config.h 55.9 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
#include <LightGBM/export.h>
#include <LightGBM/meta.h>
Guolin Ke's avatar
Guolin Ke committed
16
17
18
19
20
#include <LightGBM/utils/common.h>
#include <LightGBM/utils/log.h>

#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
103
104
  // 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``
  // desc = ``convert_model``, for converting model file into if-else format, see more information in `IO Parameters <#io-parameters>`__
  // 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
Guolin Ke's avatar
Guolin Ke committed
148
149
  std::string boosting = "gbdt";

150
  // alias = train, train_data, train_data_file, data_filename
151
  // desc = path of training data, LightGBM will train from this data
152
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
153
154
  std::string data = "";

155
  // alias = test, valid_data, valid_data_file, test_data, test_data_file, valid_filenames
156
  // default = ""
157
  // desc = path(s) of validation/test data, LightGBM will output metrics for these data
158
  // desc = support multiple validation data, separated by ``,``
159
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
160
161
  std::vector<std::string> valid;

162
  // alias = num_iteration, n_iter, num_tree, num_trees, num_round, num_rounds, num_boost_round, n_estimators
163
164
165
  // 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
166
  int num_iterations = 100;
Guolin Ke's avatar
Guolin Ke committed
167

168
  // alias = shrinkage_rate, eta
169
  // check = >0.0
170
171
  // desc = shrinkage rate
  // desc = in ``dart``, it also affects on normalization weights of dropped trees
Guolin Ke's avatar
Guolin Ke committed
172
173
  double learning_rate = 0.1;

174
  // default = 31
175
  // alias = num_leaf, max_leaves, max_leaf
176
  // check = >1
177
  // check = <=131072
178
  // desc = max number of leaves in one tree
Guolin Ke's avatar
Guolin Ke committed
179
180
181
  int num_leaves = kDefaultNumLeaves;

  // [doc-only]
182
183
  // type = enum
  // options = serial, feature, data, voting
184
  // alias = tree, tree_type, tree_learner_type
185
186
187
188
189
  // 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
190
191
  std::string tree_learner = "serial";

192
  // alias = num_thread, nthread, nthreads, n_jobs
Guolin Ke's avatar
Guolin Ke committed
193
  // desc = number of threads for LightGBM
194
195
196
197
198
  // 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
199
  // 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
200
201
202
  int num_threads = 0;

  // [doc-only]
203
204
  // type = enum
  // options = cpu, gpu
205
  // alias = device
206
207
208
209
  // 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
210
211
212
  std::string device_type = "cpu";

  // [doc-only]
213
  // alias = random_seed, random_state
214
215
216
217
  // 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
218
219
220
221
222
223
  int seed = 0;

  #pragma endregion

  #pragma region Learning Control Parameters

224
225
226
227
228
229
230
231
  // 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
  // descl2 = ``num_threads`` is large, e.g. ``>20``
  // 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
232
233
  bool force_col_wise = false;

234
235
236
237
238
239
240
241
242
  // 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
  // descl2 = ``num_threads`` is relatively small, e.g. ``<=16``
  // 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
243
244
  bool force_row_wise = false;

245
246
247
248
249
  // alias = hist_pool_size
  // desc = max cache size in MB for historical histogram
  // desc = ``< 0`` means no limit
  double histogram_pool_size = -1.0;

250
  // 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
251
  // desc = ``<= 0`` means no limit
Guolin Ke's avatar
Guolin Ke committed
252
253
254
  int max_depth = -1;

  // alias = min_data_per_leaf, min_data, min_child_samples
255
256
  // 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
257
258
  int min_data_in_leaf = 20;

259
260
261
  // 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
262
263
  double min_sum_hessian_in_leaf = 1e-3;

264
265
266
267
268
269
270
  // 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
271
272
  double bagging_fraction = 1.0;

Guolin Ke's avatar
Guolin Ke committed
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
  // 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;

297
298
299
300
  // 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
301
302
303
304
305
306
307
  int bagging_freq = 0;

