Custom classifier with quasi-randomized layer

Parameters' description can be found at https://techtonique.github.io/nnetsauce/

CustomClassifier(
  obj,
  n_hidden_features = 5L,
  activation_name = "relu",
  a = 0.01,
  nodes_sim = "sobol",
  bias = TRUE,
  dropout = 0,
  direct_link = TRUE,
  n_clusters = 2L,
  cluster_encode = TRUE,
  type_clust = "kmeans",
  col_sample = 1,
  row_sample = 1,
  seed = 123L,
  backend = c("cpu", "gpu", "tpu")
)

Examples

library(datasets)

set.seed(123)
X <- as.matrix(iris[, 1:4])
y <- as.integer(iris$Species) - 1L

(index_train <- base::sample.int(n = nrow(X),
                                 size = floor(0.8*nrow(X)),
                                 replace = FALSE))
#>   [1]  14  50 118  43 150 148  90  91 143  92 137  99  72  26   7  78  81 147
#>  [19] 103 117  76  32 106 109 136   9  41  74  23  27  60  53 126 119 121  96
#>  [37]  38  89  34  93  69 138 130  63  13  82  97 142  25 114  21  79 124  47
#>  [55] 144 120  16   6 127  86 132  39  31 134 149 112   4 128 110 102  52  22
#>  [73] 129  87  35  40  30  12  88 123  64 146  67 122  37   8  51  10 115  42
#>  [91]  44  85 107 139  73  20  46  17  54 108  75  80  71  15  24  68 133 145
#> [109]  29 104  45 140 101 135  95 116   5 111  94  49
X_train <- X[index_train, ]
y_train <- y[index_train]
X_test <- X[-index_train, ]
y_test <- y[-index_train]

obj <- sklearn$tree$DecisionTreeClassifier()
obj2 <- CustomClassifier(obj)
obj2$fit(X_train, y_train)
#> Error in py_call_impl(callable, call_args$unnamed, call_args$named): AttributeError: 'list' object has no attribute 'dtype'
#> Run `reticulate::py_last_error()` for details.
print(obj2$score(X_test, y_test))
#> Error in py_call_impl(callable, call_args$unnamed, call_args$named): sklearn.exceptions.NotFittedError: This DecisionTreeClassifier instance is not fitted yet. Call 'fit' with appropriate arguments before using this estimator.
#> Run `reticulate::py_last_error()` for details.