config.h 59.3 KB
Newer Older
1
2
3
4
5
/*!
 * Copyright (c) 2016 Microsoft Corporation. All rights reserved.
 * Licensed under the MIT License. See LICENSE file in the project root for license information.
 *
 * \note
6
7
8
9
 * - desc and descl2 fields must be written in reStructuredText format;
 * - nested sections can be placed only at the bottom of parent's section;
 * - [doc-only] tag indicates that only documentation for this param should be generated and all other actions are performed manually;
 * - [no-save] tag indicates that this param should not be saved into a model text representation.
10
 */
Guolin Ke's avatar
Guolin Ke committed
11
12
13
#ifndef LIGHTGBM_CONFIG_H_
#define LIGHTGBM_CONFIG_H_

14
15
16
17
18
#include <LightGBM/export.h>
#include <LightGBM/meta.h>
#include <LightGBM/utils/common.h>
#include <LightGBM/utils/log.h>

Guolin Ke's avatar
Guolin Ke committed
19
20
#include <string>
#include <algorithm>
Guolin Ke's avatar
Guolin Ke committed
21
#include <memory>
22
23
24
#include <unordered_map>
#include <unordered_set>
#include <vector>
Guolin Ke's avatar
Guolin Ke committed
25
26
27

namespace LightGBM {

Guolin Ke's avatar
Guolin Ke committed
28
29
30
31
/*! \brief Types of tasks */
enum TaskType {
  kTrain, kPredict, kConvertModel, KRefitTree
};
32
const int kDefaultNumLeaves = 31;
Guolin Ke's avatar
Guolin Ke committed
33

Guolin Ke's avatar
Guolin Ke committed
34
struct Config {
Nikita Titov's avatar
Nikita Titov committed
35
 public:
Guolin Ke's avatar
Guolin Ke committed
36
  std::string ToString() const;
Guolin Ke's avatar
Guolin Ke committed
37
38
39
40
  /*!
  * \brief Get string value by specific name of key
  * \param params Store the key and value for params
  * \param name Name of key
Hui Xue's avatar
Hui Xue committed
41
  * \param out Value will assign to out if key exists
Guolin Ke's avatar
Guolin Ke committed
42
43
  * \return True if key exists
  */
Guolin Ke's avatar
Guolin Ke committed
44
  inline static bool GetString(
Guolin Ke's avatar
Guolin Ke committed
45
46
47
48
49
50
51
    const std::unordered_map<std::string, std::string>& params,
    const std::string& name, std::string* out);

  /*!
  * \brief Get int value by specific name of key
  * \param params Store the key and value for params
  * \param name Name of key
Hui Xue's avatar
Hui Xue committed
52
  * \param out Value will assign to out if key exists
Guolin Ke's avatar
Guolin Ke committed
53
54
  * \return True if key exists
  */
Guolin Ke's avatar
Guolin Ke committed
55
  inline static bool GetInt(
Guolin Ke's avatar
Guolin Ke committed
56
57
58
59
    const std::unordered_map<std::string, std::string>& params,
    const std::string& name, int* out);

  /*!
60
  * \brief Get double value by specific name of key
Guolin Ke's avatar
Guolin Ke committed
61
62
  * \param params Store the key and value for params
  * \param name Name of key
Hui Xue's avatar
Hui Xue committed
63
  * \param out Value will assign to out if key exists
Guolin Ke's avatar
Guolin Ke committed
64
65
  * \return True if key exists
  */
Guolin Ke's avatar
Guolin Ke committed
66
  inline static bool GetDouble(
Guolin Ke's avatar
Guolin Ke committed
67
    const std::unordered_map<std::string, std::string>& params,
68
    const std::string& name, double* out);
Guolin Ke's avatar
Guolin Ke committed
69
70
71
72
73

  /*!
  * \brief Get bool value by specific name of key
  * \param params Store the key and value for params
  * \param name Name of key
Hui Xue's avatar
Hui Xue committed
74
  * \param out Value will assign to out if key exists
Guolin Ke's avatar
Guolin Ke committed
75
76
  * \return True if key exists
  */
Guolin Ke's avatar
Guolin Ke committed
77
  inline static bool GetBool(
Guolin Ke's avatar
Guolin Ke committed
78
79
    const std::unordered_map<std::string, std::string>& params,
    const std::string& name, bool* out);
80

Guolin Ke's avatar
Guolin Ke committed
81
  static void KV2Map(std::unordered_map<std::string, std::string>* params, const char* kv);
82
  static std::unordered_map<std::string, std::string> Str2Map(const char* parameters);
Guolin Ke's avatar
Guolin Ke committed
83

Guolin Ke's avatar
Guolin Ke committed
84
  #pragma region Parameters
85

Guolin Ke's avatar
Guolin Ke committed
86
87
  #pragma region Core Parameters

88
  // [no-save]
Guolin Ke's avatar
Guolin Ke committed
89
  // [doc-only]
90
91
  // alias = config_file
  // desc = path of config file
92
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
93
94
  std::string config = "";

95
  // [no-save]
Guolin Ke's avatar
Guolin Ke committed
96
  // [doc-only]
97
98
99
100
101
102
  // type = enum
  // default = train
  // options = train, predict, convert_model, refit
  // alias = task_type
  // desc = ``train``, for training, aliases: ``training``
  // desc = ``predict``, for prediction, aliases: ``prediction``, ``test``
Nikita Titov's avatar
Nikita Titov committed
103
  // desc = ``convert_model``, for converting model file into if-else format, see more information in `Convert Parameters <#convert-parameters>`__
104
  // desc = ``refit``, for refitting existing models with new data, aliases: ``refit_tree``
Guolin Ke's avatar
Guolin Ke committed
105
  // desc = **Note**: can be used only in CLI version; for language-specific packages you can use the correspondent functions
Guolin Ke's avatar
Guolin Ke committed
106
107
108
  TaskType task = TaskType::kTrain;

  // [doc-only]
109
  // type = enum
110
  // options = regression, regression_l1, huber, fair, poisson, quantile, mape, gamma, tweedie, binary, multiclass, multiclassova, cross_entropy, cross_entropy_lambda, lambdarank, rank_xendcg
111
112
  // alias = objective_type, app, application
  // desc = regression application
Guolin Ke's avatar
Guolin Ke committed
113
114
  // descl2 = ``regression``, L2 loss, aliases: ``regression_l2``, ``l2``, ``mean_squared_error``, ``mse``, ``l2_root``, ``root_mean_squared_error``, ``rmse``
  // descl2 = ``regression_l1``, L1 loss, aliases: ``l1``, ``mean_absolute_error``, ``mae``
115
116
117
118
119
  // descl2 = ``huber``, `Huber loss <https://en.wikipedia.org/wiki/Huber_loss>`__
  // descl2 = ``fair``, `Fair loss <https://www.kaggle.com/c/allstate-claims-severity/discussion/24520>`__
  // descl2 = ``poisson``, `Poisson regression <https://en.wikipedia.org/wiki/Poisson_regression>`__
  // descl2 = ``quantile``, `Quantile regression <https://en.wikipedia.org/wiki/Quantile_regression>`__
  // descl2 = ``mape``, `MAPE loss <https://en.wikipedia.org/wiki/Mean_absolute_percentage_error>`__, aliases: ``mean_absolute_percentage_error``
120
  // descl2 = ``gamma``, Gamma regression with log-link. It might be useful, e.g., for modeling insurance claims severity, or for any target that might be `gamma-distributed <https://en.wikipedia.org/wiki/Gamma_distribution#Occurrence_and_applications>`__
121
  // descl2 = ``tweedie``, Tweedie regression with log-link. It might be useful, e.g., for modeling total loss in insurance, or for any target that might be `tweedie-distributed <https://en.wikipedia.org/wiki/Tweedie_distribution#Occurrence_and_applications>`__
122
123
124
  // desc = binary classification application
  // descl2 = ``binary``, binary `log loss <https://en.wikipedia.org/wiki/Cross_entropy>`__ classification (or logistic regression)
  // descl2 = requires labels in {0, 1}; see ``cross-entropy`` application for general probability labels in [0, 1]
125
126
127
128
129
  // desc = multi-class classification application
  // descl2 = ``multiclass``, `softmax <https://en.wikipedia.org/wiki/Softmax_function>`__ objective function, aliases: ``softmax``
  // descl2 = ``multiclassova``, `One-vs-All <https://en.wikipedia.org/wiki/Multiclass_classification#One-vs.-rest>`__ binary objective function, aliases: ``multiclass_ova``, ``ova``, ``ovr``
  // descl2 = ``num_class`` should be set as well
  // desc = cross-entropy application
Guolin Ke's avatar
Guolin Ke committed
130
131
  // descl2 = ``cross_entropy``, objective function for cross-entropy (with optional linear weights), aliases: ``xentropy``
  // descl2 = ``cross_entropy_lambda``, alternative parameterization of cross-entropy, aliases: ``xentlambda``
132
  // descl2 = label is anything in interval [0, 1]
133
  // desc = ranking application
134
  // descl2 = ``lambdarank``, `lambdarank <https://papers.nips.cc/paper/2971-learning-to-rank-with-nonsmooth-cost-functions.pdf>`__ objective. `label_gain <#label_gain>`__ can be used to set the gain (weight) of ``int`` label and all values in ``label`` must be smaller than number of elements in ``label_gain``
135
136
  // descl2 = ``rank_xendcg``, `XE_NDCG_MART <https://arxiv.org/abs/1911.09798>`__ ranking objective function, aliases: ``xendcg``, ``xe_ndcg``, ``xe_ndcg_mart``, ``xendcg_mart``
  // descl2 = ``rank_xendcg`` is faster than and achieves the similar performance as ``lambdarank``
137
  // descl2 = label should be ``int`` type, and larger number represents the higher relevance (e.g. 0:bad, 1:fair, 2:good, 3:perfect)
Guolin Ke's avatar
Guolin Ke committed
138
139
140
  std::string objective = "regression";

