dask.py 34.5 KB
Newer Older
1
# coding: utf-8
2
"""Distributed training with LightGBM and dask.distributed.
3

4
This module enables you to perform distributed training with LightGBM on
5
dask.Array and dask.DataFrame collections.
6
7

It is based on dask-lightgbm, which was based on dask-xgboost.
8
"""
9
import socket
10
from collections import defaultdict
11
from copy import deepcopy
12
from typing import Any, Callable, Dict, Iterable, List, Optional, Type, Union, Set
13
14
15
from urllib.parse import urlparse

import numpy as np
16
17
import scipy.sparse as ss

18
19
20
21
22
23
24
25
from .basic import (_LIB, LightGBMError, _choose_param_value, _ConfigAliases,
                    _log_warning, _safe_call)
from .compat import (DASK_INSTALLED, PANDAS_INSTALLED, SKLEARN_INSTALLED,
                     Client, LGBMNotFittedError, concat, dask_Array,
                     dask_DataFrame, dask_Series, default_client, delayed,
                     get_worker, pd_DataFrame, pd_Series, wait)
from .sklearn import (LGBMClassifier, LGBMModel, LGBMRanker, LGBMRegressor,
                      _lgbmmodel_doc_fit, _lgbmmodel_doc_predict)
26
27
28
29
30

_DaskCollection = Union[dask_Array, dask_DataFrame, dask_Series]
_DaskMatrixLike = Union[dask_Array, dask_DataFrame]
_DaskPart = Union[np.ndarray, pd_DataFrame, pd_Series, ss.spmatrix]
_PredictionDtype = Union[Type[np.float32], Type[np.float64], Type[np.int32], Type[np.int64]]
31
32


33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
def _get_dask_client(client: Optional[Client]) -> Client:
    """Choose a Dask client to use.

    Parameters
    ----------
    client : dask.distributed.Client or None
        Dask client.

    Returns
    -------
    client : dask.distributed.Client
        A Dask client.
    """
    if client is None:
        return default_client()
    else:
        return client


52
53
def _find_open_port(worker_ip: str, local_listen_port: int, ports_to_skip: Iterable[int]) -> int:
    """Find an open port.
54

55
56
    This function tries to find a free port on the machine it's run on. It is intended to
    be run once on each Dask worker, sequentially.
57

58
59
60
61
62
63
64
65
66
67
68
69
70
71
    Parameters
    ----------
    worker_ip : str
        IP address for the Dask worker.
    local_listen_port : int
        First port to try when searching for open ports.
    ports_to_skip: Iterable[int]
        An iterable of integers referring to ports that should be skipped. Since multiple Dask
        workers can run on the same physical machine, this method may be called multiple times
        on the same machine. ``ports_to_skip`` is used to ensure that LightGBM doesn't try to use
        the same port for two worker processes running on the same machine.

    Returns
    -------
72
    port : int
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
        A free port on the machine referenced by ``worker_ip``.
    """
    max_tries = 1000
    found_port = False
    for i in range(max_tries):
        out_port = local_listen_port + i
        if out_port in ports_to_skip:
            continue
        try:
            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
                s.bind((worker_ip, out_port))
            found_port = True
            break
        # if unavailable, you'll get OSError: Address already in use
        except OSError:
            continue
    if not found_port:
        msg = "LightGBM tried %s:%d-%d and could not create a connection. Try setting local_listen_port to a different value."
        raise RuntimeError(msg % (worker_ip, local_listen_port, out_port))
    return out_port


def _find_ports_for_workers(client: Client, worker_addresses: Iterable[str], local_listen_port: int) -> Dict[str, int]:
    """Find an open port on each worker.

    LightGBM distributed training uses TCP sockets by default, and this method is used to
    identify open ports on each worker so LightGBM can reliable create those sockets.
100
101
102

    Parameters
    ----------
103
104
105
    client : dask.distributed.Client
        Dask client.
    worker_addresses : Iterable[str]
106
        An iterable of addresses for workers in the cluster. These are strings of the form ``<protocol>://<host>:port``.
107
    local_listen_port : int
108
        First port to try when searching for open ports.
109
110
111

