Python keras.backend.int_shape() Examples

def Highway(x, num_layers=1, activation='relu', name_prefix=''):
    '''
    Layer wrapper function for Highway network
    # Arguments:
        x: tensor, shape = (B, input_size)
    # Optional Arguments:
        num_layers: int, dafault is 1, the number of Highway network layers
        activation: keras activation, default is 'relu'
        name_prefix: str, default is '', layer name prefix
    # Returns:
        out: tensor, shape = (B, input_size)
    '''
    input_size = K.int_shape(x)[1]
    for i in range(num_layers):
        gate_ratio_name = '{}Highway/Gate_ratio_{}'.format(name_prefix, i)
        fc_name = '{}Highway/FC_{}'.format(name_prefix, i)
        gate_name = '{}Highway/Gate_{}'.format(name_prefix, i)

        gate_ratio = Dense(input_size, activation='sigmoid', name=gate_ratio_name)(x)
        fc = Dense(input_size, activation=activation, name=fc_name)(x)
        x = Lambda(lambda args: args[0] * args[2] + args[1] * (1 - args[2]), name=gate_name)([fc, x, gate_ratio])
    return x 

def pool_block(feats, pool_factor):

    if IMAGE_ORDERING == 'channels_first':
        h = K.int_shape(feats)[2]
        w = K.int_shape(feats)[3]
    elif IMAGE_ORDERING == 'channels_last':
        h = K.int_shape(feats)[1]
        w = K.int_shape(feats)[2]

    pool_size = strides = [
        int(np.round(float(h) / pool_factor)),
        int(np.round(float(w) / pool_factor))]

    x = AveragePooling2D(pool_size, data_format=IMAGE_ORDERING,
                         strides=strides, padding='same')(feats)
    x = Conv2D(512, (1, 1), data_format=IMAGE_ORDERING,
               padding='same', use_bias=False)(x)
    x = BatchNormalization()(x)
    x = Activation('relu')(x)

    x = resize_image(x, strides, data_format=IMAGE_ORDERING)

    return x 

def resize_image(inp,  s, data_format):

    try:

        return Lambda(lambda x: K.resize_images(x,
                                                height_factor=s[0],
                                                width_factor=s[1],
                                                data_format=data_format,
                                                interpolation='bilinear'))(inp)

    except Exception as e:
        # if keras is old, then rely on the tf function
        # Sorry theano/cntk users!!!
        assert data_format == 'channels_last'
        assert IMAGE_ORDERING == 'channels_last'

        import tensorflow as tf

        return Lambda(
            lambda x: tf.image.resize_images(
                x, (K.int_shape(x)[1]*s[0], K.int_shape(x)[2]*s[1]))
        )(inp) 

def residual(_x, out_dim, name, stride=1):
  shortcut = _x
  num_channels = K.int_shape(shortcut)[-1]
  _x = ZeroPadding2D(padding=1, name=name + '.pad1')(_x)
  _x = Conv2D(out_dim, 3, strides=stride, use_bias=False, name=name + '.conv1')(_x)
  _x = BatchNormalization(epsilon=1e-5, name=name + '.bn1')(_x)
  _x = Activation('relu', name=name + '.relu1')(_x)

  _x = Conv2D(out_dim, 3, padding='same', use_bias=False, name=name + '.conv2')(_x)
  _x = BatchNormalization(epsilon=1e-5, name=name + '.bn2')(_x)

  if num_channels != out_dim or stride != 1:
    shortcut = Conv2D(out_dim, 1, strides=stride, use_bias=False, name=name + '.shortcut.0')(
        shortcut)
    shortcut = BatchNormalization(epsilon=1e-5, name=name + '.shortcut.1')(shortcut)

  _x = Add(name=name + '.add')([_x, shortcut])
  _x = Activation('relu', name=name + '.relu')(_x)
  return _x 

def call(self, inputs, training=None):
        input_shape = K.int_shape(inputs)
        reduction_axes = list(range(0, len(input_shape)))

        if (self.axis is not None):
            del reduction_axes[self.axis]

        del reduction_axes[0]

        mean = K.mean(inputs, reduction_axes, keepdims=True)
        stddev = K.std(inputs, reduction_axes, keepdims=True) + self.epsilon
        normed = (inputs - mean) / stddev

        broadcast_shape = [1] * len(input_shape)
        if self.axis is not None:
            broadcast_shape[self.axis] = input_shape[self.axis]