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

  // alias = sub_feature, colsample_bytree
308
309
  // check = >0.0
  // check = <=1.0
310
  // 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
311
312
  // desc = can be used to speed up training
  // desc = can be used to deal with over-fitting
Guolin Ke's avatar
Guolin Ke committed
313
314
  double feature_fraction = 1.0;

315
316
317
  // alias = sub_feature_bynode, colsample_bynode
  // check = >0.0
  // check = <=1.0
Nikita Titov's avatar
Nikita Titov committed
318
  // 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
319
320
321
322
323
  // 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;

324
  // desc = random seed for ``feature_fraction``
Guolin Ke's avatar
Guolin Ke committed
325
326
  int feature_fraction_seed = 2;

327
328
329
330
331
332
333
334
  // 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;

335
  // alias = early_stopping_rounds, early_stopping, n_iter_no_change
336
337
  // 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
338
339
  int early_stopping_round = 0;

340
341
342
  // desc = set this to ``true``, if you want to use only the first metric for early stopping
  bool first_metric_only = false;

343
344
345
346
  // 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
347
348
  double max_delta_step = 0.0;

349
350
351
  // alias = reg_alpha
  // check = >=0.0
  // desc = L1 regularization
Guolin Ke's avatar
Guolin Ke committed
352
353
  double lambda_l1 = 0.0;

354
  // alias = reg_lambda, lambda
355
  // check = >=0.0
Guolin Ke's avatar
Guolin Ke committed
356
357
358
  // desc = L2 regularization
  double lambda_l2 = 0.0;

359
360
361
  // alias = min_split_gain
  // check = >=0.0
  // desc = the minimal gain to perform split
Guolin Ke's avatar
Guolin Ke committed
362
363
  double min_gain_to_split = 0.0;

364
  // alias = rate_drop
365
366
367
  // check = >=0.0
  // check = <=1.0
  // desc = used only in ``dart``
368
  // desc = dropout rate: a fraction of previous trees to drop during the dropout
Guolin Ke's avatar
Guolin Ke committed
369
370
  double drop_rate = 0.1;

371
  // desc = used only in ``dart``
372
  // desc = max number of dropped trees during one boosting iteration
373
  // desc = ``<=0`` means no limit
Guolin Ke's avatar
Guolin Ke committed
374
375
  int max_drop = 50;

376
377
378
  // check = >=0.0
  // check = <=1.0
  // desc = used only in ``dart``
379
  // desc = probability of skipping the dropout procedure during a boosting iteration
Guolin Ke's avatar
Guolin Ke committed
380
381
  double skip_drop = 0.5;

382
383
  // 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
384
385
  bool xgboost_dart_mode = false;

386
387
  // desc = used only in ``dart``
  // desc = set this to ``true``, if you want to use uniform drop
Guolin Ke's avatar
Guolin Ke committed
388
389
  bool uniform_drop = false;

390
391
  // desc = used only in ``dart``
  // desc = random seed to choose dropping models
Guolin Ke's avatar
Guolin Ke committed
392
393
  int drop_seed = 4;

394
395
396
397
  // 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
398
399
  double top_rate = 0.2;

400
401
402
403
  // 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
404
405
  double other_rate = 0.1;

406
407
  // check = >0
  // desc = minimal number of data per categorical group
Guolin Ke's avatar
Guolin Ke committed
408
409
  int min_data_per_group = 100;

410
411
412
  // check = >0
  // desc = used for the categorical features
  // desc = limit the max threshold points in categorical features
Guolin Ke's avatar
Guolin Ke committed
413
414
  int max_cat_threshold = 32;

415
416
  // check = >=0.0
  // desc = used for the categorical features
417
  // desc = L2 regularization in categorical split
418
  double cat_l2 = 10.0;
Guolin Ke's avatar
Guolin Ke committed
419

420
421
422
423
  // 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;
424

425
426
  // 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
427
428
429
  int max_cat_to_onehot = 4;

