Merge pull request #1118 from khotilov/parsing_speedup
[R-package] xgb.model.dt.tree up to x100 faster
This commit is contained in:
Michaël Benesty
commit
51154f42fe
@ -43,19 +43,14 @@ importFrom(Matrix,sparse.model.matrix)
|
||||
importFrom(Matrix,sparseVector)
|
||||
importFrom(data.table,":=")
|
||||
importFrom(data.table,as.data.table)
|
||||
importFrom(data.table,copy)
|
||||
importFrom(data.table,data.table)
|
||||
importFrom(data.table,fread)
|
||||
importFrom(data.table,rbindlist)
|
||||
importFrom(data.table,set)
|
||||
importFrom(data.table,setnames)
|
||||
importFrom(magrittr,"%>%")
|
||||
importFrom(magrittr,add)
|
||||
importFrom(magrittr,not)
|
||||
importFrom(stringr,str_detect)
|
||||
importFrom(stringr,str_extract)
|
||||
importFrom(stringr,str_extract_all)
|
||||
importFrom(stringr,str_match)
|
||||
importFrom(stringr,str_replace)
|
||||
importFrom(stringr,str_split)
|
||||
importFrom(stringr,str_trim)
|
||||
|
||||
@ -1,153 +1,127 @@
|
||||
#' Parse boosted tree model text dump
|
||||
#' Parse a boosted tree model text dump
|
||||
#'
|
||||
#' Parse a boosted tree model text dump and return a \code{data.table}.
|
||||
#' Parse a boosted tree model text dump into a \code{data.table} structure.
|
||||
#'
|
||||
#' @importFrom data.table data.table
|
||||
#' @importFrom data.table set
|
||||
#' @importFrom data.table rbindlist
|
||||
#' @importFrom data.table copy
|
||||
#' @importFrom data.table :=
|
||||
#' @importFrom magrittr %>%
|
||||
#' @importFrom magrittr not
|
||||
#' @importFrom magrittr add
|
||||
#' @importFrom stringr str_extract
|
||||
#' @importFrom stringr str_split
|
||||
#' @importFrom stringr str_trim
|
||||
#' @param feature_names names of each feature as a character vector. Can be extracted from a sparse matrix (see example). If the model already contains feature names, this argument should be \code{NULL} (default value).
|
||||
#' @param model object created by the \code{xgb.train} function.
|
||||
#' @param text \code{character} vector generated by the \code{xgb.dump} function. Model dump must include the gain per feature and per tree (parameter \code{with.stats = TRUE} in function \code{xgb.dump}).
|
||||
#' @param n_first_tree limit the plot to the \code{n} first trees. If set to \code{NULL}, all trees of the model are plotted. Performance can be low depending of the size of the model.
|
||||
#'
|
||||
#' @return A \code{data.table} of the features used in the model with their gain, cover and few other information.
|
||||
#'
|
||||
#' @details
|
||||
#' General function to convert a text dump of tree model to a \code{data.table}.
|
||||
#'
|
||||
#' The purpose is to help user to explore the model and get a better understanding of it.
|
||||
#' @importFrom stringr str_match
|
||||
#'
|
||||
#' @param feature_names character vector of feature names. If the model already
|
||||
#' contains feature names, this argument should be \code{NULL} (default value)
|
||||
#' @param model object of class \code{xgb.Booster}
|
||||
#' @param text \code{character} vector previously generated by the \code{xgb.dump}
|
||||
#' function (where parameter \code{with.stats = TRUE} should have been set).
|
||||
#' @param n_first_tree limit the parsing to the \code{n} first trees.
|
||||
#' If set to \code{NULL}, all trees of the model are parsed.
|
||||
#'
|
||||
#' @return
|
||||
#' A \code{data.table} with detailed information about model trees' nodes.
