Modified all tests to not depend on data generated in "auxiliar" vign…

…ettes.

This allowed to delete said vignettes the 4 data examples stored that were not being used. To remplace this data in tests, helper-testing_data.R has been with 2 functions to generate data for tests.

R/nn2poly_algorithm.R

@@ -27,7 +27,7 @@ nn2poly_algorithm <- function(weights_list,
                               all_partitions = NULL
                               ) {
 
-  if (!check_weights_dimensions(object)) {
+  if (!check_weights_dimensions(weights_list)) {
     stop("The list of weights has incorrect dimensions.
          Please, check the  right dimmensions in the documentation.",
          call. = FALSE

tests/testthat/helper-luz.R

@@ -28,8 +28,8 @@ luz_test_data <- function(example) {
   data_val   <- torch::dataset_subset(data_full, val_indices)
 
   list(
-    train = torch::dataloader(data_train, batch_size = 32, shuffle = TRUE),
-    valid = torch::dataloader(data_val, batch_size = 32)
+    train = torch::dataloader(data_train, batch_size = 5, shuffle = TRUE),
+    valid = torch::dataloader(data_val, batch_size = 5)
   )
 }
 

tests/testthat/helper-testing_data.R

@@ -0,0 +1,36 @@
+
+# Testing data 1 ----
+# Created to test with all weights equal to 1, not training any NN.
+# With these values we have some tests with manually computed coefficients.
+testing_helper_1 <- function(){
+  testing_data_1 <- vector(mode="list", length= 0)
+  testing_data_1$weights_list <- vector(mode="list", length= 3)
+  testing_data_1$weights_list[[1]] <- matrix(1,3,2)
+  testing_data_1$weights_list[[2]] <- matrix(1,3,2)
+  testing_data_1$weights_list[[3]] <- matrix(1,3,1)
+  testing_data_1$af_string_list <- list("softplus", "softplus", "linear")
+  testing_data_1$taylor_orders <- c(2, 2, 1)
+  return(testing_data_1)
+}
+
+
+# Testing data 2 ----
+# In this case we will generate some polynomial data with linear output
+# This data will be used in testing keras/tensorflow and luz/torch
+# Note that the data is random so the NN will not learn properly but
+# we can still test if the coefficients are as expected by nn2poly.
+
+testing_helper_2 <- function(){
+  set.seed(42)
+  n_sample <- 100
+  p <- 2
+  testing_data_2 <- vector(mode="list", length= 0)
+  testing_data_2$train_x <- matrix(rnorm(n_sample*p),ncol = p)
+  testing_data_2$train_y <- matrix(rnorm(n_sample),ncol = 1)
+  testing_data_2$taylor_orders <- c(2, 2, 1)
+  return(testing_data_2)
+}
+
+
+
+

tests/testthat/test-add_constraints.R

@@ -50,6 +50,9 @@ test_that("The function works over an already trained network and the connstrain
   skip_if_not_installed("tensorflow")
   skip_on_cran()
 
+  testing_data <- testing_helper_2()
+  testing_data
+
   nn <- keras_test_model()
 
   # compile the model
@@ -60,10 +63,10 @@ test_that("The function works over an already trained network and the connstrain
 
   # train the model
   keras::fit(nn,
-             nn2poly_example0$train_x,
-             nn2poly_example0$train_y,
+             x = testing_data$train_x,
+             y = testing_data$train_y,
              verbose = 0,
-             epochs = 5,
+             epochs = 3,
              validation_split = 0.2
   )
 
@@ -79,10 +82,10 @@ test_that("The function works over an already trained network and the connstrain
 
   # train the model
   keras::fit(constrained_nn,
-             nn2poly_example0$train_x,
-             nn2poly_example0$train_y,
+             testing_data$train_x,
+             testing_data$train_y,
              verbose = 0,
-             epochs = 5,
+             epochs = 3,
              validation_split = 0.2
   )
 

tests/testthat/test-luz_constraint.R

@@ -3,7 +3,9 @@ test_that("The constranied training works using the l1 and l2 constraints", {
   skip_if_not_installed("torch")
   skip_on_cran()
 
-  data <- luz_test_data(nn2poly_example0)
+  testing_data <- testing_helper_2()
+
+  data <- luz_test_data(testing_data)
 
   # Do the constrained training with the l1 constraint
   fitted <- luz_test_model() %>%
@@ -12,7 +14,7 @@ test_that("The constranied training works using the l1 and l2 constraints", {
       optimizer = torch::optim_adam,
     ) %>%
     add_constraints("l1_norm") %>%
-    fit(data$train, epochs = 3, valid_data = data$valid)
+    fit(data$train, epochs = 3, valid_data = data$valid, verbose = FALSE)
 
   wb <- torch::torch_tensor(
     rbind(
@@ -36,7 +38,7 @@ test_that("The constranied training works using the l1 and l2 constraints", {
       optimizer = torch::optim_adam,
     ) %>%
     add_constraints("l2_norm") %>%
-    fit(data$train, epochs = 3, valid_data = data$valid)
+    fit(data$train, epochs = 3, valid_data = data$valid, verbose = FALSE)
 
   wb <- torch::torch_tensor(
     rbind(

tests/testthat/test-nn2poly.R

@@ -1,13 +1,11 @@
 test_that("nn2poly with list input against precomputed example with single
           output, check also that nn2poly class is given to the output", {
-  nn2poly_example <- nn2poly_example0
+  testing_data <- testing_helper_1()
 
