updates to vignettes + examples

Author: t-kalinowski

vignettes-src/distributed_training_with_tensorflow.Rmd

@@ -123,20 +123,16 @@ get_compiled_model <- function() {
   model |> compile(
     optimizer = optimizer_adam(),
     loss = loss_sparse_categorical_crossentropy(from_logits = TRUE),
-    metrics = list(metric_sparse_categorical_accuracy()),
-
-    # XLA compilation is temporarily disabled due to a bug
-    # https://github.com/keras-team/keras/issues/19005
-    jit_compile = FALSE
+    metrics = list(metric_sparse_categorical_accuracy())
   )
   model
 }
 
 get_dataset <- function(batch_size = 64) {
 
   c(c(x_train, y_train), c(x_test, y_test)) %<-% dataset_mnist()
-  x_train <- array_reshape(x_train, c(-1, 784))
-  x_test <- array_reshape(x_test, c(-1, 784))
+  x_train <- array_reshape(x_train, c(-1, 784)) / 255
+  x_test <- array_reshape(x_test, c(-1, 784)) / 255
 
   # Reserve 10,000 samples for validation.
   val_i <- sample.int(nrow(x_train), 10000)
@@ -145,18 +141,25 @@ get_dataset <- function(batch_size = 64) {
   x_train = x_train[-val_i,]
   y_train = y_train[-val_i]
 
+  y_train <- array_reshape(y_train, c(-1, 1))
+  y_val <- array_reshape(y_val, c(-1, 1))
+  y_test <- array_reshape(y_test, c(-1, 1))
+
   # Prepare the training dataset.
   train_dataset <- list(x_train, y_train) |>
+    lapply(np_array, "float32") |>
     tensor_slices_dataset() |>
     dataset_batch(batch_size)
 
   # Prepare the validation dataset.
   val_dataset <- list(x_val, y_val) |>
+    lapply(np_array, "float32") |>
     tensor_slices_dataset() |>
     dataset_batch(batch_size)
 
   # Prepare the test dataset.
   test_dataset <- list(x_test, y_test) |>
+    lapply(np_array, "float32") |>
     tensor_slices_dataset() |>
     dataset_batch(batch_size)
 

vignettes-src/examples/nlp/text_classification_from_scratch.Rmd

@@ -25,7 +25,6 @@ word splitting & indexing.
 library(tensorflow, exclude = c("shape", "set_random_seed"))
 library(tfdatasets, exclude = "shape")
 library(keras3)
-use_virtualenv("r-keras")
 ```
 
 ## Load the data: IMDB movie review sentiment classification

vignettes-src/getting_started.Rmd

@@ -44,7 +44,7 @@ install.packages("keras3")
 or install the development version with:
 
 ```{r, eval=FALSE}
-remotes::install_github("rstudio/keras")
+remotes::install_github("rstudio/keras3")
 ```
 
 The Keras R interface requires that a backend engine be installed. This is [TensorFlow](https://www.tensorflow.org/) by default.

vignettes/custom_train_step_in_tensorflow.Rmd

@@ -127,11 +127,11 @@ model |> fit(x, y, epochs = 3)
 
 ```
 ## Epoch 1/3
-## 32/32 - 1s - 23ms/step - loss: 3.2271 - mae: 1.4339
+## 32/32 - 1s - 24ms/step - mae: 1.4339 - loss: 3.2271
 ## Epoch 2/3
-## 32/32 - 0s - 1ms/step - loss: 2.9034 - mae: 1.3605
+## 32/32 - 0s - 2ms/step - mae: 1.3605 - loss: 2.9034
 ## Epoch 3/3
-## 32/32 - 0s - 1ms/step - loss: 2.6272 - mae: 1.2960
+## 32/32 - 0s - 2ms/step - mae: 1.2960 - loss: 2.6272
 ```
 
 ## Going lower-level
@@ -212,11 +212,11 @@ model |> fit(x, y, epochs = 3)
 
 ```
 ## Epoch 1/3
-## 32/32 - 1s - 22ms/step - loss: 2.5170 - mae: 1.2923
+## 32/32 - 1s - 23ms/step - loss: 2.5170 - mae: 1.2923
 ## Epoch 2/3
-## 32/32 - 0s - 1ms/step - loss: 2.2689 - mae: 1.2241
+## 32/32 - 0s - 2ms/step - loss: 2.2689 - mae: 1.2241
 ## Epoch 3/3
-## 32/32 - 0s - 1ms/step - loss: 2.0578 - mae: 1.1633
+## 32/32 - 0s - 3ms/step - loss: 2.0578 - mae: 1.1633
 ```
 
 ## Supporting `sample_weight` & `class_weight`
@@ -282,11 +282,11 @@ model |> fit(x, y, sample_weight = sw, epochs = 3)
 
 ```
 ## Epoch 1/3
-## 32/32 - 1s - 26ms/step - loss: 0.1681 - mae: 1.3434
+## 32/32 - 1s - 26ms/step - mae: 1.3434 - loss: 0.1681
 ## Epoch 2/3
-## 32/32 - 0s - 9ms/step - loss: 0.1394 - mae: 1.3364
+## 32/32 - 0s - 2ms/step - mae: 1.3364 - loss: 0.1394
 ## Epoch 3/3
-## 32/32 - 0s - 1ms/step - loss: 0.1148 - mae: 1.3286
+## 32/32 - 0s - 2ms/step - mae: 1.3286 - loss: 0.1148
 ```
 
 ## Providing your own evaluation step
@@ -332,15 +332,16 @@ model |> evaluate(x, y)
 ```
 
 ```
-## 32/32 - 0s - 10ms/step - loss: 0.0000e+00 - mae: 1.3871
+## 32/32 - 0s - 10ms/step - mae: 1.3871 - loss: 0.0000e+00
 ```
 
 ```
 ## $loss
-## [1] 0
+## tf.Tensor(0.0, shape=(), dtype=float32)
 ##
-## $mae
-## [1] 1.387149
+## $compile_metrics
+## $compile_metrics$mae
+## tf.Tensor(1.3871489, shape=(), dtype=float32)
 ```
 