        if self.scale:
            broadcast_gamma = K.reshape(self.gamma, broadcast_shape)
            normed = normed * broadcast_gamma
        if self.center:
            broadcast_beta = K.reshape(self.beta, broadcast_shape)
            normed = normed + broadcast_beta
        return normed 

def test_vgg_conv():
    if K.image_data_format() == 'channels_first':
        x = Input(shape=(3, 224, 224))
        y1_shape = (None, 64, 112, 112)
        y2_shape = (None, 128, 56, 56)
    else:
        x = Input(shape=(224, 224, 3))
        y1_shape = (None, 112, 112, 64)
        y2_shape = (None, 56, 56, 128)

    block1 = vgg_conv(filters=64, convs=2, block_name='block1')
    y = block1(x)
    assert K.int_shape(y) == y1_shape

    block2 = vgg_conv(filters=128, convs=2, block_name='block2')
    y = block2(y)
    assert K.int_shape(y) == y2_shape 

def timeception_layers(tensor, n_layers=4, n_groups=8, is_dilated=True):
    input_shape = K.int_shape(tensor)
    assert len(input_shape) == 5

    expansion_factor = 1.25
    _, n_timesteps, side_dim, side_dim, n_channels_in = input_shape

    # how many layers of timeception
    for i in range(n_layers):
        layer_num = i + 1

        # get details about grouping
        n_channels_per_branch, n_channels_out = __get_n_channels_per_branch(n_groups, expansion_factor, n_channels_in)

        # temporal conv per group
        tensor = __grouped_convolutions(tensor, n_groups, n_channels_per_branch, is_dilated, layer_num)

        # downsample over time
        tensor = MaxPooling3D(pool_size=(2, 1, 1), name='maxpool_tc%d' % (layer_num))(tensor)
        n_channels_in = n_channels_out

    return tensor 

def __call_timeception_layers(self, tensor, n_layers, n_groups, expansion_factor):
        input_shape = K.int_shape(tensor)
        assert len(input_shape) == 5

        _, n_timesteps, side_dim, side_dim, n_channels_in = input_shape

        # how many layers of timeception
        for i in range(n_layers):
            layer_num = i + 1

            # get details about grouping
            n_channels_per_branch, n_channels_out = self.__get_n_channels_per_branch(n_groups, expansion_factor, n_channels_in)

            # temporal conv per group
            tensor = self.__call_grouped_convolutions(tensor, n_groups, n_channels_per_branch, layer_num)

            # downsample over time
            tensor = getattr(self, 'maxpool_tc%d' % (layer_num))(tensor)
            n_channels_in = n_channels_out

        return tensor 

def sampling(args: tuple):
    """
    Reparameterization trick by sampling z from unit Gaussian
    :param args: (tensor, tensor) mean and log of variance of q(z|x)
    :returns tensor: sampled latent vector z
    """

    # unpack the input tuple
    z_mean, z_log_var = args

    # mini-batch size
    mb_size = K.shape(z_mean)[0]

    # latent space size
    dim = K.int_shape(z_mean)[1]

    # random normal vector with mean=0 and std=1.0
    epsilon = K.random_normal(shape=(mb_size, dim))

    return z_mean + K.exp(0.5 * z_log_var) * epsilon 

def test_vgg_decoder():
    if K.image_data_format() == 'channels_last':
        inputs = Input(shape=(500, 500, 3))
        pool3 = Input(shape=(88, 88, 256))
        pool4 = Input(shape=(44, 44, 512))
        drop7 = Input(shape=(16, 16, 4096))
        score_shape = (None, 500, 500, 21)
    else:
        inputs = Input(shape=(3, 500, 500))
        pool3 = Input(shape=(256, 88, 88))
        pool4 = Input(shape=(512, 44, 44))
        drop7 = Input(shape=(4096, 16, 16))
        score_shape = (None, 21, 500, 500)
    pyramid = [drop7, pool4, pool3, inputs]
    scales = [1., 1e-2, 1e-4]
    score = VGGDecoder(pyramid, scales, classes=21)
    assert K.int_shape(score) == score_shape 

def VGGUpsampler(pyramid, scales, classes, weight_decay=0.):
    """A Functional upsampler for the VGG Nets.

