tensorflow - 尝试使用 Keras 功能 API 构建 CNN 模型时图断开连接

Question

我正在尝试使用 Keras 功能 API 构建 CNN 模型，但每当我尝试执行这行代码时: model = CNN(settings, np.expand_dims(x_train, axis = 3)).build_network() 我一直遇到以下问题:

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_11:0", shape=(?, 28, 28, 1), dtype=float32) at layer "input_11". The following previous layers were accessed without issue: []

这是我的代码:

class CNN:

    def __init__(self, settings, data):

        self.flag = False
        self.settings = settings

        if self.check_network_settings() == False:
            self.flag = True
            return

        self.data = data

        if K.image_data_format() == "channels_first":
            self.data = self.data.reshape(data.shape[0], data.shape[3], data.shape[2], data.shape[1])

        self.build_network()

    def show_model_chart(self):

        if not os.path.isfile('model.png'):
            plot_model(self.model, to_file = 'model.png')

        model_pic = cv2.imread('model.png')
        plt.imshow(model_pic)
        plt.show()


    def build_network(self):
        print('Bulding Convolutional Neural Network ...')

        inputs = Input(shape = (self.data.shape[1], self.data.shape[2], self.data.shape[3]))
        final_output = None

        for layer_idx in range(self.settings['conv']['layers']):
            inputs = Conv2D(
                              filters = self.settings['conv']['filters'][layer_idx],
                              kernel_size = self.settings['conv']['kernel_size'][layer_idx],
                              strides = self.settings['conv']['strides'][layer_idx],
                              padding = self.settings['conv']['padding']
                            )(inputs)

            if self.settings['pooling']['apply'] == True:
                inputs = MaxPooling2D(
                                  pool_size = self.settings['pooling']['pool_size'][layer_idx],
                                  strides = self.settings['pooling']['strides'][layer_idx],
                                  padding = self.settings['pooling']['padding']
                                )(inputs)

            inputs = Activation(
                                activation = self.settings['detector_stage'][layer_idx]
                            )(inputs)

        inputs = Flatten()(inputs)

        for dense_layer_idx in range(self.settings['dense']['layers']):

            if self.settings['dense']['activations'][dense_layer_idx] != 'softmax':
                inputs = Dense(
                                units = self.settings['dense']['output_units'][dense_layer_idx],
                                activation = self.settings['dense']['activations'][dense_layer_idx]
                             )(inputs)
            else:
                final_output = Dense(
                                units = self.settings['dense']['output_units'][dense_layer_idx],
                                activation = self.settings['dense']['activations'][dense_layer_idx]
                             )(inputs)

        self.model = Model(inputs = inputs, outputs = final_output)

    def check_network_settings(self):

        for key in self.settings:

            if key == 'conv':

                if set(self.settings['conv'].keys()) != {'layers', 'filters', 'kernel_size', 'strides', 'padding'}:
                    print('[INCORRECT SETTINGS]: Convolutional layers ...')
                    return False

            elif key == 'pooling':

                if set(self.settings['pooling'].keys()) != {'apply', 'pool_size', 'strides', 'padding'}:
                    print('[INCORRECT SETTINGS]: Pooling layers ...')
                    return False

                if len(self.settings['pooling']['apply']) != self.settings['conv']['layers']:
                    print('Please specify wether or not to apply pooling for each convolutional layer')
                    return False

            elif key == 'detector_stage':

                if self.settings['conv']['layers'] != len(self.settings['detector_stage']):
                    print('Number of activation functions you have specified does not match the number of convolutional layers inside the network ...')
                    return False

            elif key == 'dense':

                if set(self.settings['dense'].keys()) != {'layers', 'output_units', 'activations'}:
                    print('[INCORRECT SETTINGS]: Dense layers ...')
                    return False

                if 'softmax' != self.settings['dense']['activations'][len(self.settings['dense']['activations'])-1]:
                    print('Your network must contain Softmax activation function at the last Dense layer in order to produce class probabilities ...')
                    return False

        print('Network settings have been correctly specified ...')
        return True

以下是我作为类构造函数的参数提供的设置:

settings = {
    'conv':
        {
         'layers': 3,
         'filters': [32, 64, 128],
         'kernel_size':[(3,3), (5,5), (5,5)],
         'strides': [1, 1, 1],
         'padding': 'same',
        },
    'pooling':
        {
         'apply': [True, True, True],
         'pool_size': [(2,2), (3,3), (3,3)],
         'strides': [1, 1, 1],
         'padding': 'same'
        },
    'detector_stage': ['relu', 'relu', 'relu'],
    'dense':
        {
          'layers': 2,
          'output_units': [500, 10],
          'activations': ['relu', 'softmax'],
        },
    'batch_norm': [False, False, False, False]
}

Answer 1

问题是 inputs 变量具有第一个 Dense 层的输出张量，而不是实际输入。