dask.py 22.1 KB
Newer Older
1
2
3
# coding: utf-8
"""Distributed training with LightGBM and Dask.distributed.

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

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, Dict, Iterable, List, Optional, Type, Union
13
14
15
from urllib.parse import urlparse

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

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

_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]]
28
29


30
31
def _find_open_port(worker_ip: str, local_listen_port: int, ports_to_skip: Iterable[int]) -> int:
    """Find an open port.
32

33
34
    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.
35

36
37
38
39
40
41
42
43
44
45
46
47
48
49
    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
    -------
50
    port : int
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
        A free port on the machine referenced by ``worker_ip``.
    """
    max_tries = 1000
    out_port = None
    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.
79
80
81

    Parameters
    ----------
82
83
84
    client : dask.distributed.Client
        Dask client.
    worker_addresses : Iterable[str]
85
        An iterable of addresses for workers in the cluster. These are strings of the form ``<protocol>://<host>:port``.
86
    local_listen_port : int
87
        First port to try when searching for open ports.
88
89
90

    Returns
    -------
91
92
    result : Dict[str, int]
        Dictionary where keys are worker addresses and values are an open port for LightGBM to use.
93
    """
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    lightgbm_ports = set()
    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
108
109


110
def _concat(seq: List[_DaskPart]) -> _DaskPart:
111
112
    if isinstance(seq[0], np.ndarray):
        return np.concatenate(seq, axis=0)
113
    elif isinstance(seq[0], (pd_DataFrame, pd_Series)):
114
        return concat(seq, axis=0)
115
116
117
118
119
120
    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])))


121
122
123
124
125
126
127
128
129
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]:
130
131
132
133
134
135
136
137
138
139
140
141
    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)
    }
142
143
    params.update(network_params)

144
145
    is_ranker = issubclass(model_factory, LGBMRanker)

146
    # Concatenate many parts into one
147
148
149
150
151
152
153
154
155
156
157
158
    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
159
160
161

    try:
        model = model_factory(**params)
162
        if is_ranker:
163
            model.fit(data, label, sample_weight=weight, group=group, **kwargs)
164
        else:
165
            model.fit(data, label, sample_weight=weight, **kwargs)
166

167
168
169
170
171
172
    finally:
        _safe_call(_LIB.LGBM_NetworkFree())

    return model if return_model else None


173
def _split_to_parts(data: _DaskCollection, is_matrix: bool) -> List[_DaskPart]:
174
175
    parts = data.to_delayed()
    if isinstance(parts, np.ndarray):
176
177
178
179
        if is_matrix:
            assert parts.shape[1] == 1
        else:
            assert parts.ndim == 1 or parts.shape[1] == 1
180
181
182
183
        parts = parts.flatten().tolist()
    return parts


184
185
186
187
188
189
190
191
192
193
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:
194
195
196
197
    """Inner train routine.

    Parameters
    ----------
198
199
    client : dask.distributed.Client
        Dask client.
200
    data : dask Array or dask DataFrame of shape = [n_samples, n_features]
201
        Input feature matrix.
202
    label : dask Array, dask DataFrame or dask Series of shape = [n_samples]
203
204
        The target values (class labels in classification, real numbers in regression).
    params : dict
205
        Parameters passed to constructor of the local underlying model.
206
    model_factory : lightgbm.LGBMClassifier, lightgbm.LGBMRegressor, or lightgbm.LGBMRanker class
207
        Class of the local underlying model.
208
    sample_weight : dask Array, dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)
209
        Weights of training data.
210
    group : dask Array, dask DataFrame, Dask Series of shape = [n_samples] or None, optional (default=None)
211
212
213
214
215
        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.
216
217
218
219
220
221
222
    **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.
223
    """
224
225
    params = deepcopy(params)

226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
    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:
246
        _log_warning('Parameter tree_learner set to %s, which is not allowed. Using "data" as default' % params['tree_learner'])
247
248
249
250
251
252
253
254
255
256
257
258
        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
259
    for param_alias in _ConfigAliases.get('machines', 'num_machines', 'num_threads'):
260
        params.pop(param_alias, None)
261

