/
lgb.cv.R
568 lines (470 loc) · 17.3 KB
/
lgb.cv.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
#' @importFrom R6 R6Class
CVBooster <- R6::R6Class(
classname = "lgb.CVBooster",
cloneable = FALSE,
public = list(
best_iter = -1L,
best_score = NA,
record_evals = list(),
boosters = list(),
initialize = function(x) {
self$boosters <- x
},
reset_parameter = function(new_params) {
for (x in boosters) { x$reset_parameter(new_params) }
self
}
)
)
#' @title Main CV logic for LightGBM
#' @description Cross validation logic used by LightGBM
#' @name lgb.cv
#' @inheritParams lgb_shared_params
#' @param nfold the original dataset is randomly partitioned into \code{nfold} equal size subsamples.
#' @param label vector of response values. Should be provided only when data is an R-matrix.
#' @param weight vector of response values. If not NULL, will set to dataset
#' @param obj objective function, can be character or custom objective function. Examples include
#' \code{regression}, \code{regression_l1}, \code{huber},
#' \code{binary}, \code{lambdarank}, \code{multiclass}, \code{multiclass}
#' @param eval evaluation function, can be (list of) character or custom eval function
#' @param record Boolean, TRUE will record iteration message to \code{booster$record_evals}
#' @param showsd \code{boolean}, whether to show standard deviation of cross validation
#' @param stratified a \code{boolean} indicating whether sampling of folds should be stratified
#' by the values of outcome labels.
#' @param folds \code{list} provides a possibility to use a list of pre-defined CV folds
#' (each element must be a vector of test fold's indices). When folds are supplied,
#' the \code{nfold} and \code{stratified} parameters are ignored.
#' @param colnames feature names, if not null, will use this to overwrite the names in dataset
#' @param categorical_feature list of str or int
#' type int represents index,
#' type str represents feature names
#' @param callbacks List of callback functions that are applied at each iteration.
#' @param reset_data Boolean, setting it to TRUE (not the default value) will transform the booster model
#' into a predictor model which frees up memory and the original datasets
#' @param ... other parameters, see Parameters.rst for more information. A few key parameters:
#' \itemize{
#' \item{boosting}{Boosting type. \code{"gbdt"} or \code{"dart"}}
#' \item{num_leaves}{number of leaves in one tree. defaults to 127}
#' \item{max_depth}{Limit the max depth for tree model. This is used to deal with
#' overfit when #data is small. Tree still grow by leaf-wise.}
#' \item{num_threads}{Number of threads for LightGBM. For the best speed, set this to
#' the number of real CPU cores, not the number of threads (most
#' CPU using hyper-threading to generate 2 threads per CPU core).}
#' }
#'
#' @return a trained model \code{lgb.CVBooster}.
#'
#' @examples
#' library(lightgbm)
#' data(agaricus.train, package = "lightgbm")
#' train <- agaricus.train
#' dtrain <- lgb.Dataset(train$data, label = train$label)
#' params <- list(objective = "regression", metric = "l2")
#' model <- lgb.cv(
#' params = params
#' , data = dtrain
#' , nrounds = 10L
#' , nfold = 3L
#' , min_data = 1L
#' , learning_rate = 1.0
#' , early_stopping_rounds = 5L
#' )
#' @export
lgb.cv <- function(params = list()
, data
, nrounds = 10L
, nfold = 3L
, label = NULL
, weight = NULL
, obj = NULL
, eval = NULL
, verbose = 1L
, record = TRUE
, eval_freq = 1L
, showsd = TRUE
, stratified = TRUE
, folds = NULL
, init_model = NULL
, colnames = NULL
, categorical_feature = NULL
, early_stopping_rounds = NULL
, callbacks = list()
, reset_data = FALSE
, ...
