def cnn_model_003(ctx, x, act=F.relu, test=False):
    with nn.context_scope(ctx):
        # Convblock0
        h = conv_unit(x, "conv00", 128, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv01", 128, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv02", 128, k=3, s=1, p=1, act=act, test=test)
        h = F.max_pooling(h, (2, 2))  # 32 -> 16
        with nn.parameter_scope("bn0"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test:
            h = F.dropout(h)

        # Convblock 1
        h = conv_unit(h, "conv10", 256, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv11", 256, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv12", 256, k=3, s=1, p=1, act=act, test=test)
        h = F.max_pooling(h, (2, 2))  # 16 -> 8
        with nn.parameter_scope("bn1"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test:
            h = F.dropout(h)

        # Convblock 2
        h = conv_unit(h, "conv20", 512, k=3, s=1, p=0, act=act, test=test)  # 8 -> 6
        h = conv_unit(h, "conv21", 256, k=1, s=1, p=0, act=act, test=test)
        h = conv_unit(h, "conv22", 128, k=1, s=1, p=0, act=act, test=test)
        h = conv_unit(h, "conv23", 10, k=1, s=1, p=0, act=act, test=test)

        # Convblock 3
        h = F.average_pooling(h, (6, 6))
        with nn.parameter_scope("bn2"):
            h = PF.batch_normalization(h, batch_stat=not test)
        h = F.reshape(h, (h.shape[0], np.prod(h.shape[1:])))
        return h

def resnet_model(ctx, x, inmaps=64, act=F.relu, test=False):
    # Conv -> BN -> Relu
    with nn.context_scope(ctx):
        with nn.parameter_scope("conv1"):
            h = PF.convolution(x, inmaps, kernel=(3, 3), pad=(1, 1), with_bias=False)
            h = PF.batch_normalization(h, decay_rate=0.9, batch_stat=not test)
            h = act(h)
        
        h = res_unit(h, "conv2", act, False) # -> 32x32
        h = res_unit(h, "conv3", act, True)  # -> 16x16
        with nn.parameter_scope("bn0"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test:
            h = F.dropout(h)
        h = res_unit(h, "conv4", act, False) # -> 16x16
        h = res_unit(h, "conv5", act, True)  # -> 8x8
        with nn.parameter_scope("bn1"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test:
            h = F.dropout(h)
        h = res_unit(h, "conv6", act, False) # -> 8x8
        h = res_unit(h, "conv7", act, True)  # -> 4x4
        with nn.parameter_scope("bn2"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test:
            h = F.dropout(h)
        h = res_unit(h, "conv8", act, False) # -> 4x4
        h = F.average_pooling(h, kernel=(4, 4))  # -> 1x1
        
        pred = PF.affine(h, 10)
    return pred

    def res_unit(x, scope_name, dn=False, test=False):
        C = x.shape[1]
        with nn.parameter_scope(scope_name):

            # Conv -> BN -> Relu
            with nn.parameter_scope("conv1"):
                h = PF.convolution(x, C / 2, kernel=(1, 1), pad=(0, 0),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
                h = F.relu(h)
            # Conv -> BN -> Relu
            with nn.parameter_scope("conv2"):
                h = PF.convolution(h, C / 2, kernel=(3, 3), pad=(1, 1),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
                h = F.relu(h)
            # Conv -> BN
            with nn.parameter_scope("conv3"):
                h = PF.convolution(h, C, kernel=(1, 1), pad=(0, 0),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
            # Residual -> Relu
            h = F.relu(h + x)

            # Maxpooling
            if dn:
                h = F.max_pooling(h, kernel=(2, 2), stride=(2, 2))
            return h

def cifar10_resnet23_prediction(ctx, scope, image, test=False):
    """
    Construct ResNet 23
    """
    # Residual Unit
    def res_unit(x, scope_name, dn=False, test=False):
        C = x.shape[1]
        with nn.parameter_scope(scope_name):

