Update internal functions to reflect deprecated keras functions (#639)

* Add keras_predict_classes to replace use of keras::predict_classes

* use keras_predict_classes

* update mlp_keras reference calculations in tests

* move from predict_proba() to predict()

* Create conditional tensorflow checking

* use newer version of tensorflow in GHA

* Add old tensorflow version GHA

* Set seeds for tensorflow

* only set tensorflow seed when you need it

* conditionally set seed in tensorflow by tensorflow version

* do conditional check innside keras_predict_* functions as well

* Add missing set_seed to keras logistic reg test

* seperate out old-tensorflow GHA

* add last missing tensorflow set_seed

* you need tensorflow AND R seed...

* use keras version as switch

* Conditionally transform predictions depending on tensorflow version

* skip test if tensorflow version can't be found

* refactor keras_predict_classes to avoid post function

* Add keras_set_seed function

* remove remotes

* rename keras_set_seed

* add news

* rename to set_tf_seed in all tests

* adjust "R CMD Check" and "old tensorflow" GHA

* don't use keras_predict_proba anymore

Co-authored-by: Max Kuhn <mxkuhn@gmail.com>

Author: EmilHvitfeldt

.github/workflows/R-CMD-check.yaml

@@ -59,14 +59,14 @@ jobs:
 
       - name: Install Miniconda
         # conda can fail at downgrading python, so we specify python version in advance
-        env: 
-          RETICULATE_MINICONDA_PYTHON_VERSION: "3.7"  
+        env:
+          RETICULATE_MINICONDA_PYTHON_VERSION: "3.7"
         run: reticulate::install_miniconda() # creates r-reticulate conda env by default
         shell: Rscript {0}
 
       - name: Install TensorFlow
         run: |
-          tensorflow::install_tensorflow(version='1.15', conda_python_version = NULL)
+          tensorflow::install_tensorflow(version='2.7', conda_python_version = NULL)
         shell: Rscript {0}
 
       - uses: r-lib/actions/check-r-package@v2

.github/workflows/old-tensorflow.yaml

@@ -0,0 +1,79 @@
+# Workflow derived from https://github.com/r-lib/actions/tree/master/examples
+# Need help debugging build failures? Start at https://github.com/r-lib/actions#where-to-find-help
+#
+# NOTE: This workflow is overkill for most R packages and
+# check-standard.yaml is likely a better choice.
+# usethis::use_github_action("check-standard") will install it.
+on:
+  push:
+    branches: [main, master]
+  pull_request:
+    branches: [main, master]
+  workflow_dispatch:
+
+name: old-tensorflow
+
+jobs:
+  old-tensorflow:
+    runs-on: ${{ matrix.config.os }}
+
+    name: ${{ matrix.config.os }} (${{ matrix.config.r }})
+
+    strategy:
+      fail-fast: false
+      matrix:
+        config:
+          - {os: windows-latest, r: 'release'}
+          # Use older ubuntu to maximise backward compatibility
+          - {os: ubuntu-18.04,   r: 'devel', http-user-agent: 'release'}
+    env:
+      GITHUB_PAT: ${{ secrets.GITHUB_TOKEN }}
+      R_KEEP_PKG_SOURCE: yes
+      CXX14: g++
+      CXX14STD: -std=c++1y
+      CXX14FLAGS: -Wall -g -02
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - uses: r-lib/actions/setup-pandoc@v2
+
+      - uses: r-lib/actions/setup-r@v2
+        with:
+          r-version: ${{ matrix.config.r }}
+          http-user-agent: ${{ matrix.config.http-user-agent }}
+          use-public-rspm: true
+
+      - uses: r-lib/actions/setup-r-dependencies@v2
+        with:
+          extra-packages: rcmdcheck
+
+      - name: Install dev reticulate
+        run: pak::pkg_install('rstudio/reticulate')
+        shell: Rscript {0}
+
+      - name: Install Miniconda
+        # conda can fail at downgrading python, so we specify python version in advance
+        env:
+          RETICULATE_MINICONDA_PYTHON_VERSION: "3.7"
+        run: reticulate::install_miniconda() # creates r-reticulate conda env by default
+        shell: Rscript {0}
+
+      - name: Install TensorFlow
+        run: |
+          tensorflow::install_tensorflow(version='1.15', conda_python_version = NULL)
+        shell: Rscript {0}
+
+      - uses: r-lib/actions/check-r-package@v2
+
+      - name: Show testthat output
+        if: always()
+        run: find check -name 'testthat.Rout*' -exec cat '{}' \; || true
+        shell: bash
+
+      - name: Upload check results
+        if: failure()
+        uses: actions/upload-artifact@main
+        with:
+          name: ${{ runner.os }}-r${{ matrix.config.r }}-results
+          path: check