 ## Wrapping up: an end-to-end GAN example
@@ -507,7 +508,7 @@ gan |> fit(
 ```
 
 ```
-## 100/100 - 5s - 54ms/step - d_loss: 0.0000e+00 - g_loss: 0.0000e+00
+## 100/100 - 5s - 55ms/step - d_loss: 0.0000e+00 - g_loss: 0.0000e+00
 ```
 
 The ideas behind deep learning are simple, so why should their implementation be painful?

vignettes/distribution.Rmd

@@ -184,24 +184,24 @@ model |> fit(dataset, epochs = 3)
 
 ```
 ## Epoch 1/3
-## 8/8 - 0s - 38ms/step - loss: 1.0768
+## 8/8 - 0s - 40ms/step - loss: 1.1398
 ## Epoch 2/3
-## 8/8 - 0s - 6ms/step - loss: 0.9754
+## 8/8 - 0s - 7ms/step - loss: 1.0593
 ## Epoch 3/3
-## 8/8 - 0s - 5ms/step - loss: 0.9347
+## 8/8 - 0s - 5ms/step - loss: 1.0071
 ```
 
 ``` r
 model |> evaluate(dataset)
 ```
 
 ```
-## 8/8 - 0s - 7ms/step - loss: 0.8936
+## 8/8 - 0s - 6ms/step - loss: 0.9609
 ```
 
 ```
 ## $loss
-## [1] 0.8935966
+## [1] 0.9609067
 ```
 
 
@@ -278,24 +278,24 @@ model |> fit(dataset, epochs = 3)
 
 ```
 ## Epoch 1/3
-## 8/8 - 0s - 29ms/step - loss: 1.0836
+## 8/8 - 0s - 25ms/step - loss: 1.1612
 ## Epoch 2/3
-## 8/8 - 0s - 4ms/step - loss: 1.0192
+## 8/8 - 0s - 4ms/step - loss: 1.0702
 ## Epoch 3/3
-## 8/8 - 0s - 4ms/step - loss: 0.9821
+## 8/8 - 0s - 5ms/step - loss: 1.0184
 ```
 
 ``` r
 model |> evaluate(dataset)
 ```
 
 ```
-## 8/8 - 0s - 8ms/step - loss: 0.9576
+## 8/8 - 0s - 7ms/step - loss: 0.9844
 ```
 
 ```
 ## $loss
-## [1] 0.9576273
+## [1] 0.9844129
 ```
 
 

vignettes/functional_api.Rmd

@@ -208,17 +208,17 @@ history <- model |> fit(
 
 ```
 ## Epoch 1/2
-## 750/750 - 2s - 3ms/step - accuracy: 0.8979 - loss: 0.3540 - val_accuracy: 0.9448 - val_loss: 0.1903
+## 750/750 - 2s - 3ms/step - accuracy: 0.8980 - loss: 0.3540 - val_accuracy: 0.9444 - val_loss: 0.1898
 ## Epoch 2/2
-## 750/750 - 1s - 784us/step - accuracy: 0.9509 - loss: 0.1635 - val_accuracy: 0.9597 - val_loss: 0.1397
+## 750/750 - 1s - 1ms/step - accuracy: 0.9512 - loss: 0.1633 - val_accuracy: 0.9605 - val_loss: 0.1387
 ```
 
 ``` r
 test_scores <- model |> evaluate(x_test, y_test, verbose=2)
 ```
 
 ```
-## 313/313 - 0s - 1ms/step - accuracy: 0.9595 - loss: 0.1328
+## 313/313 - 1s - 2ms/step - accuracy: 0.9598 - loss: 0.1322
 ```
 
 
@@ -229,8 +229,8 @@ cat("Test accuracy:", test_scores$accuracy, "\n")
 ```
 
 ```
-## Test loss: 0.132778
-## Test accuracy: 0.9595
+## Test loss: 0.1321879
+## Test accuracy: 0.9598
 ```
 
 For further reading, see the [training and evaluation](training_with_built_in_methods.html) guide.
@@ -643,9 +643,9 @@ model |> fit(
 
 ```
 ## Epoch 1/2
-## 40/40 - 3s - 66ms/step - department_loss: 381.0319 - loss: 498.7886 - priority_loss: 117.7568
+## 40/40 - 3s - 65ms/step - department_loss: 2.8465 - loss: 0.7669 - priority_loss: 0.1976
 ## Epoch 2/2
-## 40/40 - 0s - 10ms/step - department_loss: 358.5705 - loss: 438.4252 - priority_loss: 79.8546
+## 40/40 - 0s - 6ms/step - department_loss: 2.8554 - loss: 0.7538 - priority_loss: 0.1828
 ```
 
 When calling fit with a `Dataset` object, it should yield either a
@@ -776,7 +776,7 @@ model |> fit(
 ```
 
 ```
-## 13/13 - 5s - 375ms/step - acc: 0.1250 - loss: 2.3001 - val_acc: 0.1400 - val_loss: 2.2938
+## 13/13 - 5s - 367ms/step - acc: 0.1250 - loss: 2.3003 - val_acc: 0.1250 - val_loss: 2.2946
 ```
 
 ## Shared layers