操纵神经网络的输出

Question

我有一个神经网络，输入 (m, 2, 3, 96, 96) 和输出 (m, 2, 128)。我试图通过减去 output[m][0][0] - output[m][0][1] 将输出转换为 (m, 1, 128)，然后通过输入1x128 输出到密集层

我在网络和包装器中尝试了 Lambda 和 keras.backend.Subtract 层

def faceRecoModel(input_shape):

    """

    Implementation of the Inception model used for FaceNet

    Arguments:

    input_shape -- shape of the images of the dataset

    Returns:

    model -- a Model() instance in Keras

    """

    # Define the input as a tensor with shape input_shape

    X_input = Input(input_shape)

    # Zero-Padding

    X = ZeroPadding2D((3, 3))(X_input)

    # First Block

    X = Conv2D(64, (7, 7), strides=(2, 2), name='conv1')(X)

    X = BatchNormalization(axis=1, name='bn1')(X)

    X = Activation('relu')(X)

    # Zero-Padding + MAXPOOL

    X = ZeroPadding2D((1, 1))(X)

    X = MaxPooling2D((3, 3), strides=2)(X)

    # Second Block

    X = Conv2D(64, (1, 1), strides=(1, 1), name='conv2')(X)

    X = BatchNormalization(axis=1, epsilon=0.00001, name='bn2')(X)

    X = Activation('relu')(X)

    # Zero-Padding + MAXPOOL

    X = ZeroPadding2D((1, 1))(X)

    # Second Block

    X = Conv2D(192, (3, 3), strides=(1, 1), name='conv3')(X)

    X = BatchNormalization(axis=1, epsilon=0.00001, name='bn3')(X)

    X = Activation('relu')(X)

    # Zero-Padding + MAXPOOL

    X = ZeroPadding2D((1, 1))(X)

    X = MaxPooling2D(pool_size=3, strides=2)(X)

    # Inception 1: a/b/c

    X = inception_block_1a(X)

    X = inception_block_1b(X)

    X = inception_block_1c(X)

    # Inception 2: a/b

    X = inception_block_2a(X)

    X = inception_block_2b(X)

    # Inception 3: a/b

    X = inception_block_3a(X)

    X = inception_block_3b(X)

    # Top layer

    X = AveragePooling2D(pool_size=(3, 3), strides=(1, 1), data_format='channels_first')(X)

    X = Flatten()(X)

    X = Dense(128, name='dense_layer')(X)

    # L2 normalization

    X = Lambda(lambda x: K.l2_normalize(x, axis=1))(X)

Answer 1

X = Lambda(lambda x: return x[:,0] - x[:,1])(X) X = Dense(...)(X)