## https://tensorflow.rstudio.com/install/
## remotes::install_github("rstudio/tensorflow")
## reticulate::install_python()
library(tensorflow)
### install_tensorflow(envname = "r-tensorflow")
## install.packages("keras")
library(keras)
### install_keras()
library(ggplot2)
# Load the Fashion MNIST dataset
mnist <- dataset_fashion_mnist()
# Create training and testing datasets
train_images <- mnist$train$x # 60,000x28x28 tensor for training images
train_labels <- mnist$train$y # 60,000-element vector for training labels
test_images <- mnist$test$x # 10,000x28x28 tensor for test images
test_labels <- mnist$test$y # 10,000-element vector for test labels
# Plot the 100th training image
digit <- train_images[100,,]
plot(as.raster(digit, max = 255))
# Reshape input data for CNN (adding a channel dimension: 28x28x1)
train_images <- array_reshape(train_images, c(60000, 28, 28, 1))
test_images <- array_reshape(test_images, c(10000, 28, 28, 1))
# Normalize the pixel values to [0,1]
train_images <- train_images / 255
test_images <- test_images / 255
# One-hot encode the labels
train_labels <- to_categorical(train_labels, 10)
test_labels <- to_categorical(test_labels, 10)
# Build the CNN model
network <- keras_model_sequential() %>%
# First convolutional layer with 32 filters, 3x3 kernel, ReLU activation, and same padding
layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = 'relu',
input_shape = c(28, 28, 1), padding = 'same') %>%
# Max-pooling layer with 2x2 pool size
layer_max_pooling_2d(pool_size = c(2, 2)) %>%
# Second convolutional layer with 64 filters and ReLU activation
layer_conv_2d(filters = 64, kernel_size = c(3, 3), activation = 'relu', padding = 'same') %>%
layer_max_pooling_2d(pool_size = c(2, 2)) %>%
# Flatten the feature maps into a 1D vector for the dense layers
layer_flatten() %>%
# Fully connected dense layer with 128 units and ReLU activation
layer_dense(units = 128, activation = 'relu') %>%
# Output layer with 10 units (one for each class) and softmax activation
layer_dense(units = 10, activation = 'softmax')
# Compile the model
network %>% compile(
optimizer = "rmsprop",
loss = "categorical_crossentropy",
metrics = c("accuracy")
)
# Check the dimensions to verify correctness
dim(train_images)
## [1] 60000 28 28 1
## [1] 10000 28 28 1
## [1] 60000 10
# Train the CNN model
history <- network %>% fit(
train_images, train_labels,
epochs = 10,
batch_size = 1000,
validation_split = 0.1
)
## Epoch 1/10
## 54/54 - 44s - loss: 0.9846 - accuracy: 0.6490 - val_loss: 0.5675 - val_accuracy: 0.7882 - 44s/epoch - 817ms/step
## Epoch 2/10
## 54/54 - 37s - loss: 0.5570 - accuracy: 0.7929 - val_loss: 0.5161 - val_accuracy: 0.7982 - 37s/epoch - 677ms/step
## Epoch 3/10
## 54/54 - 39s - loss: 0.4529 - accuracy: 0.8346 - val_loss: 0.4689 - val_accuracy: 0.8152 - 39s/epoch - 731ms/step
## Epoch 4/10
## 54/54 - 35s - loss: 0.4012 - accuracy: 0.8537 - val_loss: 0.3832 - val_accuracy: 0.8610 - 35s/epoch - 653ms/step
## Epoch 5/10
## 54/54 - 35s - loss: 0.3631 - accuracy: 0.8677 - val_loss: 0.3668 - val_accuracy: 0.8668 - 35s/epoch - 653ms/step
## Epoch 6/10
## 54/54 - 35s - loss: 0.3432 - accuracy: 0.8749 - val_loss: 0.3332 - val_accuracy: 0.8775 - 35s/epoch - 657ms/step
## Epoch 7/10
## 54/54 - 38s - loss: 0.3218 - accuracy: 0.8825 - val_loss: 0.3234 - val_accuracy: 0.8812 - 38s/epoch - 709ms/step
## Epoch 8/10
## 54/54 - 39s - loss: 0.3037 - accuracy: 0.8883 - val_loss: 0.3324 - val_accuracy: 0.8773 - 39s/epoch - 728ms/step
## Epoch 9/10
## 54/54 - 35s - loss: 0.2907 - accuracy: 0.8923 - val_loss: 0.3006 - val_accuracy: 0.8922 - 35s/epoch - 653ms/step
## Epoch 10/10
## 54/54 - 43s - loss: 0.2787 - accuracy: 0.8968 - val_loss: 0.3080 - val_accuracy: 0.8838 - 43s/epoch - 799ms/step
# Plot the training history
plot(history) + theme_bw()
# Evaluate the model on test data
network %>% evaluate(test_images, test_labels)
## 313/313 - 2s - loss: 0.3292 - accuracy: 0.8800 - 2s/epoch - 8ms/step
## loss accuracy
## 0.3291659 0.8800000