For the decoder network, we keep the same structure, except that, instead of pooling layers, we use upsampling layers. We can use the following code to do this:

# Decoder
decoder <- encoder %>% 
         layer_conv_2d(filters = 32, 
                       kernel_size = c(3,3), 
                       activation = 'relu',
                       padding = 'same') %>%   
         layer_upsampling_2d(c(2,2)) %>% 
         layer_conv_2d(filters = 32, 
                       kernel_size = c(3,3), 
                       activation = 'relu',
                       padding = 'same') %>%  
         layer_upsampling_2d(c(2,2)) %>% 
         layer_conv_2d(filters = 1, 
                       kernel_size = c(3,3), 
                       activation = 'sigmoid',
                       padding = 'same')
summary(decoder)
Output
Tensor("conv2d_15/Sigmoid:0", shape=(?, 28, 28, 1), dtype=float32)

In the preceding code, the first upsampling layer changes the height and width to 14 x 14 and the second upsampling layer restores it to the original height and width of 28 x 28. In the last layer, we use a sigmoid activation function, which ensures that the output values remain between 0 and 1.