  // [doc-only]
141
142
  // type = enum
  // alias = boosting_type, boost
143
  // options = gbdt, rf, dart, goss
144
145
  // desc = ``gbdt``, traditional Gradient Boosting Decision Tree, aliases: ``gbrt``
  // desc = ``rf``, Random Forest, aliases: ``random_forest``
146
147
  // desc = ``dart``, `Dropouts meet Multiple Additive Regression Trees <https://arxiv.org/abs/1505.01866>`__
  // desc = ``goss``, Gradient-based One-Side Sampling
Nikita Titov's avatar
Nikita Titov committed
148
  // descl2 = **Note**: internally, LightGBM uses ``gbdt`` mode for the first ``1 / learning_rate`` iterations
Guolin Ke's avatar
Guolin Ke committed
149
150
  std::string boosting = "gbdt";

151
  // alias = train, train_data, train_data_file, data_filename
152
  // desc = path of training data, LightGBM will train from this data
153
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
154
155
  std::string data = "";

156
  // alias = test, valid_data, valid_data_file, test_data, test_data_file, valid_filenames
157
  // default = ""
158
  // desc = path(s) of validation/test data, LightGBM will output metrics for these data
159
  // desc = support multiple validation data, separated by ``,``
160
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
161
162
  std::vector<std::string> valid;

163
  // alias = num_iteration, n_iter, num_tree, num_trees, num_round, num_rounds, num_boost_round, n_estimators
164
165
166
  // check = >=0
  // desc = number of boosting iterations
  // desc = **Note**: internally, LightGBM constructs ``num_class * num_iterations`` trees for multi-class classification problems
Guolin Ke's avatar
Guolin Ke committed
167
  int num_iterations = 100;
Guolin Ke's avatar
Guolin Ke committed
168

169
  // alias = shrinkage_rate, eta
170
  // check = >0.0
171
172
  // desc = shrinkage rate
  // desc = in ``dart``, it also affects on normalization weights of dropped trees
Guolin Ke's avatar
Guolin Ke committed
173
174
  double learning_rate = 0.1;

175
  // default = 31
176
  // alias = num_leaf, max_leaves, max_leaf
177
  // check = >1
178
  // check = <=131072
179
  // desc = max number of leaves in one tree
Guolin Ke's avatar
Guolin Ke committed
180
181
182
  int num_leaves = kDefaultNumLeaves;

  // [doc-only]
183
184
  // type = enum
  // options = serial, feature, data, voting
185
  // alias = tree, tree_type, tree_learner_type
186
187
188
189
190
  // desc = ``serial``, single machine tree learner
  // desc = ``feature``, feature parallel tree learner, aliases: ``feature_parallel``
  // desc = ``data``, data parallel tree learner, aliases: ``data_parallel``
  // desc = ``voting``, voting parallel tree learner, aliases: ``voting_parallel``
  // desc = refer to `Parallel Learning Guide <./Parallel-Learning-Guide.rst>`__ to get more details
Guolin Ke's avatar
Guolin Ke committed
191
192
  std::string tree_learner = "serial";

193
  // alias = num_thread, nthread, nthreads, n_jobs
Guolin Ke's avatar
Guolin Ke committed
194
  // desc = number of threads for LightGBM
195
196
197
198
199
  // desc = ``0`` means default number of threads in OpenMP
  // desc = for the best speed, set this to the number of **real CPU cores**, not the number of threads (most CPUs use `hyper-threading <https://en.wikipedia.org/wiki/Hyper-threading>`__ to generate 2 threads per CPU core)
  // desc = do not set it too large if your dataset is small (for instance, do not use 64 threads for a dataset with 10,000 rows)
  // desc = be aware a task manager or any similar CPU monitoring tool might report that cores not being fully utilized. **This is normal**
  // desc = for parallel learning, do not use all CPU cores because this will cause poor performance for the network communication
200
  // desc = **Note**: please **don't** change this during training, especially when running multiple jobs simultaneously by external packages, otherwise it may cause undesirable errors
Guolin Ke's avatar
Guolin Ke committed
201
202
203
  int num_threads = 0;

  // [doc-only]
204
205
  // type = enum
  // options = cpu, gpu
206
  // alias = device
207
208
209
210
  // desc = device for the tree learning, you can use GPU to achieve the faster learning
  // desc = **Note**: it is recommended to use the smaller ``max_bin`` (e.g. 63) to get the better speed up
  // desc = **Note**: for the faster speed, GPU uses 32-bit float point to sum up by default, so this may affect the accuracy for some tasks. You can set ``gpu_use_dp=true`` to enable 64-bit float point, but it will slow down the training
  // desc = **Note**: refer to `Installation Guide <./Installation-Guide.rst#build-gpu-version>`__ to build LightGBM with GPU support
Guolin Ke's avatar
Guolin Ke committed
211
212
213
  std::string device_type = "cpu";

  // [doc-only]
214
  // alias = random_seed, random_state
215
216
217
218
  // default = None
  // desc = this seed is used to generate other seeds, e.g. ``data_random_seed``, ``feature_fraction_seed``, etc.
  // desc = by default, this seed is unused in favor of default values of other seeds
  // desc = this seed has lower priority in comparison with other seeds, which means that it will be overridden, if you set other seeds explicitly
Guolin Ke's avatar
Guolin Ke committed
219
220
221
222
223
224
  int seed = 0;

  #pragma endregion

  #pragma region Learning Control Parameters

225
226
227
228
  // desc = used only with ``cpu`` device type
  // desc = set this to ``true`` to force col-wise histogram building
  // desc = enabling this is recommended when:
  // descl2 = the number of columns is large, or the total number of bins is large
Nikita Titov's avatar
Nikita Titov committed
229
  // descl2 = ``num_threads`` is large, e.g. ``> 20``
230
231
232
  // descl2 = you want to reduce memory cost
  // desc = **Note**: when both ``force_col_wise`` and ``force_row_wise`` are ``false``, LightGBM will firstly try them both, and then use the faster one. To remove the overhead of testing set the faster one to ``true`` manually
  // desc = **Note**: this parameter cannot be used at the same time with ``force_row_wise``, choose only one of them
233
234
  bool force_col_wise = false;

235
236
237
238
  // desc = used only with ``cpu`` device type
  // desc = set this to ``true`` to force row-wise histogram building
  // desc = enabling this is recommended when:
  // descl2 = the number of data points is large, and the total number of bins is relatively small
Nikita Titov's avatar
Nikita Titov committed
239
  // descl2 = ``num_threads`` is relatively small, e.g. ``<= 16``
240
241
242
243
  // descl2 = you want to use small ``bagging_fraction`` or ``goss`` boosting to speed up
  // desc = **Note**: setting this to ``true`` will double the memory cost for Dataset object. If you have not enough memory, you can try setting ``force_col_wise=true``
  // desc = **Note**: when both ``force_col_wise`` and ``force_row_wise`` are ``false``, LightGBM will firstly try them both, and then use the faster one. To remove the overhead of testing set the faster one to ``true`` manually
  // desc = **Note**: this parameter cannot be used at the same time with ``force_col_wise``, choose only one of them
244
245
  bool force_row_wise = false;

246
247
248
249
250
  // alias = hist_pool_size
  // desc = max cache size in MB for historical histogram
  // desc = ``< 0`` means no limit
  double histogram_pool_size = -1.0;

251
  // desc = limit the max depth for tree model. This is used to deal with over-fitting when ``#data`` is small. Tree still grows leaf-wise
252
  // desc = ``<= 0`` means no limit
Guolin Ke's avatar
Guolin Ke committed
253
254
255
  int max_depth = -1;

  // alias = min_data_per_leaf, min_data, min_child_samples
256
257
  // check = >=0
  // desc = minimal number of data in one leaf. Can be used to deal with over-fitting
Guolin Ke's avatar
Guolin Ke committed
258
259
  int min_data_in_leaf = 20;

260
261
262
  // alias = min_sum_hessian_per_leaf, min_sum_hessian, min_hessian, min_child_weight
  // check = >=0.0
  // desc = minimal sum hessian in one leaf. Like ``min_data_in_leaf``, it can be used to deal with over-fitting
Guolin Ke's avatar
Guolin Ke committed
263
264
  double min_sum_hessian_in_leaf = 1e-3;

265
266
267
268
269
270
271
  // alias = sub_row, subsample, bagging
  // check = >0.0
  // check = <=1.0
  // desc = like ``feature_fraction``, but this will randomly select part of data without resampling
  // desc = can be used to speed up training
  // desc = can be used to deal with over-fitting
  // desc = **Note**: to enable bagging, ``bagging_freq`` should be set to a non zero value as well
Guolin Ke's avatar
Guolin Ke committed
272
273
  double bagging_fraction = 1.0;