.github/workflows/pkgdown.yaml

@@ -34,7 +34,7 @@ jobs:
           pak::pkg_install('rstudio/reticulate')
           reticulate::install_miniconda()
           reticulate::conda_create('r-reticulate', packages = c('python==3.6.9'))
-          tensorflow::install_tensorflow(version='1.14.0')
+          tensorflow::install_tensorflow(version='2.7.0')
         shell: Rscript {0}
 
       - name: Install package

.github/workflows/test-coverage.yaml

@@ -31,7 +31,7 @@ jobs:
           pak::pkg_install('rstudio/reticulate')
           reticulate::install_miniconda()
           reticulate::conda_create('r-reticulate', packages = c('python==3.6.9'))
-          tensorflow::install_tensorflow(version='1.14.0')
+          tensorflow::install_tensorflow(version='2.7.0')
         shell: Rscript {0}
 
       - name: Test coverage

DESCRIPTION

@@ -40,6 +40,7 @@ Suggests:
     covr,
     dials (>= 0.0.10.9001),
     earth,
+    tensorflow,
     ggplot2,
     keras,
     kernlab,
@@ -86,6 +87,3 @@ Encoding: UTF-8
 LazyData: true
 Roxygen: list(markdown = TRUE)
 RoxygenNote: 7.1.2
-Remotes:  
-    tidymodels/dials,
-    tidymodels/hardhat

NAMESPACE

@@ -206,6 +206,7 @@ export(glance)
 export(has_multi_predict)
 export(is_varying)
 export(keras_mlp)
+export(keras_predict_classes)
 export(knit_engine_docs)
 export(linear_reg)
 export(list_md_problems)
@@ -271,6 +272,7 @@ export(set_model_engine)
 export(set_model_mode)
 export(set_new_model)
 export(set_pred)
+export(set_tf_seed)
 export(show_call)
 export(show_engines)
 export(show_fit)

NEWS.md

@@ -35,11 +35,13 @@
 * Argument `interval` was added for prediction: For types "survival" and "quantile", estimates for the confidence or prediction interval can be added if available (#615).
 
 * `set_dependency()` now allows developers to create package requirements that are specific to the model's mode (#604). 
-
+* 
 * `varying()` is soft-deprecated in favor of `tune()`.
 
 * `varying_args()` is soft-deprecated in favor of `tune_args()`.
 
+* parsnip is now more robust working with keras and tensorflow for a larger range of versions (#596).
+
 # parsnip 0.1.7
 
 ## Model Specification Changes

R/logistic_reg_data.R

@@ -438,13 +438,11 @@ set_pred(
   type = "class",
   value = list(
     pre = NULL,
-    post = function(x, object) {
-      object$lvl[x + 1]
-    },
-    func = c(pkg = "keras", fun = "predict_classes"),
+    post = NULL,
+    func = c(pkg = "parsnip", fun = "keras_predict_classes"),
     args =
       list(
-        object = quote(object$fit),
+        object = quote(object),
         x = quote(as.matrix(new_data))
       )
   )
@@ -462,7 +460,7 @@ set_pred(
       x <- as_tibble(x)
       x
     },
-    func = c(pkg = "keras", fun = "predict_proba"),
+    func = c(fun = "predict"),
     args =
       list(
         object = quote(object$fit),