    Returns
    -------
112
113
    result : Dict[str, int]
        Dictionary where keys are worker addresses and values are an open port for LightGBM to use.
114
    """
115
    lightgbm_ports: Set[int] = set()
116
117
118
119
120
121
122
123
124
125
126
127
128
    worker_ip_to_port = {}
    for worker_address in worker_addresses:
        port = client.submit(
            func=_find_open_port,
            workers=[worker_address],
            worker_ip=urlparse(worker_address).hostname,
            local_listen_port=local_listen_port,
            ports_to_skip=lightgbm_ports
        ).result()
        lightgbm_ports.add(port)
        worker_ip_to_port[worker_address] = port

    return worker_ip_to_port
129
130


131
def _concat(seq: List[_DaskPart]) -> _DaskPart:
132
133
    if isinstance(seq[0], np.ndarray):
        return np.concatenate(seq, axis=0)
134
    elif isinstance(seq[0], (pd_DataFrame, pd_Series)):
135
        return concat(seq, axis=0)
136
137
138
139
140
141
    elif isinstance(seq[0], ss.spmatrix):
        return ss.vstack(seq, format='csr')
    else:
        raise TypeError('Data must be one of: numpy arrays, pandas dataframes, sparse matrices (from scipy). Got %s.' % str(type(seq[0])))


142
143
144
145
146
147
148
149
150
def _train_part(
    params: Dict[str, Any],
    model_factory: Type[LGBMModel],
    list_of_parts: List[Dict[str, _DaskPart]],
    worker_address_to_port: Dict[str, int],
    return_model: bool,
    time_out: int = 120,
    **kwargs: Any
) -> Optional[LGBMModel]:
151
152
153
154
155
156
157
158
159
160
161
162
    local_worker_address = get_worker().address
    machine_list = ','.join([
        '%s:%d' % (urlparse(worker_address).hostname, port)
        for worker_address, port
        in worker_address_to_port.items()
    ])
    network_params = {
        'machines': machine_list,
        'local_listen_port': worker_address_to_port[local_worker_address],
        'time_out': time_out,
        'num_machines': len(worker_address_to_port)
    }
163
164
    params.update(network_params)

165
166
    is_ranker = issubclass(model_factory, LGBMRanker)

167
    # Concatenate many parts into one
168
169
170
171
172
173
174
175
176
177
178
179
    data = _concat([x['data'] for x in list_of_parts])
    label = _concat([x['label'] for x in list_of_parts])

    if 'weight' in list_of_parts[0]:
        weight = _concat([x['weight'] for x in list_of_parts])
    else:
        weight = None

    if 'group' in list_of_parts[0]:
        group = _concat([x['group'] for x in list_of_parts])
    else:
        group = None
180
181
182

    try:
        model = model_factory(**params)
183
        if is_ranker:
184
            model.fit(data, label, sample_weight=weight, group=group, **kwargs)
185
        else:
186
            model.fit(data, label, sample_weight=weight, **kwargs)
187

188
189
190
191
192
193
    finally:
        _safe_call(_LIB.LGBM_NetworkFree())

    return model if return_model else None


194
def _split_to_parts(data: _DaskCollection, is_matrix: bool) -> List[_DaskPart]:
195
196
    parts = data.to_delayed()
    if isinstance(parts, np.ndarray):
197
198
199
200
        if is_matrix:
            assert parts.shape[1] == 1
        else:
            assert parts.ndim == 1 or parts.shape[1] == 1
201
202
203
204
        parts = parts.flatten().tolist()
    return parts


205
206
207
208
209
210
211
212
213
214
def _train(
    client: Client,
    data: _DaskMatrixLike,
    label: _DaskCollection,
    params: Dict[str, Any],
    model_factory: Type[LGBMModel],
    sample_weight: Optional[_DaskCollection] = None,
    group: Optional[_DaskCollection] = None,
    **kwargs: Any
) -> LGBMModel:
215
216
217
218
    """Inner train routine.

    Parameters
    ----------
219
220
    client : dask.distributed.Client
        Dask client.
221
    data : Dask Array or Dask DataFrame of shape = [n_samples, n_features]
222
        Input feature matrix.
223
    label : Dask Array, Dask DataFrame or Dask Series of shape = [n_samples]
224
225
        The target values (class labels in classification, real numbers in regression).
    params : dict
226
        Parameters passed to constructor of the local underlying model.
227
    model_factory : lightgbm.LGBMClassifier, lightgbm.LGBMRegressor, or lightgbm.LGBMRanker class
228
        Class of the local underlying model.
229
    sample_weight : Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)
230
        Weights of training data.
231
    group : Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)
232
233
234
235
236
        Group/query data.
        Only used in the learning-to-rank task.
        sum(group) = n_samples.
        For example, if you have a 100-document dataset with ``group = [10, 20, 40, 10, 10, 10]``, that means that you have 6 groups,
        where the first 10 records are in the first group, records 11-30 are in the second group, records 31-70 are in the third group, etc.
237
238
239
240
241
242
243
    **kwargs
        Other parameters passed to ``fit`` method of the local underlying model.

