Merge pull request #218 from tidymodels/reticulate-to-suggests

Move keras and tensorflow imports to suggests
tidymodels · Mar 20, 2024 · 3d119e2 · 3d119e2
2 parents 1651d1a + 8c33289
commit 3d119e2
Show file tree

Hide file tree

Showing 10 changed files with 39 additions and 55 deletions.
diff --git a/.github/workflows/R-CMD-check-hard.yaml b/.github/workflows/R-CMD-check-hard.yaml
@@ -52,17 +52,6 @@ jobs:
             any::rmarkdown
           needs: check
           pak-version: devel
-
-      - uses: actions/setup-python@v4
-        with:
-          python-version: 3.9
-
-      - name: Install TensorFlow
-        run: |
-          reticulate::virtualenv_create('r-reticulate', python='3.9')
-          reticulate::use_virtualenv('r-reticulate')
-          tensorflow::install_tensorflow(version='2.11.0')
-        shell: Rscript {0}
 
       - uses: r-lib/actions/check-r-package@v2
         with:

diff --git a/DESCRIPTION b/DESCRIPTION
@@ -23,13 +23,11 @@ Imports:
     glue,
     dplyr (>= 1.1.0),
     generics (>= 0.1.0),
-    keras,
     lifecycle,
     purrr,
     rlang (>= 0.4.10),
     rsample,
     stats,
-    tensorflow,
     tibble,
     tidyr,
     utils,
@@ -42,13 +40,15 @@ Suggests:
     ggplot2,
     hardhat,
     irlba,
+    keras,
     knitr,
     lme4,
     modeldata,
     rmarkdown,
     rpart,
     rstanarm,
     stringdist,
+    tensorflow,
     testthat (>= 3.0.0),
     VBsparsePCA,
     xgboost

diff --git a/NAMESPACE b/NAMESPACE
@@ -113,17 +113,6 @@ importFrom(generics,required_pkgs)
 importFrom(generics,tidy)
 importFrom(generics,tunable)
 importFrom(glue,glue)
-importFrom(keras,backend)
-importFrom(keras,compile)
-importFrom(keras,fit)
-importFrom(keras,get_layer)
-importFrom(keras,keras_model)
-importFrom(keras,keras_model_sequential)
-importFrom(keras,layer_concatenate)
-importFrom(keras,layer_dense)
-importFrom(keras,layer_embedding)
-importFrom(keras,layer_flatten)
-importFrom(keras,layer_input)
 importFrom(lifecycle,deprecated)
 importFrom(recipes,remove_original_cols)
 importFrom(stats,as.formula)

diff --git a/NEWS.md b/NEWS.md
@@ -6,6 +6,8 @@
 
 * Documentation for tidy methods for all steps has been improved to describe the return value more accurately. (#217)
 
+* {keras} and {tensorflow} have been moved to Suggests instead of Imports. (#218)
+
 # embed 1.1.3
 
 * `step_collapse_stringdist()` will now return predictors as factors. (#204)

diff --git a/R/embed-package.R b/R/embed-package.R
@@ -14,17 +14,6 @@
 #' @importFrom dplyr mutate
 #' @importFrom dplyr one_of
 #' @importFrom glue glue
-#' @importFrom keras backend
-#' @importFrom keras compile
-#' @importFrom keras fit
-#' @importFrom keras get_layer
-#' @importFrom keras keras_model
-#' @importFrom keras keras_model_sequential
-#' @importFrom keras layer_concatenate
-#' @importFrom keras layer_dense
-#' @importFrom keras layer_embedding
-#' @importFrom keras layer_flatten
-#' @importFrom keras layer_input
 #' @importFrom lifecycle deprecated
 #' @importFrom recipes remove_original_cols
 #' @importFrom stats as.formula

diff --git a/R/embed.R b/R/embed.R
@@ -128,7 +128,7 @@
 #' "Concatenate Embeddings for Categorical Variables with Keras"
 #' \url{https://flovv.github.io/Embeddings_with_keras_part2/}
 #'
-#' @examplesIf !embed:::is_cran_check() && rlang::is_installed("modeldata")
+#' @examplesIf !embed:::is_cran_check() && rlang::is_installed(c("modeldata", "keras"))
 #' data(grants, package = "modeldata")
 #'
 #' set.seed(1)
@@ -273,6 +273,8 @@ is_tf_2 <- function() {
 
 tf_coefs2 <- function(x, y, z, opt, num, lab, h, seeds = sample.int(10000, 4),
                       ...) {
+  rlang::check_installed("keras")
+
   vars <- names(x)
   p <- length(vars)
 
