Using keras in R to perform neural network, my model has very low accuracy but the prediction is good and I don't know why

I used the classic dataset – mnist dataset that has 784 columns of pixels and 1 column of the label (from 0 to 9), and I was going to transform the images into their corresponding seven segment display representation. The following is my code.

# Convert the labels(digits) in train_set and test_set to seven-segment display
train_set$a <- ifelse(train_set$V1 %in% c(0,2,3,5,6,7,8,9),1,0)
train_set$b <- ifelse(train_set$V1 %in% c(0,1,2,3,4,7,8,9),1,0)
train_set$c <- ifelse(train_set$V1 %in% c(0,1,3,4,5,6,7,8,9),1,0)
train_set$d <- ifelse(train_set$V1 %in% c(0,2,3,5,6,8),1,0)
train_set$e <- ifelse(train_set$V1 %in% c(0,2,6,8),1,0)
train_set$f <- ifelse(train_set$V1 %in% c(0,4,5,6,8,9),1,0)
train_set$g <- ifelse(train_set$V1 %in% c(2,3,4,5,6,8,9),1,0)

test_set$a <- ifelse(test_set$V1 %in% c(0,2,3,5,6,7,8,9),1,0)
test_set$b <- ifelse(test_set$V1 %in% c(0,1,2,3,4,7,8,9),1,0)
test_set$c <- ifelse(test_set$V1 %in% c(0,1,3,4,5,6,7,8,9),1,0)
test_set$d <- ifelse(test_set$V1 %in% c(0,2,3,5,6,8),1,0)
test_set$e <- ifelse(test_set$V1 %in% c(0,2,6,8),1,0)
test_set$f <- ifelse(test_set$V1 %in% c(0,4,5,6,8,9),1,0)
test_set$g <- ifelse(test_set$V1 %in% c(2,3,4,5,6,8,9),1,0)

# Split the given train data to train_x and train_y
# Reshaping the training pixels and labels data to arrays
train_x <- as.matrix(train_set[, 2:785])
train_x <- array_reshape(train_x, c(nrow(train_x), 784))
train_y <- as.matrix(train_set[, 786:792])
train_y <- array_reshape(train_y, c(nrow(train_y), 7))

# Split the given test data to test_x and test_y
# Reshaping the testing pixels and labels data to arrays
test_x <- as.matrix(test_set[, 2:785])
test_x <- array_reshape(test_x, c(nrow(test_x), 784))
test_y <- as.matrix(test_set[, 786:792])
test_y <- array_reshape(test_y, c(nrow(test_y), 7))

# Normalize inputs from 0-255 to 0-1
train_x <- train_x / 255
test_x <- test_x / 255

# Build the Model
image_size <- 784 # 28*28
num_classes <- 7 #7 segment display of the digits
model <- keras_model_sequential() 
model %>%
  #Hidden Layers
  layer_dense(units = 512, activation = 'relu', input_shape = c(image_size)) %>%
  layer_dropout(rate = 0.25) %>%
  layer_dense(units = 256, activation = 'relu') %>%
  layer_dropout(rate = 0.5) %>%
  # Output Layer
  layer_dense(units = num_classes, activation = 'sigmoid')

# Summary of the model
summary(model)

# Compile the neural network
model %>% compile(
  loss = 'binary_crossentropy',
  optimizer = 'adam',
  metrics = c('accuracy')
)

# Modeling on Training Dataset
model %>% fit(
  train_x, train_y,
  epochs = 5, batch_size = 128,
  validation_data = list(test_x, test_y)
)
# Prediction
pred <- predict_proba(model, test_x)
pred <- round(as.data.frame(pred))

test_set$predict <- ifelse(pred$V1==1 & pred$V2==1 & pred$V3==1 & pred$V4==1
                           & pred$V5==1 & pred$V6==1 & pred$V7==0,0,NA)

test_set$predict <- ifelse(pred$V1==0 & pred$V2==1 & pred$V3==1 & pred$V4==0
                           & pred$V5==0 & pred$V6==0 & pred$V7==0,1,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==1 & pred$V3==0 & pred$V4==1
                           & pred$V5==1 & pred$V6==0 & pred$V7==1,2,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==1 & pred$V3==1 & pred$V4==1
                           & pred$V5==0 & pred$V6==0 & pred$V7==1,3,test_set$predict)

test_set$predict <- ifelse(pred$V1==0 & pred$V2==1 & pred$V3==1 & pred$V4==0
                           & pred$V5==0 & pred$V6==1 & pred$V7==1,4,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==0 & pred$V3==1 & pred$V4==1
                           & pred$V5==0 & pred$V6==1 & pred$V7==1,5,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==0 & pred$V3==1 & pred$V4==1
                           & pred$V5==1 & pred$V6==1 & pred$V7==1,6,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==1 & pred$V3==1 & pred$V4==0
                           & pred$V5==0 & pred$V6==0 & pred$V7==0,7,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==1 & pred$V3==1 & pred$V4==1
                           & pred$V5==1 & pred$V6==1 & pred$V7==1,8,test_set$predict)

test_set$predict <- ifelse(pred$V1==1 & pred$V2==1 & pred$V3==1 & pred$V4==0
                           & pred$V5==0 & pred$V6==1 & pred$V7==1,9,test_set$predict)

confusionMatrix(factor(test_set$predict), factor(test_set$V1))

It turned out that my model always had only around 20% or 30% accuracy. However, when I used the model to do the prediction and transformed the outputs back into labels, the accuracy was quite good like roughly 85% every time.
I don’t know what part is wrong with my model. Can someone help me out? Really appreciated!
The accuracy of my model:

The accuracy of my prediction:

The dataset can be downloaded here:
https://www.kaggle.com/zalando-research/fashionmnist
Here is the seven-segment display chart:
enter image description here