) {
# Setup temporary variables
addiction_params <- list(...)
params <- append(params, addiction_params)
params$verbose <- verbose
params <- lgb.check.obj(params, obj)
params <- lgb.check.eval(params, eval)
fobj <- NULL
feval <- NULL
if (nrounds <= 0L) {
stop("nrounds should be greater than zero")
}
# Check for objective (function or not)
if (is.function(params$objective)) {
fobj <- params$objective
params$objective <- "NONE"
}
# Check for loss (function or not)
if (is.function(eval)) {
feval <- eval
}
# Check for parameters
lgb.check.params(params)
# Init predictor to empty
predictor <- NULL
# Check for boosting from a trained model
if (is.character(init_model)) {
predictor <- Predictor$new(init_model)
} else if (lgb.is.Booster(init_model)) {
predictor <- init_model$to_predictor()
}
# Set the iteration to start from / end to (and check for boosting from a trained model, again)
begin_iteration <- 1L
if (!is.null(predictor)) {
begin_iteration <- predictor$current_iter() + 1L
}
# Check for number of rounds passed as parameter - in case there are multiple ones, take only the first one
n_trees <- .PARAMETER_ALIASES()[["num_iterations"]]
if (any(names(params) %in% n_trees)) {
end_iteration <- begin_iteration + params[[which(names(params) %in% n_trees)[1L]]] - 1L
} else {
end_iteration <- begin_iteration + nrounds - 1L
}
# Check for training dataset type correctness
if (!lgb.is.Dataset(data)) {
if (is.null(label)) {
stop("Labels must be provided for lgb.cv")
}
data <- lgb.Dataset(data, label = label)
}
# Check for weights
if (!is.null(weight)) {
data$setinfo("weight", weight)
}
# Update parameters with parsed parameters
data$update_params(params)
# Create the predictor set
data$.__enclos_env__$private$set_predictor(predictor)
# Write column names
if (!is.null(colnames)) {
data$set_colnames(colnames)
}
# Write categorical features
if (!is.null(categorical_feature)) {
data$set_categorical_feature(categorical_feature)
}
# Construct datasets, if needed
data$construct()
# Check for folds
if (!is.null(folds)) {
# Check for list of folds or for single value
if (!is.list(folds) || length(folds) < 2L) {
stop(sQuote("folds"), " must be a list with 2 or more elements that are vectors of indices for each CV-fold")
}
# Set number of folds
nfold <- length(folds)
} else {
# Check fold value
if (nfold <= 1L) {
stop(sQuote("nfold"), " must be > 1")
}
# Create folds
folds <- generate.cv.folds(
nfold
, nrow(data)
, stratified
, getinfo(data, "label")
, getinfo(data, "group")
, params
)
}
# Add printing log callback
if (verbose > 0L && eval_freq > 0L) {
callbacks <- add.cb(callbacks, cb.print.evaluation(eval_freq))
}
# Add evaluation log callback
if (record) {
callbacks <- add.cb(callbacks, cb.record.evaluation())
}
# If early stopping was passed as a parameter in params(), prefer that to keyword argument
# early_stopping_rounds by overwriting the value in 'early_stopping_rounds'
early_stop <- .PARAMETER_ALIASES()[["early_stopping_round"]]
early_stop_param_indx <- names(params) %in% early_stop
if (any(early_stop_param_indx)) {
first_early_stop_param <- which(early_stop_param_indx)[[1L]]
first_early_stop_param_name <- names(params)[[first_early_stop_param]]
early_stopping_rounds <- params[[first_early_stop_param_name]]
}
# Did user pass parameters that indicate they want to use early stopping?
using_early_stopping_via_args <- !is.null(early_stopping_rounds)
boosting_param_names <- .PARAMETER_ALIASES()[["boosting"]]
using_dart <- any(
sapply(
X = boosting_param_names
, FUN = function(param) {
identical(params[[param]], "dart")
}
)
)
# Cannot use early stopping with 'dart' boosting
if (using_dart) {
warning("Early stopping is not available in 'dart' mode.")