262
    # Split arrays/dataframes into parts. Arrange parts into dicts to enforce co-locality
263
264
    data_parts = _split_to_parts(data=data, is_matrix=True)
    label_parts = _split_to_parts(data=label, is_matrix=False)
265
    parts = [{'data': x, 'label': y} for (x, y) in zip(data_parts, label_parts)]
266
267
268

    if sample_weight is not None:
        weight_parts = _split_to_parts(data=sample_weight, is_matrix=False)
269
270
        for i in range(len(parts)):
            parts[i]['weight'] = weight_parts[i]
271
272
273

    if group is not None:
        group_parts = _split_to_parts(data=group, is_matrix=False)
274
275
        for i in range(len(parts)):
            parts[i]['group'] = group_parts[i]
276
277

    # Start computation in the background
278
    parts = list(map(delayed, parts))
279
280
281
282
283
284
285
286
    parts = client.compute(parts)
    wait(parts)

    for part in parts:
        if part.status == 'error':
            return part  # trigger error locally

    # Find locations of all parts and map them to particular Dask workers
287
    key_to_part_dict = {part.key: part for part in parts}
288
289
290
291
292
293
294
295
    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()

296
297
298
299
300
301
    # 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(),
302
        local_listen_port=params["local_listen_port"]
303
304
    )

305
    # Tell each worker to train on the parts that it has locally
306
307
308
309
310
311
312
313
314
315
316
317
318
    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()
    ]
319
320
321
322
323
324

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


325
326
327
328
329
330
331
332
333
def _predict_part(
    part: _DaskPart,
    model: LGBMModel,
    raw_score: bool,
    pred_proba: bool,
    pred_leaf: bool,
    pred_contrib: bool,
    **kwargs: Any
) -> _DaskPart:
334
    data = part.values if isinstance(part, pd_DataFrame) else part
335
336
337

    if data.shape[0] == 0:
        result = np.array([])
338
339
340
341
342
343
344
345
    elif pred_proba:
        result = model.predict_proba(
            data,
            raw_score=raw_score,
            pred_leaf=pred_leaf,
            pred_contrib=pred_contrib,
            **kwargs
        )
346
    else:
347
348
349
350
351
352
353
        result = model.predict(
            data,
            raw_score=raw_score,
            pred_leaf=pred_leaf,
            pred_contrib=pred_contrib,
            **kwargs
        )
354

355
    if isinstance(part, pd_DataFrame):
356
        if pred_proba or pred_contrib:
357
            result = pd_DataFrame(result, index=part.index)
358
        else:
359
            result = pd_Series(result, index=part.index, name='predictions')
360
361
362
363

    return result


364
365
366
367
368
369
370
371
372
373
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:
374
375
376
377
    """Inner predict routine.

    Parameters
    ----------
378
    model : lightgbm.LGBMClassifier, lightgbm.LGBMRegressor, or lightgbm.LGBMRanker class
379
        Fitted underlying model.
380
    data : dask Array or dask DataFrame of shape = [n_samples, n_features]
381
        Input feature matrix.
382
383
    raw_score : bool, optional (default=False)
        Whether to predict raw scores.
384
385
386
387
388
389
    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.
390
    dtype : np.dtype, optional (default=np.float32)
391
        Dtype of the output.
392
    **kwargs
393
        Other parameters passed to ``predict`` or ``predict_proba`` method.
394
395
396