  // alias = topk
430
  // check = >0
431
  // desc = used only in ``voting`` tree learner, refer to `Voting parallel <./Parallel-Learning-Guide.rst#choose-appropriate-parallel-algorithm>`__
432
  // 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
433
434
435
  int top_k = 20;

  // type = multi-int
436
437
438
439
440
  // 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
441
  std::vector<int8_t> monotone_constraints;
Guolin Ke's avatar
Guolin Ke committed
442

443
444
445
446
447
448
449
  // alias = monotone_constraining_method, mc_method
  // 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
  std::string monotone_constraints_method = "basic";

Guolin Ke's avatar
Guolin Ke committed
450
  // type = multi-double
451
  // alias = feature_contrib, fc, fp, feature_penalty
Guolin Ke's avatar
Guolin Ke committed
452
453
454
455
  // 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;
456

457
458
459
460
  // 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
461
  // desc = **Note**: the forced split logic will be ignored, if the split makes gain worse
462
  // 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
463
464
  std::string forcedsplits_filename = "";

Guolin Ke's avatar
Guolin Ke committed
465
466
467
468
469
470
  // 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;

471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
  // 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
489
  std::vector<double> cegb_penalty_feature_coupled;
490

491
492
  // alias = verbose
  // desc = controls the level of LightGBM's verbosity
493
  // desc = ``< 0``: Fatal, ``= 0``: Error (Warning), ``= 1``: Info, ``> 1``: Debug
494
495
  int verbosity = 1;

496
  // [no-save]
497
498
499
500
501
502
503
  // 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 = "";

504
  // [no-save]
505
506
507
508
509
  // 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";

510
  // [no-save]
511
512
513
514
515
516
517
518
519
520
521
522
  // 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

523
524
525
526
  // 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``
527
  int max_bin = 255;
Guolin Ke's avatar
Guolin Ke committed
528

Belinda Trotta's avatar
Belinda Trotta committed
529
530
531
532
  // type = multi-int
  // default = None
  // desc = max number of bins for each feature
  // desc = if not specified, will use ``max_bin`` for all features
533
  std::vector<int32_t> max_bin_by_feature;
Belinda Trotta's avatar
Belinda Trotta committed
534

535
536
537
  // 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
538
539
  int min_data_in_bin = 3;

540
541
542
543
544
545
546
  // 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;

547
  // alias = data_seed
548
  // desc = random seed for sampling data to construct histogram bins
Guolin Ke's avatar
Guolin Ke committed
549
  int data_random_seed = 1;
Guolin Ke's avatar
Guolin Ke committed
550

551
552
553
  // 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
554

555
556
557
558
559
560
561
562
  // 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;

563
  // desc = set this to ``true`` to treat all zero as missing values (including the unshown values in LibSVM / sparse matrices)
564
565
566
  // desc = set this to ``false`` to use ``na`` for representing missing values
  bool zero_as_missing = false;

567
568
569
570
571
572
573
574
575
576
  // 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;

577
578
579
  // 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
580
  // desc = **Note**: works only in case of loading data directly from file
Guolin Ke's avatar
Guolin Ke committed
581
582
583
  bool two_round = false;

  // alias = has_header
584
  // desc = set this to ``true`` if input data has header
585
  // desc = **Note**: works only in case of loading data directly from file
Guolin Ke's avatar
Guolin Ke committed
586
587
  bool header = false;

588
589
590
591
592
  // 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``
593
  // desc = **Note**: works only in case of loading data directly from file
Guolin Ke's avatar
Guolin Ke committed
594
  std::string label_column = "";
Guolin Ke's avatar
Guolin Ke committed
595

596
597
598
599
600
  // 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``
601
  // desc = **Note**: works only in case of loading data directly from file
602
  // 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
603
  std::string weight_column = "";
Guolin Ke's avatar
Guolin Ke committed
604

605
606
607
608
609
  // 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``
610
  // desc = **Note**: works only in case of loading data directly from file
611
612
  // 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
613
  std::string group_column = "";
Guolin Ke's avatar
Guolin Ke committed
614

