실습: 딥러닝으로 MNIST 98%이상 해보기 PART 5 (DROP OUT)

---

개요(Abstract)
아래의 내용은 hunkim님의 모두를 위한 머신러닝/딥러닝 강의에서 tensorFlow 사용 예제를 R로 변경한 내용이다.

hunkim님의 강좌 주소는 http://hunkim.github.io/ml/ 이다.

참조 : hunkim님의 모두를 위한 머신러닝/딥러닝 강의

---

아래의 내용은 훈련 중 뉴런을 비활성화하여 학습하는 예제이다.
훈련 단계에서 tensorflow에 에러가 발생할 경우는 R session을 재시작 해야 한다.

library(tensorflow)
library(ggplot2)
library(reshape2)
library(stringr)

# Check out https://www.tensorflow.org/get_started/mnist/beginners for
# more information about the mnist dataset
datasets <- tf$contrib$learn$datasets
mnist <- datasets$mnist$read_data_sets("MNIST-data", one_hot = TRUE)

# parameters
learning_rate = 0.001
batch_size = 100L
num_epochs = 50
num_iterations = as.integer(mnist$train$num_examples / batch_size)


# input place holders
X = tf$placeholder(tf$float32, shape=list(NULL,784L))
Y = tf$placeholder(tf$float32, shape=list(NULL, 10L))

# weights & bias for nn layers
drop_rate = tf$placeholder(tf$float32)

W1 = tf$get_variable(name="weight1", shape=list(784L, 512L), 
                     initializer = tf$contrib$layers$xavier_initializer())
b1 = tf$Variable(initial_value=tf$random_normal(shape=list(512L)), name="biase1")

W2 = tf$get_variable(name="weight2", shape=list(512L, 512L), 
                     initializer = tf$contrib$layers$xavier_initializer())
b2 = tf$Variable(initial_value=tf$random_normal(shape=list(512L)), name="biase2")

W3 = tf$get_variable(name="weight3", shape=list(512L, 512L), 
                     initializer = tf$contrib$layers$xavier_initializer())
b3 = tf$Variable(initial_value=tf$random_normal(shape=list(512L)), name="biase3")

W4 = tf$get_variable(name="weight4", shape=list(512L, 512L), 
                     initializer = tf$contrib$layers$xavier_initializer())
b4 = tf$Variable(initial_value=tf$random_normal(shape=list(512L)), name="biase4")

W5 = tf$get_variable(name="weight5", shape=list(512L, 10L), 
                     initializer = tf$contrib$layers$xavier_initializer())
b5 = tf$Variable(initial_value=tf$random_normal(shape=list(10L)), name="biase5")

L1 = tf$nn$relu(tf$matmul(X, W1) + b1)
L1 = tf$nn$dropout(L1, rate = drop_rate)

L2 = tf$nn$relu(tf$matmul(L1, W2)) + b2
L2 = tf$nn$dropout(L2, rate = drop_rate)

L3 = tf$nn$relu(tf$matmul(L2, W3)) + b3
L3 = tf$nn$dropout(L3, rate = drop_rate)

L4 = tf$nn$relu(tf$matmul(L3, W4)) + b4
L4 = tf$nn$dropout(L4, rate = drop_rate)

hypothesis = tf$matmul(L4, W5) + b5

# softmax_cross_entropy_with_logits_v2
cost = tf$reduce_mean(
    tf$nn$softmax_cross_entropy_with_logits_v2(logits=hypothesis, labels=tf$stop_gradient(Y))
)
train = tf$train$AdamOptimizer(learning_rate=learning_rate)$minimize(cost)

correct_prediction = tf$equal(x=tf$argmax(input=hypothesis, axis=1L), y=tf$argmax(input=Y, axis=1L))
accuracy = tf$reduce_mean(tf$cast(x=correct_prediction,dtype=tf$float32))


# train my model
sess <- tf$Session()

# initialize
sess$run(fetches=tf$global_variables_initializer())

for (epoch in 1:num_epochs) {
    avg_cost = 0
    
    for (interation in 1:num_iterations) {
        
        batches = mnist$train$next_batch(batch_size)
        batch_xs <- batches[[1]]
        batch_ys <- batches[[2]]
        
        stage = sess$run(fetches=list(train, cost), 
                         feed_dict=dict(X = batch_xs, Y = batch_ys, drop_rate = 0.3))
        avg_cost = avg_cost + stage[[2]] / num_iterations
    }
    
    cat("Epoch:", str_pad((epoch + 1), 4, pad = "0"), "Cost:", avg_cost, "\n")
}

cat("\nLearning Finished!\n")

# Test model and check accuracy
accur <- sess$run(fetches=list(accuracy), 
                  feed_dict=dict(X = mnist$test$images, Y = mnist$test$labels, drop_rate = 0))
cat("Accuracy:", accur[[1]], "\n")

# Learning Finished!
# Accuracy: 0.9204

pred <- sess$run(tf$argmax(hypothesis, axis=1L), feed_dict=dict(X = mnist$test$images, drop_rate = 0))

lbls <- apply(mnist$test$labels, 1, function(row){ (which.max(row) - 1) })
table(PRED = pred, OBSRV = lbls)

sess$close()


# 이미지를 그래프로 보는 방법은
# https://tensorflow.rstudio.com/tfestimators/articles/examples/mnist.html 를 참조

# 0 인 글자 중 예측 오류 글자 확인
tmp.df <- as.data.frame(cbind(pred=pred, lbls = lbls))

indices <- which(tmp.df$pred != 0 & tmp.df$lbls == 0)
ifelse(length(indices) <= 36, indices <- indices, indices <- sample(indices, 36))
indices <- indices[order(indices)]

n <- length(indices)
data <- array(mnist$test$images[indices, ], dim = c(n, 28, 28))
melted <- melt(data, varnames = c("image", "x", "y"), value.name = "intensity")
ggplot(melted, aes(x = x, y = y, fill = intensity)) +
    geom_tile() +
    scale_fill_continuous(name = "Pixel Intensity") +
    scale_y_reverse() +
    facet_wrap(~ image, nrow = sqrt(n), ncol = sqrt(n)) +
    theme(
        strip.background = element_blank(),
        strip.text.x = element_blank(),
        panel.spacing = unit(0, "lines"),
        axis.text = element_blank(),
        axis.ticks = element_blank()
    ) +
    labs(
        title = "MNIST Image Data",
        subtitle = "Visualization of a sample of images contained in MNIST data set.",
        x = NULL,
        y = NULL
    )


print("END OF CODE")