set.seed(100)
library(keras)
use_condaenv("r-tensorflow")
use_session_with_seed(100)
library(dplyr)
library(caret)
library(plotly)
# test library
model <- keras_model_sequential()train <- read.csv("fashionmnist/train.csv")
test <- read.csv("fashionmnist/test.csv")# mengecek jumlah kolom (variabel)
ncol(train)## [1] 785# mengecek nama kolom ke 1-5 dan ke 782-785
colnames(train)[c(1:5, 782:785)]## [1] "label" "pixel1" "pixel2" "pixel3" "pixel4" "pixel781"
## [7] "pixel782" "pixel783" "pixel784"# mengecek dimensi data (jumlah baris dan kolom)
dim(train)## [1] 60000 785# mengecek interval nilai data train
range(train)## [1] 0 255# mengecek proporsi setiap levels pada target
prop.table(table(train$label))##
## 0 1 2 3 4 5 6 7 8 9
## 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1Data yang disediakan sudah dalam bentuk training set dan testing set (sudah dilakukan cross validation)
train_x <- train %>%
select(-label) %>%
data.matrix()
train_x_keras <- array_reshape(train_x, dim = dim(train_x))/255
train_y <- train %>%
select(label) %>%
data.matrix()
train_y_keras <- to_categorical(train_y, num_classes = 10)test_x <- test %>%
select(-label) %>%
data.matrix()
test_x_keras <- array_reshape(test_x, dim = dim(test_x))/255
test_y <- test %>%
select(label) %>%
data.matrix()
test_y_keras <- to_categorical(train_y, num_classes = 10)# set-up reproducible weight initializer
initializer <- initializer_random_uniform(minval = -0.5, maxval = 0.5, seed = 100)
# menunjukkan bahwa objek model berupa sequence (membuat tempat arsitektur)
model <- keras_model_sequential()
# membuat arsitektur
model %>%
layer_dense(input_shape = 784, units = 256, activation = "relu", kernel_initializer = initializer, bias_initializer = initializer) %>%
layer_dense(units = 128, activation = "relu", kernel_initializer = initializer, bias_initializer = initializer) %>%
layer_dense(units = 68, activation = "relu", kernel_initializer = initializer, bias_initializer = initializer) %>%
layer_dense(units = 10, activation = "softmax", kernel_initializer = initializer, bias_initializer = initializer) %>%
compile(loss = "categorical_crossentropy",
optimizer = optimizer_adam(),
metrics = "accuracy")
summary(model)## ___________________________________________________________________________
## Layer (type) Output Shape Param #
## ===========================================================================
## dense (Dense) (None, 256) 200960
## ___________________________________________________________________________
## dense_1 (Dense) (None, 128) 32896
## ___________________________________________________________________________
## dense_2 (Dense) (None, 68) 8772
## ___________________________________________________________________________
## dense_3 (Dense) (None, 10) 690
## ===========================================================================
## Total params: 243,318
## Trainable params: 243,318
## Non-trainable params: 0
## ___________________________________________________________________________history <- model %>%
fit(train_x_keras, train_y_keras,
epoch = 20,
batch_size = 128,shuffle=F)ggplotly(plot(history))prediction <- model %>%
predict_classes(test_x_keras)
prediction[1:5]## [1] 0 1 2 2 3confusionMatrix(as.factor(prediction), as.factor(test_y))## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1 2 3 4 5 6 7 8 9
## 0 834 2 12 19 1 0 139 0 7 0
## 1 0 979 1 19 3 1 2 0 1 0
## 2 13 3 711 5 39 1 54 0 6 0
## 3 50 12 16 916 35 1 40 0 5 0
## 4 6 0 160 24 866 0 84 0 5 0
## 5 1 1 0 0 0 901 0 6 1 3
## 6 83 2 92 15 54 0 667 1 9 0
## 7 0 0 3 0 0 58 0 939 3 30
## 8 12 1 5 1 2 8 14 1 960 1
## 9 1 0 0 1 0 30 0 53 3 966
##
## Overall Statistics
##
## Accuracy : 0.8739
## 95% CI : (0.8672, 0.8803)
## No Information Rate : 0.1
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.8599
##
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: 0 Class: 1 Class: 2 Class: 3 Class: 4 Class: 5
## Sensitivity 0.8340 0.9790 0.7110 0.9160 0.8660 0.9010
## Specificity 0.9800 0.9970 0.9866 0.9823 0.9690 0.9987
## Pos Pred Value 0.8225 0.9732 0.8546 0.8521 0.7563 0.9869
## Neg Pred Value 0.9815 0.9977 0.9685 0.9906 0.9849 0.9891
## Prevalence 0.1000 0.1000 0.1000 0.1000 0.1000 0.1000
## Detection Rate 0.0834 0.0979 0.0711 0.0916 0.0866 0.0901
## Detection Prevalence 0.1014 0.1006 0.0832 0.1075 0.1145 0.0913
## Balanced Accuracy 0.9070 0.9880 0.8488 0.9492 0.9175 0.9498
## Class: 6 Class: 7 Class: 8 Class: 9
## Sensitivity 0.6670 0.9390 0.9600 0.9660
## Specificity 0.9716 0.9896 0.9950 0.9902
## Pos Pred Value 0.7226 0.9090 0.9552 0.9165
## Neg Pred Value 0.9633 0.9932 0.9956 0.9962
## Prevalence 0.1000 0.1000 0.1000 0.1000
## Detection Rate 0.0667 0.0939 0.0960 0.0966
## Detection Prevalence 0.0923 0.1033 0.1005 0.1054
## Balanced Accuracy 0.8193 0.9643 0.9775 0.9781Setelah melakukan beberapa percobaan variasi jumlah layer (1 sampai dengan 4 layer), units (mulai dari 512 kebawah), dan optimizer (sgd, rmsprop, adam, adamax, adadelta), nilai accuracy yang tertinggi adalah dengan 3 layer, dengan units 256 pada layer pertama, 128 pada layer kedua, dan 64 pada layer ketiga, serta optimizer menggunakan optimizer adam.