615
  // type = multi-int or string
Guolin Ke's avatar
Guolin Ke committed
616
  // alias = ignore_feature, blacklist
617
618
619
620
621
  // 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``
622
  // 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
623
  std::string ignore_column = "";
624

625
626
627
628
629
  // 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
630
  // desc = **Note**: only supports categorical with ``int`` type (not applicable for data represented as pandas DataFrame in Python-package)
631
632
  // 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)
633
  // 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
634
  // desc = **Note**: all negative values will be treated as **missing values**
635
  // desc = **Note**: the output cannot be monotonically constrained with respect to a categorical feature
Guolin Ke's avatar
Guolin Ke committed
636
637
  std::string categorical_feature = "";

638
639
640
641
642
  // 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 = "";

643
  // [no-save]
644
645
646
647
648
649
650
651
652
653
  // 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

654
  // [no-save]
655
656
657
658
659
  // 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;

660
  // [no-save]
661
662
663
664
  // 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
665
666
  bool predict_raw_score = false;

667
  // [no-save]
668
669
670
  // 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
671
672
  bool predict_leaf_index = false;

673
  // [no-save]
674
675
  // alias = is_predict_contrib, contrib
  // desc = used only in ``prediction`` task
676
  // desc = set this to ``true`` to estimate `SHAP values <https://arxiv.org/abs/1706.06060>`__, which represent how each feature contributes to each prediction
677
  // desc = produces ``#features + 1`` values where the last value is the expected value of the model output over the training data
678
  // 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
679
  // 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
680
681
  bool predict_contrib = false;

682
  // [no-save]
683
  // desc = used only in ``prediction`` task
684
685
686
687
688
  // 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
689

690
  // [no-save]
691
692
  // desc = used only in ``prediction`` task
  // desc = if ``true``, will use early-stopping to speed up the prediction. May affect the accuracy
693
  bool pred_early_stop = false;
694

695
  // [no-save]
696
697
  // desc = used only in ``prediction`` task
  // desc = the frequency of checking early-stopping prediction
698
  int pred_early_stop_freq = 10;
Guolin Ke's avatar
Guolin Ke committed
699

700
  // [no-save]
701
702
  // desc = used only in ``prediction`` task
  // desc = the threshold of margin in early-stopping prediction
Guolin Ke's avatar
Guolin Ke committed
703
  double pred_early_stop_margin = 10.0;
Guolin Ke's avatar
Guolin Ke committed
704

705
  // [no-save]
706
  // alias = predict_result, prediction_result, predict_name, prediction_name, pred_name, name_pred
707
  // desc = used only in ``prediction`` task
708
709
710
711
712
713
714
  // 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
715

716
  // [no-save]
717
  // desc = used only in ``convert_model`` task
718
  // desc = only ``cpp`` is supported yet; for conversion model to other languages consider using `m2cgen <https://github.com/BayesWitnesses/m2cgen>`__ utility
719
  // desc = if ``convert_model_language`` is set and ``task=train``, the model will be also converted
720
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
721
722
  std::string convert_model_language = "";

723
  // [no-save]
724
725
726
  // alias = convert_model_file
  // desc = used only in ``convert_model`` task
  // desc = output filename of converted model
727
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
728
729
  std::string convert_model = "gbdt_prediction.cpp";

730
  #pragma endregion
Guolin Ke's avatar
Guolin Ke committed
731

732
733
  #pragma endregion

Guolin Ke's avatar
Guolin Ke committed
734
735
  #pragma region Objective Parameters

736
  // desc = used only in ``rank_xendcg`` objective
737
738
739
  // desc = random seed for objectives, if random process is needed
  int objective_seed = 5;

740
741
742
743
  // check = >0
  // alias = num_classes
  // desc = used only in ``multi-class`` classification application
  int num_class = 1;
Guolin Ke's avatar
Guolin Ke committed
744