|
||||
#'
|
||||
#' The columns of the \code{data.table} are:
|
||||
#'
|
||||
#' \itemize{
|
||||
#' \item \code{ID}: unique identifier of a node ;
|
||||
#' \item \code{Feature}: feature used in the tree to operate a split. When Leaf is indicated, it is the end of a branch ;
|
||||
#' \item \code{Split}: value of the chosen feature where is operated the split ;
|
||||
#' \item \code{Yes}: ID of the feature for the next node in the branch when the split condition is met ;
|
||||
#' \item \code{No}: ID of the feature for the next node in the branch when the split condition is not met ;
|
||||
#' \item \code{Missing}: ID of the feature for the next node in the branch for observation where the feature used for the split are not provided ;
|
||||
#' \item \code{Quality}: it's the gain related to the split in this specific node ;
|
||||
#' \item \code{Cover}: metric to measure the number of observation affected by the split ;
|
||||
#' \item \code{Tree}: ID of the tree. It is included in the main ID ;
|
||||
#' \item \code{Yes.Feature}, \code{No.Feature}, \code{Yes.Cover}, \code{No.Cover}, \code{Yes.Quality} and \code{No.Quality}: data related to the pointer in \code{Yes} or \code{No} column ;
|
||||
#' \item \code{Tree}: ID of a tree in a model
|
||||
#' \item \code{Node}: ID of a node in a tree
|
||||
#' \item \code{ID}: unique identifier of a node in a model
|
||||
#' \item \code{Feature}: for a branch node, it's a feature id or name (when available);
|
||||
#' for a leaf note, it simply labels it as \code{'Leaf'}
|
||||
#' \item \code{Split}: location of the split for a branch node (split condition is always "less than")
|
||||
#' \item \code{Yes}: ID of the next node when the split condition is met
|
||||
#' \item \code{No}: ID of the next node when the split condition is not met
|
||||
#' \item \code{Missing}: ID of the next node when branch value is missing
|
||||
#' \item \code{Quality}: either the split gain (change in loss) or the leaf value
|
||||
#' \item \code{Cover}: metric related to the number of observation either seen by a split
|
||||
#' or collected by a leaf during training.
|
||||
#' }
|
||||
#'
|
||||
#'
|
||||
#' @examples
|
||||
#' # Basic use:
|
||||
#'
|
||||
#' data(agaricus.train, package='xgboost')
|
||||
#'
|
||||
#' bst <- xgboost(data = agaricus.train$data, label = agaricus.train$label, max.depth = 2,
|
||||
#' eta = 1, nthread = 2, nround = 2,objective = "binary:logistic")
|
||||
#'
|
||||
#' # agaricus.train$data@@Dimnames[[2]] represents the column names of the sparse matrix.
|
||||
#' xgb.model.dt.tree(feature_names = agaricus.train$data@@Dimnames[[2]], model = bst)
|
||||
#' (dt <- xgb.model.dt.tree(colnames(agaricus.train$data), bst))
|
||||
#'
|
||||
#'
|
||||
#' # How to match feature names of splits that are following a current 'Yes' branch:
|
||||
#'
|
||||
#' merge(dt, dt[, .(ID, Y.Feature=Feature)], by.x='Yes', by.y='ID', all.x=TRUE)[order(Tree,Node)]
|
||||
#'
|
||||
#' @export
|
||||
xgb.model.dt.tree <- function(feature_names = NULL, model = NULL, text = NULL, n_first_tree = NULL){
|
||||
|
||||
xgb.model.dt.tree <- function(feature_names = NULL, model = NULL, text = NULL,
|
||||
n_first_tree = NULL){
|
||||
|
||||
if (!class(feature_names) %in% c("character", "NULL")) {
|
||||
stop("feature_names: Has to be a vector of character or NULL if the model dump already contains feature name. Look at this function documentation to see where to get feature names.")
|
||||
stop("feature_names: Has to be a vector of character\n",
|
||||
" or NULL if the model dump already contains feature names.\n",
|
||||
" Look at this function documentation to see where to get feature names.")