            # Conv -> BN -> Relu
            with nn.parameter_scope("conv1"):
                h = PF.convolution(x, C / 2, kernel=(1, 1), pad=(0, 0),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
                h = F.relu(h)
            # Conv -> BN -> Relu
            with nn.parameter_scope("conv2"):
                h = PF.convolution(h, C / 2, kernel=(3, 3), pad=(1, 1),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
                h = F.relu(h)
            # Conv -> BN
            with nn.parameter_scope("conv3"):
                h = PF.convolution(h, C, kernel=(1, 1), pad=(0, 0),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
            # Residual -> Relu
            h = F.relu(h + x)

            # Maxpooling
            if dn:
                h = F.max_pooling(h, kernel=(2, 2), stride=(2, 2))
            return h

    # Random generator for using the same init parameters in all devices
    nmaps = 64
    ncls = 10

    # Conv -> BN -> Relu
    with nn.context_scope(ctx):
        with nn.parameter_scope(scope):
            with nn.parameter_scope("conv1"):
                h = PF.convolution(image, nmaps, kernel=(3, 3), pad=(1, 1),
                                   with_bias=False)
                h = PF.batch_normalization(h, batch_stat=not test)
                h = F.relu(h)
             
            h = res_unit(h, "conv2", False)    # -> 32x32
            h = res_unit(h, "conv3", True)     # -> 16x16
            h = res_unit(h, "conv4", False)    # -> 16x16
            h = res_unit(h, "conv5", True)     # -> 8x8
            h = res_unit(h, "conv6", False)    # -> 8x8
            h = res_unit(h, "conv7", True)     # -> 4x4
            h = res_unit(h, "conv8", False)    # -> 4x4
            h = F.average_pooling(h, kernel=(4, 4))  # -> 1x1
            pred = PF.affine(h, ncls)

    return pred

 def res_unit(x, scope):
     C = x.shape[1]
     with nn.parameter_scope(scope):
         with nn.parameter_scope('conv1'):
             h = F.elu(bn(PF.convolution(x, C / 2, (1, 1), with_bias=False)))
         with nn.parameter_scope('conv2'):
             h = F.elu(
                 bn(PF.convolution(h, C / 2, (3, 3), pad=(1, 1), with_bias=False)))
         with nn.parameter_scope('conv3'):
             h = bn(PF.convolution(h, C, (1, 1), with_bias=False))
     return F.elu(F.add2(h, x, inplace=True))

 def res_unit(x, scope):
     C = x.shape[1]
     with nn.parameter_scope(scope):
         with nn.parameter_scope('conv1'):
             h = F.elu(bn(PF.binary_connect_convolution(
                 x, C / 2, (1, 1), with_bias=False)))
         with nn.parameter_scope('conv2'):
             h = F.elu(
                 bn(PF.binary_connect_convolution(h, C / 2, (3, 3), pad=(1, 1), with_bias=False)))
         with nn.parameter_scope('conv3'):
             h = bn(PF.binary_connect_convolution(
                 h, C, (1, 1), with_bias=False))
     return F.elu(x + h)

def cnn_model_003(ctx, x, act=F.elu, do=True, test=False):
    with nn.context_scope(ctx):
        # Convblock0
        h = conv_unit(x, "conv00", 128, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv01", 128, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv02", 128, k=3, s=1, p=1, act=act, test=test)
        h = F.max_pooling(h, (2, 2))  # 32 -> 16
        with nn.parameter_scope("bn0"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test and do:
            h = F.dropout(h)

        # Convblock 1
        h = conv_unit(h, "conv10", 256, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv11", 256, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv12", 256, k=3, s=1, p=1, act=act, test=test)
        h = F.max_pooling(h, (2, 2))  # 16 -> 8
        with nn.parameter_scope("bn1"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test and do:
            h = F.dropout(h)

        # Convblock 2
        h = conv_unit(h, "conv20", 512, k=3, s=1, p=0, act=act, test=test)  # 8 -> 6
        h = conv_unit(h, "conv21", 256, k=1, s=1, p=0, act=act, test=test)
        h = conv_unit(h, "conv22", 128, k=1, s=1, p=0, act=act, test=test)
        h_branch = h

        # Convblock 3
        h = conv_unit(h_branch, "conv23", 10, k=1, s=1, p=0, act=act, test=test)
        h = F.average_pooling(h, (6, 6))
        with nn.parameter_scope("bn2"):
            h = PF.batch_normalization(h, batch_stat=not test)
        pred = F.reshape(h, (h.shape[0], np.prod(h.shape[1:])))