Guolin Ke's avatar
Guolin Ke committed
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
  // alias = pos_sub_row, pos_subsample, pos_bagging
  // check = >0.0
  // check = <=1.0
  // desc = used only in ``binary`` application
  // desc = used for imbalanced binary classification problem, will randomly sample ``#pos_samples * pos_bagging_fraction`` positive samples in bagging
  // desc = should be used together with ``neg_bagging_fraction``
  // desc = set this to ``1.0`` to disable
  // desc = **Note**: to enable this, you need to set ``bagging_freq`` and ``neg_bagging_fraction`` as well
  // desc = **Note**: if both ``pos_bagging_fraction`` and ``neg_bagging_fraction`` are set to ``1.0``,  balanced bagging is disabled
  // desc = **Note**: if balanced bagging is enabled, ``bagging_fraction`` will be ignored
  double pos_bagging_fraction = 1.0;

  // alias = neg_sub_row, neg_subsample, neg_bagging
  // check = >0.0
  // check = <=1.0
  // desc = used only in ``binary`` application
  // desc = used for imbalanced binary classification problem, will randomly sample ``#neg_samples * neg_bagging_fraction`` negative samples in bagging
  // desc = should be used together with ``pos_bagging_fraction``
  // desc = set this to ``1.0`` to disable
  // desc = **Note**: to enable this, you need to set ``bagging_freq`` and ``pos_bagging_fraction`` as well
  // desc = **Note**: if both ``pos_bagging_fraction`` and ``neg_bagging_fraction`` are set to ``1.0``,  balanced bagging is disabled
  // desc = **Note**: if balanced bagging is enabled, ``bagging_fraction`` will be ignored
  double neg_bagging_fraction = 1.0;

298
299
300
301
  // alias = subsample_freq
  // desc = frequency for bagging
  // desc = ``0`` means disable bagging; ``k`` means perform bagging at every ``k`` iteration
  // desc = **Note**: to enable bagging, ``bagging_fraction`` should be set to value smaller than ``1.0`` as well
Guolin Ke's avatar
Guolin Ke committed
302
303
304
305
306
307
308
  int bagging_freq = 0;

  // alias = bagging_fraction_seed
  // desc = random seed for bagging
  int bagging_seed = 3;

  // alias = sub_feature, colsample_bytree
309
310
  // check = >0.0
  // check = <=1.0
311
  // desc = LightGBM will randomly select part of features on each iteration (tree) if ``feature_fraction`` smaller than ``1.0``. For example, if you set it to ``0.8``, LightGBM will select 80% of features before training each tree
312
313
  // desc = can be used to speed up training
  // desc = can be used to deal with over-fitting
Guolin Ke's avatar
Guolin Ke committed
314
315
  double feature_fraction = 1.0;

316
317
318
  // alias = sub_feature_bynode, colsample_bynode
  // check = >0.0
  // check = <=1.0
Nikita Titov's avatar
Nikita Titov committed
319
  // desc = LightGBM will randomly select part of features on each tree node if ``feature_fraction_bynode`` smaller than ``1.0``. For example, if you set it to ``0.8``, LightGBM will select 80% of features at each tree node
320
321
322
323
324
  // desc = can be used to deal with over-fitting
  // desc = **Note**: unlike ``feature_fraction``, this cannot speed up training
  // desc = **Note**: if both ``feature_fraction`` and ``feature_fraction_bynode`` are smaller than ``1.0``, the final fraction of each node is ``feature_fraction * feature_fraction_bynode``
  double feature_fraction_bynode = 1.0;

325
  // desc = random seed for ``feature_fraction``
Guolin Ke's avatar
Guolin Ke committed
326
327
  int feature_fraction_seed = 2;

328
329
330
331
332
333
334
335
  // desc = use extremely randomized trees
  // desc = if set to ``true``, when evaluating node splits LightGBM will check only one randomly-chosen threshold for each feature
  // desc = can be used to deal with over-fitting
  bool extra_trees = false;

  // desc = random seed for selecting thresholds when ``extra_trees`` is true
  int extra_seed = 6;

336
  // alias = early_stopping_rounds, early_stopping, n_iter_no_change
337
338
  // desc = will stop training if one metric of one validation data doesn't improve in last ``early_stopping_round`` rounds
  // desc = ``<= 0`` means disable
Guolin Ke's avatar
Guolin Ke committed
339
340
  int early_stopping_round = 0;

341
342
343
  // desc = set this to ``true``, if you want to use only the first metric for early stopping
  bool first_metric_only = false;

344
345
346
347
  // alias = max_tree_output, max_leaf_output
  // desc = used to limit the max output of tree leaves
  // desc = ``<= 0`` means no constraint
  // desc = the final max output of leaves is ``learning_rate * max_delta_step``
Guolin Ke's avatar
Guolin Ke committed
348
349
  double max_delta_step = 0.0;

350
351
352
  // alias = reg_alpha
  // check = >=0.0
  // desc = L1 regularization
Guolin Ke's avatar
Guolin Ke committed
353
354
  double lambda_l1 = 0.0;

355
  // alias = reg_lambda, lambda
356
  // check = >=0.0
Guolin Ke's avatar
Guolin Ke committed
357
358
359
  // desc = L2 regularization
  double lambda_l2 = 0.0;

360
361
362
  // alias = min_split_gain
  // check = >=0.0
  // desc = the minimal gain to perform split
Guolin Ke's avatar
Guolin Ke committed
363
364
  double min_gain_to_split = 0.0;

365
  // alias = rate_drop
366
367
368
  // check = >=0.0
  // check = <=1.0
  // desc = used only in ``dart``
369
  // desc = dropout rate: a fraction of previous trees to drop during the dropout
Guolin Ke's avatar
Guolin Ke committed
370
371
  double drop_rate = 0.1;

372
  // desc = used only in ``dart``
373
  // desc = max number of dropped trees during one boosting iteration
374
  // desc = ``<=0`` means no limit
Guolin Ke's avatar
Guolin Ke committed
375
376
  int max_drop = 50;

377
378
379
  // check = >=0.0
  // check = <=1.0
  // desc = used only in ``dart``
380
  // desc = probability of skipping the dropout procedure during a boosting iteration
Guolin Ke's avatar
Guolin Ke committed
381
382
  double skip_drop = 0.5;

383
384
  // desc = used only in ``dart``
  // desc = set this to ``true``, if you want to use xgboost dart mode
Guolin Ke's avatar
Guolin Ke committed
385
386
  bool xgboost_dart_mode = false;

387
388
  // desc = used only in ``dart``
  // desc = set this to ``true``, if you want to use uniform drop
Guolin Ke's avatar
Guolin Ke committed
389
390
  bool uniform_drop = false;

391
392
  // desc = used only in ``dart``
  // desc = random seed to choose dropping models
Guolin Ke's avatar
Guolin Ke committed
393
394
  int drop_seed = 4;

395
396
397
398
  // check = >=0.0
  // check = <=1.0
  // desc = used only in ``goss``
  // desc = the retain ratio of large gradient data
Guolin Ke's avatar
Guolin Ke committed
399
400
  double top_rate = 0.2;

401
402
403
404
  // check = >=0.0
  // check = <=1.0
  // desc = used only in ``goss``
  // desc = the retain ratio of small gradient data
Guolin Ke's avatar
Guolin Ke committed
405
406
  double other_rate = 0.1;

407
408
  // check = >0
  // desc = minimal number of data per categorical group
Guolin Ke's avatar
Guolin Ke committed
409
410
  int min_data_per_group = 100;

411
412
413
  // check = >0
  // desc = used for the categorical features
  // desc = limit the max threshold points in categorical features
Guolin Ke's avatar
Guolin Ke committed
414
415
  int max_cat_threshold = 32;

416
417
  // check = >=0.0
  // desc = used for the categorical features
418
  // desc = L2 regularization in categorical split
419
  double cat_l2 = 10.0;
Guolin Ke's avatar
Guolin Ke committed
420

421
422
423
424
  // check = >=0.0
  // desc = used for the categorical features
  // desc = this can reduce the effect of noises in categorical features, especially for categories with few data
  double cat_smooth = 10.0;
425

426
427
  // check = >0
  // desc = when number of categories of one feature smaller than or equal to ``max_cat_to_onehot``, one-vs-other split algorithm will be used
Guolin Ke's avatar
Guolin Ke committed
428
429
430
  int max_cat_to_onehot = 4;

  // alias = topk
431
  // check = >0
432
  // desc = used only in ``voting`` tree learner, refer to `Voting parallel <./Parallel-Learning-Guide.rst#choose-appropriate-parallel-algorithm>`__
433
  // desc = set this to larger value for more accurate result, but it will slow down the training speed
Guolin Ke's avatar
Guolin Ke committed
434
435
436
  int top_k = 20;

  // type = multi-int
437
438
439
440
441
  // alias = mc, monotone_constraint
  // default = None
  // desc = used for constraints of monotonic features
  // desc = ``1`` means increasing, ``-1`` means decreasing, ``0`` means non-constraint
  // desc = you need to specify all features in order. For example, ``mc=-1,0,1`` means decreasing for 1st feature, non-constraint for 2nd feature and increasing for the 3rd feature
Guolin Ke's avatar
Guolin Ke committed
442
  std::vector<int8_t> monotone_constraints;
Guolin Ke's avatar
Guolin Ke committed
443