    :param: pyramid: A list of features in pyramid, scaling from large
                    receptive field to small receptive field.
                    The bottom of the pyramid is the input image.
    :param: scales: A list of weights for each of the feature map in the
                    pyramid, sorted in the same order as the pyramid.
    :param: classes: Integer, number of classes.
    """
    if len(scales) != len(pyramid) - 1:
        raise ValueError('`scales` needs to match the length of'
                         '`pyramid` - 1.')
    blocks = []

    for i in range(len(pyramid) - 1):
        block_name = 'feat{}'.format(i + 1)
        block = vgg_upsampling(classes=classes,
                               target_shape=K.int_shape(pyramid[i + 1]),
                               scale=scales[i],
                               weight_decay=weight_decay,
                               block_name=block_name)
        blocks.append(block)

    return Decoder(pyramid=pyramid[:-1], blocks=blocks) 

def timeception_layers(tensor, n_layers=4, n_groups=8, is_dilated=True):
    input_shape = K.int_shape(tensor)
    assert len(input_shape) == 5

    expansion_factor = 1.25
    _, n_timesteps, side_dim, side_dim, n_channels_in = input_shape

    # how many layers of timeception
    for i in range(n_layers):
        layer_num = i + 1

        # get details about grouping
        n_channels_per_branch, n_channels_out = __get_n_channels_per_branch(n_groups, expansion_factor, n_channels_in)

        # temporal conv per group
        tensor = __grouped_convolutions(tensor, n_groups, n_channels_per_branch, is_dilated, layer_num)

        # downsample over time
        tensor = MaxPooling3D(pool_size=(2, 1, 1), name='maxpool_tc%d' % (layer_num))(tensor)
        n_channels_in = n_channels_out

    return tensor 

def call(self, inputs, training=None):
        input_shape = K.int_shape(inputs)
        reduction_axes = list(range(0, len(input_shape)))

        if (self.axis is not None):
            del reduction_axes[self.axis]

        del reduction_axes[0]

        mean = K.mean(inputs, reduction_axes, keepdims=True)
        stddev = K.std(inputs, reduction_axes, keepdims=True) + self.epsilon
        normed = (inputs - mean) / stddev

        broadcast_shape = [1] * len(input_shape)
        if self.axis is not None:
            broadcast_shape[self.axis] = input_shape[self.axis]

        if self.scale:
            broadcast_gamma = K.reshape(self.gamma, broadcast_shape)
            normed = normed * broadcast_gamma
        if self.center:
            broadcast_beta = K.reshape(self.beta, broadcast_shape)
            normed = normed + broadcast_beta
        return normed 

def sampling(self, args):
        """Reparametrisation by sampling from Gaussian, N(0,I)
        To sample from epsilon = Norm(0,I) instead of from likelihood Q(z|X)
        with latent variables z: z = z_mean + sqrt(var) * epsilon

        Parameters
        ----------
        args : tensor
            Mean and log of variance of Q(z|X).
    
        Returns
        -------
        z : tensor
            Sampled latent variable.
        """

        z_mean, z_log = args
        batch = K.shape(z_mean)[0]  # batch size
        dim = K.int_shape(z_mean)[1]  # latent dimension
        epsilon = K.random_normal(shape=(batch, dim))  # mean=0, std=1.0

        return z_mean + K.exp(0.5 * z_log) * epsilon 

def get_constants(self, x):
		constants = []
		if 0 < self.dropout_U < 1:
			ones = K.ones_like(K.reshape(x[:, 0, 0], (-1, 1)))
			ones = K.tile(ones, (1, self.input_dim))
			B_U = [K.in_train_phase(K.dropout(ones, self.dropout_U), ones) for _ in range(4)]
			constants.append(B_U)
		else:
			constants.append([K.cast_to_floatx(1.) for _ in range(4)])

		if 0 < self.dropout_W < 1:
			input_shape = K.int_shape(x)
			input_dim = input_shape[-1]
			ones = K.ones_like(K.reshape(x[:, 0, 0], (-1, 1)))
			ones = K.tile(ones, (1, int(input_dim)))
			B_W = [K.in_train_phase(K.dropout(ones, self.dropout_W), ones) for _ in range(4)]
			constants.append(B_W)
		else:
			constants.append([K.cast_to_floatx(1.) for _ in range(4)])
		return constants 

def __call_timeception_layers(self, tensor, n_layers, n_groups, expansion_factor):
        input_shape = K.int_shape(tensor)
        assert len(input_shape) == 5

        _, n_timesteps, side_dim, side_dim, n_channels_in = input_shape

        # how many layers of timeception
        for i in range(n_layers):
            layer_num = i + 1

            # get details about grouping
            n_channels_per_branch, n_channels_out = self.__get_n_channels_per_branch(n_groups, expansion_factor, n_channels_in)