        # Uncertainty
        u0 = conv_unit(h_branch, "u0", 10, k=1, s=1, p=0, act=act, test=test)
        u0 = F.average_pooling(u0, (6, 6))
        with nn.parameter_scope("u0bn"):
            u0 = PF.batch_normalization(u0, batch_stat=not test)
            log_var = F.reshape(u0, (u0.shape[0], np.prod(u0.shape[1:])))

        # Uncertainty for uncertainty
        u1 = conv_unit(h_branch, "u1", 10, k=1, s=1, p=0, act=act, test=test)
        u1 = F.average_pooling(u1, (6, 6))
        with nn.parameter_scope("u1bn"):
            u1 = PF.batch_normalization(u1, batch_stat=not test)
            log_s = F.reshape(u1, (u1.shape[0], np.prod(u1.shape[1:])))

        return pred, log_var, log_s

def cnn_model_003(ctx, h, act=F.elu, do=True, test=False):
    with nn.context_scope(ctx):
        if not test:
            b, c, s, s = h.shape
            h = F.image_augmentation(h, (c, s, s),
                                     min_scale=1.0, max_scale=1.5,
                                     angle=0.5, aspect_ratio=1.3, distortion=0.2,
                                     flip_lr=True)
        # Convblock0
        h = conv_unit(h, "conv00", 128, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv01", 128, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv02", 128, k=3, s=1, p=1, act=act, test=test)
        h = F.max_pooling(h, (2, 2))  # 32 -> 16
        with nn.parameter_scope("bn0"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test and do:
            h = F.dropout(h)

        # Convblock 1
        h = conv_unit(h, "conv10", 256, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv11", 256, k=3, s=1, p=1, act=act, test=test)
        h = conv_unit(h, "conv12", 256, k=3, s=1, p=1, act=act, test=test)
        h = F.max_pooling(h, (2, 2))  # 16 -> 8
        with nn.parameter_scope("bn1"):
            h = PF.batch_normalization(h, batch_stat=not test)
        if not test and do:
            h = F.dropout(h)

        # Convblock 2
        h = conv_unit(h, "conv20", 512, k=3, s=1, p=0, act=act, test=test)  # 8 -> 6
        h = conv_unit(h, "conv21", 256, k=1, s=1, p=0, act=act, test=test)
        h = conv_unit(h, "conv22", 128, k=1, s=1, p=0, act=act, test=test)
        u = h

        # Convblock 3
        h = conv_unit(h, "conv23", 10, k=1, s=1, p=0, act=act, test=test)
        h = F.average_pooling(h, (6, 6))
        with nn.parameter_scope("bn2"):
            h = PF.batch_normalization(h, batch_stat=not test)
        pred = F.reshape(h, (h.shape[0], np.prod(h.shape[1:])))

        # Uncertainty
        u = conv_unit(u, "u0", 10, k=1, s=1, p=0, act=act, test=test)
        u = F.average_pooling(u, (6, 6))
        with nn.parameter_scope("u0bn"):
            u = PF.batch_normalization(u, batch_stat=not test)
            log_var = F.reshape(u, (u.shape[0], np.prod(u.shape[1:])))

        return pred, log_var

def mnist_binary_weight_lenet_prediction(image, test=False):
    """
    Construct LeNet for MNIST (Binary Weight Network version).
    """
    with nn.parameter_scope("conv1"):
        c1 = PF.binary_weight_convolution(image, 16, (5, 5))
        c1 = F.elu(F.average_pooling(c1, (2, 2)))
    with nn.parameter_scope("conv2"):
        c2 = PF.binary_weight_convolution(c1, 16, (5, 5))
        c2 = F.elu(F.average_pooling(c2, (2, 2)))
    with nn.parameter_scope("fc3"):
        c3 = F.elu(PF.binary_weight_affine(c2, 50))
    with nn.parameter_scope("fc4"):
        c4 = PF.binary_weight_affine(c3, 10)
    return c4

def cnn_ae_model_000(ctx, x, act=F.relu, test=False):
    with nn.parameter_scope("ae"):
        with nn.context_scope(ctx):
            # Convblock0
            h = conv_unit(x, "conv00", 32, k=3, s=1, p=1, act=act, test=test)
            h = conv_unit(h, "conv01", 32, k=3, s=1, p=1, act=act, test=test)
            h = conv_unit(h, "conv02", 32, k=3, s=1, p=1, act=act, test=test)
            h = conv_unit(h, "conv03", 32, k=4, s=2, p=1, act=act, test=test)  # 32 -> 16
            if not test:
                h = F.dropout(h)
     
