使用 R 中的 keras 使用 Pima Indian 糖尿病数据集来拟合 Conv1d 以对糖尿病患者进行分类

Question

我是机器学习领域的新手。我在 R 中使用 keras 包对糖尿病个体进行分类，使用 Pima Indian 糖尿病数据集并拟合 Conv1d。我不确定我在哪里做错了，但会出现一些错误。

1:8 列是特征，第 9 列是我们的标签，编码为 0 和 1。

数据维度如下：

dim (data)=(768,9)#Main data dimension

dim (training)=(534, 8)#Training set dimension

dim (test)=(234,8)#Testing set dimension

请帮我找出错误。提前致谢。

data=read.csv(file.choose(), header=T)


data=as.matrix(data)
head(data)
dimnames(data)=NULL
str(data)

#Data Normalization
data[, 1:8]=normalize(data[, 1:8])
summary(data)


#Data Partition 
set.seed(1234) 

ind=sample(2, nrow(data), replace=T, prob=c(0.7,0.3))
training=data[ind==1, 1:8]
test=data[ind==2, 1:8]
trainingtarget=data[ind==1, 9]
testtarget=data[ind==2, 9]


# reshape
dim(training) <- c(nrow(training), 8, 1)
dim(test) <- c(nrow(test), 8, 1)

dim(trainingtarget) <- c(nrow(trainingtarget), 1, 1)
dim(testtarget) <- c(nrow(testtarget), 1, 1)



cnn_bc_model <- keras_model_sequential()

cnn_bc_model %>%
  layer_conv_1d(filter=32, kernel_size=5, padding="same", input_shape=c(8,1)) %>%
  layer_activation("relu") %>%

  layer_conv_1d(filter=32, kernel_size=3) %>%
  layer_activation("relu") %>%

  layer_max_pooling_1d(pool_size=2) %>%
  layer_dropout(0.2) %>%

  layer_dense(units=1, activation='sigmoid')

cnn_bc_model %>% compile(loss='binary_crossentropy', optimizer='nadam', metrics = c('accuracy'))
cnn_bc_history <- cnn_bc_model %>% fit(training, trainingtarget, epochs=100, 
                                       batch_size=8, validation_split=0.2)

#Now evaluate the model with our test data set
cnn_bc_model %>%
  evaluate(test, testLabels)

错误：

Error in py_call_impl(callable, dots$args, dots$keywords) : 
  ValueError: A target array with shape (534, 1, 1) was passed for output of shape (None, 3, 1) while using as loss `binary_crossentropy`. This loss expects targets to have the same shape as the output.

Answer 1

问题在于max_pooling_1d 层的形状。对于您的特定模型，您有以下图层形状。

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv1d_3 (Conv1D)            (None, 8, 32)             192       
_________________________________________________________________
activation_4 (Activation)    (None, 8, 32)             0         
_________________________________________________________________
conv1d_4 (Conv1D)            (None, 6, 32)             3104      
_________________________________________________________________
activation_5 (Activation)    (None, 6, 32)             0         
_________________________________________________________________
max_pooling1d_2 (MaxPooling1 (None, 3, 32)             0         
_________________________________________________________________
dropout_2 (Dropout)          (None, 3, 32)             0         
_________________________________________________________________
dense_2 (Dense)              (None, 3, 1)              33        
_________________________________________________________________
activation_6 (Activation)    (None, 3, 1)              0         
=================================================================
Total params: 3,329
Trainable params: 3,329
Non-trainable params: 0
_________________________________________________________________

您网络的输出形状为(None,3,1)，与先前定义的目标形状(None,1,1) 不匹配。为了解决这个问题，您可以在max_pooling_1dlayer 和dropoutlayer 之间为您的模型添加一个layer_flatten 层。

max_pooling1d_3 (MaxPooling1 (None, 3, 32)             0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 96)                0         
_________________________________________________________________
dropout_3 (Dropout)          (None, 96)                0         
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 97        
_________________________________________________________________
activation_9 (Activation)    (None, 1)                 0         

或者您可以在最后一个dense 层之后添加layer_flatten，然后再添加一个dense层。

activation_12 (Activation)   (None, 3, 1)              0         
_________________________________________________________________
flatten_2 (Flatten)          (None, 3)                 0         
_________________________________________________________________
dense_5 (Dense)              (None, 1)                 4         
_________________________________________________________________
activation_13 (Activation)   (None, 1)                 0         

如果您难以跟踪所有图层形状，请在开始训练模型之前尝试添加summary(cnn_bc_model)，以便调试任何与形状相关的问题。希望对你有帮助

Answer 2

我按照您的建议尝试了两种方法，但仍然出现错误。我使用了以下代码。

data=read.csv(file.choose(), header=T)


data=as.matrix(data)
head(data)
dimnames(data)=NULL
str(data)

#Data Normalization
data[, 1:8]=normalize(data[, 1:8])
summary(data)


#Data Partition 
set.seed(1234) 

ind=sample(2, nrow(data), replace=T, prob=c(0.7,0.3))
training=data[ind==1, 1:8]
test=data[ind==2, 1:8]
trainingtarget=data[ind==1, 9]
testtarget=data[ind==2, 9]


# reshape
dim(training) <- c(nrow(training), 8, 1)
dim(test) <- c(nrow(test), 8, 1)

dim(trainingtarget) <- c(534, 1, 1)#534 is the length of this series
dim(testtarget) <- c(234, 1, 1)#234 is the length of this series


cnn_bc_model <- keras_model_sequential()
cnn_bc_model %>%
  layer_conv_1d(filter=32, kernel_size=5, input_shape=c(8,1)) %>%
  layer_activation("relu") %>%
  layer_conv_1d(filter=32, kernel_size=3) %>%
  layer_activation("relu") %>%
  layer_max_pooling_1d(pool_size=2) %>%
  layer_flatten() %>%
  layer_dropout(0.2) %>%
  layer_dense(units=1, activation='sigmoid') %>%
  layer_activation("softmax")

summary(cnn_bc_model)

cnn_bc_model %>% compile(loss='binary_crossentropy', optimizer='nadam', metrics = c('accuracy'))
cnn_bc_history <- cnn_bc_model %>% fit(training, trainingtarget, epochs=100, 
                                       batch_size=8, validation_split=0.2)

#Now evaluate the model with our test data set
cnn_bc_model %>%
  evaluate(test, testLabels)

错误现在看起来不同了，如下所示：

py_call_impl(callable, dots$args, dots$keywords) 中的错误： InvalidArgumentError：无法使用形状为 [8] 的张量更新形状为 [] 的变量，形状必须相等。 [[{{node metrics_22/acc/AssignAddVariableOp}}]]