Release agtboost 0.9.2

See NEWS.md for details
- Patch fixing slow gbt.save() functionality.
- Obtain XGBoost and LightGBM hyperparameters from gbt.complexity().
- Include attribute "offset" in gbt.train() and predict().
- Throw error when gb-loss-approximation deviates from true loss. Suggest lower learning_rate.
- Solves $\arg\min_\eta \sum_i l(y_i, g^{-1}(offset_i+\eta))$ numerically instead of simple average to obtain initial prediction
- Deprecate warnings for zero-inflation
- Decrease package size drastically by stripping debugging information

Author: Blunde1

.gitignore

@@ -83,4 +83,6 @@ rsconnect/
 
 *.html
 
-*.zip
+*.zip
+
+.DS_Store

R-package/.Rbuildignore

@@ -2,4 +2,5 @@
 ^\.Rproj\.user$
 ^\.travis\.yml$
 ^NEWS\.md$          # A news file written in Markdown
+^cran-comments\.md$
 ^LICENSE

R-package/DESCRIPTION

@@ -2,7 +2,7 @@ Package: agtboost
 Type: Package
 Title: Adaptive and Automatic Gradient Boosting Computations
 Version: 0.9.2
-Date: 2020-09-06
+Date: 2021-11-09
 Author: Berent Ånund Strømnes Lunde
 Maintainer: Berent Ånund Strømnes Lunde <lundeberent@gmail.com>
 Description: Fast and automatic gradient tree boosting designed

R-package/NEWS.md

@@ -1,8 +1,14 @@
 ------------------------------------------------------------------------
-agtboost 0.9.0 (2020-08-12)
+agtboost 0.9.2 (2021-11-12)
 ------------------------------------------------------------------------
 
-- Initial release.
+- Patch fixing slow gbt.save() functionality.
+- Obtain XGBoost and LightGBM hyperparameters from gbt.complexity().
+- Include attribute "offset" in gbt.train() and predict().
+- Throw error when gb-loss-approximation deviates from true loss. Suggest lower learning_rate.
+- Solves $\arg\min_\eta \sum_i l(y_i, g^{-1}(offset_i+\eta))$ numerically instead of simple average to obtain initial prediction
+- Deprecate warnings for zero-inflation
+- Decrease package size drastically by stripping debugging information 
 
 ------------------------------------------------------------------------
 agtboost 0.9.1 (2020-10-04)
@@ -13,11 +19,7 @@ agtboost 0.9.1 (2020-10-04)
 - Checked with rhub::check_with_sanitizers() and rhub::check_on_solaris()
 
 ------------------------------------------------------------------------
-agtboost 0.9.2 (2021-11-01)
+agtboost 0.9.0 (2020-08-12)
 ------------------------------------------------------------------------
 
-- Patch fixing slow gbt.save() functionality.
-- Obtain XGBoost and LightGBM hyperparameters from gbt.complexity().
-- Include attribute "offset" in gbt.train() and predict().
-- Throw error when gb-loss-approximation deviates from true loss. Suggest lower learning_rate.
-- Solves $\arg\min_\eta \sum_i l(y_i, g^{-1}(offset_i+\eta))$ numerically instead of simple average to obtain initial prediction
+- Initial release.

R-package/R/gbt.complexity.R

@@ -68,22 +68,22 @@ gbt.model.complexity <- function(model){
     initial_raw_prediction = model$initialPred
 