|
||||
}
|
||||
|
||||
|
||||
if (class(model) != "xgb.Booster" & class(text) != "character") {
|
||||
"model: Has to be an object of class xgb.Booster model generaged by the xgb.train function.\n" %>%
|
||||
paste0("text: Has to be a vector of character or NULL if a path to the model dump has already been provided.") %>%
|
||||
stop()
|
||||
stop("Either 'model' has to be an object of class xgb.Booster\n",
|
||||
" or 'text' has to be a character vector with the result of xgb.dump\n",
|
||||
" (or NULL if the model was provided).")
|
||||
}
|
||||
|
||||
|
||||
if (!class(n_first_tree) %in% c("numeric", "NULL") | length(n_first_tree) > 1) {
|
||||
stop("n_first_tree: Has to be a numeric vector of size 1.")
|
||||
}
|
||||
|
||||
if(is.null(text)){
|
||||
|
||||
if(is.null(text)){
|
||||
text <- xgb.dump(model = model, with.stats = T)
|
||||
}
|
||||
|
||||
position <- str_match(text, "booster") %>% is.na %>% not %>% which %>% c(length(text) + 1)
|
||||
|
||||
extract <- function(x, pattern) str_extract(x, pattern) %>% str_split("=") %>% lapply(function(x) x[2] %>% as.numeric) %>% unlist
|
||||
|
||||
n_round <- min(length(position) - 1, n_first_tree)
|
||||
|
||||
position <- which(!is.na(str_match(text, "booster")))
|
||||
|
||||
addTreeId <- function(x, i) paste(i,x,sep = "-")
|
||||
|
||||
allTrees <- data.table()
|
||||
|
||||
|
||||
anynumber_regex <- "[-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?"
|
||||
for (i in 1:n_round){
|
||||
|
||||
td <- data.table(t=text)
|
||||
td[position, Tree := 1L]
|
||||
td[, Tree := cumsum(ifelse(is.na(Tree), 0L, Tree)) - 1L]
|
||||
|
||||
n_first_tree <- min(max(td$Tree), n_first_tree)
|
||||
td <- td[Tree <= n_first_tree & !grepl('^booster', t)]
|
||||
|
||||
td[, Node := str_match(t, "(\\d+):")[,2] %>% as.numeric ]
|
||||
td[, ID := addTreeId(Node, Tree)]
|
||||
td[, isLeaf := !is.na(str_match(t, "leaf"))]
|
||||
|
||||
tree <- text[(position[i] + 1):(position[i + 1] - 1)]
|
||||
|
||||
# avoid tree made of a leaf only (no split)
|
||||
if(length(tree) < 2) next
|
||||
|
||||
treeID <- i - 1
|
||||
|
||||
notLeaf <- str_match(tree, "leaf") %>% is.na
|
||||
leaf <- notLeaf %>% not %>% tree[.]
|
||||
branch <- notLeaf %>% tree[.]
|
||||
idBranch <- str_extract(branch, "\\d*:") %>% str_replace(":", "") %>% addTreeId(treeID)
|
||||
idLeaf <- str_extract(leaf, "\\d*:") %>% str_replace(":", "") %>% addTreeId(treeID)
|
||||
featureBranch <- str_extract(branch, "f\\d*<") %>% str_replace("<", "") %>% str_replace("f", "") %>% as.numeric
|
||||
if(!is.null(feature_names)){
|
||||
featureBranch <- feature_names[featureBranch + 1]
|
||||
}
|
||||
featureLeaf <- rep("Leaf", length(leaf))
|
||||
splitBranch <- str_extract(branch, paste0("<",anynumber_regex,"\\]")) %>% str_replace("<", "") %>% str_replace("\\]", "")
|
||||
splitLeaf <- rep(NA, length(leaf))
|
||||
yesBranch <- extract(branch, "yes=\\d*") %>% addTreeId(treeID)
|
||||
yesLeaf <- rep(NA, length(leaf))
|
||||
noBranch <- extract(branch, "no=\\d*") %>% addTreeId(treeID)