            # Convblock 1
            h = conv_unit(h, "conv10", 64, k=3, s=1, p=1, act=act, test=test)
            h = conv_unit(h, "conv11", 64, k=3, s=1, p=1, act=act, test=test)
            h = conv_unit(h, "conv12", 64, k=3, s=1, p=1, act=act, test=test)
            h = conv_unit(h, "conv13", 64, k=4, s=2, p=1, act=act, test=test) # 16 -> 8
            if not test:
                h = F.dropout(h)
     
            # Deconvblock0
            h = deconv_unit(h, "deconv00", 64, k=4, s=2, p=1, act=act, test=test) # 8 -> 16
            h = deconv_unit(h, "deconv01", 64, k=3, s=1, p=1, act=act, test=test)
     
            h = deconv_unit(h, "deconv02", 64, k=3, s=1, p=1, act=act, test=test)
            h = deconv_unit(h, "deconv03", 64, k=3, s=1, p=1, act=act, test=test)  
     
            # Deconvblock 1
            h = deconv_unit(h, "deconv10", 32, k=4, s=2, p=1, act=act, test=test)  # 16 -> 32
            h = deconv_unit(h, "deconv11", 32, k=3, s=1, p=1, act=act, test=test)
            h = deconv_unit(h, "deconv12", 32, k=3, s=1, p=1, act=act, test=test)
            h = deconv_unit(h, "deconv13", 3, k=3, s=1, p=1, act=None, test=test)

        return h

def test_parameter_scope_slash():
    """Testing if parameter_scope('aaa/bbb') works.
    """
    import nnabla as nn
    from nnabla.parameter import get_parameter_or_create
    nn.clear_parameters()
    with nn.parameter_scope('aaa/bbb'):
        param = get_parameter_or_create('ccc', (2, 3, 4, 5))
    ref = np.random.randn(*param.shape).astype(np.float32)
    param.d = ref

    with nn.parameter_scope('aaa'):
        with nn.parameter_scope('bbb'):
            param = get_parameter_or_create('ccc', (2, 3, 4, 5))
    assert np.all(param.d == ref)
    nn.clear_parameters()

def test_parametric_function_api():
    """
    Testing :function:`nnabla.parametric_functions.parametric_function_api`.
    """
    import nnabla as nn
    import inspect
    nn.clear_parameters()
    shape = (2, 3, 4)

    # Signature check
    spec = inspect.getargspec(dummy_parametric_function)
    assert spec.args == ['shape', 'f', 'i', 's', 'name']
    assert spec.defaults == (10, 1, 'dummy', None)
    assert dummy_parametric_function.__doc__.splitlines()[0] == 'Doc'

    # Verify two different ways does the same thing.
    # Using name argument
    v = dummy_parametric_function(shape, name='group1')
    # Using parameter_scope
    with nn.parameter_scope('group1'):
        v = dummy_parametric_function(shape)

    params = nn.get_parameters()
    assert len(params) == 2
    assert list(iterkeys(params)) == ['group1/dummy/p1', 'group1/dummy/p2']

    # No scope
    v = dummy_parametric_function(shape)

    params = nn.get_parameters()
    len(params) == 4
    assert list(iterkeys(params)) == ['group1/dummy/p1', 'group1/dummy/p2',
                                      'dummy/p1', 'dummy/p2']
    nn.clear_parameters()

def conv_unit(x, scope, maps, k=4, s=2, p=1, act=F.relu, test=False):
    with nn.parameter_scope(scope):
        h = PF.convolution(x, maps, kernel=(k, k), stride=(s, s), pad=(p, p))
        if act is None:
            return h
        h = PF.batch_normalization(h, batch_stat=not test)
        h = act(h)
        return h

