basic.py 169 KB
Newer Older
wxchan's avatar
wxchan committed
1
# coding: utf-8
2
"""Wrapper for C API of LightGBM."""
3
import abc
wxchan's avatar
wxchan committed
4
import ctypes
5
import json
wxchan's avatar
wxchan committed
6
import warnings
7
from collections import OrderedDict
8
from copy import deepcopy
9
from enum import Enum
10
from functools import wraps
11
from os import SEEK_END, environ
12
13
from os.path import getsize
from pathlib import Path
14
from tempfile import NamedTemporaryFile
15
from typing import Any, Callable, Dict, Iterable, List, Optional, Set, Tuple, Union
wxchan's avatar
wxchan committed
16
17
18
19

import numpy as np
import scipy.sparse

20
from .compat import PANDAS_INSTALLED, concat, dt_DataTable, pd_CategoricalDtype, pd_DataFrame, pd_Series
wxchan's avatar
wxchan committed
21
22
from .libpath import find_lib_path

23
24
25
_LGBM_EvalFunctionResultType = Tuple[str, float, bool]
_LGBM_BoosterEvalMethodResultType = Tuple[str, str, float, bool]

26
27
28
ZERO_THRESHOLD = 1e-35


29
30
31
32
def _is_zero(x: float) -> bool:
    return -ZERO_THRESHOLD <= x <= ZERO_THRESHOLD


33
def _get_sample_count(total_nrow: int, params: str) -> int:
34
35
36
37
38
39
40
41
    sample_cnt = ctypes.c_int(0)
    _safe_call(_LIB.LGBM_GetSampleCount(
        ctypes.c_int32(total_nrow),
        c_str(params),
        ctypes.byref(sample_cnt),
    ))
    return sample_cnt.value

wxchan's avatar
wxchan committed
42

43
44
45
46
47
48
class _MissingType(Enum):
    NONE = 'None'
    NAN = 'NaN'
    ZERO = 'Zero'


49
class _DummyLogger:
50
    def info(self, msg: str) -> None:
51
52
        print(msg)

53
    def warning(self, msg: str) -> None:
54
55
56
        warnings.warn(msg, stacklevel=3)


57
58
59
_LOGGER: Any = _DummyLogger()
_INFO_METHOD_NAME = "info"
_WARNING_METHOD_NAME = "warning"
60
61


62
63
64
def register_logger(
    logger: Any, info_method_name: str = "info", warning_method_name: str = "warning"
) -> None:
65
66
67
68
    """Register custom logger.

    Parameters
    ----------
69
    logger : Any
70
        Custom logger.
71
72
73
74
    info_method_name : str, optional (default="info")
        Method used to log info messages.
    warning_method_name : str, optional (default="warning")
        Method used to log warning messages.
75
    """
76
77
78
79
80
81
82
83
84
    def _has_method(logger: Any, method_name: str) -> bool:
        return callable(getattr(logger, method_name, None))

    if not _has_method(logger, info_method_name) or not _has_method(logger, warning_method_name):
        raise TypeError(
            f"Logger must provide '{info_method_name}' and '{warning_method_name}' method"
        )

    global _LOGGER, _INFO_METHOD_NAME, _WARNING_METHOD_NAME
85
    _LOGGER = logger
86
87
    _INFO_METHOD_NAME = info_method_name
    _WARNING_METHOD_NAME = warning_method_name
88
89


90
def _normalize_native_string(func: Callable[[str], None]) -> Callable[[str], None]:
91
    """Join log messages from native library which come by chunks."""
92
    msg_normalized: List[str] = []
93
94

    @wraps(func)
95
    def wrapper(msg: str) -> None:
96
97
98
99
100
101
102
103
104
105
106
        nonlocal msg_normalized
        if msg.strip() == '':
            msg = ''.join(msg_normalized)
            msg_normalized = []
            return func(msg)
        else:
            msg_normalized.append(msg)

    return wrapper


107
def _log_info(msg: str) -> None:
108
    getattr(_LOGGER, _INFO_METHOD_NAME)(msg)
109
110


111
def _log_warning(msg: str) -> None:
112
    getattr(_LOGGER, _WARNING_METHOD_NAME)(msg)
113
114
115


@_normalize_native_string
116
def _log_native(msg: str) -> None:
117
    getattr(_LOGGER, _INFO_METHOD_NAME)(msg)
118
119


120
def _log_callback(msg: bytes) -> None:
121
    """Redirect logs from native library into Python."""
122
    _log_native(str(msg.decode('utf-8')))
123
124


125
def _load_lib() -> ctypes.CDLL:
126
    """Load LightGBM library."""
wxchan's avatar
wxchan committed
127
128
129
    lib_path = find_lib_path()
    lib = ctypes.cdll.LoadLibrary(lib_path[0])
    lib.LGBM_GetLastError.restype = ctypes.c_char_p
130
131
132
    callback = ctypes.CFUNCTYPE(None, ctypes.c_char_p)
    lib.callback = callback(_log_callback)
    if lib.LGBM_RegisterLogCallback(lib.callback) != 0:
133
        raise LightGBMError(lib.LGBM_GetLastError().decode('utf-8'))
wxchan's avatar
wxchan committed
134
135
    return lib

wxchan's avatar
wxchan committed
136

137
138
139
140
141
142
143
# we don't need lib_lightgbm while building docs
_LIB: ctypes.CDLL
if environ.get('LIGHTGBM_BUILD_DOC', False):
    from unittest.mock import Mock  # isort: skip
    _LIB = Mock(ctypes.CDLL)  # type: ignore
else:
    _LIB = _load_lib()
wxchan's avatar
wxchan committed
144

wxchan's avatar
wxchan committed
145

146
_NUMERIC_TYPES = (int, float, bool)
147
_ArrayLike = Union[List, np.ndarray, pd_Series]
148
149


150
def _safe_call(ret: int) -> None:
151
152
    """Check the return value from C API call.

