深度神经网络分类问题 MNIST

如何解决深度神经网络分类问题 MNIST

嗨，我正在尝试开发一个能够读取手写数字的神经网络 我从网上复制了这个模型（参考：https://www.youtube.com/watch?v=5bso_5X7Zu4&t=643s） 我的问题来自新数据，错误分类率很高（见帖子末尾）

    #Youtube link https://www.youtube.com/watch?v=5bso_5X7Zu4&t=785s
# MNIST data
library(keras)
mnist <- dataset_mnist()
trainx <- mnist$train$x
trainy <- mnist$train$y
testx <- mnist$test$x
testy <- mnist$test$y

# Plot images
par(mfrow = c(3,3))
for (i in 1:9) plot(as.raster(trainx[i,],max = 255))
par(mfrow= c(1,1))

# Five 
a <- c(1,12,36,48,66,101,133,139,146)
par(mfrow = c(3,3))
for (i in a) plot(as.raster(trainx[i,1))

# Reshape & rescale
trainx <- array_reshape(trainx,c(nrow(trainx),784))
testx <- array_reshape(testx,c(nrow(testx),784))
trainx <- trainx / 255
testx <- testx /255

# One hot encoding
trainy <- to_categorical(trainy,10)
testy <- to_categorical(testy,10)

# Model
model <- keras_model_sequential()
model %>% 
  layer_dense(units = 512,activation = 'relu',input_shape = c(784)) %>% 
  layer_dropout(rate = 0.4) %>% 
  layer_dense(units= 256,activation = 'relu') %>% 
  layer_dropout(rate = 0.3) %>% 
  layer_dense(units = 10,activation = 'softmax')

# Compile
model %>% 
  compile(loss = 'categorical_crossentropy',optimizer = optimizer_rmsprop(),metrics = 'accuracy')

# Fit model
history <- model %>% 
  fit(trainx,trainy,epochs = 30,batch_size = 32,validation_split = 0.2)

# Evaluation and Prediction - Test data
model %>% evaluate(testx,testy)
pred <- model %>% predict_classes(testx)
table(Predicted = pred,Actual = mnist$test$y)

prob <- model %>% predict_proba(testx)
cbind(prob,Predicted_class = pred,Actual = mnist$test$y)[1:5,]

# New data
library(EBImage)

temp = list.files(pattern = "*.jpg")
mypic <- list()
for (i in 1:length(temp)) {mypic[[i]] <- readImage(temp[[i]])}

par(mfrow = c(4,4))
for (i in 1:length(temp)) plot(mypic[[i]])


for (i in 1:length(temp)) {colorMode(mypic[[i]]) <- Grayscale}
for (i in 1:length(temp)) {mypic[[i]] <- 1-mypic[[i]]}
for (i in 1:length(temp)) {mypic[[i]] <- resize(mypic[[i]],28,28)}
str(mypic)
par(mfrow = c(4,5))
for (i in 1:length(temp)) plot(mypic[[i]])
for (i in 1:length(temp)) {mypic[[i]] <- array_reshape(mypic[[i]],c(28,3))}
new <- NULL

for (i in 1:length(temp)) {new <- rbind(new,mypic[[i]])}
newx <- new[,1:784]
newy <- c(7,5,2,3,4,7,6,8,6)

# Prediction
pred <- model %>% predict_classes(newx)
pred
table(Predicted = pred,Actual = newy)

prob <- model %>% predict_proba(newx)
cbind(prob,Predicted = pred,Actual = newy)

问题是新数据的分类

我已经获得了下一个预测（您可以从here下载图像）

7 5 2 0 5 3 8 2 2 6 6 6 3 5 5

前六个数字分类正确（视频中的数字）（7 5 2 0 5 3）但接下来的 9 个（我写的数字）显示出非常糟糕的结果。 我还尝试用训练数据数字测试我的 nnet，但它继续失败，我不明白为什么 :(

有什么想法吗？为什么会发生这种情况？ 谢谢