    Returns
    -------
397
    predicted_result : dask Array of shape = [n_samples] or shape = [n_samples, n_classes]
398
        The predicted values.
399
    X_leaves : dask Array of shape = [n_samples, n_trees] or shape = [n_samples, n_trees * n_classes]
400
        If ``pred_leaf=True``, the predicted leaf of every tree for each sample.
401
    X_SHAP_values : dask Array of shape = [n_samples, n_features + 1] or shape = [n_samples, (n_features + 1) * n_classes]
402
        If ``pred_contrib=True``, the feature contributions for each sample.
403
    """
404
405
    if not all((DASK_INSTALLED, PANDAS_INSTALLED, SKLEARN_INSTALLED)):
        raise LightGBMError('dask, pandas and scikit-learn are required for lightgbm.dask')
406
    if isinstance(data, dask_DataFrame):
407
408
409
410
411
412
413
414
415
        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
416
    elif isinstance(data, dask_Array):
417
        if pred_proba:
418
419
420
            kwargs['chunks'] = (data.chunks[0], (model.n_classes_,))
        else:
            kwargs['drop_axis'] = 1
421
422
423
424
425
426
427
428
429
430
        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
        )
431
    else:
432
        raise TypeError('Data must be either dask Array or dask DataFrame. Got %s.' % str(type(data)))
433
434


435
class _DaskLGBMModel:
436
437
438
439
440
441
442
443
444
445
446

    def _fit(
        self,
        model_factory: Type[LGBMModel],
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        group: Optional[_DaskCollection] = None,
        client: Optional[Client] = None,
        **kwargs: Any
    ) -> "_DaskLGBMModel":
447
448
        if not all((DASK_INSTALLED, PANDAS_INSTALLED, SKLEARN_INSTALLED)):
            raise LightGBMError('dask, pandas and scikit-learn are required for lightgbm.dask')
449
450
451
452
        if client is None:
            client = default_client()

        params = self.get_params(True)
453
454
455
456
457
458
459
460
461
462
463

        model = _train(
            client=client,
            data=X,
            label=y,
            params=params,
            model_factory=model_factory,
            sample_weight=sample_weight,
            group=group,
            **kwargs
        )
464
465
466
467
468
469

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

        return self

470
    def _to_local(self, model_factory: Type[LGBMModel]) -> LGBMModel:
471
472
473
474
475
        model = model_factory(**self.get_params())
        self._copy_extra_params(self, model)
        return model

    @staticmethod
476
    def _copy_extra_params(source: Union["_DaskLGBMModel", LGBMModel], dest: Union["_DaskLGBMModel", LGBMModel]) -> None:
477
478
479
480
481
482
483
        params = source.get_params()
        attributes = source.__dict__
        extra_param_names = set(attributes.keys()).difference(params.keys())
        for name in extra_param_names:
            setattr(dest, name, attributes[name])


484
class DaskLGBMClassifier(LGBMClassifier, _DaskLGBMModel):
485
486
    """Distributed version of lightgbm.LGBMClassifier."""

487
488
489
490
491
492
493
494
    def fit(
        self,
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        client: Optional[Client] = None,
        **kwargs: Any
    ) -> "DaskLGBMClassifier":
495
        """Docstring is inherited from the lightgbm.LGBMClassifier.fit."""
496
497
498
499
500
501
502
503
504
        return self._fit(
            model_factory=LGBMClassifier,
            X=X,
            y=y,
            sample_weight=sample_weight,
            client=client,
            **kwargs
        )

505
506
507
508
509
510
    _base_doc = LGBMClassifier.fit.__doc__
    _before_init_score, _init_score, _after_init_score = _base_doc.partition('init_score :')
    fit.__doc__ = (_before_init_score
                   + 'client : dask.distributed.Client or None, optional (default=None)\n'
                   + ' ' * 12 + 'Dask client.\n'
                   + ' ' * 8 + _init_score + _after_init_score)
511

512
    def predict(self, X: _DaskMatrixLike, **kwargs: Any) -> dask_Array:
513
        """Docstring is inherited from the lightgbm.LGBMClassifier.predict."""
514
515
516
517
518
519
520
        return _predict(
            model=self.to_local(),
            data=X,
            dtype=self.classes_.dtype,
            **kwargs
        )

521
522
    predict.__doc__ = LGBMClassifier.predict.__doc__

523
    def predict_proba(self, X: _DaskMatrixLike, **kwargs: Any) -> dask_Array:
524
        """Docstring is inherited from the lightgbm.LGBMClassifier.predict_proba."""
525
526
527
528
529
530
531
        return _predict(
            model=self.to_local(),
            data=X,
            pred_proba=True,
            **kwargs
        )

532
533
    predict_proba.__doc__ = LGBMClassifier.predict_proba.__doc__

534
    def to_local(self) -> LGBMClassifier:
535
536
537
538
539
        """Create regular version of lightgbm.LGBMClassifier from the distributed version.