    Returns
    -------
    model : lightgbm.LGBMClassifier, lightgbm.LGBMRegressor, or lightgbm.LGBMRanker class
        Returns fitted underlying model.
244
    """
245
246
    params = deepcopy(params)

247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
    params = _choose_param_value(
        main_param_name="local_listen_port",
        params=params,
        default_value=12400
    )

    params = _choose_param_value(
        main_param_name="tree_learner",
        params=params,
        default_value="data"
    )
    allowed_tree_learners = {
        'data',
        'data_parallel',
        'feature',
        'feature_parallel',
        'voting',
        'voting_parallel'
    }
    if params["tree_learner"] not in allowed_tree_learners:
267
        _log_warning('Parameter tree_learner set to %s, which is not allowed. Using "data" as default' % params['tree_learner'])
268
269
270
271
272
273
274
275
276
277
278
279
        params['tree_learner'] = 'data'

    if params['tree_learner'] not in {'data', 'data_parallel'}:
        _log_warning(
            'Support for tree_learner %s in lightgbm.dask is experimental and may break in a future release. \n'
            'Use "data" for a stable, well-tested interface.' % params['tree_learner']
        )

    # Some passed-in parameters can be removed:
    #   * 'machines': constructed automatically from Dask worker list
    #   * 'num_machines': set automatically from Dask worker list
    #   * 'num_threads': overridden to match nthreads on each Dask process
280
    for param_alias in _ConfigAliases.get('machines', 'num_machines', 'num_threads'):
281
        params.pop(param_alias, None)
282

283
    # Split arrays/dataframes into parts. Arrange parts into dicts to enforce co-locality
284
285
    data_parts = _split_to_parts(data=data, is_matrix=True)
    label_parts = _split_to_parts(data=label, is_matrix=False)
286
    parts = [{'data': x, 'label': y} for (x, y) in zip(data_parts, label_parts)]
287
    n_parts = len(parts)
288
289
290

    if sample_weight is not None:
        weight_parts = _split_to_parts(data=sample_weight, is_matrix=False)
291
        for i in range(n_parts):
292
            parts[i]['weight'] = weight_parts[i]
293
294
295

    if group is not None:
        group_parts = _split_to_parts(data=group, is_matrix=False)
296
        for i in range(n_parts):
297
            parts[i]['group'] = group_parts[i]
298
299

    # Start computation in the background
300
    parts = list(map(delayed, parts))
301
302
303
304
    parts = client.compute(parts)
    wait(parts)

    for part in parts:
305
        if part.status == 'error':  # type: ignore
306
307
308
            return part  # trigger error locally

    # Find locations of all parts and map them to particular Dask workers
309
    key_to_part_dict = {part.key: part for part in parts}  # type: ignore
310
311
312
313
314
315
316
317
    who_has = client.who_has(parts)
    worker_map = defaultdict(list)
    for key, workers in who_has.items():
        worker_map[next(iter(workers))].append(key_to_part_dict[key])

    master_worker = next(iter(worker_map))
    worker_ncores = client.ncores()

318
319
320
321
322
323
    # find an open port on each worker. note that multiple workers can run
    # on the same machine, so this needs to ensure that each one gets its
    # own port
    worker_address_to_port = _find_ports_for_workers(
        client=client,
        worker_addresses=worker_map.keys(),
324
        local_listen_port=params["local_listen_port"]
325
326
    )

327
    # Tell each worker to train on the parts that it has locally
328
329
330
331
332
333
334
335
336
337
338
339
340
    futures_classifiers = [
        client.submit(
            _train_part,
            model_factory=model_factory,
            params={**params, 'num_threads': worker_ncores[worker]},
            list_of_parts=list_of_parts,
            worker_address_to_port=worker_address_to_port,
            time_out=params.get('time_out', 120),
            return_model=(worker == master_worker),
            **kwargs
        )
        for worker, list_of_parts in worker_map.items()
    ]
341
342
343
344
345
346

    results = client.gather(futures_classifiers)
    results = [v for v in results if v]
    return results[0]


347
348
349
350
351
352
353
354
355
def _predict_part(
    part: _DaskPart,
    model: LGBMModel,
    raw_score: bool,
    pred_proba: bool,
    pred_leaf: bool,
    pred_contrib: bool,
    **kwargs: Any
) -> _DaskPart:
356