R/mlp.R

@@ -436,5 +436,40 @@ reformat_torch_num <- function(results, object) {
   results
 }
 
+#' Wrapper for keras class predictions
+#' @param object A keras model fit
+#' @param x A data set.
+#' @export
+#' @keywords internal
+keras_predict_classes <- function(object, x) {
+  if (utils::packageVersion("keras") >= package_version("2.6")) {
+    preds <- predict(object$fit, x)
+    if (tensorflow::tf_version() <= package_version("2.0.0")) {
+      # -1 to assign with keras' zero indexing
+      index <- apply(preds, 1, which.max) - 1
+    } else {
+      index <- preds %>% keras::k_argmax() %>% as.integer()
+    }
+  } else {
+    index <- keras::predict_classes(object$fit, x)
+  }
+  object$lvl[index + 1]
+}
 
-
+#' Set seed in R and TensorFlow at the same time
+#'
+#' Some Keras models requires seeds to be set in both R and TensorFlow to
+#' achieve reproducible results. This function sets these seeds at the same
+#' time using version appropriate functions.
+#'
+#' @param seed 1 integer value.
+#' @export
+#' @keywords internal
+set_tf_seed <- function(seed) {
+  set.seed(seed)
+  if (tensorflow::tf_version() >= package_version("2.0")) {
+    tensorflow::tf$random$set_seed(seed)
+  } else {
+    tensorflow::tf$random$set_random_seed(seed)
+  }
+}

R/mlp_data.R

@@ -146,13 +146,11 @@ set_pred(
   type = "class",
   value = list(
     pre = NULL,
-    post = function(x, object) {
-      object$lvl[x + 1]
-    },
-    func = c(pkg = "keras", fun = "predict_classes"),
+    post = NULL,
+    func = c(pkg = "parsnip", fun = "keras_predict_classes"),
     args =
       list(
-        object = quote(object$fit),
+        object = quote(object),
         x = quote(as.matrix(new_data))
       )
   )
@@ -170,7 +168,7 @@ set_pred(
       x <- as_tibble(x)
       x
     },
-    func = c(pkg = "keras", fun = "predict_proba"),
+    func = c(fun = "predict"),
     args =
       list(
         object = quote(object$fit),

R/multinom_reg_data.R

@@ -234,12 +234,10 @@ set_pred(
   type = "class",
   value = list(
     pre = NULL,
-    post = function(x, object) {
-      object$lvl[x + 1]
-    },
-    func = c(pkg = "keras", fun = "predict_classes"),
+    post = NULL,
+    func = c(pkg = "parsnip", fun = "keras_predict_classes"),
     args =
-      list(object = quote(object$fit),
+      list(object = quote(object),
            x = quote(as.matrix(new_data)))
   )
 )
@@ -256,7 +254,7 @@ set_pred(
       x <- as_tibble(x)
       x
     },
-    func = c(pkg = "keras", fun = "predict_proba"),
+    func = c(fun = "predict"),
     args =
       list(object = quote(object$fit),
            x = quote(as.matrix(new_data)))

man/keras_predict_classes.Rd

@@ -27,8 +27,10 @@ ctrl <- control_parsnip(verbosity = 0, catch = FALSE)
 test_that('model fitting', {
   skip_on_cran()
   skip_if_not_installed("keras")
+  skip_if(is.null(tensorflow::tf_version()))
+
+  set_tf_seed(257)
 
-  set.seed(257)
   expect_error(
     fit1 <-
       fit_xy(
@@ -40,7 +42,8 @@ test_that('model fitting', {
     regexp = NA
   )
 
-  set.seed(257)
+  set_tf_seed(257)
+
   expect_error(
     fit2 <-
       fit_xy(
@@ -94,6 +97,7 @@ test_that('model fitting', {
 test_that('regression prediction', {
   skip_on_cran()
   skip_if_not_installed("keras")
+  skip_if(is.null(tensorflow::tf_version()))
 
   library(keras)