Nikita Titov's avatar
Nikita Titov committed
444
  // type = enum
445
  // alias = monotone_constraining_method, mc_method
446
  // options = basic, intermediate, advanced
447
448
449
450
  // desc = used only if ``monotone_constraints`` is set
  // desc = monotone constraints method
  // descl2 = ``basic``, the most basic monotone constraints method. It does not slow the library at all, but over-constrains the predictions
  // descl2 = ``intermediate``, a `more advanced method <https://github.com/microsoft/LightGBM/files/3457826/PR-monotone-constraints-report.pdf>`__, which may slow the library very slightly. However, this method is much less constraining than the basic method and should significantly improve the results
451
  // descl2 = ``advanced``, an `even more advanced method <https://github.com/microsoft/LightGBM/files/3457826/PR-monotone-constraints-report.pdf>`__, which may slow the library. However, this method is even less constraining than the intermediate method and should again significantly improve the results
452
453
  std::string monotone_constraints_method = "basic";

454
455
456
457
458
459
460
  // alias = monotone_splits_penalty, ms_penalty, mc_penalty
  // check = >=0.0
  // desc = used only if ``monotone_constraints`` is set
  // desc = `monotone penalty <https://github.com/microsoft/LightGBM/files/3457826/PR-monotone-constraints-report.pdf>`__: a penalization parameter X forbids any monotone splits on the first X (rounded down) level(s) of the tree. The penalty applied to monotone splits on a given depth is a continuous, increasing function the penalization parameter
  // desc = if ``0.0`` (the default), no penalization is applied
  double monotone_penalty = 0.0;

Guolin Ke's avatar
Guolin Ke committed
461
  // type = multi-double
462
  // alias = feature_contrib, fc, fp, feature_penalty
Guolin Ke's avatar
Guolin Ke committed
463
464
465
466
  // default = None
  // desc = used to control feature's split gain, will use ``gain[i] = max(0, feature_contri[i]) * gain[i]`` to replace the split gain of i-th feature
  // desc = you need to specify all features in order
  std::vector<double> feature_contri;
467

468
469
470
471
  // alias = fs, forced_splits_filename, forced_splits_file, forced_splits
  // desc = path to a ``.json`` file that specifies splits to force at the top of every decision tree before best-first learning commences
  // desc = ``.json`` file can be arbitrarily nested, and each split contains ``feature``, ``threshold`` fields, as well as ``left`` and ``right`` fields representing subsplits
  // desc = categorical splits are forced in a one-hot fashion, with ``left`` representing the split containing the feature value and ``right`` representing other values
472
  // desc = **Note**: the forced split logic will be ignored, if the split makes gain worse
473
  // desc = see `this file <https://github.com/microsoft/LightGBM/tree/master/examples/binary_classification/forced_splits.json>`__ as an example
Guolin Ke's avatar
Guolin Ke committed
474
475
  std::string forcedsplits_filename = "";

Guolin Ke's avatar
Guolin Ke committed
476
477
478
479
480
481
  // check = >=0.0
  // check = <=1.0
  // desc = decay rate of ``refit`` task, will use ``leaf_output = refit_decay_rate * old_leaf_output + (1.0 - refit_decay_rate) * new_leaf_output`` to refit trees
  // desc = used only in ``refit`` task in CLI version or as argument in ``refit`` function in language-specific package
  double refit_decay_rate = 0.9;

482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
  // check = >=0.0
  // desc = cost-effective gradient boosting multiplier for all penalties
  double cegb_tradeoff = 1.0;

  // check = >=0.0
  // desc = cost-effective gradient-boosting penalty for splitting a node
  double cegb_penalty_split = 0.0;

  // type = multi-double
  // default = 0,0,...,0
  // desc = cost-effective gradient boosting penalty for using a feature
  // desc = applied per data point
  std::vector<double> cegb_penalty_feature_lazy;

  // type = multi-double
  // default = 0,0,...,0
  // desc = cost-effective gradient boosting penalty for using a feature
  // desc = applied once per forest
500
  std::vector<double> cegb_penalty_feature_coupled;
501

Belinda Trotta's avatar
Belinda Trotta committed
502
503
504
505
506
507
508
509
510
511
  // check = >= 0.0
  // desc = controls smoothing applied to tree nodes
  // desc = helps prevent overfitting on leaves with few samples
  // desc = if set to zero, no smoothing is applied
  // desc = if ``path_smooth > 0`` then ``min_data_in_leaf`` must be at least ``2``
  // desc = larger values give stronger regularisation
  // descl2 = the weight of each node is ``(n / path_smooth) * w + w_p / (n / path_smooth + 1)``, where ``n`` is the number of samples in the node, ``w`` is the optimal node weight to minimise the loss (approximately ``-sum_gradients / sum_hessians``), and ``w_p`` is the weight of the parent node
  // descl2 = note that the parent output ``w_p`` itself has smoothing applied, unless it is the root node, so that the smoothing effect accumulates with the tree depth
  double path_smooth = 0;

512
513
514
515
  // desc = controls which features can appear in the same branch
  // desc = by default interaction constraints are disabled, to enable them you can specify
  // descl2 = for CLI, lists separated by commas, e.g. ``[0,1,2],[2,3]``
  // descl2 = for Python-package, list of lists, e.g. ``[[0, 1, 2], [2, 3]]``
516
  // descl2 = for R-package, list of character or numeric vectors, e.g. ``list(c("var1", "var2", "var3"), c("var3", "var4"))`` or ``list(c(1L, 2L, 3L), c(3L, 4L))``. Numeric vectors should use 1-based indexing, where ``1L`` is the first feature, ``2L`` is the second feature, etc
517
518
519
  // desc = any two features can only appear in the same branch only if there exists a constraint containing both features
  std::string interaction_constraints = "";

520
521
  // alias = verbose
  // desc = controls the level of LightGBM's verbosity
522
  // desc = ``< 0``: Fatal, ``= 0``: Error (Warning), ``= 1``: Info, ``> 1``: Debug
523
524
  int verbosity = 1;

525
  // [no-save]
526
527
528
529
530
531
532
  // alias = model_input, model_in
  // desc = filename of input model
  // desc = for ``prediction`` task, this model will be applied to prediction data
  // desc = for ``train`` task, training will be continued from this model
  // desc = **Note**: can be used only in CLI version
  std::string input_model = "";

533
  // [no-save]
534
535
536
537
538
  // alias = model_output, model_out
  // desc = filename of output model in training
  // desc = **Note**: can be used only in CLI version
  std::string output_model = "LightGBM_model.txt";

539
540
541
542
543
  // desc = the feature importance type in the saved model file
  // desc = ``0``: count-based feature importance (numbers of splits are counted); ``1``: gain-based feature importance (values of gain are counted)
  // desc = **Note**: can be used only in CLI version
  int saved_feature_importance_type = 0;

544
  // [no-save]
545
546
547
548
549
550
551
552
553
554
555
556
  // alias = save_period
  // desc = frequency of saving model file snapshot
  // desc = set this to positive value to enable this function. For example, the model file will be snapshotted at each iteration if ``snapshot_freq=1``
  // desc = **Note**: can be used only in CLI version
  int snapshot_freq = -1;

  #pragma endregion

  #pragma region IO Parameters

  #pragma region Dataset Parameters

557
558
559
560
  // check = >1
  // desc = max number of bins that feature values will be bucketed in
  // desc = small number of bins may reduce training accuracy but may increase general power (deal with over-fitting)
  // desc = LightGBM will auto compress memory according to ``max_bin``. For example, LightGBM will use ``uint8_t`` for feature value if ``max_bin=255``
561
  int max_bin = 255;
Guolin Ke's avatar
Guolin Ke committed
562

Belinda Trotta's avatar
Belinda Trotta committed
563
564
565
566
  // type = multi-int
  // default = None
  // desc = max number of bins for each feature
  // desc = if not specified, will use ``max_bin`` for all features
567
  std::vector<int32_t> max_bin_by_feature;
Belinda Trotta's avatar
Belinda Trotta committed
568

569
570
571
  // check = >0
  // desc = minimal number of data inside one bin
  // desc = use this to avoid one-data-one-bin (potential over-fitting)
Guolin Ke's avatar
Guolin Ke committed
572
573
  int min_data_in_bin = 3;

574
575
576
577
578
579
580
  // alias = subsample_for_bin
  // check = >0
  // desc = number of data that sampled to construct histogram bins
  // desc = setting this to larger value will give better training result, but will increase data loading time
  // desc = set this to larger value if data is very sparse
  int bin_construct_sample_cnt = 200000;

581
  // alias = data_seed
582
  // desc = random seed for sampling data to construct histogram bins
Guolin Ke's avatar
Guolin Ke committed
583
  int data_random_seed = 1;
Guolin Ke's avatar
Guolin Ke committed
584

585
586
587
  // alias = is_sparse, enable_sparse, sparse
  // desc = used to enable/disable sparse optimization
  bool is_enable_sparse = true;
Guolin Ke's avatar
Guolin Ke committed
588