def conv_unit(x, scope, maps, k=4, s=2, p=1, act=F.prelu, test=False):
    with nn.parameter_scope(scope):
        h = PF.convolution(x, maps, kernel=(k, k), stride=(s, s), pad=(p, p))
        h = PF.batch_normalization(h, batch_stat=not test)
        shape = h.shape
        w = nn.Variable()
        w.d = 0.3
        h = act(h, w)
        return h

def res_block(x, scope_name, act=F.relu, dn=False, test=False):
    C = x.shape[1]
    with nn.parameter_scope(scope_name):
        # Conv -> BN -> Relu
        with nn.parameter_scope("conv1"):
            h = PF.convolution(x, C/2, kernel=(1, 1), pad=(0, 0), with_bias=False)
            h = PF.batch_normalization(h, decay_rate=0.9, batch_stat=not test)
            h = act(h)
        # Conv -> BN -> Relu
        with nn.parameter_scope("conv2"):
            h = PF.convolution(h, C/2, kernel=(3, 3), pad=(1, 1), with_bias=False)
            h = PF.batch_normalization(h, decay_rate=0.9, batch_stat=not test)
            h = act(h)
        # Conv -> BN
        with nn.parameter_scope("conv3"): 
            h = PF.convolution(h, C, kernel=(1, 1), pad=(0, 0), with_bias=False)
            h = PF.batch_normalization(h, decay_rate=0.9, batch_stat=not test)
    return h

def test_graph_model(model, seed):
    np.random.seed(313)
    rng = np.random.RandomState(seed)
    x = nn.Variable([2, 3, 4, 4], need_grad=True)
    t = nn.Variable([2, 1])
    x.d = rng.randn(*x.shape)
    t.d = rng.randint(0, 5, size=t.shape)

    nn.set_default_context(nn.Context())

    # Forwardprop by definintion
    nn.clear_parameters()
    if model == "mlp":
        with nn.parameter_scope('fc1'):
            z = PF.affine(x, 3)
        z2 = F.relu(z, inplace=True)
        with nn.parameter_scope('fc2'):
            z3 = PF.affine(z2, 5)
    elif model == "recurrent":
        with nn.parameter_scope('fc1'):
            z = PF.affine(x, 3)
            z2 = F.relu(z, inplace=True)
        h = z2
        for _ in range(2):
            with nn.parameter_scope('fc2'):
                h = PF.affine(h, 3)
                h = F.relu(h, inplace=True)
        with nn.parameter_scope('fc3'):
            z3 = PF.affine(h, 5)
    elif model == "convolution":
        with nn.parameter_scope('conv1'):
            z = PF.convolution(x, 3, (2, 2))
            z2 = F.relu(z, inplace=True)
        with nn.parameter_scope('fc2'):
            z3 = PF.affine(z2, 5)
    else:
        raise ValueError()
    l = F.softmax_cross_entropy(z3, t, 1)
    L = F.mean(l)

    # Forwardprop
    L.forward(clear_no_need_grad=True)

    # Backprop
    # Diff should be initialized since they are always accumulated
    x.grad.zero()
    L.backward(clear_buffer=True)
    x.g = rng.randn(*x.shape)
    parameters = nn.get_parameters()
    for param in parameters.values():
        param.grad.zero()
    inputs = [x] + list(parameters.values())

    from nbla_test_utils import \
        compute_analytical_and_numerical_grad_graph as grads
    agrad, ngrad = grads(L, inputs, 1e-3)
    assert np.allclose(ngrad, agrad, atol=1.05e-2)

def test_graph_unlink_backward(seed):
    rng = np.random.RandomState(seed)
    x0 = nn.Variable([2, 4], need_grad=True)
    x1 = nn.Variable([2, 4], need_grad=True)
    x0.d = rng.randn(*x0.shape)
    x1.d = rng.randn(*x1.shape)
    x0.grad.zero()
    x1.grad.zero()
    with nn.auto_forward():
        with nn.parameter_scope("fc0"):
            h0 = PF.affine(x0, 2)
        with nn.parameter_scope("fc1"):
            h1 = PF.affine(x1, 2)
        h0.need_grad = False
        h = h0 + h1
        with nn.parameter_scope("fc"):
            y = PF.affine(h, 1)
    y.backward(clear_buffer=True)
    assert np.all(x0.g == 0)
    assert not np.all(x1.g == 0)

