compatibility with tf2

Author: Turgut Abdullayev

vignettes/examples/eager_cvae.R

@@ -5,17 +5,14 @@
 #' https://blogs.rstudio.com/tensorflow/posts/2018-10-22-mmd-vae/
 
 
+
 library(keras)
-use_implementation("tensorflow")
 library(tensorflow)
-tfe_enable_eager_execution(device_policy = "silent")
-
 library(tfdatasets)
 library(dplyr)
 library(ggplot2)
 library(glue)
 
-
 # Setup and preprocessing -------------------------------------------------
 
 fashion <- dataset_fashion_mnist()
@@ -24,9 +21,9 @@ c(train_images, train_labels) %<-% fashion$train
 c(test_images, test_labels) %<-% fashion$test
 
 train_x <-
-  train_images %>% `/`(255) %>% k_reshape(c(60000, 28, 28, 1))
+  train_images %>% `/`(255) %>% k_reshape(c(60000, 28, 28, 1)) %>% k_cast(dtype='float32')
 test_x <-
-  test_images %>% `/`(255) %>% k_reshape(c(10000, 28, 28, 1))
+  test_images %>% `/`(255) %>% k_reshape(c(10000, 28, 28, 1)) %>% k_cast(dtype='float32')
 
 class_names = c('T-shirt/top',
                 'Trouser',
@@ -63,7 +60,8 @@ encoder_model <- function(name = NULL) {
         filters = 32,
         kernel_size = 3,
         strides = 2,
-        activation = "relu"
+        activation = "relu",
+        dtype='float32'
       )
     self$conv2 <-
       layer_conv_2d(
@@ -88,7 +86,7 @@ encoder_model <- function(name = NULL) {
 
 decoder_model <- function(name = NULL) {
   keras_model_custom(name = name, function(self) {
-    self$dense <- layer_dense(units = 7 * 7 * 32, activation = "relu")
+    self$dense <- layer_dense(units = 7 * 7 * 32, activation = "relu",dtype='float32')
     self$reshape <- layer_reshape(target_shape = c(7, 7, 32))
     self$deconv1 <-
       layer_conv_2d_transpose(
@@ -126,22 +124,22 @@ decoder_model <- function(name = NULL) {
 }
 
 reparameterize <- function(mean, logvar) {
-  eps <- k_random_normal(shape = mean$shape, dtype = tf$float64)
+  eps <- k_random_normal(shape = mean$shape, dtype = tf$float32)
   eps * k_exp(logvar * 0.5) + mean
 }
 
 
 # Loss and optimizer ------------------------------------------------------
 
 normal_loglik <- function(sample, mean, logvar, reduce_axis = 2) {
-  loglik <- k_constant(0.5, dtype = tf$float64) * 
-    (k_log(2 * k_constant(pi, dtype = tf$float64)) +
-     logvar +
-     k_exp(-logvar) * (sample - mean) ^ 2)
+  loglik <- k_constant(0.5, dtype = tf$float32) * 
+    (k_log(2 * k_constant(pi, dtype = tf$float32)) +
+       logvar +
+       k_exp(-logvar) * (sample - mean) ^ 2)
   - k_sum(loglik, axis = reduce_axis)
 }
 
-optimizer <- tf$train$AdamOptimizer(1e-4)
+optimizer <- tf$keras$optimizers$Adam(1e-4)
 
 
 
@@ -151,7 +149,7 @@ num_examples_to_generate <- 64
 
 random_vector_for_generation <-
   k_random_normal(shape = list(num_examples_to_generate, latent_dim),
-                  dtype = tf$float64)
+                  dtype = tf$float32)
 
 generate_random_clothes <- function(epoch) {
   predictions <-
@@ -267,8 +265,8 @@ for (epoch in seq_len(num_epochs)) {
         -k_sum(crossentropy_loss)
       logpz <-
         normal_loglik(z,
-                      k_constant(0, dtype = tf$float64),
-                      k_constant(0, dtype = tf$float64))
+                      k_constant(0, dtype = tf$float32),
+                      k_constant(0, dtype = tf$float32))
       logqz_x <- normal_loglik(z, mean, logvar)
       loss <- -k_mean(logpx_z + logpz - logqz_x)
 
@@ -284,12 +282,10 @@ for (epoch in seq_len(num_epochs)) {
 
     optimizer$apply_gradients(purrr::transpose(list(
       encoder_gradients, encoder$variables
-    )),
-    global_step = tf$train$get_or_create_global_step())
+    )))
     optimizer$apply_gradients(purrr::transpose(list(
       decoder_gradients, decoder$variables
-    )),
-    global_step = tf$train$get_or_create_global_step())
+    )))
 
   })
 
@@ -313,3 +309,5 @@ for (epoch in seq_len(num_epochs)) {
   }
 }
 
+
+