|
||||
noLeaf <- rep(NA, length(leaf))
|
||||
missingBranch <- extract(branch, "missing=\\d+") %>% addTreeId(treeID)
|
||||
missingLeaf <- rep(NA, length(leaf))
|
||||
qualityBranch <- extract(branch, paste0("gain=",anynumber_regex))
|
||||
qualityLeaf <- extract(leaf, paste0("leaf=",anynumber_regex))
|
||||
coverBranch <- extract(branch, "cover=\\d*\\.*\\d*")
|
||||
coverLeaf <- extract(leaf, "cover=\\d*\\.*\\d*")
|
||||
dt <- data.table(ID = c(idBranch, idLeaf), Feature = c(featureBranch, featureLeaf), Split = c(splitBranch, splitLeaf), Yes = c(yesBranch, yesLeaf), No = c(noBranch, noLeaf), Missing = c(missingBranch, missingLeaf), Quality = c(qualityBranch, qualityLeaf), Cover = c(coverBranch, coverLeaf))[order(ID)][,Tree := treeID]
|
||||
|
||||
allTrees <- rbindlist(list(allTrees, dt), use.names = T, fill = F)
|
||||
}
|
||||
|
||||
yes <- allTrees[!is.na(Yes), Yes]
|
||||
|
||||
set(allTrees, i = which(allTrees[, Feature] != "Leaf"),
|
||||
j = "Yes.Feature",
|
||||
value = allTrees[ID %in% yes, Feature])
|
||||
|
||||
set(allTrees, i = which(allTrees[, Feature] != "Leaf"),
|
||||
j = "Yes.Cover",
|
||||
value = allTrees[ID %in% yes, Cover])
|
||||
|
||||
set(allTrees, i = which(allTrees[, Feature] != "Leaf"),
|
||||
j = "Yes.Quality",
|
||||
value = allTrees[ID %in% yes, Quality])
|
||||
no <- allTrees[!is.na(No), No]
|
||||
|
||||
set(allTrees, i = which(allTrees[, Feature] != "Leaf"),
|
||||
j = "No.Feature",
|
||||
value = allTrees[ID %in% no, Feature])
|
||||
|
||||
set(allTrees, i = which(allTrees[, Feature] != "Leaf"),
|
||||
j = "No.Cover",
|
||||
value = allTrees[ID %in% no, Cover])
|
||||
|
||||
set(allTrees, i = which(allTrees[, Feature] != "Leaf"),
|
||||
j = "No.Quality",
|
||||
value = allTrees[ID %in% no, Quality])
|
||||
|
||||
allTrees
|
||||
# parse branch lines
|
||||
td[isLeaf==FALSE, c("Feature", "Split", "Yes", "No", "Missing", "Quality", "Cover") := {
|
||||
rx <- paste0("f(\\d+)<(", anynumber_regex, ")\\] yes=(\\d+),no=(\\d+),missing=(\\d+),",
|
||||
"gain=(", anynumber_regex, "),cover=(", anynumber_regex, ")")
|
||||
# skip some indices with spurious capture groups from anynumber_regex
|
||||
xtr <- str_match(t, rx)[, c(2,3,5,6,7,8,10)]
|
||||
xtr[, 3:5] <- addTreeId(xtr[, 3:5], Tree)
|
||||
lapply(1:ncol(xtr), function(i) xtr[,i])
|
||||
}]
|
||||
# assign feature_names when available
|
||||
td[isLeaf==FALSE & !is.null(feature_names),
|
||||
Feature := feature_names[as.numeric(Feature) + 1] ]
|
||||
|
||||
# parse leaf lines
|
||||
td[isLeaf==TRUE, c("Feature", "Quality", "Cover") := {
|
||||
rx <- paste0("leaf=(", anynumber_regex, "),cover=(", anynumber_regex, ")")
|
||||
xtr <- str_match(t, rx)[, c(2,4)]
|
||||
c("Leaf", lapply(1:ncol(xtr), function(i) xtr[,i]))
|
||||
}]
|
||||
|
||||
# convert some columns to numeric
|
||||
numeric_cols <- c("Quality", "Cover")
|
||||
td[, (numeric_cols) := lapply(.SD, as.numeric), .SDcols=numeric_cols]
|
||||
|
||||
td[, t := NULL]
|
||||
td[, isLeaf := NULL]
|
||||
|
||||
td[order(Tree, Node)]