def mnist_binary_connect_lenet_prediction(image, test=False):
    """
    Construct LeNet for MNIST (BinaryNet version).
    """
    with nn.parameter_scope("conv1"):
        c1 = PF.binary_connect_convolution(image, 16, (5, 5))
        c1 = PF.batch_normalization(c1, batch_stat=not test)
        c1 = F.elu(F.average_pooling(c1, (2, 2)))
    with nn.parameter_scope("conv2"):
        c2 = PF.binary_connect_convolution(c1, 16, (5, 5))
        c2 = PF.batch_normalization(c2, batch_stat=not test)
        c2 = F.elu(F.average_pooling(c2, (2, 2)))
    with nn.parameter_scope("fc3"):
        c3 = PF.binary_connect_affine(c2, 50)
        c3 = PF.batch_normalization(c3, batch_stat=not test)
        c3 = F.elu(c3)
    with nn.parameter_scope("fc4"):
        c4 = PF.binary_connect_affine(c3, 10)
        c4 = PF.batch_normalization(c4, batch_stat=not test)
    return c4

def test_graph_clear_buffer(seed):
    np.random.seed(313)
    rng = np.random.RandomState(seed)
    x = nn.Variable([2, 3, 4, 4])
    t = nn.Variable([2, 1])
    x.d = rng.randn(*x.shape)
    t.d = rng.randint(0, 5, size=t.shape)

    # Network definition
    nn.set_default_context(nn.Context())
    nn.clear_parameters()
    x1 = x + 1
    x2 = x1 - 1
    with nn.parameter_scope('conv1'):
        z = PF.convolution(x2, 3, (2, 2))
        z2 = F.relu(z, inplace=True)
    with nn.parameter_scope('fc2'):
        z3 = PF.affine(z2, 5)
    l = F.softmax_cross_entropy(z3, t, 1)
    L = F.mean(l)

    # Forwardprop
    import tempfile
    import os
    tmpd = tempfile.mkdtemp()
    nn.save_parameters(os.path.join(tmpd, 'parameter.h5'))
    first = False
    for cnng in [False, True]:
        for cb in [False, True]:
            _ = nn.load_parameters(os.path.join(tmpd, 'parameter.h5'))
            for v in nn.get_parameters().values():
                v.grad.zero()
            L.forward(clear_no_need_grad=cnng)
            L.backward(clear_buffer=cb)
            if not first:
                first = True
                g = list(nn.get_parameters().values())[0].g.copy()
            else:
                g2 = list(nn.get_parameters().values())[0].g.copy()
                assert np.all(g == g2)

def discriminator(x, maxh=256, test=False, output_hidden=False):
    """
    Building discriminator network which maps a (B, 1, 28, 28) input to
    a (B, 1).
    """
    # Define shortcut functions
    def bn(xx):
        # Batch normalization
        return PF.batch_normalization(xx, batch_stat=not test)

    def downsample2(xx, c):
        return PF.convolution(xx, c, (3, 3), pad=(1, 1), stride=(2, 2), with_bias=False)

    assert maxh / 8 > 0
    with nn.parameter_scope("dis"):
        # (1, 28, 28) --> (32, 16, 16)
        with nn.parameter_scope("conv1"):
            c1 = F.elu(bn(PF.convolution(x, maxh / 8,
                                         (3, 3), pad=(3, 3), stride=(2, 2), with_bias=False)))
        # (32, 16, 16) --> (64, 8, 8)
        with nn.parameter_scope("conv2"):
            c2 = F.elu(bn(downsample2(c1, maxh / 4)))
        # (64, 8, 8) --> (128, 4, 4)
        with nn.parameter_scope("conv3"):
            c3 = F.elu(bn(downsample2(c2, maxh / 2)))
        # (128, 4, 4) --> (256, 4, 4)
        with nn.parameter_scope("conv4"):
            c4 = bn(PF.convolution(c3, maxh, (3, 3),
                                   pad=(1, 1), with_bias=False))
        # (256, 4, 4) --> (1,)
        with nn.parameter_scope("fc1"):
            f = PF.affine(c4, 1)
    if output_hidden:
        return f, [c1, c2, c3, c4]
    return f