357
    if part.shape[0] == 0:
358
        result = np.array([])
359
360
    elif pred_proba:
        result = model.predict_proba(
361
            part,
362
363
364
365
366
            raw_score=raw_score,
            pred_leaf=pred_leaf,
            pred_contrib=pred_contrib,
            **kwargs
        )
367
    else:
368
        result = model.predict(
369
            part,
370
371
372
373
374
            raw_score=raw_score,
            pred_leaf=pred_leaf,
            pred_contrib=pred_contrib,
            **kwargs
        )
375

376
    # dask.DataFrame.map_partitions() expects each call to return a pandas DataFrame or Series
377
    if isinstance(part, pd_DataFrame):
378
        if pred_proba or pred_contrib or pred_leaf:
379
            result = pd_DataFrame(result, index=part.index)
380
        else:
381
            result = pd_Series(result, index=part.index, name='predictions')
382
383
384
385

    return result


386
387
388
389
390
391
392
393
394
395
def _predict(
    model: LGBMModel,
    data: _DaskMatrixLike,
    raw_score: bool = False,
    pred_proba: bool = False,
    pred_leaf: bool = False,
    pred_contrib: bool = False,
    dtype: _PredictionDtype = np.float32,
    **kwargs: Any
) -> dask_Array:
396
397
398
399
    """Inner predict routine.

    Parameters
    ----------
400
    model : lightgbm.LGBMClassifier, lightgbm.LGBMRegressor, or lightgbm.LGBMRanker class
401
        Fitted underlying model.
402
    data : Dask Array or Dask DataFrame of shape = [n_samples, n_features]
403
        Input feature matrix.
404
405
    raw_score : bool, optional (default=False)
        Whether to predict raw scores.
406
407
408
409
410
411
    pred_proba : bool, optional (default=False)
        Should method return results of ``predict_proba`` (``pred_proba=True``) or ``predict`` (``pred_proba=False``).
    pred_leaf : bool, optional (default=False)
        Whether to predict leaf index.
    pred_contrib : bool, optional (default=False)
        Whether to predict feature contributions.
412
    dtype : np.dtype, optional (default=np.float32)
413
        Dtype of the output.
414
    **kwargs
415
        Other parameters passed to ``predict`` or ``predict_proba`` method.
416
417
418

    Returns
    -------
419
    predicted_result : Dask Array of shape = [n_samples] or shape = [n_samples, n_classes]
420
        The predicted values.
421
    X_leaves : Dask Array of shape = [n_samples, n_trees] or shape = [n_samples, n_trees * n_classes]
422
        If ``pred_leaf=True``, the predicted leaf of every tree for each sample.
423
    X_SHAP_values : Dask Array of shape = [n_samples, n_features + 1] or shape = [n_samples, (n_features + 1) * n_classes]
424
        If ``pred_contrib=True``, the feature contributions for each sample.
425
    """
426
427
    if not all((DASK_INSTALLED, PANDAS_INSTALLED, SKLEARN_INSTALLED)):
        raise LightGBMError('dask, pandas and scikit-learn are required for lightgbm.dask')
428
    if isinstance(data, dask_DataFrame):
429
430
431
432
433
434
435
436
437
        return data.map_partitions(
            _predict_part,
            model=model,
            raw_score=raw_score,
            pred_proba=pred_proba,
            pred_leaf=pred_leaf,
            pred_contrib=pred_contrib,
            **kwargs
        ).values
438
    elif isinstance(data, dask_Array):
439
        if pred_proba:
440
441
442
            kwargs['chunks'] = (data.chunks[0], (model.n_classes_,))
        else:
            kwargs['drop_axis'] = 1
443
444
445
446
447
448
449
450
451
452
        return data.map_blocks(
            _predict_part,
            model=model,
            raw_score=raw_score,
            pred_proba=pred_proba,
            pred_leaf=pred_leaf,
            pred_contrib=pred_contrib,
            dtype=dtype,
            **kwargs
        )
453
    else:
454
        raise TypeError('Data must be either Dask Array or Dask DataFrame. Got %s.' % str(type(data)))
455
456


457
class _DaskLGBMModel:
458

459
460
    @property
    def client_(self) -> Client:
461
        """:obj:`dask.distributed.Client`: Dask client.
462
463
464
465
466
467
468
469
470
471
472
473
474
475

        This property can be passed in the constructor or updated
        with ``model.set_params(client=client)``.
        """
        if not getattr(self, "fitted_", False):
            raise LGBMNotFittedError('Cannot access property client_ before calling fit().')

        return _get_dask_client(client=self.client)

    def _lgb_getstate(self) -> Dict[Any, Any]:
        """Remove un-picklable attributes before serialization."""
        client = self.__dict__.pop("client", None)
        self._other_params.pop("client", None)
        out = deepcopy(self.__dict__)
476
        out.update({"client": None})
477
478
479
        self.client = client
        return out