     res <- list(
-      loss_function = loss_function,
-      nrounds = model$get_num_trees(),
-      learning_rate = model$get_learning_rate(),
-      initial_raw_prediction = initial_raw_prediction,
-      initial_prediction = transform_prediction(initial_raw_prediction, loss_function),
+      "loss_function" = loss_function,
+      "nrounds" = model$get_num_trees(),
+      "learning_rate" = model$get_learning_rate(),
+      "initial_raw_prediction" = initial_raw_prediction,
+      "initial_prediction" = transform_prediction(initial_raw_prediction, loss_function),
       # # tree parameters
-      max_depth = max(model$get_tree_depths()),
-      min_loss_reductions = model$get_max_node_optimism(),
-      sum_hessian_weights = model$get_min_hessian_weights(),
-      number_of_leaves = max(model$get_num_leaves()),
+      "max_depth" = max(model$get_tree_depths()),
+      "min_loss_reductions" = model$get_max_node_optimism(),
+      "sum_hessian_weights" = model$get_min_hessian_weights(),
+      "number_of_leaves" = max(model$get_num_leaves()),
       # # objective
-      l1_regularization = 0.0,
-      l2_regularization = 0.0,
+      "l1_regularization" = 0.0,
+      "l2_regularization" = 0.0,
       # # sampling
-      row_subsampling = 1.0,
-      column_subsampling = 1.0
+      "row_subsampling" = 1.0,
+      "column_subsampling" = 1.0
     )
     return(res)
 }
@@ -114,8 +114,9 @@ gbt.model.complexity <- function(model){
 #' yte <- rnorm(n, xte, 1)
 #' model <- gbt.train(ytr, xtr, learning_rate = 0.1)
 #' gbt.complexity(model, type="xgboost")
-#' gbt.complexity(model, type="lightgbm)
+#' gbt.complexity(model, type="lightgbm")
 #' ## See demo(topic="gbt-complexity", package="agtboost")
+#' }
 #' 
 #' @importFrom graphics barplot mtext par
 #' @rdname gbt.complexity
@@ -129,51 +130,50 @@ gbt.complexity <- function(model, type){
     # Setup or get agtboost implicit parameters/complexity measures
     model_complexity <- gbt.model.complexity(model)
 
-    
     # Transform parameters/complexity into library-specific parameters
-    attach(model_complexity)
-    if(type=="xgboost"){
-      # Transform agtboost parameters/complexity measures to xgboost parameters
-      parameters = list(
-        # ensemble param
-        "base_score" = initial_prediction,
-        "nrounds" = nrounds,
-        "learning_rate" = learning_rate,
-        # tree param
-        "max_depth" = max_depth,
-        "gamma" = min_loss_reductions,
-        "min_child_weight" = sum_hessian_weights,
-        "max_leaves" = number_of_leaves,
-        "grow_policy" = "lossguide",
-        # objective
-        "objective" = loss_to_xgbloss(loss_function),
-        "alpha" = 0.0,
-        "lambda" = 0.0,
-        # subsampling
-        "subsample" = 1.0,
-        "colsample_bytree" = 1.0
-      )
-    }else if(type=="lightgbm"){
-      # Transform agtboost parameters/complexity measures to lightgbm parameters
-      parameters = list(
-        # ensemble param
-        "init_score" = initial_prediction,
-        "nrounds" = nrounds,
-        "learning_rate" = learning_rate,
-        # tree param
-        "max_depth" = max_depth,
-        "min_gain_to_split" = min_loss_reductions,
-        "min_sum_hessian_in_leaf" = sum_hessian_weights,
-        "num_leaves" = number_of_leaves,
-        # objective
-        "objective" = loss_to_lgbloss(loss_function),
-        "lambda_l1" = 0.0,
-        "lambda_l2" = 0.0,
-        # subsampling
-        "bagging_fraction" = 1.0,
-        "feature_fraction" = 1.0
-      ) 
-    }
-    detach(model_complexity)
+    parameters = with(model_complexity,
+         if(type=="xgboost"){
+           # Transform agtboost parameters/complexity measures to xgboost parameters
+           list(
+             # ensemble param
+             "base_score" = initial_prediction,
+             "nrounds" = nrounds,
+             "learning_rate" = learning_rate,
+             # tree param
+             "max_depth" = max_depth,
+             "gamma" = min_loss_reductions,
+             "min_child_weight" = sum_hessian_weights,
+             "max_leaves" = number_of_leaves,
+             "grow_policy" = "lossguide",
+             # objective
+             "objective" = loss_to_xgbloss(loss_function),
+             "alpha" = 0.0,
+             "lambda" = 0.0,
+             # subsampling
+             "subsample" = 1.0,
+             "colsample_bytree" = 1.0
+           )
+         }else if(type=="lightgbm"){
+           # Transform agtboost parameters/complexity measures to lightgbm parameters
+           list(
+             # ensemble param
+             "init_score" = initial_prediction,
+             "nrounds" = nrounds,
+             "learning_rate" = learning_rate,
+             # tree param
+             "max_depth" = max_depth,
+             "min_gain_to_split" = min_loss_reductions,
+             "min_sum_hessian_in_leaf" = sum_hessian_weights,
+             "num_leaves" = number_of_leaves,
+             # objective
+             "objective" = loss_to_lgbloss(loss_function),
+             "lambda_l1" = 0.0,
+             "lambda_l2" = 0.0,
+             # subsampling
+             "bagging_fraction" = 1.0,
+             "feature_fraction" = 1.0
+             )
+         }
+    )
     return(parameters)
 }