745
  // alias = unbalance, unbalanced_sets
746
  // desc = used only in ``binary`` and ``multiclassova`` applications
747
  // desc = set this to ``true`` if training data are unbalanced
748
  // 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
749
750
  // 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
751

752
  // check = >0.0
753
  // desc = used only in ``binary`` and ``multiclassova`` applications
754
  // desc = weight of labels with positive class
755
  // 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
756
757
  // 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
758

759
760
761
762
  // 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
763

764
  // desc = used only in ``regression``, ``binary``, ``multiclassova`` and ``cross-entropy`` applications
765
  // desc = adjusts initial score to the mean of labels for faster convergence
Guolin Ke's avatar
Guolin Ke committed
766
767
  bool boost_from_average = true;

768
769
770
771
  // 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
772

773
774
775
776
  // 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
777

778
779
780
781
  // 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
782

783
784
785
786
  // 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
787

788
789
790
791
792
793
794
  // 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
795

796
797
  // check = >0
  // desc = used only in ``lambdarank`` application
798
  // 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>`__
799
  int lambdarank_truncation_level = 20;
Guolin Ke's avatar
Guolin Ke committed
800

801
802
  // 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
803
804
  // desc = set this to ``false`` to enforce the original lambdarank algorithm
  bool lambdarank_norm = true;
805

806
807
808
809
810
811
812
  // 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
813
814
815
  #pragma endregion

  #pragma region Metric Parameters
816

Guolin Ke's avatar
Guolin Ke committed
817
  // [doc-only]
818
819
820
  // alias = metrics, metric_types
  // default = ""
  // type = multi-enum
821
  // desc = metric(s) to be evaluated on the evaluation set(s)
822
  // 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)
823
  // descl2 = ``"None"`` (string, **not** a ``None`` value) means that no metric will be registered, aliases: ``na``, ``null``, ``custom``
824
825
  // descl2 = ``l1``, absolute loss, aliases: ``mean_absolute_error``, ``mae``, ``regression_l1``
  // descl2 = ``l2``, square loss, aliases: ``mean_squared_error``, ``mse``, ``regression_l2``, ``regression``
826
  // descl2 = ``rmse``, root square loss, aliases: ``root_mean_squared_error``, ``l2_root``
827
828
829
830
831
832
833
834
  // 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
835
  // 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``
836
837
838
839
  // 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
840
  // descl2 = ``auc_mu``, `AUC-mu <http://proceedings.mlr.press/v97/kleiman19a/kleiman19a.pdf>`__
841
842
  // 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
843
844
845
  // 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``
846
  // desc = support multiple metrics, separated by ``,``
Guolin Ke's avatar
Guolin Ke committed
847
848
  std::vector<std::string> metric;

849
  // [no-save]
850
  // check = >0
Guolin Ke's avatar
Guolin Ke committed
851
852
  // alias = output_freq
  // desc = frequency for metric output
853
  // desc = **Note**: can be used only in CLI version
Guolin Ke's avatar
Guolin Ke committed
854
855
  int metric_freq = 1;

856
  // [no-save]
857
858
  // alias = training_metric, is_training_metric, train_metric
  // desc = set this to ``true`` to output metric result over training dataset
859
  // desc = **Note**: can be used only in CLI version
860
  bool is_provide_training_metric = false;
861

862
863
  // type = multi-int
  // default = 1,2,3,4,5
864
  // alias = ndcg_eval_at, ndcg_at, map_eval_at, map_at
865
  // desc = used only with ``ndcg`` and ``map`` metrics
866
  // 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
867
  std::vector<int> eval_at;
Guolin Ke's avatar
Guolin Ke committed
868

Belinda Trotta's avatar
Belinda Trotta committed
869
870
871
872
873
874
875
876
  // 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
877
878
879
880
881
882
883
884
885
  // 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
886
887
888
889
  #pragma endregion

  #pragma region Network Parameters

890
891
892
893
  // 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
894
  int num_machines = 1;
Guolin Ke's avatar
Guolin Ke committed
895

