Function test_Deconv2D

numpy_ml/tests/test_nn.py:1348–1427

(N=15)

def test_Deconv2D(N=15):
    from numpy_ml.neural_nets.layers import Deconv2D
    from numpy_ml.neural_nets.activations import Tanh, ReLU, Sigmoid, Affine

    N = np.inf if N is None else N

    np.random.seed(12345)

    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, 10)
        in_rows = np.random.randint(1, 10)
        in_cols = np.random.randint(1, 10)
        n_in, n_out = np.random.randint(1, 3), np.random.randint(1, 3)
        f_shape = (
            min(in_rows, np.random.randint(1, 5)),
            min(in_cols, np.random.randint(1, 5)),
        )
        p, s = np.random.randint(0, 5), np.random.randint(1, 3)

        out_rows = s * (in_rows - 1) - 2 * p + f_shape[0]
        out_cols = s * (in_cols - 1) - 2 * p + f_shape[1]

        if out_rows <= 0 or out_cols <= 0:
            continue

        X = random_tensor((n_ex, in_rows, in_cols, n_in), standardize=True)

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

        # initialize Deconv2D layer
        L1 = Deconv2D(
            out_ch=n_out, kernel_shape=f_shape, act_fn=act_fn, pad=p, stride=s
        )

        # forward prop
        try:
            y_pred = L1.forward(X)
        except ValueError:
            print("Improper dimensions; retrying")
            continue

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

        # get gold standard gradients
        gold_mod = TorchDeconv2DLayer(
            n_in, n_out, torch_fn, L1.parameters, L1.hyperparameters
        )