butcher: Model Butcher

# NOTE:
# spark has proven to be very difficult to test in CI, due to installation
# issues across all platforms. Because of this, we choose not to test it at
# all. It is also not listed as a Suggest-ed package anymore.

#
# source("utils.R")
#
# test_that("spark ml_model + tbl_spark input + predict() works", {
#   skip_on_cran()
#   skip_on_os(c("windows", "solaris"))
#   skip_if_not_installed("sparklyr")
#   skip_if_not_installed("dplyr")
#   # Load
#   suppressPackageStartupMessages(library(sparklyr))
#   suppressPackageStartupMessages(library(dplyr))
#   # Create connection
#   sc <- testthat_spark_connection()
#   # Data
#   iris_tbls <- sdf_copy_to(sc, iris, overwrite = TRUE) %>%
#     sdf_random_split(train = 2/3, validation = 2/3, seed = 2018)
#   train <- iris_tbls$train
#   validation <- iris_tbls$validation
#   # Model
#   spark_fit <- ml_logistic_regression(train, Species ~ .)
#   expected_output <- ml_predict(spark_fit, validation) %>%
#     select(predicted_label) %>%
#     collect()
#   # Test sub axe functions
#   x <- axe_call(spark_fit)
#   expect_equal(x$formula, "")
#   x <- axe_ctrl(spark_fit)
#   expect_equal(x$label_col, "")
#   expect_equal(x$features_col, "")
#   expect_equal(x$feature_names, "")
#   expect_equal(x$response, "")
#   expect_equal(x$index_labels, "")
#   x <- axe_data(spark_fit)
#   expect_equal(x$dataset, NULL)
#   expect_equal(spark_fit$dataset, train)
#   x <- axe_fitted(spark_fit)
#   expect_equal(x$coefficients, matrix(NA))
#   expect_null(x$model)
#   expect_null(x$summary)
#   # Butcher
#   x <- butcher(spark_fit)
#   version <- x %>%
#     spark_jobj() %>%
#     spark_connection() %>%
#     spark_version()
#   expect_false(version < "2.0.0")
#   output <- ml_predict(x$pipeline_model, validation) %>%
#     select(predicted_label) %>%
#     collect()
#   expect_gt(lobstr::obj_size(spark_fit), lobstr::obj_size(x))
#   expect_equal(output, expected_output)
#   expect_equal(expected_output$predicted_label[1], "setosa")
#   expect_null(ml_stages(x))
#   # Now test on temp object
#   path <- tempfile()
#   ml_save(x, path, overwrite = TRUE)
#   x_loaded <- ml_load(sc, path)
#   path <- tempfile()
#   ml_save(spark_fit, path, overwrite = TRUE)
#   spark_fit_loaded <- ml_load(sc, path)
#   expected_output <- spark_fit_loaded %>%
#     ml_stages() %>%
#     lapply(ml_param_map)
#   new_output <- x_loaded %>%
#     ml_stages() %>%
#     lapply(ml_param_map)
#   expect_equal(new_output, expected_output)
#   expect_equal(x_loaded$uid, spark_fit_loaded$uid)
#   # Test sub axe functions
#   axed_loaded <- axe_call(spark_fit_loaded)
#   expect_null(axed_loaded$formula)
#   axed_loaded <- axe_ctrl(spark_fit_loaded)
#   expect_null(axed_loaded$label_col)
#   expect_null(axed_loaded$features_col)
#   expect_null(axed_loaded$response)
#   expect_null(axed_loaded$index_labels)
#   axed_loaded <- axe_data(spark_fit_loaded)
#   expect_null(axed_loaded$dataset)
#   axed_loaded <- axe_fitted(spark_fit_loaded)
#   expect_null(axed_loaded$coefficients)
#   # Butcher
#   axed_loaded <- butcher(spark_fit_loaded)
#   expect_equal(axed_loaded, spark_fit_loaded)
# })
#
# test_that("spark decision_tree + predict() works", {
#   skip_on_cran()
#   skip_if_not_installed("sparklyr")
#   skip_if_not_installed("dplyr")
#   # Load
#   suppressPackageStartupMessages(library(sparklyr))
#   suppressPackageStartupMessages(library(dplyr))
#   # Create connection
#   sc <- testthat_spark_connection()
#   # Data
#   iris_tbls <- sdf_copy_to(sc, iris, overwrite = TRUE) %>%
#     sdf_random_split(train = 2/3, validation = 2/3, seed = 2018)
#   train <- iris_tbls$train
#   validation <- iris_tbls$validation
#   spark_fit <- ml_decision_tree_classifier(train, Species ~ .)