wxchan's avatar
wxchan committed
153
154
155
    Parameters
    ----------
    ret : int
156
        The return value from C API calls.
wxchan's avatar
wxchan committed
157
158
    """
    if ret != 0:
159
        raise LightGBMError(_LIB.LGBM_GetLastError().decode('utf-8'))
wxchan's avatar
wxchan committed
160

wxchan's avatar
wxchan committed
161

162
def _is_numeric(obj: Any) -> bool:
163
    """Check whether object is a number or not, include numpy number, etc."""
wxchan's avatar
wxchan committed
164
165
166
    try:
        float(obj)
        return True
wxchan's avatar
wxchan committed
167
168
169
    except (TypeError, ValueError):
        # TypeError: obj is not a string or a number
        # ValueError: invalid literal
wxchan's avatar
wxchan committed
170
171
        return False

wxchan's avatar
wxchan committed
172

173
def is_numpy_1d_array(data: Any) -> bool:
174
    """Check whether data is a numpy 1-D array."""
175
    return isinstance(data, np.ndarray) and len(data.shape) == 1
wxchan's avatar
wxchan committed
176

wxchan's avatar
wxchan committed
177

178
def is_numpy_column_array(data: Any) -> bool:
179
180
181
182
183
184
185
    """Check whether data is a column numpy array."""
    if not isinstance(data, np.ndarray):
        return False
    shape = data.shape
    return len(shape) == 2 and shape[1] == 1


186
187
def cast_numpy_array_to_dtype(array, dtype):
    """Cast numpy array to given dtype."""
188
189
190
191
192
    if array.dtype == dtype:
        return array
    return array.astype(dtype=dtype, copy=False)


193
def is_1d_list(data: Any) -> bool:
194
    """Check whether data is a 1-D list."""
195
    return isinstance(data, list) and (not data or _is_numeric(data[0]))
wxchan's avatar
wxchan committed
196

wxchan's avatar
wxchan committed
197

198
199
200
201
202
203
204
205
206
207
def _is_1d_collection(data: Any) -> bool:
    """Check whether data is a 1-D collection."""
    return (
        is_numpy_1d_array(data)
        or is_numpy_column_array(data)
        or is_1d_list(data)
        or isinstance(data, pd_Series)
    )


208
def list_to_1d_numpy(data, dtype=np.float32, name='list'):
209
    """Convert data to numpy 1-D array."""
wxchan's avatar
wxchan committed
210
    if is_numpy_1d_array(data):
211
        return cast_numpy_array_to_dtype(data, dtype)
212
213
214
    elif is_numpy_column_array(data):
        _log_warning('Converting column-vector to 1d array')
        array = data.ravel()
215
        return cast_numpy_array_to_dtype(array, dtype)
wxchan's avatar
wxchan committed
216
217
    elif is_1d_list(data):
        return np.array(data, dtype=dtype, copy=False)
218
    elif isinstance(data, pd_Series):
219
        _check_for_bad_pandas_dtypes(data.to_frame().dtypes)
220
        return np.array(data, dtype=dtype, copy=False)  # SparseArray should be supported as well
wxchan's avatar
wxchan committed
221
    else:
222
223
        raise TypeError(f"Wrong type({type(data).__name__}) for {name}.\n"
                        "It should be list, numpy 1-D array or pandas Series")
wxchan's avatar
wxchan committed
224

wxchan's avatar
wxchan committed
225

226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
def _is_numpy_2d_array(data: Any) -> bool:
    """Check whether data is a numpy 2-D array."""
    return isinstance(data, np.ndarray) and len(data.shape) == 2 and data.shape[1] > 1


def _is_2d_list(data: Any) -> bool:
    """Check whether data is a 2-D list."""
    return isinstance(data, list) and len(data) > 0 and is_1d_list(data[0])


def _is_2d_collection(data: Any) -> bool:
    """Check whether data is a 2-D collection."""
    return (
        _is_numpy_2d_array(data)
        or _is_2d_list(data)
        or isinstance(data, pd_DataFrame)
    )


def _data_to_2d_numpy(data: Any, dtype: type = np.float32, name: str = 'list') -> np.ndarray:
    """Convert data to numpy 2-D array."""
    if _is_numpy_2d_array(data):
        return cast_numpy_array_to_dtype(data, dtype)
    if _is_2d_list(data):
        return np.array(data, dtype=dtype)
    if isinstance(data, pd_DataFrame):
252
        _check_for_bad_pandas_dtypes(data.dtypes)
253
254
255
256
257
        return cast_numpy_array_to_dtype(data.values, dtype)
    raise TypeError(f"Wrong type({type(data).__name__}) for {name}.\n"
                    "It should be list of lists, numpy 2-D array or pandas DataFrame")


258
def cfloat32_array_to_numpy(cptr: ctypes.POINTER, length: int) -> np.ndarray:
259
    """Convert a ctypes float pointer array to a numpy array."""
wxchan's avatar
wxchan committed
260
    if isinstance(cptr, ctypes.POINTER(ctypes.c_float)):
261
        return np.ctypeslib.as_array(cptr, shape=(length,)).copy()
wxchan's avatar
wxchan committed
262
    else:
263
        raise RuntimeError('Expected float pointer')
wxchan's avatar
wxchan committed
264

Guolin Ke's avatar
Guolin Ke committed
265

266
def cfloat64_array_to_numpy(cptr: ctypes.POINTER, length: int) -> np.ndarray:
267
    """Convert a ctypes double pointer array to a numpy array."""
Guolin Ke's avatar
Guolin Ke committed
268
    if isinstance(cptr, ctypes.POINTER(ctypes.c_double)):
269
        return np.ctypeslib.as_array(cptr, shape=(length,)).copy()
Guolin Ke's avatar
Guolin Ke committed
270
271
272
    else:
        raise RuntimeError('Expected double pointer')

wxchan's avatar
wxchan committed
273

274
def cint32_array_to_numpy(cptr: ctypes.POINTER, length: int) -> np.ndarray:
275
    """Convert a ctypes int pointer array to a numpy array."""
wxchan's avatar
wxchan committed
276
    if isinstance(cptr, ctypes.POINTER(ctypes.c_int32)):
277
        return np.ctypeslib.as_array(cptr, shape=(length,)).copy()
wxchan's avatar
wxchan committed
278
    else:
279
280
281
        raise RuntimeError('Expected int32 pointer')


