使用R语言neuralnet包训练神经网络时遇维度错误求助

报错信息 #

Error in x[0, , drop = FALSE] : número incorreto de dimensiones

原问题代码 #

set.seed(123)
library(neuralnet)
library(caret)

# data1: training data
sex <- factor(sample(c("male", "female"), size = 1000, replace = TRUE))
height <- rnorm(1000, mean = c(1.8, 1.6)[sex])
weight <- rnorm(1000, mean = c(80, 60)[sex])
data1 <- data.frame(sex, height, weight)

# data2: population data
sex <- factor(sample(c("male", "female"), size = 10000, replace = TRUE))
height <- rnorm(1000, mean = c(1.81, 1.61)[sex])
weight <- rnorm(1000, mean = c(81, 61)[sex])
data2 <- data.frame(sex, height, weight)

# training 

net  <- neuralnet(sex ~ height + weight, 
                  data = data1, hidden = c(10, 10), 
                  linear.output = FALSE, 
                  threshold = 0.1, stepmax = 10000)

trained_net <- train(net, data, error.fun = "ce", stepmax = 1e+07)

# Prediction and results

predictions <- compute(trained_net, data2[, c("height", "weight")])
predictions_df <- data.frame(predictions$net.result)
colnames(predictions_df) <- c("probfemale")
result <- cbind(datos2, predictions_df)
                  result$prediccion_sex <- ifelse(resul$probfemale >= 0.5, "female", "male")

问题分析与修正点 #

• 数据维度不匹配：data2中height和weight生成时用了rnorm(1000,...)，但sex是10000行，导致列维度不一致，这是核心错误来源。
• train()函数误用：caret的train()不能直接传入neuralnet训练好的对象，且代码中data变量未定义，无需用caret二次训练。
• 变量名拼写错误：预测部分的datos2应为data2，resul应为result，拼写错误会导致变量找不到。
• 因变量格式问题：neuralnet对因子类型的因变量支持不佳，建议转为二元数值（如female=1，male=0），更适合分类概率预测。

修正后完整代码 #

set.seed(123)
library(neuralnet)

# 生成训练数据：将sex转为二元数值
sex <- sample(c(0, 1), size = 1000, replace = TRUE)  # 0=male, 1=female
height <- rnorm(1000, mean = c(1.8, 1.6)[sex + 1])  # R索引从1开始，+1匹配mean的两个值
weight <- rnorm(1000, mean = c(80, 60)[sex + 1])
data1 <- data.frame(sex, height, weight)

# 生成测试数据：修正行数匹配问题
sex <- sample(c(0, 1), size = 10000, replace = TRUE)
height <- rnorm(10000, mean = c(1.81, 1.61)[sex + 1])
weight <- rnorm(10000, mean = c(81, 61)[sex + 1])
data2 <- data.frame(sex, height, weight)

# 训练神经网络：直接用neuralnet，无需caret的train
net <- neuralnet(sex ~ height + weight, 
                 data = data1, 
                 hidden = c(10, 10), 
                 linear.output = FALSE,  # 分类问题设为FALSE，输出概率
                 threshold = 0.1, 
                 stepmax = 1e5)  # 增大stepmax避免训练提前终止

# 预测并整理结果
predictions <- compute(net, data2[, c("height", "weight")])
predictions_df <- data.frame(probfemale = predictions$net.result)
result <- cbind(data2, predictions_df)
# 将原始sex转回因子方便对比，生成预测性别
result$sex <- factor(result$sex, levels = c(0,1), labels = c("male", "female"))
result$prediccion_sex <- ifelse(result$probfemale >= 0.5, "female", "male")

# 查看前几行结果
head(result)

关键说明 #

• 把因变量转为二元数值后，neuralnet能更稳定处理分类任务，linear.output=FALSE确保输出是sigmoid激活后的概率值。
• 修正了data2的行数问题，保证所有列维度一致。
• 移除了不必要的caret包调用，简化训练流程，避免函数误用。
• 修正了变量拼写错误，确保代码可运行。

内容的提问来源于stack exchange，提问作者Fran