R-package/R/gbt.train.R

@@ -26,7 +26,7 @@
 #' @param previous_pred prediction vector for training. Boosted training given predictions from another model.
 #' @param weights weights vector for scaling contributions of individual observations. Default \code{NULL} (the unit vector).
 #' @param force_continued_learning Boolean: \code{FALSE} (default) stops at information stopping criterion, \code{TRUE} stops at \code{nround} iterations.
-#' @param offset add offset to the model \code(g(mu) = offset + F(x)).
+#' @param offset add offset to the model g(mu) = offset + F(x).
 #' @param ... additional parameters passed.
 #'   \itemize{
 #'   \item if loss_function is 'negbinom', dispersion must be provided in \code{...}
@@ -102,6 +102,9 @@ gbt.train <- function(y, x, learning_rate = 0.01,
             algorithm = "vanilla"
         }
     }
+    if(loss_function %in% c("count::auto")){
+        warning("The `count::auto` option to the `loss_function` argument of `gbt.train()` will be deprecated as of 0.9.3")
+    }
 
     error_messages <- c()
     error_messages_type <- c(
@@ -151,8 +154,6 @@ gbt.train <- function(y, x, learning_rate = 0.01,
         if(
             loss_function %in% c("mse", "logloss", "poisson", "gamma::neginv", 
                                  "gamma::log", "negbinom", 
-                                 #"poisson::zip", "zero_inflation", "zero_inflation::poisson",
-                                 #"zero_inflation::negbinom", "zero_inflation::auto",
                                  "count::auto")
         ){}else{
             error_messages <- c(error_messages, error_messages_type["loss_fun"])
@@ -271,8 +272,15 @@ gbt.train <- function(y, x, learning_rate = 0.01,
         # train ensemble
         if(is.null(previous_pred)){
             # train from scratch
-            mod$train(y,x, verbose, gsub_compare, force_continued_learning, 
-                      weights, offset, !all(offset==0))   
+            mod$train(
+                y, 
+                x, 
+                verbose, 
+                gsub_compare, 
+                force_continued_learning,
+                weights, 
+                offset
+                )   
         }else{
 
             # train from previous predictions

R-package/README.md

@@ -0,0 +1,18 @@
+------------------------------------------------------------------------
+agtboost 0.9.2 (2021-11-09)
+------------------------------------------------------------------------
+## Test environments
+* local darwin17 install, R 4.1.1
+* winbuilder
+* devtools check
+* rhub check_for_cran, check_on_solaris, check_with_sanitizers
+
+## R CMD check results
+
+0 errors | 0 warnings | 1 notes
+
+* checking CRAN incoming feasibility ... NOTE
+Maintainer: 'Berent Ånund Strømnes Lunde <lundeberent@gmail.com>'
+
+## Changes in agtboost 0.9.2
+Some patching of serialization/deserialization of model objects, along with new features for model training. A deprecation notice is included, and package should be smaller on some OS due to stripping debugging information.

R-package/cran-comments.md

@@ -25,4 +25,4 @@
 #include "initial_prediction.hpp"
 
 
-#endif // __GMGTB_HPP_INCLUDED__
+#endif // __GMGTB_HPP_INCLUDED__

R-package/inst/include/agtboost.hpp

@@ -251,4 +251,4 @@ double simpson(Tvec<double>& fval, Tvec<double>& grid){
 
 }
 
-#endif
+#endif