480
481
482
483
484
485
486
487
488
    def _fit(
        self,
        model_factory: Type[LGBMModel],
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        group: Optional[_DaskCollection] = None,
        **kwargs: Any
    ) -> "_DaskLGBMModel":
489
490
        if not all((DASK_INSTALLED, PANDAS_INSTALLED, SKLEARN_INSTALLED)):
            raise LightGBMError('dask, pandas and scikit-learn are required for lightgbm.dask')
491
492

        params = self.get_params(True)
493
        params.pop("client", None)
494
495

        model = _train(
496
            client=_get_dask_client(self.client),
497
498
499
500
501
502
503
504
            data=X,
            label=y,
            params=params,
            model_factory=model_factory,
            sample_weight=sample_weight,
            group=group,
            **kwargs
        )
505
506
507
508
509
510

        self.set_params(**model.get_params())
        self._copy_extra_params(model, self)

        return self

511
    def _to_local(self, model_factory: Type[LGBMModel]) -> LGBMModel:
512
513
514
        params = self.get_params()
        params.pop("client", None)
        model = model_factory(**params)
515
        self._copy_extra_params(self, model)
516
        model._other_params.pop("client", None)
517
518
519
        return model

    @staticmethod
520
    def _copy_extra_params(source: Union["_DaskLGBMModel", LGBMModel], dest: Union["_DaskLGBMModel", LGBMModel]) -> None:
521
522
523
524
        params = source.get_params()
        attributes = source.__dict__
        extra_param_names = set(attributes.keys()).difference(params.keys())
        for name in extra_param_names:
525
            setattr(dest, name, attributes[name])
526
527


528
class DaskLGBMClassifier(LGBMClassifier, _DaskLGBMModel):
529
530
    """Distributed version of lightgbm.LGBMClassifier."""

531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
    def __init__(
        self,
        boosting_type: str = 'gbdt',
        num_leaves: int = 31,
        max_depth: int = -1,
        learning_rate: float = 0.1,
        n_estimators: int = 100,
        subsample_for_bin: int = 200000,
        objective: Optional[Union[Callable, str]] = None,
        class_weight: Optional[Union[dict, str]] = None,
        min_split_gain: float = 0.,
        min_child_weight: float = 1e-3,
        min_child_samples: int = 20,
        subsample: float = 1.,
        subsample_freq: int = 0,
        colsample_bytree: float = 1.,
        reg_alpha: float = 0.,
        reg_lambda: float = 0.,
        random_state: Optional[Union[int, np.random.RandomState]] = None,
        n_jobs: int = -1,
        silent: bool = True,
        importance_type: str = 'split',
        client: Optional[Client] = None,
        **kwargs: Any
    ):
        """Docstring is inherited from the lightgbm.LGBMClassifier.__init__."""
        self.client = client
        super().__init__(
            boosting_type=boosting_type,
            num_leaves=num_leaves,
            max_depth=max_depth,
            learning_rate=learning_rate,
            n_estimators=n_estimators,
            subsample_for_bin=subsample_for_bin,
            objective=objective,
            class_weight=class_weight,
            min_split_gain=min_split_gain,
            min_child_weight=min_child_weight,
            min_child_samples=min_child_samples,
            subsample=subsample,
            subsample_freq=subsample_freq,
            colsample_bytree=colsample_bytree,
            reg_alpha=reg_alpha,
            reg_lambda=reg_lambda,
            random_state=random_state,
            n_jobs=n_jobs,
            silent=silent,
            importance_type=importance_type,
            **kwargs
        )

    _base_doc = LGBMClassifier.__init__.__doc__
    _before_kwargs, _kwargs, _after_kwargs = _base_doc.partition('**kwargs')
584
    _base_doc = (
585
586
587
588
589
590
        _before_kwargs
        + 'client : dask.distributed.Client or None, optional (default=None)\n'
        + ' ' * 12 + 'Dask client. If ``None``, ``distributed.default_client()`` will be used at runtime. The Dask client used by this class will not be saved if the model object is pickled.\n'
        + ' ' * 8 + _kwargs + _after_kwargs
    )

591
592
593
594
    # the note on custom objective functions in LGBMModel.__init__ is not
    # currently relevant for the Dask estimators
    __init__.__doc__ = _base_doc[:_base_doc.find('Note\n')]

595
596
597
    def __getstate__(self) -> Dict[Any, Any]:
        return self._lgb_getstate()

