keras linear regression

Author: topepo

R/linear_reg.R

@@ -2,11 +2,11 @@
 #'
 #' `linear_reg` is a way to generate a _specification_ of a model
 #'  before fitting and allows the model to be created using
-#'  different packages in R, Stan, or via Spark. The main arguments for the
-#'  model are:
+#'  different packages in R, Stan, keras, or via Spark. The main
+#'  arguments for the model are:
 #' \itemize{
 #'   \item \code{penalty}: The total amount of regularization
-#'  in the model. Note that this must be zero for some engines .
+#'  in the model. Note that this must be zero for some engines.
 #'   \item \code{mixture}: The proportion of L1 regularization in
 #'  the model. Note that this will be ignored for some engines.
 #' }
@@ -19,8 +19,11 @@
 #' @inheritParams boost_tree
 #' @param mode A single character string for the type of model.
 #'  The only possible value for this model is "regression".
-#' @param penalty An non-negative number representing the
-#'  total amount of regularization (`glmnet` and `spark` only).
+#' @param penalty An non-negative number representing the total
+#'  amount of regularization (`glmnet`, `keras`, and `spark` only).
+#'  For `keras` models, this corresponds to purely L2 regularization
+#'  (aka weight decay) while the other models can be a combination
+#'  of L1 and L2 (depending on the value of `mixture`).
 #' @param mixture A number between zero and one (inclusive) that
 #'  represents the proportion of regularization that is used for the
 #'  L2 penalty (i.e. weight decay, or ridge regression) versus L1
@@ -36,6 +39,7 @@
 #' \item \pkg{R}:  `"lm"` or `"glmnet"`
 #' \item \pkg{Stan}:  `"stan"`
 #' \item \pkg{Spark}: `"spark"`
+#' \item \pkg{keras}: `"keras"`
 #' }
 #'
 #' @section Engine Details:
@@ -59,6 +63,10 @@
 #' \pkg{spark}
 #'
 #' \Sexpr[results=rd]{parsnip:::show_fit(parsnip:::linear_reg(), "spark")}
+#' 
+#' \pkg{keras}
+#'
+#' \Sexpr[results=rd]{parsnip:::show_fit(parsnip:::linear_reg(), "keras")}
 #'
 #' When using `glmnet` models, there is the option to pass
 #'  multiple values (or no values) to the `penalty` argument.

R/linear_reg_data.R

@@ -4,17 +4,19 @@ linear_reg_arg_key <- data.frame(
   glmnet =  c(   "lambda",             "alpha"),
   spark  =  c("reg_param", "elastic_net_param"),
   stan   =  c(        NA,                  NA),
+  keras  =  c(   "decay",                  NA),
   stringsAsFactors = FALSE,
   row.names =  c("penalty", "mixture")
 )
 
 linear_reg_modes <- "regression"
 
 linear_reg_engines <- data.frame(
-  lm    = TRUE,
+  lm     = TRUE,
   glmnet = TRUE,
   spark  = TRUE,
   stan   = TRUE,
+  keras  = TRUE,
   row.names =  c("regression")
 )
 
@@ -240,5 +242,24 @@ linear_reg_spark_data <-
     )
   )
 
-
+linear_reg_keras_data <-
+  list(
+    libs = c("keras", "magrittr"),
+    fit = list(
+      interface = "matrix",
+      protect = c("x", "y"),
+      func = c(pkg = "parsnip", fun = "keras_mlp"),
+      defaults = list(hidden_units = 1, act = "linear")
+    ),
+    numeric = list(
+      pre = NULL,
+      post = maybe_multivariate,
+      func = c(fun = "predict"),
+      args =
+        list(
+          object = quote(object$fit),
+          x = quote(as.matrix(new_data))
+        )
+    )
+  )
 

docs/articles/articles/Classification.html

@@ -0,0 +1,126 @@
+library(testthat)
+library(parsnip)
+library(rlang)
+library(tibble)
+
+# ------------------------------------------------------------------------------
+
+context("keras linear regression")
+source("helpers.R")
+
+# ------------------------------------------------------------------------------
+
+basic_mod <- 
+  linear_reg() %>% 
+  set_engine("keras", epochs = 50, verbose = 0)
+
+ridge_mod <- 
+  linear_reg(penalty = 0.1) %>% 
+  set_engine("keras", epochs = 50, verbose = 0)
+
+ctrl <- fit_control(verbosity = 0, catch = FALSE)
+
+# ------------------------------------------------------------------------------
+
+test_that('model fitting', {
+  
+  skip_if_not_installed("keras")
+  
+  set.seed(257)
+  expect_error(
+    fit1 <- 
+      fit_xy(
+      basic_mod,
+      control = ctrl,
+      x = iris[,2:4],
+      y = iris$Sepal.Length
+    ),
+    regexp = NA
+  )
+  
+  set.seed(257)
+  expect_error(
+    fit2 <- 
+      fit_xy(
+        basic_mod,
+        control = ctrl,
+        x = iris[,2:4],
+        y = iris$Sepal.Length
+      ),
+    regexp = NA
+  )
+  expect_equal(fit1, fit2)
+  
+  expect_error(
+    fit(
+      basic_mod,
+      Sepal.Length ~ .,
+      data = iris[, -5],
+      control = ctrl
+    ),
+    regexp = NA
+  )
+  
+  expect_error(
+    fit1 <- 
+      fit_xy(
+        ridge_mod,
+        control = ctrl,
+        x = iris[,2:4],
+        y = iris$Sepal.Length
+      ),
+    regexp = NA
+  )
+  
+  expect_error(
+    fit(
+      ridge_mod,
+      Sepal.Length ~ .,
+      data = iris[, -5],
+      control = ctrl
+    ),
+    regexp = NA
+  )
+  
+})
+
+
+test_that('regression prediction', {
+  
+  skip_if_not_installed("keras")
+  
+  library(keras)
+  
+  set.seed(257)
+  lm_fit <- 
+    fit_xy(
+      basic_mod,
+      control = ctrl,
+      x = iris[,2:4],
+      y = iris$Sepal.Length
+    )
+  
+  keras_pred <-
+    predict(lm_fit$fit, as.matrix(iris[1:3,2:4])) %>%
+    as_tibble() %>%
+    setNames(".pred")
+  parsnip_pred <- predict(lm_fit, iris[1:3,2:4])
+  expect_equal(as.data.frame(keras_pred), as.data.frame(parsnip_pred))
+  
+  set.seed(257)
+  rr_fit <- 
+    fit_xy(
+      ridge_mod,
+      control = ctrl,
+      x = iris[,2:4],
+      y = iris$Sepal.Length
+    )
+  
+  keras_pred <-
+    predict(rr_fit$fit, as.matrix(iris[1:3,2:4])) %>%
+    as_tibble() %>%
+    setNames(".pred")
+  parsnip_pred <- predict(rr_fit, iris[1:3,2:4])
+  expect_equal(as.data.frame(keras_pred), as.data.frame(parsnip_pred))
+  
+})