Keras功能CNN模型出错：图形在主输入层断开连接
我正在为CNN开发一个功能性Keras模型，在R中具有1d输入层。
当我运行keras_model函数来构建模型时，我得到以下错误：
py\u call\u impl（可调用，dots$args，dots$keywords）中出错：值错误：图形断开连接：无法获取“main\u input”层的张量张量（“main\u input\u 15:0”，shape=（4201，1024），dtype=float32）的值。访问了以下以前的层，没有问题：[]
详细回溯：文件“/库/框架/Python.framework/Versions/3.7/lib/python3.7/site-packages/keras/legacy/interfaces.py”，第91行，在wrapper return func（*args，**kwargs）文件“/库/框架/Python.framework/Versions/3.7/lib/python3.7/site-packages/keras/engine/network.py”，第93行，在init self中_init\u graph\u network（*args，**kwargs）文件“/库/框架/Python.framework/Versions/3.7/lib/python3.7/site packages/keras/engine/network.py”，第231行，在init\u graph\u network self中。输入，自我。输出）文件“/库/框架/Python.framew
我附上我的代码，任何帮助都将不胜感激。
main_input = layer_input(shape = c(1024), batch_shape = c(4201,1024), dtype = 'float32', name = 'main_input') %>%
layer_reshape( target_shape = list(1024,1), input_shape = c(1024),dtype = 'float32', batch_input_shape = c(4201, 1024), batch_size = 4201)
conv1 = layer_conv_1d(filters = 64, kernel_size = 10, strides = 5, dtype = 'float32', activation = 'relu' )
max1 = layer_max_pooling_1d(pool_size = 10)
first_conv = main_input %>% conv1%>%max1
conv2 = layer_conv_1d(filters = 32, kernel_size = 5, strides = 3, dtype = 'float32', activation = 'relu' )
max2 = layer_max_pooling_1d(pool_size = 5)
second_conv = first_conv %>% conv2 %>% max2
conc1 = second_conv %>% layer_flatten()
semantic_input = layer_input(shape = c(2074), dtype = 'float32', batch_shape = c(4201,2074), name = 'semantic_input')%>%
layer_reshape(target_shape = list(2074,1), input_shape = c (2074), dtype = 'float32')
conc2 = semantic_input %>% layer_flatten()
output = layer_concatenate(c(conc1, conc2)) %>%
layer_dense( units = 100, activation = 'relu', use_bias = TRUE) %>%
layer_dense(units = 50, activation = 'relu', use_bias = TRUE) %>%
layer_dense(units = 25, activation = 'relu', use_bias = TRUE)%>%
layer_dense(units = 10, activation = 'relu', use_bias = TRUE)%>%
layer_dense(units = 1, activation = 'softmax', name = 'output')
cnn1_model = keras_model(
inputs = c(main_input,semantic_input),
outputs = c(output)
)
当我尝试构建模型时，在代码的最后一行中出现了上述错误。
共有1个答案
我已经想通了，经过2天的封锁！
定义两个输入层时，不应重塑它们。重塑可以在下一步中进行，输入层应该独立声明。
以下是固定代码：
main_input = layer_input(shape = c(1024), batch_shape = c(4201,1024), dtype = 'float32', name = 'main_input')
main_reshaped = main_input %>% layer_reshape( target_shape = list(1024,1), input_shape = c(1024),dtype = 'float32', batch_input_shape = c(4201, 1024), batch_size = 4201)
conv1 = layer_conv_1d(filters = 64, kernel_size = 10, strides = 5, dtype = 'float32', activation = 'relu' )
max1 = layer_max_pooling_1d(pool_size = 10)
conv2 = layer_conv_1d(filters = 32, kernel_size = 5, strides = 3, dtype = 'float32', activation = 'relu' )
max2 = layer_max_pooling_1d(pool_size = 5)
conv = reshaped %>% conv1%>%max1%>%conv2 %>% max2 %>% layer_flatten()
semantic_input = layer_input(shape = c(2074), dtype = 'float32', batch_shape = c(4201,2074), name = 'semantic_input')
sem_reshaped = semantic_input %>% layer_reshape(target_shape = list(2074,1), input_shape = c (2074), dtype = 'float32')
conc = sem_reshaped %>% layer_flatten()
output = layer_concatenate(c(conv, conc)) %>%
layer_dense( units = 100, activation = 'relu', use_bias = TRUE) %>%
layer_dense(units = 50, activation = 'relu', use_bias = TRUE) %>%
layer_dense(units = 25, activation = 'relu', use_bias = TRUE)%>%
layer_dense(units = 10, activation = 'relu', use_bias = TRUE)%>%
layer_dense(units = 1, activation = 'softmax', name = 'output')
cnn1_model = keras_model(
inputs = c(main_input,semantic_input),
outputs = c(output)
)
所以模型是这样的
summary (cnn1_model)
_______________________________________________________________________________________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
===============================================================================================================================================================================
main_input (InputLayer) (4201, 1024) 0