            # temporal conv per group
            tensor = self.__call_grouped_convolutions(tensor, n_groups, n_channels_per_branch, layer_num)

            # downsample over time
            tensor = getattr(self, 'maxpool_tc%d' % (layer_num))(tensor)
            n_channels_in = n_channels_out

        return tensor 

def _sampling(args):
    """Reparameterization trick by sampling fr an isotropic unit Gaussian.
       Instead of sampling from Q(z|X), sample eps = N(0,I)

    # Arguments:
        args (tensor): mean and log of variance of Q(z|X)
    # Returns:
        z (tensor): sampled latent vector
    """
    z_mean, z_log_var = args
    batch = K.shape(z_mean)[0]
    dim = K.int_shape(z_mean)[1]
    # by default, random_normal has mean=0 and std=1.0
    epsilon = K.random_normal(shape=(batch, dim))
    return z_mean + K.exp(0.5 * z_log_var) * epsilon 

def __call__(self, inputs):
        filters = K.int_shape(inputs)[self._channel_axis]
        grouped_op = GroupConvolution(filters, self.kernels, groups=len(self.kernels),
                                      type='depthwise_conv', conv_kwargs=self._conv_kwargs)
        x = grouped_op(inputs)
        return x


# Obtained from https://github.com/tensorflow/tpu/blob/master/models/official/mnasnet/mixnet/mixnet_model.py 

def mean_gaussian_negative_log_likelihood(y_true, y_pred):
    nll = 0.5 * np.log(2 * np.pi) + 0.5 * K.square(y_pred - y_true)
    axis = tuple(range(1, len(K.int_shape(y_true))))
    return K.mean(K.sum(nll, axis=axis), axis=-1) 

def _preprocess_weights_for_loading(layer,
                                    weights):
    """
    Converts layers weights.

    Parameters
    ----------
    layer : Layer
        Layer instance.
    weights : list of np.array
        List of weights values.

    Returns
    -------
    list of np.array
        A list of weights values.
    """
    is_channels_first = (K.image_data_format() == "channels_first")
    if ((K.backend() == "mxnet") and (not is_channels_first)) or (K.backend() == "tensorflow"):
        if layer.__class__.__name__ == "Conv2D":
            weights[0] = np.transpose(weights[0], (2, 3, 1, 0))
        elif layer.__class__.__name__ == "DepthwiseConv2D":
            weights[0] = np.transpose(weights[0], (2, 3, 0, 1))
    for i in range(len(weights)):
        assert (K.int_shape(layer.weights[i]) == weights[i].shape)
    return weights 

def mrcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox):
    """Loss for Mask R-CNN bounding box refinement.

    target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))]
    target_class_ids: [batch, num_rois]. Integer class IDs.
    pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))]
    """
    # Reshape to merge batch and roi dimensions for simplicity.
    target_class_ids = K.reshape(target_class_ids, (-1,))
    target_bbox = K.reshape(target_bbox, (-1, 4))
    pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4))

    # Only positive ROIs contribute to the loss. And only
    # the right class_id of each ROI. Get their indicies.
    positive_roi_ix = tf.where(target_class_ids > 0)[:, 0]
    positive_roi_class_ids = tf.cast(
        tf.gather(target_class_ids, positive_roi_ix), tf.int64)
    indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1)

    # Gather the deltas (predicted and true) that contribute to loss
    target_bbox = tf.gather(target_bbox, positive_roi_ix)
    pred_bbox = tf.gather_nd(pred_bbox, indices)

    # Smooth-L1 Loss
    loss = K.switch(tf.size(target_bbox) > 0,
                    smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox),
                    tf.constant(0.0))
    loss = K.mean(loss)
    return loss 

def call(self, x, mask=None):

        assert self.built, 'Layer must be built before being called'
        input_shape = K.int_shape(x)

        reduction_axes = list(range(len(input_shape)))
        del reduction_axes[self.axis]
        broadcast_shape = [1] * len(input_shape)
        broadcast_shape[self.axis] = input_shape[self.axis]

        if sorted(reduction_axes) == range(K.ndim(x))[:-1]:
            x_normed = K.batch_normalization(
                x, self.running_mean, self.running_std,
                self.beta, self.gamma,
                epsilon=self.epsilon)
        else:
            # need broadcasting
            broadcast_running_mean = K.reshape(self.running_mean, broadcast_shape)
            broadcast_running_std = K.reshape(self.running_std, broadcast_shape)
            broadcast_beta = K.reshape(self.beta, broadcast_shape)
            broadcast_gamma = K.reshape(self.gamma, broadcast_shape)
            x_normed = K.batch_normalization(
                x, broadcast_running_mean, broadcast_running_std,
                broadcast_beta, broadcast_gamma,
                epsilon=self.epsilon)

        return x_normed 

def mrcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox):
    """Loss for Mask R-CNN bounding box refinement.