282
def cint64_array_to_numpy(cptr: ctypes.POINTER, length: int) -> np.ndarray:
283
284
    """Convert a ctypes int pointer array to a numpy array."""
    if isinstance(cptr, ctypes.POINTER(ctypes.c_int64)):
285
        return np.ctypeslib.as_array(cptr, shape=(length,)).copy()
286
287
    else:
        raise RuntimeError('Expected int64 pointer')
wxchan's avatar
wxchan committed
288

wxchan's avatar
wxchan committed
289

290
def c_str(string: str) -> ctypes.c_char_p:
291
    """Convert a Python string to C string."""
wxchan's avatar
wxchan committed
292
293
    return ctypes.c_char_p(string.encode('utf-8'))

wxchan's avatar
wxchan committed
294

295
def c_array(ctype: type, values: List[Any]) -> ctypes.Array:
296
    """Convert a Python array to C array."""
wxchan's avatar
wxchan committed
297
298
    return (ctype * len(values))(*values)

wxchan's avatar
wxchan committed
299

300
def json_default_with_numpy(obj: Any) -> Any:
301
302
303
304
305
306
307
308
309
    """Convert numpy classes to JSON serializable objects."""
    if isinstance(obj, (np.integer, np.floating, np.bool_)):
        return obj.item()
    elif isinstance(obj, np.ndarray):
        return obj.tolist()
    else:
        return obj


310
def param_dict_to_str(data: Optional[Dict[str, Any]]) -> str:
311
    """Convert Python dictionary to string, which is passed to C API."""
312
    if data is None or not data:
wxchan's avatar
wxchan committed
313
314
315
        return ""
    pairs = []
    for key, val in data.items():
316
        if isinstance(val, (list, tuple, set)) or is_numpy_1d_array(val):
317
318
            def to_string(x):
                if isinstance(x, list):
319
                    return f"[{','.join(map(str, x))}]"
320
321
                else:
                    return str(x)
322
            pairs.append(f"{key}={','.join(map(to_string, val))}")
323
        elif isinstance(val, (str, Path, _NUMERIC_TYPES)) or _is_numeric(val):
324
            pairs.append(f"{key}={val}")
325
        elif val is not None:
326
            raise TypeError(f'Unknown type of parameter:{key}, got:{type(val).__name__}')
wxchan's avatar
wxchan committed
327
    return ' '.join(pairs)
328

wxchan's avatar
wxchan committed
329

330
class _TempFile:
331
332
    """Proxy class to workaround errors on Windows."""

333
334
335
    def __enter__(self):
        with NamedTemporaryFile(prefix="lightgbm_tmp_", delete=True) as f:
            self.name = f.name
336
            self.path = Path(self.name)
337
        return self
wxchan's avatar
wxchan committed
338

339
    def __exit__(self, exc_type, exc_val, exc_tb):
340
341
        if self.path.is_file():
            self.path.unlink()
342

wxchan's avatar
wxchan committed
343

344
class LightGBMError(Exception):
345
346
    """Error thrown by LightGBM."""

347
348
349
    pass


350
351
352
353
354
355
356
357
# DeprecationWarning is not shown by default, so let's create our own with higher level
class LGBMDeprecationWarning(UserWarning):
    """Custom deprecation warning."""

    pass


class _ConfigAliases:
358
359
360
361
    # lazy evaluation to allow import without dynamic library, e.g., for docs generation
    aliases = None

    @staticmethod
362
    def _get_all_param_aliases() -> Dict[str, List[str]]:
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
        buffer_len = 1 << 20
        tmp_out_len = ctypes.c_int64(0)
        string_buffer = ctypes.create_string_buffer(buffer_len)
        ptr_string_buffer = ctypes.c_char_p(*[ctypes.addressof(string_buffer)])
        _safe_call(_LIB.LGBM_DumpParamAliases(
            ctypes.c_int64(buffer_len),
            ctypes.byref(tmp_out_len),
            ptr_string_buffer))
        actual_len = tmp_out_len.value
        # if buffer length is not long enough, re-allocate a buffer
        if actual_len > buffer_len:
            string_buffer = ctypes.create_string_buffer(actual_len)
            ptr_string_buffer = ctypes.c_char_p(*[ctypes.addressof(string_buffer)])
            _safe_call(_LIB.LGBM_DumpParamAliases(
                ctypes.c_int64(actual_len),
                ctypes.byref(tmp_out_len),
                ptr_string_buffer))
        aliases = json.loads(
            string_buffer.value.decode('utf-8'),
382
            object_hook=lambda obj: {k: [k] + v for k, v in obj.items()}
383
384
        )
        return aliases
385
386

    @classmethod
387
388
389
    def get(cls, *args) -> Set[str]:
        if cls.aliases is None:
            cls.aliases = cls._get_all_param_aliases()
390
391
        ret = set()
        for i in args:
392
            ret.update(cls.get_sorted(i))
393
394
        return ret

395
396
397
398
399
400
    @classmethod
    def get_sorted(cls, name: str) -> List[str]:
        if cls.aliases is None:
            cls.aliases = cls._get_all_param_aliases()
        return cls.aliases.get(name, [name])

401
    @classmethod
402
403
404
    def get_by_alias(cls, *args) -> Set[str]:
        if cls.aliases is None:
            cls.aliases = cls._get_all_param_aliases()
405
406
407
408
        ret = set(args)
        for arg in args:
            for aliases in cls.aliases.values():
                if arg in aliases:
409
                    ret.update(aliases)
410
411
412
                    break
        return ret

413

414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
def _choose_param_value(main_param_name: str, params: Dict[str, Any], default_value: Any) -> Dict[str, Any]:
    """Get a single parameter value, accounting for aliases.

    Parameters
    ----------
    main_param_name : str
        Name of the main parameter to get a value for. One of the keys of ``_ConfigAliases``.
    params : dict
        Dictionary of LightGBM parameters.
    default_value : Any
        Default value to use for the parameter, if none is found in ``params``.

    Returns
    -------
    params : dict
        A ``params`` dict with exactly one value for ``main_param_name``, and all aliases ``main_param_name`` removed.
        If both ``main_param_name`` and one or more aliases for it are found, the value of ``main_param_name`` will be preferred.
    """
    # avoid side effects on passed-in parameters
    params = deepcopy(params)

435
436
    aliases = _ConfigAliases.get_sorted(main_param_name)
    aliases = [a for a in aliases if a != main_param_name]
437
438

    # if main_param_name was provided, keep that value and remove all aliases
439
    if main_param_name in params.keys():
440
441
442
        for param in aliases:
            params.pop(param, None)
        return params
443

444
445
446
447
448
    # if main param name was not found, search for an alias
    for param in aliases:
        if param in params.keys():
            params[main_param_name] = params[param]
            break
449

450
451
452
453
454
455
456
    if main_param_name in params.keys():
        for param in aliases:
            params.pop(param, None)
        return params