896
897
898
899
  // 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
900
  int local_listen_port = 12400;
Guolin Ke's avatar
Guolin Ke committed
901

902
903
904
  // check = >0
  // desc = socket time-out in minutes
  int time_out = 120;
Guolin Ke's avatar
Guolin Ke committed
905

906
907
908
  // 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
909
  std::string machine_list_filename = "";
Guolin Ke's avatar
Guolin Ke committed
910

911
912
  // alias = workers, nodes
  // desc = list of machines in the following format: ``ip1:port1,ip2:port2``
913
  std::string machines = "";
Guolin Ke's avatar
Guolin Ke committed
914

Guolin Ke's avatar
Guolin Ke committed
915
916
917
918
  #pragma endregion

  #pragma region GPU Parameters

919
920
  // desc = OpenCL platform ID. Usually each GPU vendor exposes one OpenCL platform
  // desc = ``-1`` means the system-wide default platform
921
  // 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
922
923
  int gpu_platform_id = -1;

924
925
  // 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
926
  // 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
927
928
  int gpu_device_id = -1;

929
  // desc = set this to ``true`` to use double precision math on GPU (by default single precision is used)
Guolin Ke's avatar
Guolin Ke committed
930
931
932
933
934
  bool gpu_use_dp = false;

  #pragma endregion

  #pragma endregion
Guolin Ke's avatar
Guolin Ke committed
935

936
937
  size_t file_load_progress_interval_bytes = size_t(10) * 1024 * 1024 * 1024;

Guolin Ke's avatar
Guolin Ke committed
938
  bool is_parallel = false;
939
  bool is_data_based_parallel = false;
Guolin Ke's avatar
Guolin Ke committed
940
  LIGHTGBM_EXPORT void Set(const std::unordered_map<std::string, std::string>& params);
jcipar's avatar
jcipar committed
941
942
  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
943
  std::vector<std::vector<double>> auc_mu_weights_matrix;
944

Nikita Titov's avatar
Nikita Titov committed
945
 private:
Guolin Ke's avatar
Guolin Ke committed
946
  void CheckParamConflict();
Guolin Ke's avatar
Guolin Ke committed
947
948
  void GetMembersFromString(const std::unordered_map<std::string, std::string>& params);
  std::string SaveMembersToString() const;
Belinda Trotta's avatar
Belinda Trotta committed
949
  void GetAucMuWeights();
Guolin Ke's avatar
Guolin Ke committed
950
951
};

Guolin Ke's avatar
Guolin Ke committed
952
inline bool Config::GetString(
Guolin Ke's avatar
Guolin Ke committed
953
954
  const std::unordered_map<std::string, std::string>& params,
  const std::string& name, std::string* out) {
955
  if (params.count(name) > 0 && !params.at(name).empty()) {
Guolin Ke's avatar
Guolin Ke committed
956
957
958
959
960
961
    *out = params.at(name);
    return true;
  }
  return false;
}

Guolin Ke's avatar
Guolin Ke committed
962
inline bool Config::GetInt(
Guolin Ke's avatar
Guolin Ke committed
963
964
  const std::unordered_map<std::string, std::string>& params,
  const std::string& name, int* out) {
965
  if (params.count(name) > 0 && !params.at(name).empty()) {
966
    if (!Common::AtoiAndCheck(params.at(name).c_str(), out)) {
967
      Log::Fatal("Parameter %s should be of type int, got \"%s\"",
Guolin Ke's avatar
Guolin Ke committed
968
                 name.c_str(), params.at(name).c_str());
969
    }
Guolin Ke's avatar
Guolin Ke committed
970
971
972
973
974
    return true;
  }
  return false;
}