    target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))]
    target_class_ids: [batch, num_rois]. Integer class IDs.
    pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))]
    """
    # Reshape to merge batch and roi dimensions for simplicity.
    target_class_ids = K.reshape(target_class_ids, (-1,))
    target_bbox = K.reshape(target_bbox, (-1, 4))
    pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4))

    # Only positive ROIs contribute to the loss. And only
    # the right class_id of each ROI. Get their indicies.
    positive_roi_ix = tf.where(target_class_ids > 0)[:, 0]
    positive_roi_class_ids = tf.cast(tf.gather(target_class_ids, positive_roi_ix), tf.int64)
    indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1)

    # Gather the deltas (predicted and true) that contribute to loss
    target_bbox = tf.gather(target_bbox, positive_roi_ix)
    pred_bbox = tf.gather_nd(pred_bbox, indices)

    # Smooth-L1 Loss
    loss = K.switch(tf.size(target_bbox) > 0,
                    smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox),
                    tf.constant(0.0))
    loss = K.mean(loss)
    loss = K.reshape(loss, [1, 1])
    return loss 

def _keras_unstack_hack(ab):
    """Implements tf.unstack(y_true_keras, num=2, axis=-1).

       Keras-hack adopted to be compatible with Theano backend.

       :param ab: stacked variables
       :return a, b: unstacked variables
    """
    ndim = len(K.int_shape(ab))
    if ndim == 0:
        print('can not unstack with ndim=0')
    else:
        a = ab[..., 0]
        b = ab[..., 1]
    return a, b 

def call(self, x, mask=None):

        assert self.built, 'Layer must be built before being called'
        input_shape = K.int_shape(x)

        reduction_axes = list(range(len(input_shape)))
        del reduction_axes[self.axis]
        broadcast_shape = [1] * len(input_shape)
        broadcast_shape[self.axis] = input_shape[self.axis]

        if sorted(reduction_axes) == range(K.ndim(x))[:-1]:
            x_normed = K.batch_normalization(
                x, self.running_mean, self.running_std,
                self.beta, self.gamma,
                epsilon=self.epsilon)
        else:
            # need broadcasting
            broadcast_running_mean = K.reshape(self.running_mean, broadcast_shape)
            broadcast_running_std = K.reshape(self.running_std, broadcast_shape)
            broadcast_beta = K.reshape(self.beta, broadcast_shape)
            broadcast_gamma = K.reshape(self.gamma, broadcast_shape)
            x_normed = K.batch_normalization(
                x, broadcast_running_mean, broadcast_running_std,
                broadcast_beta, broadcast_gamma,
                epsilon=self.epsilon)

        return x_normed 

def preprocess_input(self, x):
		if self.consume_less == 'cpu':
			input_shape = K.int_shape(x)
			input_dim = input_shape[2]
			timesteps = input_shape[1]
			return time_distributed_dense(x, self.W, self.b, self.dropout_W,
			                              input_dim, self.hidden_recurrent_dim,
			                              timesteps)
		else:
			return x 

def _spatial_flattenND(ip, rank):
    assert rank in [3, 4, 5], "Rank of input must be 3, 4 or 5"

    ip_shape = K.int_shape(ip)
    channel_dim = 1 if K.image_data_format() == 'channels_first' else -1

    if rank == 3:
        x = ip  # identity op for rank 3

    elif rank == 4:
        if channel_dim == 1:
            # [C, D1, D2] -> [C, D1 * D2]
            shape = [ip_shape[1], ip_shape[2] * ip_shape[3]]
        else:
            # [D1, D2, C] -> [D1 * D2, C]
            shape = [ip_shape[1] * ip_shape[2], ip_shape[3]]

        x = Reshape(shape)(ip)

