使用 mx.net 包 r 中的 mx.mlp 时出错

Question

我已经使用

在 R 版本 3.4.3 中安装了 mx.net 包

cran

cran["dmlc"] https://s3-us-west-2.amazonaws.com/apache-mxnet/R/CRAN/"

选项（repos = cran）

install.packages("mxnet").

估计神经网络时会出现一些问题。

在下面的代码中，我使用了可用的乳腺癌数据集 https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Original%29.

BC_data <- read.csv("Data_breastcancer.csv", sep = ";")

# generate a train and test set
trainIndex = sample(1:nrow(BC_data), size = round(0.8*nrow(BC_data)), replace=FALSE)
train_data <- BC_data[trainIndex,]
test_data <- BC_data[-trainIndex,]
X_train <- train_data[,c(-1,-11)]
y_train <- train_data[,11]   

# estimate neural network
model = mx.mlp(as.matrix(X_train), as.numeric(y_train), hidden_node = 10, out_node = 2, out_activation = "softmax", learning.rate = 0.1, num.round = 20)

但是，我得到的唯一输出是

，而不是迭代地返回精度值

Start training with 1 devices
Warning message:
In mx.model.select.layout.train(X, y) :
Auto detect layout of input matrix, use rowmajor..

所以看起来迭代过程根本没有开始。

有人知道如何解决这个问题吗？

Answer 1

您需要做一些调整：

1. MxNet 并不真正适用于任意名称的类。在您的情况下，类标签是“2”和“4”。您需要将它们转换为从 0 或更高开始的数字（参见下面代码中的示例）
2. 您需要将准确度作为指标提供给 mlp。文档中有一个错误：它说参数名称是“eval_metric”，但实际上它是“eval.metric”，如 mx.model.FeedForward.create 函数
3. 初始数据集包含“？”作为 NA 值。 Mlp 不适用于 NA，因此您需要用某些东西替换它们。我选择了 0，但如果需要，您可以使用更具体的域。

这是适用于 Breast-cancer-wisconsin.data 文件的代码，我从您的链接中找到：

library(mxnet)
BC_data <- read.csv(
  file = "breast-cancer-wisconsin.data", 
  sep = ",", 
  header = FALSE, 
  colClasses = c(rep("numeric", 11)), 
  na.strings = c("", "?") # Few records of the 7th column contain "?" - treat "?" as NA
)

BC_data[is.na(BC_data)] <- 0 # Replace NA with zeroes

# generate a train and test set
trainIndex = sample(1:nrow(BC_data), size = round(0.8*nrow(BC_data)), replace=FALSE)
train_data <- BC_data[trainIndex,]
test_data <- BC_data[-trainIndex,]
X_train <- train_data[,c(-1,-11)]
y_train <- train_data[,11]   

# estimate neural network
model = mx.mlp(
    data = as.matrix(X_train), 
    label = as.numeric(ifelse(y_train == 2, 0, 1)), # Replace classes with 0 and 1
    hidden_node = 10, 
    out_node = 2, 
    out_activation = "softmax", 
    learning.rate = 0.1, 
    num.round = 20,
    array.layout = "rowmajor", # get rid of a nasty warning
    eval.metric=mx.metric.accuracy # set Accuracy as a metric
  )

如果我运行这段代码，我会得到以下输出：

Start training with 1 devices
[1] Train-accuracy=0.64453125
[2] Train-accuracy=0.6515625
[3] Train-accuracy=0.65625
[4] Train-accuracy=0.9
[5] Train-accuracy=0.95625
[6] Train-accuracy=0.95
[7] Train-accuracy=0.94375
[8] Train-accuracy=0.9328125
[9] Train-accuracy=0.93125
[10] Train-accuracy=0.9328125
[11] Train-accuracy=0.9375
[12] Train-accuracy=0.9390625
[13] Train-accuracy=0.9484375
[14] Train-accuracy=0.95
[15] Train-accuracy=0.9453125
[16] Train-accuracy=0.946875
[17] Train-accuracy=0.9484375
[18] Train-accuracy=0.95
[19] Train-accuracy=0.9484375
[20] Train-accuracy=0.9515625