    # neither of main_param_name, aliases were found
    params[main_param_name] = default_value
457
458
459
460

    return params


461
462
MAX_INT32 = (1 << 31) - 1

463
"""Macro definition of data type in C API of LightGBM"""
wxchan's avatar
wxchan committed
464
465
466
467
C_API_DTYPE_FLOAT32 = 0
C_API_DTYPE_FLOAT64 = 1
C_API_DTYPE_INT32 = 2
C_API_DTYPE_INT64 = 3
Guolin Ke's avatar
Guolin Ke committed
468

469
"""Matrix is row major in Python"""
wxchan's avatar
wxchan committed
470
471
C_API_IS_ROW_MAJOR = 1

472
"""Macro definition of prediction type in C API of LightGBM"""
wxchan's avatar
wxchan committed
473
474
475
C_API_PREDICT_NORMAL = 0
C_API_PREDICT_RAW_SCORE = 1
C_API_PREDICT_LEAF_INDEX = 2
476
C_API_PREDICT_CONTRIB = 3
wxchan's avatar
wxchan committed
477

478
479
480
481
"""Macro definition of sparse matrix type"""
C_API_MATRIX_TYPE_CSR = 0
C_API_MATRIX_TYPE_CSC = 1

482
483
484
485
"""Macro definition of feature importance type"""
C_API_FEATURE_IMPORTANCE_SPLIT = 0
C_API_FEATURE_IMPORTANCE_GAIN = 1

486
"""Data type of data field"""
wxchan's avatar
wxchan committed
487
488
FIELD_TYPE_MAPPER = {"label": C_API_DTYPE_FLOAT32,
                     "weight": C_API_DTYPE_FLOAT32,
Guolin Ke's avatar
Guolin Ke committed
489
                     "init_score": C_API_DTYPE_FLOAT64,
490
                     "group": C_API_DTYPE_INT32}
wxchan's avatar
wxchan committed
491

492
493
494
495
"""String name to int feature importance type mapper"""
FEATURE_IMPORTANCE_TYPE_MAPPER = {"split": C_API_FEATURE_IMPORTANCE_SPLIT,
                                  "gain": C_API_FEATURE_IMPORTANCE_GAIN}

wxchan's avatar
wxchan committed
496

497
def convert_from_sliced_object(data):
498
    """Fix the memory of multi-dimensional sliced object."""
499
    if isinstance(data, np.ndarray) and isinstance(data.base, np.ndarray):
500
        if not data.flags.c_contiguous:
501
502
            _log_warning("Usage of np.ndarray subset (sliced data) is not recommended "
                         "due to it will double the peak memory cost in LightGBM.")
503
504
505
506
            return np.copy(data)
    return data


wxchan's avatar
wxchan committed
507
def c_float_array(data):
508
    """Get pointer of float numpy array / list."""
wxchan's avatar
wxchan committed
509
510
511
    if is_1d_list(data):
        data = np.array(data, copy=False)
    if is_numpy_1d_array(data):
512
513
        data = convert_from_sliced_object(data)
        assert data.flags.c_contiguous
wxchan's avatar
wxchan committed
514
515
516
517
518
519
520
        if data.dtype == np.float32:
            ptr_data = data.ctypes.data_as(ctypes.POINTER(ctypes.c_float))
            type_data = C_API_DTYPE_FLOAT32
        elif data.dtype == np.float64:
            ptr_data = data.ctypes.data_as(ctypes.POINTER(ctypes.c_double))
            type_data = C_API_DTYPE_FLOAT64
        else:
521
            raise TypeError(f"Expected np.float32 or np.float64, met type({data.dtype})")
wxchan's avatar
wxchan committed
522
    else:
523
        raise TypeError(f"Unknown type({type(data).__name__})")
524
    return (ptr_data, type_data, data)  # return `data` to avoid the temporary copy is freed
wxchan's avatar
wxchan committed
525

wxchan's avatar
wxchan committed
526

wxchan's avatar
wxchan committed
527
def c_int_array(data):
528
    """Get pointer of int numpy array / list."""
wxchan's avatar
wxchan committed
529
530
531
    if is_1d_list(data):
        data = np.array(data, copy=False)
    if is_numpy_1d_array(data):
532
533
        data = convert_from_sliced_object(data)
        assert data.flags.c_contiguous
wxchan's avatar
wxchan committed
534
535
536
537
538
539
540
        if data.dtype == np.int32:
            ptr_data = data.ctypes.data_as(ctypes.POINTER(ctypes.c_int32))
            type_data = C_API_DTYPE_INT32
        elif data.dtype == np.int64:
            ptr_data = data.ctypes.data_as(ctypes.POINTER(ctypes.c_int64))
            type_data = C_API_DTYPE_INT64
        else:
541
            raise TypeError(f"Expected np.int32 or np.int64, met type({data.dtype})")
wxchan's avatar
wxchan committed
542
    else:
543
        raise TypeError(f"Unknown type({type(data).__name__})")
544
    return (ptr_data, type_data, data)  # return `data` to avoid the temporary copy is freed
wxchan's avatar
wxchan committed
545

wxchan's avatar
wxchan committed
546

547
def _check_for_bad_pandas_dtypes(pandas_dtypes_series):
548
549
550
551
552
553
554
555
    float128 = getattr(np, 'float128', type(None))

    def is_allowed_numpy_dtype(dtype):
        return (
            issubclass(dtype, (np.integer, np.floating, np.bool_))
            and not issubclass(dtype, (np.timedelta64, float128))
        )

556
557
558
559
560
561
562
563
    bad_pandas_dtypes = [
        f'{column_name}: {pandas_dtype}'
        for column_name, pandas_dtype in pandas_dtypes_series.iteritems()
        if not is_allowed_numpy_dtype(pandas_dtype.type)
    ]
    if bad_pandas_dtypes:
        raise ValueError('pandas dtypes must be int, float or bool.\n'
                         f'Fields with bad pandas dtypes: {", ".join(bad_pandas_dtypes)}')
564
565


