.seed(12345)

    N = np.inf if N is None else N

    acts = [
        (Tanh(), nn.Tanh(), "Tanh"),
        (Sigmoid(), nn.Sigmoid(), "Sigmoid"),
        (ReLU(), nn.ReLU(), "ReLU"),
        (Affine(), TorchLinearActivation(), "Affine"),
    ]

    i = 1
    while i < N + 1:
        n_ex = np.random.randint(1, 100)
        n_in = np.random.randint(1, 100)
        n_out = np.random.randint(1, 100)
        X = random_tensor((n_ex, n_in), standardize=True)

        # randomly select an activation function
        act_fn, torch_fn, act_fn_name = acts[np.random.randint(0, len(acts))]

        # initialize FC layer
        L1 = FullyConnected(n_out=n_out, act_fn=act_fn)

        # forward prop
        y_pred = L1.forward(X)

        # backprop
        dLdy = np.ones_like(y_pred)
        dLdX = L1.backward(dLdy)

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

        params = [
            (L1.X[0], "X"),
            (y_pred, "y"),
            (L1.parameters["W"].T, "W"),
            (L1.parameters["b"], "b"),
            (dLdy, "dLdy"),
            (L1.gradients["W"].T, "dLdW"),
            (L1.gradients["b"], "dLdB"),
            (dLdX, "dLdX"),
        ]

        print("\nTrial {}\nact_fn={}".format(i, act_fn_name))
        for ix, (mine, label) in enumerate(params):
            assert_almost_equal(
                mine, golds[label], err_msg=err_fmt(params, golds, ix), decimal=3
            )
            print("\tPASSED {}".format(label))
        i += 1