using_early_stopping_via_args <- FALSE
# Remove the cb.early.stop() function if it was passed in to callbacks
callbacks <- Filter(
f = function(cb_func) {
!identical(attr(cb_func, "name"), "cb.early.stop")
}
, x = callbacks
)
}
# If user supplied early_stopping_rounds, add the early stopping callback
if (using_early_stopping_via_args) {
callbacks <- add.cb(
callbacks
, cb.early.stop(
stopping_rounds = early_stopping_rounds
, verbose = verbose
)
)
}
# Categorize callbacks
cb <- categorize.callbacks(callbacks)
# Construct booster using a list apply, check if requires group or not
if (!is.list(folds[[1L]])) {
bst_folds <- lapply(seq_along(folds), function(k) {
dtest <- slice(data, folds[[k]])
dtrain <- slice(data, seq_len(nrow(data))[-folds[[k]]])
setinfo(dtrain, "weight", getinfo(data, "weight")[-folds[[k]]])
setinfo(dtrain, "init_score", getinfo(data, "init_score")[-folds[[k]]])
setinfo(dtest, "weight", getinfo(data, "weight")[folds[[k]]])
setinfo(dtest, "init_score", getinfo(data, "init_score")[folds[[k]]])
booster <- Booster$new(params, dtrain)
booster$add_valid(dtest, "valid")
list(booster = booster)
})
} else {
bst_folds <- lapply(seq_along(folds), function(k) {
dtest <- slice(data, folds[[k]]$fold)
dtrain <- slice(data, (seq_len(nrow(data)))[-folds[[k]]$fold])
setinfo(dtrain, "weight", getinfo(data, "weight")[-folds[[k]]$fold])
setinfo(dtrain, "init_score", getinfo(data, "init_score")[-folds[[k]]$fold])
setinfo(dtrain, "group", getinfo(data, "group")[-folds[[k]]$group])
setinfo(dtest, "weight", getinfo(data, "weight")[folds[[k]]$fold])
setinfo(dtest, "init_score", getinfo(data, "init_score")[folds[[k]]$fold])
setinfo(dtest, "group", getinfo(data, "group")[folds[[k]]$group])
booster <- Booster$new(params, dtrain)
booster$add_valid(dtest, "valid")
list(booster = booster)
})
}
# Create new booster
cv_booster <- CVBooster$new(bst_folds)
# Callback env
env <- CB_ENV$new()
env$model <- cv_booster
env$begin_iteration <- begin_iteration
env$end_iteration <- end_iteration
# Start training model using number of iterations to start and end with
for (i in seq.int(from = begin_iteration, to = end_iteration)) {
# Overwrite iteration in environment
env$iteration <- i
env$eval_list <- list()
# Loop through "pre_iter" element
for (f in cb$pre_iter) {
f(env)
}
# Update one boosting iteration
msg <- lapply(cv_booster$boosters, function(fd) {
fd$booster$update(fobj = fobj)
fd$booster$eval_valid(feval = feval)
})
# Prepare collection of evaluation results
merged_msg <- lgb.merge.cv.result(msg)
# Write evaluation result in environment
env$eval_list <- merged_msg$eval_list
# Check for standard deviation requirement
if (showsd) {
env$eval_err_list <- merged_msg$eval_err_list
}
# Loop through env
for (f in cb$post_iter) {
f(env)
}
# Check for early stopping and break if needed
if (env$met_early_stop) break
}
if (record && is.na(env$best_score)) {
if (env$eval_list[[1L]]$higher_better[1L] == TRUE) {
cv_booster$best_iter <- unname(which.max(unlist(cv_booster$record_evals[[2L]][[1L]][[1L]])))
cv_booster$best_score <- cv_booster$record_evals[[2L]][[1L]][[1L]][[cv_booster$best_iter]]
} else {
cv_booster$best_iter <- unname(which.min(unlist(cv_booster$record_evals[[2L]][[1L]][[1L]])))
cv_booster$best_score <- cv_booster$record_evals[[2L]][[1L]][[1L]][[cv_booster$best_iter]]
}
}
if (reset_data) {
lapply(cv_booster$boosters, function(fd) {
# Store temporarily model data elsewhere
booster_old <- list(
best_iter = fd$booster$best_iter
, best_score = fd$booster$best_score
, record_evals = fd$booster$record_evals
)
# Reload model
fd$booster <- lgb.load(model_str = fd$booster$save_model_to_string())
fd$booster$best_iter <- booster_old$best_iter
fd$booster$best_score <- booster_old$best_score
fd$booster$record_evals <- booster_old$record_evals
})
}
# Return booster
return(cv_booster)
}
# Generates random (stratified if needed) CV folds
generate.cv.folds <- function(nfold, nrows, stratified, label, group, params) {
# Check for group existence
if (is.null(group)) {
# Shuffle
rnd_idx <- sample.int(nrows)
# Request stratified folds
if (isTRUE(stratified) && params$objective %in% c("binary", "multiclass") && length(label) == length(rnd_idx)) {
y <- label[rnd_idx]
y <- factor(y)
folds <- lgb.stratified.folds(y, nfold)
} else {
# Make simple non-stratified folds
folds <- list()
# Loop through each fold
for (i in seq_len(nfold)) {
kstep <- length(rnd_idx) %/% (nfold - i + 1L)
folds[[i]] <- rnd_idx[seq_len(kstep)]
rnd_idx <- rnd_idx[-seq_len(kstep)]
}
}
} else {
# When doing group, stratified is not possible (only random selection)
if (nfold > length(group)) {
stop("\n\tYou requested too many folds for the number of available groups.\n")
}
# Degroup the groups
ungrouped <- inverse.rle(list(lengths = group, values = seq_along(group)))
# Can't stratify, shuffle
rnd_idx <- sample.int(length(group))
# Make simple non-stratified folds
folds <- list()
# Loop through each fold
for (i in seq_len(nfold)) {
kstep <- length(rnd_idx) %/% (nfold - i + 1L)
folds[[i]] <- list(
fold = which(ungrouped %in% rnd_idx[seq_len(kstep)])
, group = rnd_idx[seq_len(kstep)]
)
rnd_idx <- rnd_idx[-seq_len(kstep)]
}
}
# Return folds
return(folds)