566
def _data_from_pandas(data, feature_name, categorical_feature, pandas_categorical):
567
    if isinstance(data, pd_DataFrame):
568
569
        if len(data.shape) != 2 or data.shape[0] < 1:
            raise ValueError('Input data must be 2 dimensional and non empty.')
570
        if feature_name == 'auto' or feature_name is None:
571
            data = data.rename(columns=str, copy=False)
572
        cat_cols = [col for col, dtype in zip(data.columns, data.dtypes) if isinstance(dtype, pd_CategoricalDtype)]
573
        cat_cols_not_ordered = [col for col in cat_cols if not data[col].cat.ordered]
574
575
576
577
578
        if pandas_categorical is None:  # train dataset
            pandas_categorical = [list(data[col].cat.categories) for col in cat_cols]
        else:
            if len(cat_cols) != len(pandas_categorical):
                raise ValueError('train and valid dataset categorical_feature do not match.')
579
            for col, category in zip(cat_cols, pandas_categorical):
580
581
                if list(data[col].cat.categories) != list(category):
                    data[col] = data[col].cat.set_categories(category)
582
        if len(cat_cols):  # cat_cols is list
583
            data = data.copy(deep=False)  # not alter origin DataFrame
584
            data[cat_cols] = data[cat_cols].apply(lambda x: x.cat.codes).replace({-1: np.nan})
585
586
587
        if categorical_feature is not None:
            if feature_name is None:
                feature_name = list(data.columns)
588
            if categorical_feature == 'auto':  # use cat cols from DataFrame
589
                categorical_feature = cat_cols_not_ordered
590
591
            else:  # use cat cols specified by user
                categorical_feature = list(categorical_feature)
592
593
        if feature_name == 'auto':
            feature_name = list(data.columns)
594
        _check_for_bad_pandas_dtypes(data.dtypes)
595
596
597
598
        df_dtypes = [dtype.type for dtype in data.dtypes]
        df_dtypes.append(np.float32)  # so that the target dtype considers floats
        target_dtype = np.find_common_type(df_dtypes, [])
        data = data.astype(target_dtype, copy=False).values
599
600
601
602
603
604
    else:
        if feature_name == 'auto':
            feature_name = None
        if categorical_feature == 'auto':
            categorical_feature = None
    return data, feature_name, categorical_feature, pandas_categorical
605
606
607


def _label_from_pandas(label):
608
    if isinstance(label, pd_DataFrame):
609
610
        if len(label.columns) > 1:
            raise ValueError('DataFrame for label cannot have multiple columns')
611
        _check_for_bad_pandas_dtypes(label.dtypes)
612
        label = np.ravel(label.values.astype(np.float32, copy=False))
613
614
615
    return label


616
def _dump_pandas_categorical(pandas_categorical, file_name=None):
617
618
    categorical_json = json.dumps(pandas_categorical, default=json_default_with_numpy)
    pandas_str = f'\npandas_categorical:{categorical_json}\n'
619
620
621
622
623
624
    if file_name is not None:
        with open(file_name, 'a') as f:
            f.write(pandas_str)
    return pandas_str


625
626
627
628
def _load_pandas_categorical(
    file_name: Optional[Union[str, Path]] = None,
    model_str: Optional[str] = None
) -> Optional[str]:
629
630
    pandas_key = 'pandas_categorical:'
    offset = -len(pandas_key)
631
    if file_name is not None:
632
        max_offset = -getsize(file_name)
633
634
635
636
        with open(file_name, 'rb') as f:
            while True:
                if offset < max_offset:
                    offset = max_offset
637
                f.seek(offset, SEEK_END)
638
639
640
641
                lines = f.readlines()
                if len(lines) >= 2:
                    break
                offset *= 2
642
        last_line = lines[-1].decode('utf-8').strip()
643
        if not last_line.startswith(pandas_key):
644
            last_line = lines[-2].decode('utf-8').strip()
645
    elif model_str is not None:
646
647
648
649
650
651
        idx = model_str.rfind('\n', 0, offset)
        last_line = model_str[idx:].strip()
    if last_line.startswith(pandas_key):
        return json.loads(last_line[len(pandas_key):])
    else:
        return None
652
653


654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
class Sequence(abc.ABC):
    """
    Generic data access interface.

    Object should support the following operations:

    .. code-block::

        # Get total row number.
        >>> len(seq)
        # Random access by row index. Used for data sampling.
        >>> seq[10]
        # Range data access. Used to read data in batch when constructing Dataset.
        >>> seq[0:100]
        # Optionally specify batch_size to control range data read size.
        >>> seq.batch_size

    - With random access, **data sampling does not need to go through all data**.
    - With range data access, there's **no need to read all data into memory thus reduce memory usage**.

674
675
    .. versionadded:: 3.3.0

676
677
678
679
680
681
682
683
684
    Attributes
    ----------
    batch_size : int
        Default size of a batch.
    """

    batch_size = 4096  # Defaults to read 4K rows in each batch.

    @abc.abstractmethod
685
    def __getitem__(self, idx: Union[int, slice, List[int]]) -> np.ndarray:
686
687
688
689
690
691
692
        """Return data for given row index.

        A basic implementation should look like this:

        .. code-block:: python

            if isinstance(idx, numbers.Integral):
693
                return self._get_one_line(idx)
694
            elif isinstance(idx, slice):
695
696
                return np.stack([self._get_one_line(i) for i in range(idx.start, idx.stop)])
            elif isinstance(idx, list):
697
                # Only required if using ``Dataset.subset()``.
698
                return np.array([self._get_one_line(i) for i in idx])
699
            else:
700
                raise TypeError(f"Sequence index must be integer, slice or list, got {type(idx).__name__}")
701
702
703

        Parameters
        ----------
704
        idx : int, slice[int], list[int]
705
706
707
708
            Item index.

        Returns
        -------
709
        result : numpy 1-D array or numpy 2-D array
710
            1-D array if idx is int, 2-D array if idx is slice or list.
711
712
713
714
715
716
717
718
719
        """
        raise NotImplementedError("Sub-classes of lightgbm.Sequence must implement __getitem__()")

    @abc.abstractmethod
    def __len__(self) -> int:
        """Return row count of this sequence."""
        raise NotImplementedError("Sub-classes of lightgbm.Sequence must implement __len__()")


720
class _InnerPredictor:
721
722
723
724
725
    """_InnerPredictor of LightGBM.

    Not exposed to user.
    Used only for prediction, usually used for continued training.