|
||||
}
|
||||
|
||||
# Avoid error messages during CRAN check.
|
||||
# The reason is that these variables are never declared
|
||||
# They are mainly column names inferred by Data.table...
|
||||
globalVariables(c("ID", "Tree", "Yes", ".", ".N", "Feature", "Cover", "Quality", "No", "Gain", "Frequency"))
|
||||
globalVariables(c("Tree", "Node", "ID", "Feature", "t", "isLeaf",".SD", ".SDcols"))
|
||||
@ -2,54 +2,60 @@
|
||||
% Please edit documentation in R/xgb.model.dt.tree.R
|
||||
\name{xgb.model.dt.tree}
|
||||
\alias{xgb.model.dt.tree}
|
||||
\title{Parse boosted tree model text dump}
|
||||
\title{Parse a boosted tree model text dump}
|
||||
\usage{
|
||||
xgb.model.dt.tree(feature_names = NULL, model = NULL, text = NULL,
|
||||
n_first_tree = NULL)
|
||||
}
|
||||
\arguments{
|
||||
\item{feature_names}{names of each feature as a character vector. Can be extracted from a sparse matrix (see example). If the model already contains feature names, this argument should be \code{NULL} (default value).}
|
||||
\item{feature_names}{character vector of feature names. If the model already
|
||||
contains feature names, this argument should be \code{NULL} (default value)}
|
||||
|
||||
\item{model}{object created by the \code{xgb.train} function.}
|
||||
\item{model}{object of class \code{xgb.Booster}}
|
||||
|
||||
\item{text}{\code{character} vector generated by the \code{xgb.dump} function. Model dump must include the gain per feature and per tree (parameter \code{with.stats = TRUE} in function \code{xgb.dump}).}
|
||||
\item{text}{\code{character} vector previously generated by the \code{xgb.dump}
|
||||
function (where parameter \code{with.stats = TRUE} should have been set).}
|
||||
|
||||
\item{n_first_tree}{limit the plot to the \code{n} first trees. If set to \code{NULL}, all trees of the model are plotted. Performance can be low depending of the size of the model.}
|
||||
\item{n_first_tree}{limit the parsing to the \code{n} first trees.
|
||||
If set to \code{NULL}, all trees of the model are parsed.}
|
||||
}
|
||||
\value{
|
||||
A \code{data.table} of the features used in the model with their gain, cover and few other information.
|
||||
}
|
||||
\description{
|
||||
Parse a boosted tree model text dump and return a \code{data.table}.
|
||||
}
|
||||
\details{
|
||||
General function to convert a text dump of tree model to a \code{data.table}.
|
||||
|
||||
The purpose is to help user to explore the model and get a better understanding of it.
|
||||
A \code{data.table} with detailed information about model trees' nodes.