598
599
600
601
602
603
604
    def fit(
        self,
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        **kwargs: Any
    ) -> "DaskLGBMClassifier":
605
        """Docstring is inherited from the lightgbm.LGBMClassifier.fit."""
606
607
608
609
610
611
612
613
        return self._fit(
            model_factory=LGBMClassifier,
            X=X,
            y=y,
            sample_weight=sample_weight,
            **kwargs
        )

614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
    _base_doc = _lgbmmodel_doc_fit.format(
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        y_shape="Dask Array, Dask DataFrame or Dask Series of shape = [n_samples]",
        sample_weight_shape="Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)",
        group_shape="Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)"
    )

    # DaskLGBMClassifier does not support init_score, evaluation data, or early stopping
    _base_doc = (_base_doc[:_base_doc.find('init_score :')]
                 + _base_doc[_base_doc.find('verbose :'):])

    # DaskLGBMClassifier support for callbacks and init_model is not tested
    fit.__doc__ = (
        _base_doc[:_base_doc.find('callbacks :')]
        + '**kwargs\n'
629
        + ' ' * 12 + 'Other parameters passed through to ``LGBMClassifier.fit()``.\n'
630
    )
631

632
    def predict(self, X: _DaskMatrixLike, **kwargs: Any) -> dask_Array:
633
        """Docstring is inherited from the lightgbm.LGBMClassifier.predict."""
634
635
636
637
638
639
640
        return _predict(
            model=self.to_local(),
            data=X,
            dtype=self.classes_.dtype,
            **kwargs
        )

641
642
643
644
645
646
647
648
    predict.__doc__ = _lgbmmodel_doc_predict.format(
        description="Return the predicted value for each sample.",
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        output_name="predicted_result",
        predicted_result_shape="Dask Array of shape = [n_samples] or shape = [n_samples, n_classes]",
        X_leaves_shape="Dask Array of shape = [n_samples, n_trees] or shape = [n_samples, n_trees * n_classes]",
        X_SHAP_values_shape="Dask Array of shape = [n_samples, n_features + 1] or shape = [n_samples, (n_features + 1) * n_classes]"
    )
649

650
    def predict_proba(self, X: _DaskMatrixLike, **kwargs: Any) -> dask_Array:
651
        """Docstring is inherited from the lightgbm.LGBMClassifier.predict_proba."""
652
653
654
655
656
657
658
        return _predict(
            model=self.to_local(),
            data=X,
            pred_proba=True,
            **kwargs
        )

659
660
661
662
    predict_proba.__doc__ = _lgbmmodel_doc_predict.format(
        description="Return the predicted probability for each class for each sample.",
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        output_name="predicted_probability",
663
        predicted_result_shape="Dask Array of shape = [n_samples] or shape = [n_samples, n_classes]",
664
665
666
        X_leaves_shape="Dask Array of shape = [n_samples, n_trees] or shape = [n_samples, n_trees * n_classes]",
        X_SHAP_values_shape="Dask Array of shape = [n_samples, n_features + 1] or shape = [n_samples, (n_features + 1) * n_classes]"
    )
667

668
    def to_local(self) -> LGBMClassifier:
669
670
671
672
673
        """Create regular version of lightgbm.LGBMClassifier from the distributed version.

        Returns
        -------
        model : lightgbm.LGBMClassifier
674
            Local underlying model.
675
676
677
678
        """
        return self._to_local(LGBMClassifier)


679
class DaskLGBMRegressor(LGBMRegressor, _DaskLGBMModel):
680
    """Distributed version of lightgbm.LGBMRegressor."""
681

682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
    def __init__(
        self,
        boosting_type: str = 'gbdt',
        num_leaves: int = 31,
        max_depth: int = -1,
        learning_rate: float = 0.1,
        n_estimators: int = 100,
        subsample_for_bin: int = 200000,
        objective: Optional[Union[Callable, str]] = None,
        class_weight: Optional[Union[dict, str]] = None,
        min_split_gain: float = 0.,
        min_child_weight: float = 1e-3,
        min_child_samples: int = 20,
        subsample: float = 1.,
        subsample_freq: int = 0,
        colsample_bytree: float = 1.,
        reg_alpha: float = 0.,
        reg_lambda: float = 0.,
        random_state: Optional[Union[int, np.random.RandomState]] = None,
        n_jobs: int = -1,
        silent: bool = True,
        importance_type: str = 'split',
        client: Optional[Client] = None,
        **kwargs: Any
    ):
        """Docstring is inherited from the lightgbm.LGBMRegressor.__init__."""
        self.client = client
        super().__init__(
            boosting_type=boosting_type,
            num_leaves=num_leaves,
            max_depth=max_depth,
            learning_rate=learning_rate,
            n_estimators=n_estimators,
            subsample_for_bin=subsample_for_bin,
            objective=objective,
            class_weight=class_weight,
            min_split_gain=min_split_gain,
            min_child_weight=min_child_weight,
            min_child_samples=min_child_samples,
            subsample=subsample,
            subsample_freq=subsample_freq,
            colsample_bytree=colsample_bytree,
            reg_alpha=reg_alpha,
            reg_lambda=reg_lambda,
            random_state=random_state,
            n_jobs=n_jobs,
            silent=silent,
            importance_type=importance_type,
            **kwargs
        )