@@ -306,40 +308,40 @@ tf_coefs2 <- function(x, y, z, opt, num, lab, h, seeds = sample.int(10000, 4),
   inputs <- vector(mode = "list", length = p)
   # For each categorical predictor, make an input layer
   for (i in 1:p) {
-    inputs[[i]] <- layer_input(shape = 1, name = paste0("input_", vars[i]))
+    inputs[[i]] <- keras::layer_input(shape = 1, name = paste0("input_", vars[i]))
   }
 
   layers <- vector(mode = "list", length = p)
   # Now add embedding to each layer and then flatten
   for (i in 1:p) {
     layers[[i]] <-
       inputs[[i]] %>%
-      layer_embedding(
+      keras::layer_embedding(
         input_dim = length(lvl[[i]]) + 1,
         output_dim = num,
         input_length = 1,
         name = paste0("layer_", vars[i])
       ) %>%
-      layer_flatten()
+      keras::layer_flatten()
   }
 
   if (is.null(z)) {
     if (p > 1) {
-      all_layers <- layer_concatenate(layers)
+      all_layers <- keras::layer_concatenate(layers)
     } else {
       all_layers <- layers[[1]]
     }
   } else {
     mats$z <- as.matrix(z)
-    pred_layer <- layer_input(shape = ncol(z), name = "other_pred")
-    all_layers <- layer_concatenate(c(layers, pred_layer))
+    pred_layer <- keras::layer_input(shape = ncol(z), name = "other_pred")
+    all_layers <- keras::layer_concatenate(c(layers, pred_layer))
     inputs <- c(inputs, pred_layer)
   }
 
   if (h > 0) {
     all_layers <-
       all_layers %>%
-      layer_dense(
+      keras::layer_dense(
         units = h, activation = "relu", name = "hidden_layer",
         kernel_initializer = keras::initializer_glorot_uniform(seed = seeds[3])
       )
@@ -348,14 +350,14 @@ tf_coefs2 <- function(x, y, z, opt, num, lab, h, seeds = sample.int(10000, 4),
   if (factor_y) {
     all_layers <-
       all_layers %>%
-      layer_dense(
+      keras::layer_dense(
         units = ncol(y), activation = "softmax", name = "output_layer",
         kernel_initializer = keras::initializer_glorot_uniform(seed = seeds[4])
       )
   } else {
     all_layers <-
       all_layers %>%
-      layer_dense(
+      keras::layer_dense(
         units = 1, activation = "linear", name = "output_layer",
         kernel_initializer = keras::initializer_glorot_uniform(seed = seeds[4])
       )
@@ -365,15 +367,15 @@ tf_coefs2 <- function(x, y, z, opt, num, lab, h, seeds = sample.int(10000, 4),
     keras::keras_model(inputs = inputs, outputs = all_layers)
 
   model %>%
-    compile(
+    keras::compile(
       loss = opt$loss,
       metrics = opt$metrics,
       optimizer = opt$optimizer
     )
 
   history <-
     model %>%
-    fit(
+    keras::fit(
       x = unname(mats),
       y = y,
       epochs = opt$epochs,
@@ -387,7 +389,7 @@ tf_coefs2 <- function(x, y, z, opt, num, lab, h, seeds = sample.int(10000, 4),
 
   for (i in 1:p) {
     layer_values[[i]] <-
-      get_layer(model, paste0("layer_", vars[i]))$get_weights() %>%
+      keras::get_layer(model, paste0("layer_", vars[i]))$get_weights() %>%
       as.data.frame() %>%
       setNames(names0(num, paste0(vars[i], "_embed_"))) %>%
       as_tibble() %>%
@@ -477,9 +479,9 @@ print.step_embed <-
 