        Returns
        -------
        model : lightgbm.LGBMClassifier
540
            Local underlying model.
541
542
543
544
        """
        return self._to_local(LGBMClassifier)


545
class DaskLGBMRegressor(LGBMRegressor, _DaskLGBMModel):
546
    """Distributed version of lightgbm.LGBMRegressor."""
547

548
549
550
551
552
553
554
555
    def fit(
        self,
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        client: Optional[Client] = None,
        **kwargs: Any
    ) -> "DaskLGBMRegressor":
556
        """Docstring is inherited from the lightgbm.LGBMRegressor.fit."""
557
558
559
560
561
562
563
564
565
        return self._fit(
            model_factory=LGBMRegressor,
            X=X,
            y=y,
            sample_weight=sample_weight,
            client=client,
            **kwargs
        )

566
567
568
569
570
571
    _base_doc = LGBMRegressor.fit.__doc__
    _before_init_score, _init_score, _after_init_score = _base_doc.partition('init_score :')
    fit.__doc__ = (_before_init_score
                   + 'client : dask.distributed.Client or None, optional (default=None)\n'
                   + ' ' * 12 + 'Dask client.\n'
                   + ' ' * 8 + _init_score + _after_init_score)
572

573
    def predict(self, X: _DaskMatrixLike, **kwargs) -> dask_Array:
574
        """Docstring is inherited from the lightgbm.LGBMRegressor.predict."""
575
576
577
578
579
580
        return _predict(
            model=self.to_local(),
            data=X,
            **kwargs
        )

581
582
    predict.__doc__ = LGBMRegressor.predict.__doc__

583
    def to_local(self) -> LGBMRegressor:
584
585
586
587
588
        """Create regular version of lightgbm.LGBMRegressor from the distributed version.

        Returns
        -------
        model : lightgbm.LGBMRegressor
589
            Local underlying model.
590
591
        """
        return self._to_local(LGBMRegressor)
592
593


594
class DaskLGBMRanker(LGBMRanker, _DaskLGBMModel):
595
    """Distributed version of lightgbm.LGBMRanker."""
596

597
598
599
600
601
602
603
604
605
606
    def fit(
        self,
        X: _DaskMatrixLike,
        y: _DaskCollection,
        sample_weight: Optional[_DaskCollection] = None,
        init_score: Optional[_DaskCollection] = None,
        group: Optional[_DaskCollection] = None,
        client: Optional[Client] = None,
        **kwargs: Any
    ) -> "DaskLGBMRanker":
607
608
609
610
        """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')

611
612
613
614
615
616
617
618
619
620
        return self._fit(
            model_factory=LGBMRanker,
            X=X,
            y=y,
            sample_weight=sample_weight,
            group=group,
            client=client,
            **kwargs
        )

621
622
623
624
625
626
    _base_doc = LGBMRanker.fit.__doc__
    _before_eval_set, _eval_set, _after_eval_set = _base_doc.partition('eval_set :')
    fit.__doc__ = (_before_eval_set
                   + 'client : dask.distributed.Client or None, optional (default=None)\n'
                   + ' ' * 12 + 'Dask client.\n'
                   + ' ' * 8 + _eval_set + _after_eval_set)
627

628
    def predict(self, X: _DaskMatrixLike, **kwargs: Any) -> dask_Array:
629
630
        """Docstring is inherited from the lightgbm.LGBMRanker.predict."""
        return _predict(self.to_local(), X, **kwargs)
631

632
633
    predict.__doc__ = LGBMRanker.predict.__doc__

634
    def to_local(self) -> LGBMRanker:
635
636
637
638
639
        """Create regular version of lightgbm.LGBMRanker from the distributed version.

        Returns
        -------
        model : lightgbm.LGBMRanker
640
            Local underlying model.
641
642
        """
        return self._to_local(LGBMRanker)