   # Get the needed data
-  object <- nn2poly_example$weights_list
-  names(object) <- nn2poly_example$af_string_list
-
-  taylor_orders <- nn2poly_example$q_taylor_vector
-
+  object <- testing_data$weights_list
+  names(object) <- testing_data$af_string_list
+  taylor_orders <- testing_data$taylor_orders
 
   result <- nn2poly(
     object = object,
@@ -34,11 +32,11 @@ test_that("nn2poly with list input against precomputed example with single
 
 test_that("nn2poly with list input against precomputed example with
           default options", {
-  nn2poly_example <- nn2poly_example0
+  testing_data <- testing_helper_1()
 
   # Get the needed data
-  object <- nn2poly_example$weights_list
-  names(object) <- nn2poly_example$af_string_list
+  object <- testing_data$weights_list
+  names(object) <- testing_data$af_string_list
 
   result <- nn2poly(
     object = object
@@ -96,7 +94,9 @@ test_that("nn2poly for a nn_module object", {
   skip_if_not_installed("torch")
   skip_on_cran()
 
-  data <- luz_test_data(nn2poly_example0)
+  testing_data <- testing_helper_2()
+
+  data <- luz_test_data(testing_data)
 
   fitted <- luz_test_model() %>%
     luz::setup(
@@ -109,21 +109,20 @@ test_that("nn2poly for a nn_module object", {
     luz::fit(data$train, epochs = 5, valid_data = data$valid, verbose = FALSE)
 
   result <- nn2poly(fitted,
-                    taylor_orders = nn2poly_example0$q_taylor_vector,
+                    taylor_orders = testing_data$taylor_orders,
                     max_order = 3)
 
-  expect_equal(round(result$values[1,1],2), 0.11)
-  # expect_equal(result$values[2,1], -0.45410551)
+  expect_equal(round(result$values[1,1],2), 0.06)
   expect_equal(result$labels[[7]], c(1,1,1))
 
 })
 
 test_that("nn2poly error when wrong dimensions are given in weights", {
-  nn2poly_example <- nn2poly_example0
+  testing_data <- testing_helper_1()
 
   # Get the needed data
-  object <- nn2poly_example$weights_list
-  names(object) <- nn2poly_example$af_string_list
+  object <- testing_data$weights_list
+  names(object) <- testing_data$af_string_list
 
   # Add weights to one layer to create the problem
   object[[2]] <- rbind(object[[2]], c(1,1))

tests/testthat/test-nn2poly_algorithm.R

@@ -2,12 +2,12 @@
 
 test_that("nn2poly_algorithm against precomputed example", {
   # Load the example:
-  nn2poly_example <- nn2poly_example0
+  testing_data <- testing_helper_1()
 
   # Get the needed data
-  weights_list <- nn2poly_example$weights_list
-  af_string_list <- nn2poly_example$af_string_list
-  taylor_orders <- nn2poly_example$q_taylor_vector
+  weights_list <- testing_data$weights_list
+  af_string_list <- testing_data$af_string_list
+  taylor_orders <- testing_data$taylor_orders
 
   result <- nn2poly_algorithm(
     weights_list = weights_list,

tests/testthat/test-parameters.R

@@ -45,5 +45,5 @@ test_that("The get_parameters function works for a torch model (nn_module)
   nn <- luz_test_model()
 
   params <- get_parameters(nn)
-  expect_equal(params$af_string_list, nn2poly_example0$af_string_list)
+  expect_equal(params$af_string_list, list("softplus", "softplus", "linear"))
 })

tests/testthat/test-plot_n_important_coeffs.R

@@ -1,9 +1,9 @@
 test_that("Test the plot for a polynomial generated `keep_layers = FALSE`", {
   # loading the example
-  nn2poly_example <- nn2poly_example0
-  object <- nn2poly_example$weights_list
-  names(object) <- nn2poly_example$af_string_list
-  taylor_orders <- nn2poly_example$q_taylor_vector
+  testing_data <- testing_helper_1()
+  object <- testing_data$weights_list
+  names(object) <- testing_data$af_string_list
+  taylor_orders <- testing_data$taylor_orders
 
   # computing the polynomial
   result <- nn2poly(
@@ -21,10 +21,10 @@ test_that("Test the plot for a polynomial generated `keep_layers = FALSE`", {
 
 test_that("Test the plot for a polynomial generated `keep_layers = TRUE`", {
   # loading the example
-  nn2poly_example <- nn2poly_example0
-  object <- nn2poly_example$weights_list
-  names(object) <- nn2poly_example$af_string_list
-  taylor_orders <- nn2poly_example$q_taylor_vector
+  testing_data <- testing_helper_1()
+  object <- testing_data$weights_list
+  names(object) <- testing_data$af_string_list
+  taylor_orders <- testing_data$taylor_orders
 
   # computing the polynomial
   result <- nn2poly(

tests/testthat/test-plot_taylor_and_activation_potentials.R

@@ -6,6 +6,8 @@ test_that("The function works as expected", {
   skip_on_covr()
   skip_on_ci()
 
+  testing_data <- testing_helper_2()
+
   nn <- keras_test_model()
 
   # compile the model
@@ -16,8 +18,8 @@ test_that("The function works as expected", {
 
   # train the model
   keras::fit(nn,
-             nn2poly_example0$train_x,
-             nn2poly_example0$train_y,
+             testing_data$train_x,
+             testing_data$train_y,
              verbose = 0,
              epochs = 30,
              validation_split = 0.3
@@ -26,8 +28,8 @@ test_that("The function works as expected", {
   # compute the different plots
   plots <- plot_taylor_and_activation_potentials(
     object = nn,
-    data = cbind(nn2poly_example0$train_x, nn2poly_example0$train_y),
-    taylor_orders = nn2poly_example0$q_taylor_vector,
+    data = cbind(testing_data$train_x, testing_data$train_y),
+    taylor_orders = testing_data$taylor_orders,
     max_order = 3,
     constraints = FALSE)