}
# Creates CV folds stratified by the values of y.
# It was borrowed from caret::lgb.stratified.folds and simplified
# by always returning an unnamed list of fold indices.
#' @importFrom stats quantile
lgb.stratified.folds <- function(y, k = 10L) {
## Group the numeric data based on their magnitudes
## and sample within those groups.
## When the number of samples is low, we may have
## issues further slicing the numeric data into
## groups. The number of groups will depend on the
## ratio of the number of folds to the sample size.
## At most, we will use quantiles. If the sample
## is too small, we just do regular unstratified CV
if (is.numeric(y)) {
cuts <- length(y) %/% k
if (cuts < 2L) {
cuts <- 2L
}
if (cuts > 5L) {
cuts <- 5L
}
y <- cut(
y
, unique(stats::quantile(y, probs = seq.int(0.0, 1.0, length.out = cuts)))
, include.lowest = TRUE
)
}
if (k < length(y)) {
## Reset levels so that the possible levels and
## the levels in the vector are the same
y <- factor(as.character(y))
numInClass <- table(y)
foldVector <- vector(mode = "integer", length(y))
## For each class, balance the fold allocation as far
## as possible, then resample the remainder.
## The final assignment of folds is also randomized.
for (i in seq_along(numInClass)) {
## Create a vector of integers from 1:k as many times as possible without
## going over the number of samples in the class. Note that if the number
## of samples in a class is less than k, nothing is producd here.
seqVector <- rep(seq_len(k), numInClass[i] %/% k)
## Add enough random integers to get length(seqVector) == numInClass[i]
if (numInClass[i] %% k > 0L) {
seqVector <- c(seqVector, sample.int(k, numInClass[i] %% k))
}
## Shuffle the integers for fold assignment and assign to this classes's data
foldVector[y == dimnames(numInClass)$y[i]] <- sample(seqVector)
}
} else {
foldVector <- seq(along = y)
}
# Return data
out <- split(seq(along = y), foldVector)
names(out) <- NULL
out
}
lgb.merge.cv.result <- function(msg, showsd = TRUE) {
# Get CV message length
if (length(msg) == 0L) {
stop("lgb.cv: size of cv result error")
}
# Get evaluation message length
eval_len <- length(msg[[1L]])
# Is evaluation message empty?
if (eval_len == 0L) {
stop("lgb.cv: should provide at least one metric for CV")
}
# Get evaluation results using a list apply
eval_result <- lapply(seq_len(eval_len), function(j) {
as.numeric(lapply(seq_along(msg), function(i) {
msg[[i]][[j]]$value }))
})
# Get evaluation
ret_eval <- msg[[1L]]
# Go through evaluation length items
for (j in seq_len(eval_len)) {
ret_eval[[j]]$value <- mean(eval_result[[j]])
}
# Preinit evaluation error
ret_eval_err <- NULL
# Check for standard deviation
if (showsd) {
# Parse standard deviation
for (j in seq_len(eval_len)) {
ret_eval_err <- c(
ret_eval_err
, sqrt(mean(eval_result[[j]] ^ 2L) - mean(eval_result[[j]]) ^ 2L)
)
}
# Convert to list
ret_eval_err <- as.list(ret_eval_err)
}
# Return errors
list(
eval_list = ret_eval
, eval_err_list = ret_eval_err
)
}