589
590
591
592
593
594
595
596
  // alias = is_enable_bundle, bundle
  // desc = set this to ``false`` to disable Exclusive Feature Bundling (EFB), which is described in `LightGBM: A Highly Efficient Gradient Boosting Decision Tree <https://papers.nips.cc/paper/6907-lightgbm-a-highly-efficient-gradient-boosting-decision-tree>`__
  // desc = **Note**: disabling this may cause the slow training speed for sparse datasets
  bool enable_bundle = true;

  // desc = set this to ``false`` to disable the special handle of missing value
  bool use_missing = true;

597
  // desc = set this to ``true`` to treat all zero as missing values (including the unshown values in LibSVM / sparse matrices)
598
599
600
  // desc = set this to ``false`` to use ``na`` for representing missing values
  bool zero_as_missing = false;

601
602
603
604
605
606
607
608
609
610
  // desc = set this to ``true`` to pre-filter the unsplittable features by ``min_data_in_leaf``
  // desc = as dataset object is initialized only once and cannot be changed after that, you may need to set this to ``false`` when searching parameters with ``min_data_in_leaf``, otherwise features are filtered by ``min_data_in_leaf`` firstly if you don't reconstruct dataset object
  // desc = **Note**: setting this to ``false`` may slow down the training
  bool feature_pre_filter = true;

  // alias = is_pre_partition
  // desc = used for parallel learning (excluding the ``feature_parallel`` mode)
  // desc = ``true`` if training data are pre-partitioned, and different machines use different partitions
  bool pre_partition = false;

611
612
613
  // alias = two_round_loading, use_two_round_loading
  // desc = set this to ``true`` if data file is too big to fit in memory
  // desc = by default, LightGBM will map data file to memory and load features from memory. This will provide faster data loading speed, but may cause run out of memory error when the data file is very big
614
  // desc = **Note**: works only in case of loading data directly from file
Guolin Ke's avatar
Guolin Ke committed
615
616
617
  bool two_round = false;

  // alias = has_header
618
  // desc = set this to ``true`` if input data has header
619
  // desc = **Note**: works only in case of loading data directly from file
Guolin Ke's avatar
Guolin Ke committed
620
621
  bool header = false;

622
623
624
625
626
  // type = int or string
  // alias = label
  // desc = used to specify the label column
  // desc = use number for index, e.g. ``label=0`` means column\_0 is the label
  // desc = add a prefix ``name:`` for column name, e.g. ``label=name:is_click``
627
  // desc = **Note**: works only in case of loading data directly from file
Guolin Ke's avatar
Guolin Ke committed
628
  std::string label_column = "";
Guolin Ke's avatar
Guolin Ke committed
629

630
631
632
633
634
  // type = int or string
  // alias = weight
  // desc = used to specify the weight column
  // desc = use number for index, e.g. ``weight=0`` means column\_0 is the weight
  // desc = add a prefix ``name:`` for column name, e.g. ``weight=name:weight``
635
  // desc = **Note**: works only in case of loading data directly from file
636
  // desc = **Note**: index starts from ``0`` and it doesn't count the label column when passing type is ``int``, e.g. when label is column\_0, and weight is column\_1, the correct parameter is ``weight=0``
Guolin Ke's avatar
Guolin Ke committed
637
  std::string weight_column = "";
Guolin Ke's avatar
Guolin Ke committed
638

639
640
641
642
643
  // type = int or string
  // alias = group, group_id, query_column, query, query_id
  // desc = used to specify the query/group id column
  // desc = use number for index, e.g. ``query=0`` means column\_0 is the query id
  // desc = add a prefix ``name:`` for column name, e.g. ``query=name:query_id``
644
  // desc = **Note**: works only in case of loading data directly from file
645
646
  // desc = **Note**: data should be grouped by query\_id
  // desc = **Note**: index starts from ``0`` and it doesn't count the label column when passing type is ``int``, e.g. when label is column\_0 and query\_id is column\_1, the correct parameter is ``query=0``
Guolin Ke's avatar
Guolin Ke committed
647
  std::string group_column = "";
Guolin Ke's avatar
Guolin Ke committed
648

649
  // type = multi-int or string
Guolin Ke's avatar
Guolin Ke committed
650
  // alias = ignore_feature, blacklist
651
652
653
654
655
  // desc = used to specify some ignoring columns in training
  // desc = use number for index, e.g. ``ignore_column=0,1,2`` means column\_0, column\_1 and column\_2 will be ignored
  // desc = add a prefix ``name:`` for column name, e.g. ``ignore_column=name:c1,c2,c3`` means c1, c2 and c3 will be ignored
  // desc = **Note**: works only in case of loading data directly from file
  // desc = **Note**: index starts from ``0`` and it doesn't count the label column when passing type is ``int``
656
  // desc = **Note**: despite the fact that specified columns will be completely ignored during the training, they still should have a valid format allowing LightGBM to load file successfully
Guolin Ke's avatar
Guolin Ke committed
657
  std::string ignore_column = "";
658

659
660
661
662
663
  // type = multi-int or string
  // alias = cat_feature, categorical_column, cat_column
  // desc = used to specify categorical features
  // desc = use number for index, e.g. ``categorical_feature=0,1,2`` means column\_0, column\_1 and column\_2 are categorical features
  // desc = add a prefix ``name:`` for column name, e.g. ``categorical_feature=name:c1,c2,c3`` means c1, c2 and c3 are categorical features
664
  // desc = **Note**: only supports categorical with ``int`` type (not applicable for data represented as pandas DataFrame in Python-package)
665
666
  // desc = **Note**: index starts from ``0`` and it doesn't count the label column when passing type is ``int``
  // desc = **Note**: all values should be less than ``Int32.MaxValue`` (2147483647)
667
  // desc = **Note**: using large values could be memory consuming. Tree decision rule works best when categorical features are presented by consecutive integers starting from zero
668
  // desc = **Note**: all negative values will be treated as **missing values**
669
  // desc = **Note**: the output cannot be monotonically constrained with respect to a categorical feature
Guolin Ke's avatar
Guolin Ke committed
670
671
  std::string categorical_feature = "";

672
673
674
675
676
  // desc = path to a ``.json`` file that specifies bin upper bounds for some or all features
  // desc = ``.json`` file should contain an array of objects, each containing the word ``feature`` (integer feature index) and ``bin_upper_bound`` (array of thresholds for binning)
  // desc = see `this file <https://github.com/microsoft/LightGBM/tree/master/examples/regression/forced_bins.json>`__ as an example
  std::string forcedbins_filename = "";

677
  // [no-save]
678
679
680
681
682
683
684
685
686
687
  // alias = is_save_binary, is_save_binary_file
  // desc = if ``true``, LightGBM will save the dataset (including validation data) to a binary file. This speed ups the data loading for the next time
  // desc = **Note**: ``init_score`` is not saved in binary file
  // desc = **Note**: can be used only in CLI version; for language-specific packages you can use the correspondent function
  bool save_binary = false;

  #pragma endregion

  #pragma region Predict Parameters

688
689
690
691
692
693
  // [no-save]
  // desc = used only in ``prediction`` task
  // desc = used to specify from which iteration to start the prediction
  // desc = ``<= 0`` means from the first iteration
  int start_iteration_predict = 0;

694
  // [no-save]
695
696
697
698
699
  // desc = used only in ``prediction`` task
  // desc = used to specify how many trained iterations will be used in prediction
  // desc = ``<= 0`` means no limit
  int num_iteration_predict = -1;

700
  // [no-save]
701
702
703
704
  // alias = is_predict_raw_score, predict_rawscore, raw_score
  // desc = used only in ``prediction`` task
  // desc = set this to ``true`` to predict only the raw scores
  // desc = set this to ``false`` to predict transformed scores
Guolin Ke's avatar
Guolin Ke committed
705
706
  bool predict_raw_score = false;

707
  // [no-save]
708
709
710
  // alias = is_predict_leaf_index, leaf_index
  // desc = used only in ``prediction`` task
  // desc = set this to ``true`` to predict with leaf index of all trees
Guolin Ke's avatar
Guolin Ke committed
711
712
  bool predict_leaf_index = false;

713
  // [no-save]
714
715
  // alias = is_predict_contrib, contrib
  // desc = used only in ``prediction`` task
716
  // desc = set this to ``true`` to estimate `SHAP values <https://arxiv.org/abs/1706.06060>`__, which represent how each feature contributes to each prediction
717
  // desc = produces ``#features + 1`` values where the last value is the expected value of the model output over the training data
718
  // desc = **Note**: if you want to get more explanation for your model's predictions using SHAP values like SHAP interaction values, you can install `shap package <https://github.com/slundberg/shap>`__
Nikita Titov's avatar
Nikita Titov committed
719
  // desc = **Note**: unlike the shap package, with ``predict_contrib`` we return a matrix with an extra column, where the last column is the expected value
Guolin Ke's avatar
Guolin Ke committed
720
721
  bool predict_contrib = false;

722
  // [no-save]
723
  // desc = used only in ``prediction`` task
724
725
726
727
728
  // desc = control whether or not LightGBM raises an error when you try to predict on data with a different number of features than the training data
  // desc = if ``false`` (the default), a fatal error will be raised if the number of features in the dataset you predict on differs from the number seen during training
  // desc = if ``true``, LightGBM will attempt to predict on whatever data you provide. This is dangerous because you might get incorrect predictions, but you could use it in situations where it is difficult or expensive to generate some features and you are very confident that they were never chosen for splits in the model
  // desc = **Note**: be very careful setting this parameter to ``true``
  bool predict_disable_shape_check = false;
Guolin Ke's avatar
Guolin Ke committed
729