|
||||
|
||||
The columns of the \code{data.table} are:
|
||||
|
||||
\itemize{
|
||||
\item \code{ID}: unique identifier of a node ;
|
||||
\item \code{Feature}: feature used in the tree to operate a split. When Leaf is indicated, it is the end of a branch ;
|
||||
\item \code{Split}: value of the chosen feature where is operated the split ;
|
||||
\item \code{Yes}: ID of the feature for the next node in the branch when the split condition is met ;
|
||||
\item \code{No}: ID of the feature for the next node in the branch when the split condition is not met ;
|
||||
\item \code{Missing}: ID of the feature for the next node in the branch for observation where the feature used for the split are not provided ;
|
||||
\item \code{Quality}: it's the gain related to the split in this specific node ;
|
||||
\item \code{Cover}: metric to measure the number of observation affected by the split ;
|
||||
\item \code{Tree}: ID of the tree. It is included in the main ID ;
|
||||
\item \code{Yes.Feature}, \code{No.Feature}, \code{Yes.Cover}, \code{No.Cover}, \code{Yes.Quality} and \code{No.Quality}: data related to the pointer in \code{Yes} or \code{No} column ;
|
||||
\item \code{Tree}: ID of a tree in a model
|
||||
\item \code{Node}: ID of a node in a tree
|
||||
\item \code{ID}: unique identifier of a node in a model
|
||||
\item \code{Feature}: for a branch node, it's a feature id or name (when available);
|
||||
for a leaf note, it simply labels it as \code{'Leaf'}
|
||||
\item \code{Split}: location of the split for a branch node (split condition is always "less than")
|
||||
\item \code{Yes}: ID of the next node when the split condition is met
|
||||
\item \code{No}: ID of the next node when the split condition is not met
|
||||
\item \code{Missing}: ID of the next node when branch value is missing
|
||||
\item \code{Quality}: either the split gain (change in loss) or the leaf value
|
||||
\item \code{Cover}: metric related to the number of observation either seen by a split
|
||||
or collected by a leaf during training.
|
||||
}
|
||||
}
|
||||
\description{
|
||||
Parse a boosted tree model text dump into a \code{data.table} structure.
|
||||
}
|
||||
\examples{
|
||||
# Basic use:
|
||||
|
||||
data(agaricus.train, package='xgboost')
|
||||
|
||||
bst <- xgboost(data = agaricus.train$data, label = agaricus.train$label, max.depth = 2,
|
||||
eta = 1, nthread = 2, nround = 2,objective = "binary:logistic")
|
||||
|
||||
# agaricus.train$data@Dimnames[[2]] represents the column names of the sparse matrix.
|
||||
xgb.model.dt.tree(feature_names = agaricus.train$data@Dimnames[[2]], model = bst)
|
||||
(dt <- xgb.model.dt.tree(colnames(agaricus.train$data), bst))
|
||||
|
||||
|
||||
# How to match feature names of splits that are following a current 'Yes' branch:
|
||||
|
||||
merge(dt, dt[, .(ID, Y.Feature=Feature)], by.x='Yes', by.y='ID', all.x=TRUE)[order(Tree,Node)]
|
||||
|
||||
}
|
||||
|
||||
|
||||
@ -20,7 +20,7 @@ bst.Tree <- xgboost(data = sparse_matrix, label = output_vector, max.depth = 9,
|
||||
bst.GLM <- xgboost(data = sparse_matrix, label = output_vector,
|
||||
eta = 1, nthread = 2, nround = 10, objective = "binary:logistic", booster = "gblinear")
|
||||
|
||||
feature.names <- agaricus.train$data@Dimnames[[2]]
|
||||
feature.names <- colnames(agaricus.train$data)
|
||||
|
||||
test_that("xgb.dump works", {
|
||||
capture.output(print(xgb.dump(bst.Tree)))
|
||||
@ -57,11 +57,10 @@ test_that("xgb-attribute functionality", {
|
||||
})
|
||||
|
||||
test_that("xgb.model.dt.tree works with and without feature names", {
|
||||
names.dt.trees <- c("ID", "Feature", "Split", "Yes", "No", "Missing", "Quality", "Cover",
|
||||
"Tree", "Yes.Feature", "Yes.Cover", "Yes.Quality", "No.Feature", "No.Cover", "No.Quality")
|
||||
names.dt.trees <- c("Tree", "Node", "ID", "Feature", "Split", "Yes", "No", "Missing", "Quality", "Cover")
|
||||
dt.tree <- xgb.model.dt.tree(feature_names = feature.names, model = bst.Tree)
|
||||
expect_equal(names.dt.trees, names(dt.tree))
|
||||
expect_equal(dim(dt.tree), c(162, 15))
|
||||
expect_equal(dim(dt.tree), c(162, 10))
|
||||
xgb.model.dt.tree(model = bst.Tree)
|
||||
})
|
||||
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user