#   expected_output <- ml_predict(spark_fit, validation) %>%
#     select(predicted_label) %>%
#     collect()
#   # Butcher
#   x <- butcher(spark_fit)
#   expect_gt(lobstr::obj_size(spark_fit), lobstr::obj_size(x))
#   output <- ml_predict(x$pipeline_model, validation) %>%
#     select(predicted_label) %>%
#     collect()
#   expect_equal(output, expected_output)
# })
#
# test_that("spark boost_tree + predict() works", {
#   skip_on_cran()
#   skip_if_not_installed("sparklyr")
#   skip_if_not_installed("dplyr")
#   # Load
#   suppressPackageStartupMessages(library(sparklyr))
#   suppressPackageStartupMessages(library(dplyr))
#   # Create connection
#   sc <- testthat_spark_connection()
#   # Binary classification data
#   iris_bin <- iris[iris$Species != "setosa", ]
#   iris_bin$Species <- factor(iris_bin$Species)
#   iris_bin_tbls <- sdf_copy_to(sc, iris_bin, overwrite = TRUE) %>%
#     sdf_random_split(train = 2/3, validation = 2/3, seed = 2018)
#   train <- iris_bin_tbls$train
#   validation <- iris_bin_tbls$validation
#   # Model
#   spark_fit <- ml_gradient_boosted_trees(train, Species ~ .)
#   expected_output <- ml_predict(spark_fit, validation) %>%
#     select(predicted_label) %>%
#     collect()
#   x <- butcher(spark_fit)
#   expect_gt(lobstr::obj_size(spark_fit), lobstr::obj_size(x))
#   output <- ml_predict(x$pipeline_model, validation) %>%
#     select(predicted_label) %>%
#     collect()
#   expect_equal(output, expected_output)
#   # Save and load
#   path <- tempfile()
#   ml_save(x, path, overwrite = TRUE)
#   x_loaded <- ml_load(sc, path)
#   path <- tempfile()
#   ml_save(spark_fit, path, overwrite = TRUE)
#   spark_fit_loaded <- ml_load(sc, path)
#   expected_output <- spark_fit_loaded %>%
#     ml_stages() %>%
#     lapply(ml_param_map)
#   new_output <- x_loaded %>%
#     ml_stages() %>%
#     lapply(ml_param_map)
#   expect_equal(new_output, expected_output)
#   expect_equal(x_loaded$uid, spark_fit_loaded$uid)
#   # Test sub axe functions
#   axed_loaded <- axe_call(spark_fit_loaded)
#   expect_null(axed_loaded$formula)
#   axed_loaded <- axe_ctrl(spark_fit_loaded)
#   expect_null(axed_loaded$label_col)
#   expect_null(axed_loaded$features_col)
#   expect_null(axed_loaded$response)
#   expect_null(axed_loaded$index_labels)
#   axed_loaded <- axe_data(spark_fit_loaded)
#   expect_null(axed_loaded$dataset)
#   axed_loaded <- axe_fitted(spark_fit_loaded)
#   expect_null(axed_loaded$coefficients)
#   # Butcher
#   axed_loaded <- butcher(spark_fit_loaded)
#   expect_equal(axed_loaded, spark_fit_loaded)
# })
#
# test_that("spark linear_reg + predict() works", {
#   skip_on_cran()
#   skip_if_not_installed("sparklyr")
#   skip_if_not_installed("dplyr")
#   skip_if_not_installed("parsnip")
#   # Load
#   suppressPackageStartupMessages(library(sparklyr))
#   suppressPackageStartupMessages(library(dplyr))
#   suppressPackageStartupMessages(library(parsnip))
#   # Create connection
#   sc <- testthat_spark_connection()
#   # Data
#   iris_tbls <- sdf_copy_to(sc, iris, overwrite = TRUE) %>%
#     sdf_random_split(train = 2/3, validation = 2/3, seed = 2018)
#   train <- iris_tbls$train
#   validation <- iris_tbls$validation
#   # Model
#   reg_spark <- linear_reg() %>%
#     set_engine("spark") %>%
#     fit(Petal_Width ~ ., data = train)
#   expected_output <- ml_predict(reg_spark$fit, validation) %>%
#     select(prediction) %>%
#     collect()
#   x <- butcher(reg_spark)
#   output <- ml_predict(x$fit$pipeline_model, validation) %>%
#     select(prediction) %>%
#     collect()
#   expect_gt(lobstr::obj_size(reg_spark), lobstr::obj_size(x))
#   expect_equal(output, expected_output)
#   # Save and load
#   path <- tempfile()
#   ml_save(x$fit, path, overwrite = TRUE)
#   x_loaded <- ml_load(sc, path)
#   path <- tempfile()
#   ml_save(reg_spark$fit, path, overwrite = TRUE)
#   spark_fit_loaded <- ml_load(sc, path)
#   expected_output <- spark_fit_loaded %>%
#     ml_stages() %>%
#     lapply(ml_param_map)
#   new_output <- x_loaded %>%
#     ml_stages() %>%
#     lapply(ml_param_map)
#   expect_equal(new_output, expected_output)
#   expect_equal(x_loaded$uid, spark_fit_loaded$uid)
#   # Test sub axe functions
#   axed_loaded <- axe_call(spark_fit_loaded)
#   expect_null(axed_loaded$formula)
#   axed_loaded <- axe_ctrl(spark_fit_loaded)
#   expect_null(axed_loaded$label_col)
#   expect_null(axed_loaded$features_col)
#   expect_null(axed_loaded$response)
#   expect_null(axed_loaded$index_labels)
#   axed_loaded <- axe_data(spark_fit_loaded)
#   expect_null(axed_loaded$dataset)
#   axed_loaded <- axe_fitted(spark_fit_loaded)
#   expect_null(axed_loaded$coefficients)
#   # Butcher
#   axed_loaded <- butcher(spark_fit_loaded)
#   expect_equal(axed_loaded, spark_fit_loaded)
# })