 #' @export
 #' @rdname step_embed
-#' @param optimizer,loss,metrics Arguments to pass to [keras::compile()]
+#' @param optimizer,loss,metrics Arguments to pass to keras::compile()
 #' @param epochs,validation_split,batch_size,verbose,callbacks Arguments to pass
-#'   to [keras::fit()]
+#'   to keras::fit()
 embed_control <- function(loss = "mse",
                           metrics = NULL,
                           optimizer = "sgd",

diff --git a/R/feature_hash.R b/R/feature_hash.R
@@ -55,7 +55,7 @@
 #' Approach for Predictive Models_. CRC/Chapman Hall
 #' \url{https://bookdown.org/max/FES/encoding-predictors-with-many-categories.html}
 #' @seealso [recipes::step_dummy()], [recipes::step_zv()]
-#' @examplesIf !embed:::is_cran_check() && rlang::is_installed("modeldata")
+#' @examplesIf !embed:::is_cran_check() && rlang::is_installed(c("modeldata", "keras"))
 #' data(grants, package = "modeldata")
 #' rec <-
 #'   recipe(class ~ sponsor_code, data = grants_other) %>%
@@ -164,6 +164,8 @@ make_hash_vars <- function(x, prefix, num_hash = 2^8) {
 
   uni_x <- unique(x)
 
+  rlang::check_installed("keras")
+
   column_int <-
     purrr::map_int(
       uni_x,

diff --git a/man/step_embed.Rd b/man/step_embed.Rd
diff --git a/man/step_feature_hash.Rd b/man/step_feature_hash.Rd
diff --git a/tests/testthat/test-embed.R b/tests/testthat/test-embed.R
@@ -6,6 +6,7 @@ withr::local_envvar(TF_CPP_MIN_LOG_LEVEL = "2")
 
 test_that("factor encoded predictor", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   class_test <- recipe(x2 ~ ., data = ex_dat) %>%
@@ -76,6 +77,7 @@ test_that("factor encoded predictor", {
 
 test_that("character encoded predictor", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   class_test <- recipe(x2 ~ ., data = ex_dat_ch) %>%
@@ -144,6 +146,7 @@ test_that("character encoded predictor", {
 
 test_that("factor encoded predictor", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   class_test <- recipe(x1 ~ ., data = ex_dat) %>%
@@ -213,6 +216,7 @@ test_that("factor encoded predictor", {
 
 test_that("character encoded predictor", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   class_test <- recipe(x1 ~ ., data = ex_dat_ch) %>%
@@ -281,6 +285,7 @@ test_that("character encoded predictor", {
 
 test_that("bad args", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   three_class <- iris
@@ -297,6 +302,7 @@ test_that("bad args", {
 
 test_that("check_name() is used", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   dat <- ex_dat
@@ -330,6 +336,7 @@ test_that("tunable", {
 
 test_that("bake method errors when needed non-standard role columns are missing", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
   rec <- recipe(x2 ~ ., data = ex_dat) %>%
     step_embed(
@@ -351,6 +358,7 @@ test_that("bake method errors when needed non-standard role columns are missing"
 
 test_that("empty printing", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   rec <- recipe(mpg ~ ., mtcars)
@@ -396,6 +404,7 @@ test_that("empty selection tidy method works", {
 
 test_that("keep_original_cols works", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   new_names <- c("x2", "x3_embed_1", "x3_embed_2")
@@ -427,6 +436,7 @@ test_that("keep_original_cols works", {
 
 test_that("keep_original_cols - can prep recipes with it missing", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   rec <- recipe(x2 ~ x3, data = ex_dat) %>%
@@ -446,6 +456,7 @@ test_that("keep_original_cols - can prep recipes with it missing", {
 
 test_that("printing", {
   skip_on_cran()
+  skip_if_not_installed("keras")
   skip_if(!embed:::is_tf_available())
 
   rec <- recipe(x2 ~ ., data = ex_dat_ch) %>%