    _base_doc = LGBMRegressor.__init__.__doc__
    _before_kwargs, _kwargs, _after_kwargs = _base_doc.partition('**kwargs')
735
    _base_doc = (
736
737
738
739
740
741
        _before_kwargs
        + 'client : dask.distributed.Client or None, optional (default=None)\n'
        + ' ' * 12 + 'Dask client. If ``None``, ``distributed.default_client()`` will be used at runtime. The Dask client used by this class will not be saved if the model object is pickled.\n'
        + ' ' * 8 + _kwargs + _after_kwargs
    )

742
743
744
745
    # the note on custom objective functions in LGBMModel.__init__ is not
    # currently relevant for the Dask estimators
    __init__.__doc__ = _base_doc[:_base_doc.find('Note\n')]

746
747
748
    def __getstate__(self) -> Dict[Any, Any]:
        return self._lgb_getstate()

749
750
751
752
753
754
755
    def fit(
        self,
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        **kwargs: Any
    ) -> "DaskLGBMRegressor":
756
        """Docstring is inherited from the lightgbm.LGBMRegressor.fit."""
757
758
759
760
761
762
763
764
        return self._fit(
            model_factory=LGBMRegressor,
            X=X,
            y=y,
            sample_weight=sample_weight,
            **kwargs
        )

765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
    _base_doc = _lgbmmodel_doc_fit.format(
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        y_shape="Dask Array, Dask DataFrame or Dask Series of shape = [n_samples]",
        sample_weight_shape="Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)",
        group_shape="Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)"
    )

    # DaskLGBMRegressor does not support init_score, evaluation data, or early stopping
    _base_doc = (_base_doc[:_base_doc.find('init_score :')]
                 + _base_doc[_base_doc.find('verbose :'):])

    # DaskLGBMRegressor support for callbacks and init_model is not tested
    fit.__doc__ = (
        _base_doc[:_base_doc.find('callbacks :')]
        + '**kwargs\n'
780
        + ' ' * 12 + 'Other parameters passed through to ``LGBMRegressor.fit()``.\n'
781
    )
782

783
    def predict(self, X: _DaskMatrixLike, **kwargs) -> dask_Array:
784
        """Docstring is inherited from the lightgbm.LGBMRegressor.predict."""
785
786
787
788
789
790
        return _predict(
            model=self.to_local(),
            data=X,
            **kwargs
        )

791
792
793
794
795
796
797
798
    predict.__doc__ = _lgbmmodel_doc_predict.format(
        description="Return the predicted value for each sample.",
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        output_name="predicted_result",
        predicted_result_shape="Dask Array of shape = [n_samples]",
        X_leaves_shape="Dask Array of shape = [n_samples, n_trees]",
        X_SHAP_values_shape="Dask Array of shape = [n_samples, n_features + 1]"
    )
799

800
    def to_local(self) -> LGBMRegressor:
801
802
803
804
805
        """Create regular version of lightgbm.LGBMRegressor from the distributed version.

        Returns
        -------
        model : lightgbm.LGBMRegressor
806
            Local underlying model.
807
808
        """
        return self._to_local(LGBMRegressor)
809
810


811
class DaskLGBMRanker(LGBMRanker, _DaskLGBMModel):
812
    """Distributed version of lightgbm.LGBMRanker."""
813