Nikita Titov's avatar
Nikita Titov committed
726
727
728
    .. note::

        Can be converted from Booster, but cannot be converted to Booster.
Guolin Ke's avatar
Guolin Ke committed
729
    """
730

731
732
733
734
735
736
    def __init__(
        self,
        model_file: Optional[Union[str, Path]] = None,
        booster_handle: Optional[ctypes.c_void_p] = None,
        pred_parameter: Optional[Dict[str, Any]] = None
    ):
737
        """Initialize the _InnerPredictor.
wxchan's avatar
wxchan committed
738
739
740

        Parameters
        ----------
741
        model_file : str, pathlib.Path or None, optional (default=None)
wxchan's avatar
wxchan committed
742
            Path to the model file.
743
744
745
        booster_handle : object or None, optional (default=None)
            Handle of Booster.
        pred_parameter: dict or None, optional (default=None)
746
            Other parameters for the prediction.
wxchan's avatar
wxchan committed
747
748
749
750
751
        """
        self.handle = ctypes.c_void_p()
        self.__is_manage_handle = True
        if model_file is not None:
            """Prediction task"""
Guolin Ke's avatar
Guolin Ke committed
752
            out_num_iterations = ctypes.c_int(0)
wxchan's avatar
wxchan committed
753
            _safe_call(_LIB.LGBM_BoosterCreateFromModelfile(
754
                c_str(str(model_file)),
wxchan's avatar
wxchan committed
755
756
                ctypes.byref(out_num_iterations),
                ctypes.byref(self.handle)))
Guolin Ke's avatar
Guolin Ke committed
757
            out_num_class = ctypes.c_int(0)
wxchan's avatar
wxchan committed
758
759
760
761
            _safe_call(_LIB.LGBM_BoosterGetNumClasses(
                self.handle,
                ctypes.byref(out_num_class)))
            self.num_class = out_num_class.value
762
            self.num_total_iteration = out_num_iterations.value
763
            self.pandas_categorical = _load_pandas_categorical(file_name=model_file)
wxchan's avatar
wxchan committed
764
        elif booster_handle is not None:
Guolin Ke's avatar
Guolin Ke committed
765
            self.__is_manage_handle = False
wxchan's avatar
wxchan committed
766
            self.handle = booster_handle
Guolin Ke's avatar
Guolin Ke committed
767
            out_num_class = ctypes.c_int(0)
wxchan's avatar
wxchan committed
768
769
770
771
            _safe_call(_LIB.LGBM_BoosterGetNumClasses(
                self.handle,
                ctypes.byref(out_num_class)))
            self.num_class = out_num_class.value
772
            self.num_total_iteration = self.current_iteration()
773
            self.pandas_categorical = None
wxchan's avatar
wxchan committed
774
        else:
775
            raise TypeError('Need model_file or booster_handle to create a predictor')
wxchan's avatar
wxchan committed
776

777
778
        pred_parameter = {} if pred_parameter is None else pred_parameter
        self.pred_parameter = param_dict_to_str(pred_parameter)
cbecker's avatar
cbecker committed
779

780
    def __del__(self) -> None:
781
782
783
784
785
        try:
            if self.__is_manage_handle:
                _safe_call(_LIB.LGBM_BoosterFree(self.handle))
        except AttributeError:
            pass
wxchan's avatar
wxchan committed
786

787
    def __getstate__(self) -> Dict[str, Any]:
788
789
790
791
        this = self.__dict__.copy()
        this.pop('handle', None)
        return this

792
793
794
795
796
797
798
799
800
801
802
    def predict(
        self,
        data,
        start_iteration: int = 0,
        num_iteration: int = -1,
        raw_score: bool = False,
        pred_leaf: bool = False,
        pred_contrib: bool = False,
        data_has_header: bool = False,
        validate_features: bool = False
    ):
803
        """Predict logic.
wxchan's avatar
wxchan committed
804
805
806

        Parameters
        ----------
807
        data : str, pathlib.Path, numpy array, pandas DataFrame, H2O DataTable's Frame or scipy.sparse
808
            Data source for prediction.
809
            If str or pathlib.Path, it represents the path to a text file (CSV, TSV, or LibSVM).
810
811
        start_iteration : int, optional (default=0)
            Start index of the iteration to predict.
812
813
814
815
816
817
818
819
820
821
822
        num_iteration : int, optional (default=-1)
            Iteration used for prediction.
        raw_score : bool, optional (default=False)
            Whether to predict raw scores.
        pred_leaf : bool, optional (default=False)
            Whether to predict leaf index.
        pred_contrib : bool, optional (default=False)
            Whether to predict feature contributions.
        data_has_header : bool, optional (default=False)
            Whether data has header.
            Used only for txt data.
823
824
825
        validate_features : bool, optional (default=False)
            If True, ensure that the features used to predict match the ones used to train.
            Used only if data is pandas DataFrame.
wxchan's avatar
wxchan committed
826
827
828

        Returns
        -------
829
        result : numpy array, scipy.sparse or list of scipy.sparse
830
            Prediction result.
831
            Can be sparse or a list of sparse objects (each element represents predictions for one class) for feature contributions (when ``pred_contrib=True``).
wxchan's avatar
wxchan committed
832
        """
wxchan's avatar
wxchan committed
833
        if isinstance(data, Dataset):
834
            raise TypeError("Cannot use Dataset instance for prediction, please use raw data instead")
835
836
837
838
839
840
841
842
843
844
845
        elif isinstance(data, pd_DataFrame) and validate_features:
            data_names = [str(x) for x in data.columns]
            ptr_names = (ctypes.c_char_p * len(data_names))()
            ptr_names[:] = [x.encode('utf-8') for x in data_names]
            _safe_call(
                _LIB.LGBM_BoosterValidateFeatureNames(
                    self.handle,
                    ptr_names,
                    ctypes.c_int(len(data_names)),
                )
            )
846
        data = _data_from_pandas(data, None, None, self.pandas_categorical)[0]
wxchan's avatar
wxchan committed
847
848
849
850
851
        predict_type = C_API_PREDICT_NORMAL
        if raw_score:
            predict_type = C_API_PREDICT_RAW_SCORE
        if pred_leaf:
            predict_type = C_API_PREDICT_LEAF_INDEX
852
853
        if pred_contrib:
            predict_type = C_API_PREDICT_CONTRIB
wxchan's avatar
wxchan committed
854
        int_data_has_header = 1 if data_has_header else 0
cbecker's avatar
cbecker committed
855