    else:
        if channel_dim == 1:
            # [C, D1, D2, D3] -> [C, D1 * D2 * D3]
            shape = [ip_shape[1], ip_shape[2] * ip_shape[3] * ip_shape[4]]
        else:
            # [D1, D2, D3, C] -> [D1 * D2 * D3, C]
            shape = [ip_shape[1] * ip_shape[2] * ip_shape[3], ip_shape[4]]

        x = Reshape(shape)(ip)

    return x 

def _inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
    channel_axis = 3
    if block_idx is None:
        prefix = None
    else:
        prefix = '_'.join((block_type, str(block_idx)))
        
    name_fmt = partial(_generate_layer_name, prefix=prefix)

    if block_type == 'Block35':
        branch_0 = conv2d_bn(x, 32, 1, name=name_fmt('Conv2d_1x1', 0))
        branch_1 = conv2d_bn(x, 32, 1, name=name_fmt('Conv2d_0a_1x1', 1))
        branch_1 = conv2d_bn(branch_1, 32, 3, name=name_fmt('Conv2d_0b_3x3', 1))
        branch_2 = conv2d_bn(x, 32, 1, name=name_fmt('Conv2d_0a_1x1', 2))
        branch_2 = conv2d_bn(branch_2, 32, 3, name=name_fmt('Conv2d_0b_3x3', 2))
        branch_2 = conv2d_bn(branch_2, 32, 3, name=name_fmt('Conv2d_0c_3x3', 2))
        branches = [branch_0, branch_1, branch_2]
    elif block_type == 'Block17':
        branch_0 = conv2d_bn(x, 128, 1, name=name_fmt('Conv2d_1x1', 0))
        branch_1 = conv2d_bn(x, 128, 1, name=name_fmt('Conv2d_0a_1x1', 1))
        branch_1 = conv2d_bn(branch_1, 128, [1, 7], name=name_fmt('Conv2d_0b_1x7', 1))
        branch_1 = conv2d_bn(branch_1, 128, [7, 1], name=name_fmt('Conv2d_0c_7x1', 1))
        branches = [branch_0, branch_1]
    elif block_type == 'Block8':
        branch_0 = conv2d_bn(x, 192, 1, name=name_fmt('Conv2d_1x1', 0))
        branch_1 = conv2d_bn(x, 192, 1, name=name_fmt('Conv2d_0a_1x1', 1))
        branch_1 = conv2d_bn(branch_1, 192, [1, 3], name=name_fmt('Conv2d_0b_1x3', 1))
        branch_1 = conv2d_bn(branch_1, 192, [3, 1], name=name_fmt('Conv2d_0c_3x1', 1))
        branches = [branch_0, branch_1]

    mixed = Concatenate(axis=channel_axis, name=name_fmt('Concatenate'))(branches)
    up = conv2d_bn(mixed,K.int_shape(x)[channel_axis],1,activation=None,use_bias=True,
                   name=name_fmt('Conv2d_1x1'))
    up = Lambda(scaling,
                output_shape=K.int_shape(up)[1:],
                arguments={'scale': scale})(up)
    x = add([x, up])
    if activation is not None:
        x = Activation(activation, name=name_fmt('Activation'))(x)
    return x 

def shape_list(x):
    if K.backend() != 'theano':
        tmp = K.int_shape(x)
    else:
        tmp = x.shape
    tmp = list(tmp)
    tmp[0] = -1
    return tmp 

def call(self, inputs):
        if not hasattr(self, "kernel"):
            embedding_layer = inputs._keras_history[0]

            if embedding_layer.name != self.embedding_name:

                def recursive_search(layer):
                    """递归向上搜索，根据名字找Embedding层
                    """
                    last_layer = layer._inbound_nodes[0].inbound_layers
                    if isinstance(last_layer, list):
                        if len(last_layer) == 0:
                            return None
                        else:
                            last_layer = last_layer[0]
                    if last_layer.name == self.embedding_name:
                        return last_layer
                    else:
                        return recursive_search(last_layer)

                embedding_layer = recursive_search(embedding_layer)
                if embedding_layer is None:
                    raise Exception("Embedding layer not found")

                self.kernel = K.transpose(embedding_layer.embeddings)
                self.units = K.int_shape(self.kernel)[1]
                self.bias = self.add_weight(name="bias",
                                            shape=(self.units, ),
                                            initializer="zeros")

        outputs = K.dot(inputs, self.kernel)
        outputs = K.bias_add(outputs, self.bias)
        outputs = self.activation(outputs)
        return outputs