814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
    def __init__(
        self,
        boosting_type: str = 'gbdt',
        num_leaves: int = 31,
        max_depth: int = -1,
        learning_rate: float = 0.1,
        n_estimators: int = 100,
        subsample_for_bin: int = 200000,
        objective: Optional[Union[Callable, str]] = None,
        class_weight: Optional[Union[dict, str]] = None,
        min_split_gain: float = 0.,
        min_child_weight: float = 1e-3,
        min_child_samples: int = 20,
        subsample: float = 1.,
        subsample_freq: int = 0,
        colsample_bytree: float = 1.,
        reg_alpha: float = 0.,
        reg_lambda: float = 0.,
        random_state: Optional[Union[int, np.random.RandomState]] = None,
        n_jobs: int = -1,
        silent: bool = True,
        importance_type: str = 'split',
        client: Optional[Client] = None,
        **kwargs: Any
    ):
        """Docstring is inherited from the lightgbm.LGBMRanker.__init__."""
        self.client = client
        super().__init__(
            boosting_type=boosting_type,
            num_leaves=num_leaves,
            max_depth=max_depth,
            learning_rate=learning_rate,
            n_estimators=n_estimators,
            subsample_for_bin=subsample_for_bin,
            objective=objective,
            class_weight=class_weight,
            min_split_gain=min_split_gain,
            min_child_weight=min_child_weight,
            min_child_samples=min_child_samples,
            subsample=subsample,
            subsample_freq=subsample_freq,
            colsample_bytree=colsample_bytree,
            reg_alpha=reg_alpha,
            reg_lambda=reg_lambda,
            random_state=random_state,
            n_jobs=n_jobs,
            silent=silent,
            importance_type=importance_type,
            **kwargs
        )

    _base_doc = LGBMRanker.__init__.__doc__
    _before_kwargs, _kwargs, _after_kwargs = _base_doc.partition('**kwargs')
867
    _base_doc = (
868
869
870
871
872
873
        _before_kwargs
        + 'client : dask.distributed.Client or None, optional (default=None)\n'
        + ' ' * 12 + 'Dask client. If ``None``, ``distributed.default_client()`` will be used at runtime. The Dask client used by this class will not be saved if the model object is pickled.\n'
        + ' ' * 8 + _kwargs + _after_kwargs
    )

874
875
876
877
    # the note on custom objective functions in LGBMModel.__init__ is not
    # currently relevant for the Dask estimators
    __init__.__doc__ = _base_doc[:_base_doc.find('Note\n')]

878
879
880
    def __getstate__(self) -> Dict[Any, Any]:
        return self._lgb_getstate()

881
882
883
884
885
886
887
888
889
    def fit(
        self,
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        init_score: Optional[_DaskCollection] = None,
        group: Optional[_DaskCollection] = None,
        **kwargs: Any
    ) -> "DaskLGBMRanker":
890
891
892
893
        """Docstring is inherited from the lightgbm.LGBMRanker.fit."""
        if init_score is not None:
            raise RuntimeError('init_score is not currently supported in lightgbm.dask')

894
895
896
897
898
899
900
901
902
        return self._fit(
            model_factory=LGBMRanker,
            X=X,
            y=y,
            sample_weight=sample_weight,
            group=group,
            **kwargs
        )

903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
    _base_doc = _lgbmmodel_doc_fit.format(
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        y_shape="Dask Array, Dask DataFrame or Dask Series of shape = [n_samples]",
        sample_weight_shape="Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)",
        group_shape="Dask Array, Dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)"
    )

    # DaskLGBMRanker does not support init_score, evaluation data, or early stopping
    _base_doc = (_base_doc[:_base_doc.find('init_score :')]
                 + _base_doc[_base_doc.find('init_score :'):])

    _base_doc = (_base_doc[:_base_doc.find('eval_set :')]
                 + _base_doc[_base_doc.find('verbose :'):])

    # DaskLGBMRanker support for callbacks and init_model is not tested
    fit.__doc__ = (
        _base_doc[:_base_doc.find('callbacks :')]
        + '**kwargs\n'
921
        + ' ' * 12 + 'Other parameters passed through to ``LGBMRanker.fit()``.\n'
922
    )
923

924
    def predict(self, X: _DaskMatrixLike, **kwargs: Any) -> dask_Array:
925
926
        """Docstring is inherited from the lightgbm.LGBMRanker.predict."""
        return _predict(self.to_local(), X, **kwargs)
927

928
929
930
931
932
933
934
935
    predict.__doc__ = _lgbmmodel_doc_predict.format(
        description="Return the predicted value for each sample.",
        X_shape="Dask Array or Dask DataFrame of shape = [n_samples, n_features]",
        output_name="predicted_result",
        predicted_result_shape="Dask Array of shape = [n_samples]",
        X_leaves_shape="Dask Array of shape = [n_samples, n_trees]",
        X_SHAP_values_shape="Dask Array of shape = [n_samples, n_features + 1]"
    )
936

937
    def to_local(self) -> LGBMRanker:
938
939
940
941
942
        """Create regular version of lightgbm.LGBMRanker from the distributed version.

        Returns
        -------
        model : lightgbm.LGBMRanker
943
            Local underlying model.
944
945
        """
        return self._to_local(LGBMRanker)