856
        if isinstance(data, (str, Path)):
857
            with _TempFile() as f:
wxchan's avatar
wxchan committed
858
859
                _safe_call(_LIB.LGBM_BoosterPredictForFile(
                    self.handle,
860
                    c_str(str(data)),
Guolin Ke's avatar
Guolin Ke committed
861
862
                    ctypes.c_int(int_data_has_header),
                    ctypes.c_int(predict_type),
863
                    ctypes.c_int(start_iteration),
Guolin Ke's avatar
Guolin Ke committed
864
                    ctypes.c_int(num_iteration),
865
                    c_str(self.pred_parameter),
wxchan's avatar
wxchan committed
866
                    c_str(f.name)))
867
868
                preds = np.loadtxt(f.name, dtype=np.float64)
                nrow = preds.shape[0]
wxchan's avatar
wxchan committed
869
        elif isinstance(data, scipy.sparse.csr_matrix):
870
            preds, nrow = self.__pred_for_csr(data, start_iteration, num_iteration, predict_type)
Guolin Ke's avatar
Guolin Ke committed
871
        elif isinstance(data, scipy.sparse.csc_matrix):
872
            preds, nrow = self.__pred_for_csc(data, start_iteration, num_iteration, predict_type)
wxchan's avatar
wxchan committed
873
        elif isinstance(data, np.ndarray):
874
            preds, nrow = self.__pred_for_np2d(data, start_iteration, num_iteration, predict_type)
875
876
877
        elif isinstance(data, list):
            try:
                data = np.array(data)
878
            except BaseException:
879
                raise ValueError('Cannot convert data list to numpy array.')
880
            preds, nrow = self.__pred_for_np2d(data, start_iteration, num_iteration, predict_type)
881
        elif isinstance(data, dt_DataTable):
882
            preds, nrow = self.__pred_for_np2d(data.to_numpy(), start_iteration, num_iteration, predict_type)
wxchan's avatar
wxchan committed
883
884
        else:
            try:
885
                _log_warning('Converting data to scipy sparse matrix.')
wxchan's avatar
wxchan committed
886
                csr = scipy.sparse.csr_matrix(data)
887
            except BaseException:
888
                raise TypeError(f'Cannot predict data for type {type(data).__name__}')
889
            preds, nrow = self.__pred_for_csr(csr, start_iteration, num_iteration, predict_type)
wxchan's avatar
wxchan committed
890
891
        if pred_leaf:
            preds = preds.astype(np.int32)
892
        is_sparse = scipy.sparse.issparse(preds) or isinstance(preds, list)
893
        if not is_sparse and preds.size != nrow:
wxchan's avatar
wxchan committed
894
            if preds.size % nrow == 0:
895
                preds = preds.reshape(nrow, -1)
wxchan's avatar
wxchan committed
896
            else:
897
                raise ValueError(f'Length of predict result ({preds.size}) cannot be divide nrow ({nrow})')
wxchan's avatar
wxchan committed
898
899
        return preds

900
    def __get_num_preds(self, start_iteration, num_iteration, nrow, predict_type):
901
        """Get size of prediction result."""
902
        if nrow > MAX_INT32:
903
            raise LightGBMError('LightGBM cannot perform prediction for data '
904
                                f'with number of rows greater than MAX_INT32 ({MAX_INT32}).\n'
905
                                'You can split your data into chunks '
906
                                'and then concatenate predictions for them')
Guolin Ke's avatar
Guolin Ke committed
907
908
909
        n_preds = ctypes.c_int64(0)
        _safe_call(_LIB.LGBM_BoosterCalcNumPredict(
            self.handle,
Guolin Ke's avatar
Guolin Ke committed
910
911
            ctypes.c_int(nrow),
            ctypes.c_int(predict_type),
912
            ctypes.c_int(start_iteration),
Guolin Ke's avatar
Guolin Ke committed
913
            ctypes.c_int(num_iteration),
Guolin Ke's avatar
Guolin Ke committed
914
915
            ctypes.byref(n_preds)))
        return n_preds.value
wxchan's avatar
wxchan committed
916

917
    def __pred_for_np2d(self, mat, start_iteration, num_iteration, predict_type):
918
        """Predict for a 2-D numpy matrix."""
wxchan's avatar
wxchan committed
919
        if len(mat.shape) != 2:
920
            raise ValueError('Input numpy.ndarray or list must be 2 dimensional')
wxchan's avatar
wxchan committed
921

922
        def inner_predict(mat, start_iteration, num_iteration, predict_type, preds=None):
923
924
            if mat.dtype == np.float32 or mat.dtype == np.float64:
                data = np.array(mat.reshape(mat.size), dtype=mat.dtype, copy=False)
925
            else:  # change non-float data to float data, need to copy
926
927
                data = np.array(mat.reshape(mat.size), dtype=np.float32)
            ptr_data, type_ptr_data, _ = c_float_array(data)
928
            n_preds = self.__get_num_preds(start_iteration, num_iteration, mat.shape[0], predict_type)
929
            if preds is None:
930
                preds = np.empty(n_preds, dtype=np.float64)
931
932
933
934
935
936
937
            elif len(preds.shape) != 1 or len(preds) != n_preds:
                raise ValueError("Wrong length of pre-allocated predict array")
            out_num_preds = ctypes.c_int64(0)
            _safe_call(_LIB.LGBM_BoosterPredictForMat(
                self.handle,
                ptr_data,
                ctypes.c_int(type_ptr_data),
938
939
                ctypes.c_int32(mat.shape[0]),
                ctypes.c_int32(mat.shape[1]),
940
941
                ctypes.c_int(C_API_IS_ROW_MAJOR),
                ctypes.c_int(predict_type),
942
                ctypes.c_int(start_iteration),
943
944
945
946
947
948
949
950
951
952
953
954
                ctypes.c_int(num_iteration),
                c_str(self.pred_parameter),
                ctypes.byref(out_num_preds),
                preds.ctypes.data_as(ctypes.POINTER(ctypes.c_double))))
            if n_preds != out_num_preds.value:
                raise ValueError("Wrong length for predict results")
            return preds, mat.shape[0]

        nrow = mat.shape[0]
        if nrow > MAX_INT32:
            sections = np.arange(start=MAX_INT32, stop=nrow, step=MAX_INT32)
            # __get_num_preds() cannot work with nrow > MAX_INT32, so calculate overall number of predictions piecemeal
955
            n_preds = [self.__get_num_preds(start_iteration, num_iteration, i, predict_type) for i in np.diff([0] + list(sections) + [nrow])]
956
            n_preds_sections = np.array([0] + n_preds, dtype=np.intp).cumsum()
957
            preds = np.empty(sum(n_preds), dtype=np.float64)
958
959
            for chunk, (start_idx_pred, end_idx_pred) in zip(np.array_split(mat, sections),
                                                             zip(n_preds_sections, n_preds_sections[1:])):
960
                # avoid memory consumption by arrays concatenation operations
961
                inner_predict(chunk, start_iteration, num_iteration, predict_type, preds[start_idx_pred:end_idx_pred])
962
            return preds, nrow
wxchan's avatar
wxchan committed
963
        else:
964
            return inner_predict(mat, start_iteration, num_iteration, predict_type)
wxchan's avatar
wxchan committed
965