Guolin Ke's avatar
Guolin Ke committed
975
inline bool Config::GetDouble(
Guolin Ke's avatar
Guolin Ke committed
976
  const std::unordered_map<std::string, std::string>& params,
977
  const std::string& name, double* out) {
978
  if (params.count(name) > 0 && !params.at(name).empty()) {
979
    if (!Common::AtofAndCheck(params.at(name).c_str(), out)) {
980
      Log::Fatal("Parameter %s should be of type double, got \"%s\"",
Guolin Ke's avatar
Guolin Ke committed
981
                 name.c_str(), params.at(name).c_str());
982
    }
Guolin Ke's avatar
Guolin Ke committed
983
984
985
986
987
    return true;
  }
  return false;
}

Guolin Ke's avatar
Guolin Ke committed
988
inline bool Config::GetBool(
Guolin Ke's avatar
Guolin Ke committed
989
990
  const std::unordered_map<std::string, std::string>& params,
  const std::string& name, bool* out) {
991
  if (params.count(name) > 0 && !params.at(name).empty()) {
Guolin Ke's avatar
Guolin Ke committed
992
    std::string value = params.at(name);
Guolin Ke's avatar
Guolin Ke committed
993
    std::transform(value.begin(), value.end(), value.begin(), Common::tolower);
994
    if (value == std::string("false") || value == std::string("-")) {
Guolin Ke's avatar
Guolin Ke committed
995
      *out = false;
996
    } else if (value == std::string("true") || value == std::string("+")) {
Guolin Ke's avatar
Guolin Ke committed
997
      *out = true;
998
    } else {
999
      Log::Fatal("Parameter %s should be \"true\"/\"+\" or \"false\"/\"-\", got \"%s\"",
Guolin Ke's avatar
Guolin Ke committed
1000
                 name.c_str(), params.at(name).c_str());
Guolin Ke's avatar
Guolin Ke committed
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
    }
    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
1011
1012
      auto alias = Config::alias_table().find(pair.first);
      if (alias != Config::alias_table().end()) {  // found alias
Guolin Ke's avatar
Guolin Ke committed
1013
        auto alias_set = tmp_map.find(alias->second);
1014
1015
        if (alias_set != tmp_map.end()) {  // alias already set
                                           // set priority by length & alphabetically to ensure reproducible behavior
wxchan's avatar
wxchan committed
1016
1017
          if (alias_set->second.size() < pair.first.size() ||
            (alias_set->second.size() == pair.first.size() && alias_set->second < pair.first)) {
1018
            Log::Warning("%s is set with %s=%s, %s=%s will be ignored. Current value: %s=%s",
Guolin Ke's avatar
Guolin Ke committed
1019
1020
                         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
1021
          } else {
1022
            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
1023
1024
                         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
1025
1026
            tmp_map[alias->second] = pair.first;
          }
1027
        } else {  // alias not set
wxchan's avatar
wxchan committed
1028
1029
          tmp_map.emplace(alias->second, pair.first);
        }
jcipar's avatar
jcipar committed
1030
      } else if (Config::parameter_set().find(pair.first) == Config::parameter_set().end()) {
wxchan's avatar
wxchan committed
1031
        Log::Warning("Unknown parameter: %s", pair.first.c_str());
Guolin Ke's avatar
Guolin Ke committed
1032
1033
1034
      }
    }
    for (const auto& pair : tmp_map) {
wxchan's avatar
wxchan committed
1035
      auto alias = params->find(pair.first);
1036
      if (alias == params->end()) {  // not find
wxchan's avatar
wxchan committed
1037
1038
1039
        params->emplace(pair.first, params->at(pair.second));
        params->erase(pair.second);
      } else {
Guolin Ke's avatar
Guolin Ke committed
1040
1041
1042
        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
1043
1044
1045
1046
1047
      }
    }
  }
};

1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
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";
1066
1067
1068
  } 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";
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
  } 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";
1084
1085
  } 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")) {
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
    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
1099
1100
  } else if (type == std::string("auc_mu")) {
    return "auc_mu";
1101
1102
1103
1104
1105
1106
  } 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
1107
1108
}   // namespace LightGBM

Belinda Trotta's avatar
Belinda Trotta committed
1109
#endif   // LightGBM_CONFIG_H_