730
  // [no-save]
731
732
  // desc = used only in ``prediction`` task
  // desc = if ``true``, will use early-stopping to speed up the prediction. May affect the accuracy
733
  bool pred_early_stop = false;
734

735
  // [no-save]
736
737
  // desc = used only in ``prediction`` task
  // desc = the frequency of checking early-stopping prediction
738
  int pred_early_stop_freq = 10;
Guolin Ke's avatar
Guolin Ke committed
739

740
  // [no-save]
741
742
  // desc = used only in ``prediction`` task
  // desc = the threshold of margin in early-stopping prediction
Guolin Ke's avatar
Guolin Ke committed
743
  double pred_early_stop_margin = 10.0;
Guolin Ke's avatar
Guolin Ke committed
744

745
  // [no-save]
746
  // alias = predict_result, prediction_result, predict_name, prediction_name, pred_name, name_pred
747
  // desc = used only in ``prediction`` task
748
749
750
751
752
753
754
  // desc = filename of prediction result
  // desc = **Note**: can be used only in CLI version
  std::string output_result = "LightGBM_predict_result.txt";

  #pragma endregion

  #pragma region Convert Parameters
755

756
  // [no-save]
757
  // desc = used only in ``convert_model`` task
758
  // desc = only ``cpp`` is supported yet; for conversion model to other languages consider using `m2cgen <https://github.com/BayesWitnesses/m2cgen>`__ utility
759
  // desc = if ``convert_model_language`` is set and ``task=train``, the model will be also converted
760
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
761
762
  std::string convert_model_language = "";

763
  // [no-save]
764
765
766
  // alias = convert_model_file
  // desc = used only in ``convert_model`` task
  // desc = output filename of converted model
767
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
768
769
  std::string convert_model = "gbdt_prediction.cpp";

770
  #pragma endregion
Guolin Ke's avatar
Guolin Ke committed
771

772
773
  #pragma endregion

Guolin Ke's avatar
Guolin Ke committed
774
775
  #pragma region Objective Parameters

776
  // desc = used only in ``rank_xendcg`` objective
777
778
779
  // desc = random seed for objectives, if random process is needed
  int objective_seed = 5;

780
781
782
783
  // check = >0
  // alias = num_classes
  // desc = used only in ``multi-class`` classification application
  int num_class = 1;
Guolin Ke's avatar
Guolin Ke committed
784

785
  // alias = unbalance, unbalanced_sets
786
  // desc = used only in ``binary`` and ``multiclassova`` applications
787
  // desc = set this to ``true`` if training data are unbalanced
788
  // desc = **Note**: while enabling this should increase the overall performance metric of your model, it will also result in poor estimates of the individual class probabilities
789
790
  // desc = **Note**: this parameter cannot be used at the same time with ``scale_pos_weight``, choose only **one** of them
  bool is_unbalance = false;
Guolin Ke's avatar
Guolin Ke committed
791

792
  // check = >0.0
793
  // desc = used only in ``binary`` and ``multiclassova`` applications
794
  // desc = weight of labels with positive class
795
  // desc = **Note**: while enabling this should increase the overall performance metric of your model, it will also result in poor estimates of the individual class probabilities
796
797
  // desc = **Note**: this parameter cannot be used at the same time with ``is_unbalance``, choose only **one** of them
  double scale_pos_weight = 1.0;
Guolin Ke's avatar
Guolin Ke committed
798

799
800
801
802
  // check = >0.0
  // desc = used only in ``binary`` and ``multiclassova`` classification and in ``lambdarank`` applications
  // desc = parameter for the sigmoid function
  double sigmoid = 1.0;
Guolin Ke's avatar
Guolin Ke committed
803

804
  // desc = used only in ``regression``, ``binary``, ``multiclassova`` and ``cross-entropy`` applications
805
  // desc = adjusts initial score to the mean of labels for faster convergence
Guolin Ke's avatar
Guolin Ke committed
806
807
  bool boost_from_average = true;

808
809
810
811
  // desc = used only in ``regression`` application
  // desc = used to fit ``sqrt(label)`` instead of original values and prediction result will be also automatically converted to ``prediction^2``
  // desc = might be useful in case of large-range labels
  bool reg_sqrt = false;
Guolin Ke's avatar
Guolin Ke committed
812

813
814
815
816
  // check = >0.0
  // desc = used only in ``huber`` and ``quantile`` ``regression`` applications
  // desc = parameter for `Huber loss <https://en.wikipedia.org/wiki/Huber_loss>`__ and `Quantile regression <https://en.wikipedia.org/wiki/Quantile_regression>`__
  double alpha = 0.9;
Guolin Ke's avatar
Guolin Ke committed
817

818
819
820
821
  // check = >0.0
  // desc = used only in ``fair`` ``regression`` application
  // desc = parameter for `Fair loss <https://www.kaggle.com/c/allstate-claims-severity/discussion/24520>`__
  double fair_c = 1.0;
Guolin Ke's avatar
Guolin Ke committed
822

823
824
825
826
  // check = >0.0
  // desc = used only in ``poisson`` ``regression`` application
  // desc = parameter for `Poisson regression <https://en.wikipedia.org/wiki/Poisson_regression>`__ to safeguard optimization
  double poisson_max_delta_step = 0.7;
Guolin Ke's avatar
Guolin Ke committed
827

828
829
830
831
832
833
834
  // check = >=1.0
  // check = <2.0
  // desc = used only in ``tweedie`` ``regression`` application
  // desc = used to control the variance of the tweedie distribution
  // desc = set this closer to ``2`` to shift towards a **Gamma** distribution
  // desc = set this closer to ``1`` to shift towards a **Poisson** distribution
  double tweedie_variance_power = 1.5;
Guolin Ke's avatar
Guolin Ke committed
835

836
837
  // check = >0
  // desc = used only in ``lambdarank`` application
838
  // desc = used for truncating the max DCG, refer to "truncation level" in the Sec. 3 of `LambdaMART paper <https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/MSR-TR-2010-82.pdf>`__
839
  int lambdarank_truncation_level = 20;
Guolin Ke's avatar
Guolin Ke committed
840

841
842
  // desc = used only in ``lambdarank`` application
  // desc = set this to ``true`` to normalize the lambdas for different queries, and improve the performance for unbalanced data
843
844
  // desc = set this to ``false`` to enforce the original lambdarank algorithm
  bool lambdarank_norm = true;
845

846
847
848
849
850
851
852
  // type = multi-double
  // default = 0,1,3,7,15,31,63,...,2^30-1
  // desc = used only in ``lambdarank`` application
  // desc = relevant gain for labels. For example, the gain of label ``2`` is ``3`` in case of default label gains
  // desc = separate by ``,``
  std::vector<double> label_gain;

Guolin Ke's avatar
Guolin Ke committed
853
854
855
  #pragma endregion

  #pragma region Metric Parameters
856

Guolin Ke's avatar
Guolin Ke committed
857
  // [doc-only]
858
859
860
  // alias = metrics, metric_types
  // default = ""
  // type = multi-enum
861
  // desc = metric(s) to be evaluated on the evaluation set(s)
862
  // descl2 = ``""`` (empty string or not specified) means that metric corresponding to specified ``objective`` will be used (this is possible only for pre-defined objective functions, otherwise no evaluation metric will be added)
863
  // descl2 = ``"None"`` (string, **not** a ``None`` value) means that no metric will be registered, aliases: ``na``, ``null``, ``custom``
864
865
  // descl2 = ``l1``, absolute loss, aliases: ``mean_absolute_error``, ``mae``, ``regression_l1``
  // descl2 = ``l2``, square loss, aliases: ``mean_squared_error``, ``mse``, ``regression_l2``, ``regression``
866
  // descl2 = ``rmse``, root square loss, aliases: ``root_mean_squared_error``, ``l2_root``
867
868
869
870
871
872
873
874
  // descl2 = ``quantile``, `Quantile regression <https://en.wikipedia.org/wiki/Quantile_regression>`__
  // descl2 = ``mape``, `MAPE loss <https://en.wikipedia.org/wiki/Mean_absolute_percentage_error>`__, aliases: ``mean_absolute_percentage_error``
  // descl2 = ``huber``, `Huber loss <https://en.wikipedia.org/wiki/Huber_loss>`__
  // descl2 = ``fair``, `Fair loss <https://www.kaggle.com/c/allstate-claims-severity/discussion/24520>`__
  // descl2 = ``poisson``, negative log-likelihood for `Poisson regression <https://en.wikipedia.org/wiki/Poisson_regression>`__
  // descl2 = ``gamma``, negative log-likelihood for **Gamma** regression
  // descl2 = ``gamma_deviance``, residual deviance for **Gamma** regression
  // descl2 = ``tweedie``, negative log-likelihood for **Tweedie** regression
875
  // descl2 = ``ndcg``, `NDCG <https://en.wikipedia.org/wiki/Discounted_cumulative_gain#Normalized_DCG>`__, aliases: ``lambdarank``, ``rank_xendcg``, ``xendcg``, ``xe_ndcg``, ``xe_ndcg_mart``, ``xendcg_mart``
876
877
878
879
  // descl2 = ``map``, `MAP <https://makarandtapaswi.wordpress.com/2012/07/02/intuition-behind-average-precision-and-map/>`__, aliases: ``mean_average_precision``
  // descl2 = ``auc``, `AUC <https://en.wikipedia.org/wiki/Receiver_operating_characteristic#Area_under_the_curve>`__
  // descl2 = ``binary_logloss``, `log loss <https://en.wikipedia.org/wiki/Cross_entropy>`__, aliases: ``binary``
  // descl2 = ``binary_error``, for one sample: ``0`` for correct classification, ``1`` for error classification
Belinda Trotta's avatar
Belinda Trotta committed
880
  // descl2 = ``auc_mu``, `AUC-mu <http://proceedings.mlr.press/v97/kleiman19a/kleiman19a.pdf>`__
881
882
  // descl2 = ``multi_logloss``, log loss for multi-class classification, aliases: ``multiclass``, ``softmax``, ``multiclassova``, ``multiclass_ova``, ``ova``, ``ovr``
  // descl2 = ``multi_error``, error rate for multi-class classification
Guolin Ke's avatar
Guolin Ke committed
883
884
885
  // descl2 = ``cross_entropy``, cross-entropy (with optional linear weights), aliases: ``xentropy``
  // descl2 = ``cross_entropy_lambda``, "intensity-weighted" cross-entropy, aliases: ``xentlambda``
  // descl2 = ``kullback_leibler``, `Kullback-Leibler divergence <https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence>`__, aliases: ``kldiv``
886
  // desc = support multiple metrics, separated by ``,``
Guolin Ke's avatar
Guolin Ke committed
887
888
  std::vector<std::string> metric;