_______________________________________________________________________________________________________________________________________________________________________________
reshape_25 (Reshape) (4201, 1024, 1) 0 main_input[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
conv1d_65 (Conv1D) (4201, 203, 64) 704 reshape_25[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
max_pooling1d_50 (MaxPooling1D) (4201, 20, 64) 0 conv1d_65[6][0]
_______________________________________________________________________________________________________________________________________________________________________________
conv1d_66 (Conv1D) (4201, 6, 32) 10272 max_pooling1d_50[6][0]
_______________________________________________________________________________________________________________________________________________________________________________
semantic_input (InputLayer) (4201, 2074) 0
_______________________________________________________________________________________________________________________________________________________________________________
max_pooling1d_51 (MaxPooling1D) (4201, 1, 32) 0 conv1d_66[5][0]
_______________________________________________________________________________________________________________________________________________________________________________
reshape_26 (Reshape) (4201, 2074, 1) 0 semantic_input[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
flatten_35 (Flatten) (4201, 32) 0 max_pooling1d_51[5][0]
_______________________________________________________________________________________________________________________________________________________________________________
flatten_36 (Flatten) (4201, 2074) 0 reshape_26[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
concatenate_38 (Concatenate) (4201, 2106) 0 flatten_35[0][0]
flatten_36[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
dense_77 (Dense) (4201, 100) 210700 concatenate_38[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
dense_78 (Dense) (4201, 50) 5050 dense_77[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
dense_79 (Dense) (4201, 25) 1275 dense_78[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
dense_80 (Dense) (4201, 10) 260 dense_79[0][0]
_______________________________________________________________________________________________________________________________________________________________________________
output (Dense) (4201, 1) 11 dense_80[0][0]
===============================================================================================================================================================================
Total params: 228,272
Trainable params: 228,272
Non-trainable params: 0
_______________________________________________________________________________________________________________________________________________________________________________
-
我正在使用Keras构建一个CNN，以下Conv1D是我的第一层： 我正在培训以下功能： 其中，train\u df是一个由两列组成的pandas数据帧，其中，对于每一行，标签是一个int（0或1），有效载荷是一个用零填充/截断为1000的浮点数组。train\U df中的培训示例总数为15641。 模型编译，但在训练期间，我得到这个错误： 我看了这篇文章，试图将输入更改为1000个浮点长列表的数
-
问题内容： 在Keras中创建顺序模型时，我知道您在第一层中提供了输入形状。然后，此输入形状会构成 隐式 输入层吗？ 例如，下面的模型明确指定了2个密集层，但这实际上是一个3层模型，即由输入形状隐含的一个输入层，一个具有32个神经元的隐藏密集层，然后一个具有10个可能输出的输出层组成的模型吗？ 问题答案： 好吧，实际上它实际上 是 一个隐式输入层，即您的模型是一个具有三层“输入，隐藏和输出”的“老
-
我用python用800个样本训练了一个CNN神经网络，并用60个样本进行了测试。输出精度是50，现在每次我使用模型。预测它会给我同样的结果。 我用了keras和tensorflow。图像为224x224像素，每个像素分为两类。我对神经网络知之甚少，这是我第一次尝试把它做得这么大。我听说它可能太合适了，或者我需要一个更重要的图层，或者我的批量大小/年代/学习率是错误的。 编辑1：种子对网络培训有何
-
我正在开发一个程序，该程序用一个无方向的Conv2D层（跨距=1）替换跨距的Conv2D层（跨距=2），然后在激活层之后添加一个AveragePooling2D层（跨距=2）。换句话说，AveragePooling2D层将减少输出维度，而不是让Conv2D层减少输出维度。 我使用本文中描述的方法将Conv2D层替换为非结构化版本，并在激活后插入averagepoolig2d层。替换Conv2D层效
-
我正试图使用Keras库实现一个序列2序列模型。模型框图如下 模型将输入序列嵌入到3D张量中。然后双向lstm创建编码层。接下来，编码序列被发送到自定义关注层，该层返回一个2D张量，该张量具有每个隐藏节点的关注权重。 解码器输入作为一个热向量注入模型中。现在在解码器（另一个bi lstm）中，解码器输入和注意权重都作为输入传递。解码器的输出通过softmax激活函数发送到时间分布密集层，以概率的方
-
在模型中添加LSTM层之前，我不确定是否需要添加密集输入层。例如，使用以下模型： LSTM层是否为输入层，密集层是否为输出层（即无隐藏层）？或者Keras是否创建了一个输入层，这意味着LSTM层将是一个隐藏层？