889
  // [no-save]
890
  // check = >0
Guolin Ke's avatar
Guolin Ke committed
891
892
  // alias = output_freq
  // desc = frequency for metric output
893
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
894
895
  int metric_freq = 1;

896
  // [no-save]
897
898
  // alias = training_metric, is_training_metric, train_metric
  // desc = set this to ``true`` to output metric result over training dataset
899
  // desc = **Note**: can be used only in CLI version
900
  bool is_provide_training_metric = false;
901

902
903
  // type = multi-int
  // default = 1,2,3,4,5
904
  // alias = ndcg_eval_at, ndcg_at, map_eval_at, map_at
905
  // desc = used only with ``ndcg`` and ``map`` metrics
906
  // desc = `NDCG <https://en.wikipedia.org/wiki/Discounted_cumulative_gain#Normalized_DCG>`__ and `MAP <https://makarandtapaswi.wordpress.com/2012/07/02/intuition-behind-average-precision-and-map/>`__ evaluation positions, separated by ``,``
Guolin Ke's avatar
Guolin Ke committed
907
  std::vector<int> eval_at;
Guolin Ke's avatar
Guolin Ke committed
908

Belinda Trotta's avatar
Belinda Trotta committed
909
910
911
912
913
914
915
916
  // check = >0
  // desc = used only with ``multi_error`` metric
  // desc = threshold for top-k multi-error metric
  // desc = the error on each sample is ``0`` if the true class is among the top ``multi_error_top_k`` predictions, and ``1`` otherwise
  // descl2 = more precisely, the error on a sample is ``0`` if there are at least ``num_classes - multi_error_top_k`` predictions strictly less than the prediction on the true class
  // desc = when ``multi_error_top_k=1`` this is equivalent to the usual multi-error metric
  int multi_error_top_k = 1;

Belinda Trotta's avatar
Belinda Trotta committed
917
918
919
920
921
922
923
924
925
  // type = multi-double
  // default = None
  // desc = used only with ``auc_mu`` metric
  // desc = list representing flattened matrix (in row-major order) giving loss weights for classification errors
  // desc = list should have ``n * n`` elements, where ``n`` is the number of classes
  // desc = the matrix co-ordinate ``[i, j]`` should correspond to the ``i * n + j``-th element of the list
  // desc = if not specified, will use equal weights for all classes
  std::vector<double> auc_mu_weights;

Guolin Ke's avatar
Guolin Ke committed
926
927
928
929
  #pragma endregion

  #pragma region Network Parameters

930
931
932
933
  // check = >0
  // alias = num_machine
  // desc = the number of machines for parallel learning application
  // desc = this parameter is needed to be set in both **socket** and **mpi** versions
Guolin Ke's avatar
Guolin Ke committed
934
  int num_machines = 1;
Guolin Ke's avatar
Guolin Ke committed
935

936
937
938
939
  // check = >0
  // alias = local_port, port
  // desc = TCP listen port for local machines
  // desc = **Note**: don't forget to allow this port in firewall settings before training
Guolin Ke's avatar
Guolin Ke committed
940
  int local_listen_port = 12400;
Guolin Ke's avatar
Guolin Ke committed
941

942
943
944
  // check = >0
  // desc = socket time-out in minutes
  int time_out = 120;
Guolin Ke's avatar
Guolin Ke committed
945

946
947
948
  // alias = machine_list_file, machine_list, mlist
  // desc = path of file that lists machines for this parallel learning application
  // desc = each line contains one IP and one port for one machine. The format is ``ip port`` (space as a separator)
Guolin Ke's avatar
Guolin Ke committed
949
  std::string machine_list_filename = "";
Guolin Ke's avatar
Guolin Ke committed
950

951
952
  // alias = workers, nodes
  // desc = list of machines in the following format: ``ip1:port1,ip2:port2``
953
  std::string machines = "";
Guolin Ke's avatar
Guolin Ke committed
954

Guolin Ke's avatar
Guolin Ke committed
955
956
957
958
  #pragma endregion

  #pragma region GPU Parameters

959
960
  // desc = OpenCL platform ID. Usually each GPU vendor exposes one OpenCL platform
  // desc = ``-1`` means the system-wide default platform
961
  // desc = **Note**: refer to `GPU Targets <./GPU-Targets.rst#query-opencl-devices-in-your-system>`__ for more details
Guolin Ke's avatar
Guolin Ke committed
962
963
  int gpu_platform_id = -1;

964
965
  // desc = OpenCL device ID in the specified platform. Each GPU in the selected platform has a unique device ID
  // desc = ``-1`` means the default device in the selected platform
966
  // desc = **Note**: refer to `GPU Targets <./GPU-Targets.rst#query-opencl-devices-in-your-system>`__ for more details
Guolin Ke's avatar
Guolin Ke committed
967
968
  int gpu_device_id = -1;

969
  // desc = set this to ``true`` to use double precision math on GPU (by default single precision is used in OpenCL implementation and double precision is used in CUDA implementation)
Guolin Ke's avatar
Guolin Ke committed
970
971
  bool gpu_use_dp = false;

972
973
974
975
976
  // check = >0
  // desc = number of GPUs
  // desc = **Note**: can be used only in CUDA implementation
  int num_gpu = 1;

Guolin Ke's avatar
Guolin Ke committed
977
978
979
  #pragma endregion

  #pragma endregion
Guolin Ke's avatar
Guolin Ke committed
980

981
982
  size_t file_load_progress_interval_bytes = size_t(10) * 1024 * 1024 * 1024;

Guolin Ke's avatar
Guolin Ke committed
983
  bool is_parallel = false;
984
  bool is_data_based_parallel = false;
Guolin Ke's avatar
Guolin Ke committed
985
  LIGHTGBM_EXPORT void Set(const std::unordered_map<std::string, std::string>& params);
jcipar's avatar
jcipar committed
986
987
  static const std::unordered_map<std::string, std::string>& alias_table();
  static const std::unordered_set<std::string>& parameter_set();
Belinda Trotta's avatar
Belinda Trotta committed
988
  std::vector<std::vector<double>> auc_mu_weights_matrix;
989
  std::vector<std::vector<int>> interaction_constraints_vector;
990

Nikita Titov's avatar
Nikita Titov committed
991
 private:
Guolin Ke's avatar
Guolin Ke committed
992
  void CheckParamConflict();
Guolin Ke's avatar
Guolin Ke committed
993
994
  void GetMembersFromString(const std::unordered_map<std::string, std::string>& params);
  std::string SaveMembersToString() const;
Belinda Trotta's avatar
Belinda Trotta committed
995
  void GetAucMuWeights();
996
  void GetInteractionConstraints();
Guolin Ke's avatar
Guolin Ke committed
997
998
};

Guolin Ke's avatar
Guolin Ke committed
999
inline bool Config::GetString(
Guolin Ke's avatar
Guolin Ke committed
1000
1001
  const std::unordered_map<std::string, std::string>& params,
  const std::string& name, std::string* out) {
1002
  if (params.count(name) > 0 && !params.at(name).empty()) {
Guolin Ke's avatar
Guolin Ke committed
1003
1004
1005
1006
1007
1008
    *out = params.at(name);
    return true;
  }
  return false;
}

Guolin Ke's avatar
Guolin Ke committed
1009
inline bool Config::GetInt(
Guolin Ke's avatar
Guolin Ke committed
1010
1011
  const std::unordered_map<std::string, std::string>& params,
  const std::string& name, int* out) {
1012
  if (params.count(name) > 0 && !params.at(name).empty()) {
1013
    if (!Common::AtoiAndCheck(params.at(name).c_str(), out)) {
1014
      Log::Fatal("Parameter %s should be of type int, got \"%s\"",
Guolin Ke's avatar
Guolin Ke committed
1015
                 name.c_str(), params.at(name).c_str());
1016
    }
Guolin Ke's avatar
Guolin Ke committed
1017
1018
1019
1020
1021
    return true;
  }
  return false;
}

Guolin Ke's avatar
Guolin Ke committed
1022
inline bool Config::GetDouble(
Guolin Ke's avatar
Guolin Ke committed
1023
  const std::unordered_map<std::string, std::string>& params,
1024
  const std::string& name, double* out) {
1025
  if (params.count(name) > 0 && !params.at(name).empty()) {
1026
    if (!Common::AtofAndCheck(params.at(name).c_str(), out)) {
1027
      Log::Fatal("Parameter %s should be of type double, got \"%s\"",
Guolin Ke's avatar
Guolin Ke committed
1028
                 name.c_str(), params.at(name).c_str());
1029
    }
Guolin Ke's avatar
Guolin Ke committed
1030
1031
1032
1033
1034
    return true;
  }
  return false;
}

Guolin Ke's avatar
Guolin Ke committed
1035
inline bool Config::GetBool(
Guolin Ke's avatar
Guolin Ke committed
1036
1037
  const std::unordered_map<std::string, std::string>& params,
  const std::string& name, bool* out) {
1038
  if (params.count(name) > 0 && !params.at(name).empty()) {
Guolin Ke's avatar
Guolin Ke committed
1039
    std::string value = params.at(name);
Guolin Ke's avatar
Guolin Ke committed
1040
    std::transform(value.begin(), value.end(), value.begin(), Common::tolower);
1041
    if (value == std::string("false") || value == std::string("-")) {
Guolin Ke's avatar
Guolin Ke committed
1042
      *out = false;
1043
    } else if (value == std::string("true") || value == std::string("+")) {
Guolin Ke's avatar
Guolin Ke committed
1044
      *out = true;
1045
    } else {
1046
      Log::Fatal("Parameter %s should be \"true\"/\"+\" or \"false\"/\"-\", got \"%s\"",
Guolin Ke's avatar
Guolin Ke committed
1047
                 name.c_str(), params.at(name).c_str());
Guolin Ke's avatar
Guolin Ke committed
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
    }
    return true;
  }
  return false;
}

struct ParameterAlias {
  static void KeyAliasTransform(std::unordered_map<std::string, std::string>* params) {
    std::unordered_map<std::string, std::string> tmp_map;
    for (const auto& pair : *params) {
jcipar's avatar
jcipar committed
1058
1059
      auto alias = Config::alias_table().find(pair.first);
      if (alias != Config::alias_table().end()) {  // found alias
Guolin Ke's avatar
Guolin Ke committed
1060
        auto alias_set = tmp_map.find(alias->second);
1061
1062
        if (alias_set != tmp_map.end()) {  // alias already set
                                           // set priority by length & alphabetically to ensure reproducible behavior
wxchan's avatar
wxchan committed
1063
1064
          if (alias_set->second.size() < pair.first.size() ||
            (alias_set->second.size() == pair.first.size() && alias_set->second < pair.first)) {
1065
            Log::Warning("%s is set with %s=%s, %s=%s will be ignored. Current value: %s=%s",
Guolin Ke's avatar
Guolin Ke committed
1066
1067
                         alias->second.c_str(), alias_set->second.c_str(), params->at(alias_set->second).c_str(),
                         pair.first.c_str(), pair.second.c_str(), alias->second.c_str(), params->at(alias_set->second).c_str());
wxchan's avatar
wxchan committed
1068
          } else {
1069
            Log::Warning("%s is set with %s=%s, will be overridden by %s=%s. Current value: %s=%s",
Guolin Ke's avatar
Guolin Ke committed
1070
1071
                         alias->second.c_str(), alias_set->second.c_str(), params->at(alias_set->second).c_str(),
                         pair.first.c_str(), pair.second.c_str(), alias->second.c_str(), pair.second.c_str());
wxchan's avatar
wxchan committed
1072
1073
            tmp_map[alias->second] = pair.first;
          }
1074
        } else {  // alias not set
wxchan's avatar
wxchan committed
1075
1076
          tmp_map.emplace(alias->second, pair.first);
        }
jcipar's avatar
jcipar committed
1077
      } else if (Config::parameter_set().find(pair.first) == Config::parameter_set().end()) {
wxchan's avatar
wxchan committed
1078
        Log::Warning("Unknown parameter: %s", pair.first.c_str());
Guolin Ke's avatar
Guolin Ke committed
1079
1080
1081
      }
    }
    for (const auto& pair : tmp_map) {
wxchan's avatar
wxchan committed
1082
      auto alias = params->find(pair.first);
1083
      if (alias == params->end()) {  // not find
wxchan's avatar
wxchan committed
1084
1085
1086
        params->emplace(pair.first, params->at(pair.second));
        params->erase(pair.second);
      } else {
Guolin Ke's avatar
Guolin Ke committed
1087
1088
1089
        Log::Warning("%s is set=%s, %s=%s will be ignored. Current value: %s=%s",
                     pair.first.c_str(), alias->second.c_str(), pair.second.c_str(), params->at(pair.second).c_str(),
                     pair.first.c_str(), alias->second.c_str());
Guolin Ke's avatar
Guolin Ke committed
1090
1091
1092
1093
1094
      }
    }
  }
};

1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
inline std::string ParseObjectiveAlias(const std::string& type) {
  if (type == std::string("regression") || type == std::string("regression_l2")
    || type == std::string("mean_squared_error") || type == std::string("mse") || type == std::string("l2")
    || type == std::string("l2_root") || type == std::string("root_mean_squared_error") || type == std::string("rmse")) {
    return "regression";
  } else if (type == std::string("regression_l1") || type == std::string("mean_absolute_error")
    || type == std::string("l1") || type == std::string("mae")) {
    return "regression_l1";
  } else if (type == std::string("multiclass") || type == std::string("softmax")) {
    return "multiclass";
  } else if (type == std::string("multiclassova") || type == std::string("multiclass_ova") || type == std::string("ova") || type == std::string("ovr")) {
    return "multiclassova";
  } else if (type == std::string("xentropy") || type == std::string("cross_entropy")) {
    return "cross_entropy";
  } else if (type == std::string("xentlambda") || type == std::string("cross_entropy_lambda")) {
    return "cross_entropy_lambda";
  } else if (type == std::string("mean_absolute_percentage_error") || type == std::string("mape")) {
    return "mape";
1113
1114
1115
  } else if (type == std::string("rank_xendcg") || type == std::string("xendcg") || type == std::string("xe_ndcg")
             || type == std::string("xe_ndcg_mart") || type == std::string("xendcg_mart")) {
    return "rank_xendcg";
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
  } else if (type == std::string("none") || type == std::string("null") || type == std::string("custom") || type == std::string("na")) {
    return "custom";
  }
  return type;
}

inline std::string ParseMetricAlias(const std::string& type) {
  if (type == std::string("regression") || type == std::string("regression_l2") || type == std::string("l2") || type == std::string("mean_squared_error") || type == std::string("mse")) {
    return "l2";
  } else if (type == std::string("l2_root") || type == std::string("root_mean_squared_error") || type == std::string("rmse")) {
    return "rmse";
  } else if (type == std::string("regression_l1") || type == std::string("l1") || type == std::string("mean_absolute_error") || type == std::string("mae")) {
    return "l1";
  } else if (type == std::string("binary_logloss") || type == std::string("binary")) {
    return "binary_logloss";
1131
1132
  } else if (type == std::string("ndcg") || type == std::string("lambdarank") || type == std::string("rank_xendcg")
             || type == std::string("xendcg") || type == std::string("xe_ndcg") || type == std::string("xe_ndcg_mart") || type == std::string("xendcg_mart")) {
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
    return "ndcg";
  } else if (type == std::string("map") || type == std::string("mean_average_precision")) {
    return "map";
  } else if (type == std::string("multi_logloss") || type == std::string("multiclass") || type == std::string("softmax") || type == std::string("multiclassova") || type == std::string("multiclass_ova") || type == std::string("ova") || type == std::string("ovr")) {
    return "multi_logloss";
  } else if (type == std::string("xentropy") || type == std::string("cross_entropy")) {
    return "cross_entropy";
  } else if (type == std::string("xentlambda") || type == std::string("cross_entropy_lambda")) {
    return "cross_entropy_lambda";
  } else if (type == std::string("kldiv") || type == std::string("kullback_leibler")) {
    return "kullback_leibler";
  } else if (type == std::string("mean_absolute_percentage_error") || type == std::string("mape")) {
    return "mape";
Belinda Trotta's avatar
Belinda Trotta committed
1146
1147
  } else if (type == std::string("auc_mu")) {
    return "auc_mu";
1148
1149
1150
1151
1152
1153
  } else if (type == std::string("none") || type == std::string("null") || type == std::string("custom") || type == std::string("na")) {
    return "custom";
  }
  return type;
}

Guolin Ke's avatar
Guolin Ke committed
1154
1155
}   // namespace LightGBM

Belinda Trotta's avatar
Belinda Trotta committed
1156
#endif   // LightGBM_CONFIG_H_