966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
    def __create_sparse_native(self, cs, out_shape, out_ptr_indptr, out_ptr_indices, out_ptr_data,
                               indptr_type, data_type, is_csr=True):
        # create numpy array from output arrays
        data_indices_len = out_shape[0]
        indptr_len = out_shape[1]
        if indptr_type == C_API_DTYPE_INT32:
            out_indptr = cint32_array_to_numpy(out_ptr_indptr, indptr_len)
        elif indptr_type == C_API_DTYPE_INT64:
            out_indptr = cint64_array_to_numpy(out_ptr_indptr, indptr_len)
        else:
            raise TypeError("Expected int32 or int64 type for indptr")
        if data_type == C_API_DTYPE_FLOAT32:
            out_data = cfloat32_array_to_numpy(out_ptr_data, data_indices_len)
        elif data_type == C_API_DTYPE_FLOAT64:
            out_data = cfloat64_array_to_numpy(out_ptr_data, data_indices_len)
        else:
            raise TypeError("Expected float32 or float64 type for data")
        out_indices = cint32_array_to_numpy(out_ptr_indices, data_indices_len)
        # break up indptr based on number of rows (note more than one matrix in multiclass case)
        per_class_indptr_shape = cs.indptr.shape[0]
        # for CSC there is extra column added
        if not is_csr:
            per_class_indptr_shape += 1
        out_indptr_arrays = np.split(out_indptr, out_indptr.shape[0] / per_class_indptr_shape)
        # reformat output into a csr or csc matrix or list of csr or csc matrices
        cs_output_matrices = []
        offset = 0
        for cs_indptr in out_indptr_arrays:
            matrix_indptr_len = cs_indptr[cs_indptr.shape[0] - 1]
            cs_indices = out_indices[offset + cs_indptr[0]:offset + matrix_indptr_len]
            cs_data = out_data[offset + cs_indptr[0]:offset + matrix_indptr_len]
            offset += matrix_indptr_len
            # same shape as input csr or csc matrix except extra column for expected value
            cs_shape = [cs.shape[0], cs.shape[1] + 1]
            # note: make sure we copy data as it will be deallocated next
            if is_csr:
                cs_output_matrices.append(scipy.sparse.csr_matrix((cs_data, cs_indices, cs_indptr), cs_shape))
            else:
                cs_output_matrices.append(scipy.sparse.csc_matrix((cs_data, cs_indices, cs_indptr), cs_shape))
        # free the temporary native indptr, indices, and data
        _safe_call(_LIB.LGBM_BoosterFreePredictSparse(out_ptr_indptr, out_ptr_indices, out_ptr_data,
                                                      ctypes.c_int(indptr_type), ctypes.c_int(data_type)))
        if len(cs_output_matrices) == 1:
            return cs_output_matrices[0]
        return cs_output_matrices

1012
    def __pred_for_csr(self, csr, start_iteration, num_iteration, predict_type):
1013
        """Predict for a CSR data."""
1014
        def inner_predict(csr, start_iteration, num_iteration, predict_type, preds=None):
1015
            nrow = len(csr.indptr) - 1
1016
            n_preds = self.__get_num_preds(start_iteration, num_iteration, nrow, predict_type)
1017
            if preds is None:
1018
                preds = np.empty(n_preds, dtype=np.float64)
1019
1020
1021
1022
1023
1024
1025
            elif len(preds.shape) != 1 or len(preds) != n_preds:
                raise ValueError("Wrong length of pre-allocated predict array")
            out_num_preds = ctypes.c_int64(0)

            ptr_indptr, type_ptr_indptr, __ = c_int_array(csr.indptr)
            ptr_data, type_ptr_data, _ = c_float_array(csr.data)

1026
            assert csr.shape[1] <= MAX_INT32
1027
            csr_indices = csr.indices.astype(np.int32, copy=False)
1028

1029
1030
1031
            _safe_call(_LIB.LGBM_BoosterPredictForCSR(
                self.handle,
                ptr_indptr,
1032
                ctypes.c_int(type_ptr_indptr),
1033
                csr_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
1034
1035
1036
1037
1038
1039
                ptr_data,
                ctypes.c_int(type_ptr_data),
                ctypes.c_int64(len(csr.indptr)),
                ctypes.c_int64(len(csr.data)),
                ctypes.c_int64(csr.shape[1]),
                ctypes.c_int(predict_type),
1040
                ctypes.c_int(start_iteration),
1041
1042
1043
1044
1045
1046
1047
                ctypes.c_int(num_iteration),
                c_str(self.pred_parameter),
                ctypes.byref(out_num_preds),
                preds.ctypes.data_as(ctypes.POINTER(ctypes.c_double))))
            if n_preds != out_num_preds.value:
                raise ValueError("Wrong length for predict results")
            return preds, nrow
wxchan's avatar
wxchan committed
1048

1049
        def inner_predict_sparse(csr, start_iteration, num_iteration, predict_type):
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
            ptr_indptr, type_ptr_indptr, __ = c_int_array(csr.indptr)
            ptr_data, type_ptr_data, _ = c_float_array(csr.data)
            csr_indices = csr.indices.astype(np.int32, copy=False)
            matrix_type = C_API_MATRIX_TYPE_CSR
            if type_ptr_indptr == C_API_DTYPE_INT32:
                out_ptr_indptr = ctypes.POINTER(ctypes.c_int32)()
            else:
                out_ptr_indptr = ctypes.POINTER(ctypes.c_int64)()
            out_ptr_indices = ctypes.POINTER(ctypes.c_int32)()
            if type_ptr_data == C_API_DTYPE_FLOAT32:
                out_ptr_data = ctypes.POINTER(ctypes.c_float)()
            else:
                out_ptr_data = ctypes.POINTER(ctypes.c_double)()
1063
            out_shape = np.empty(2, dtype=np.int64)
1064
1065
1066
            _safe_call(_LIB.LGBM_BoosterPredictSparseOutput(
                self.handle,
                ptr_indptr,
1067
                ctypes.c_int(type_ptr_indptr),
1068
1069
1070
1071
1072
1073
1074
                csr_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
                ptr_data,
                ctypes.c_int(type_ptr_data),
                ctypes.c_int64(len(csr.indptr)),
                ctypes.c_int64(len(csr.data)),
                ctypes.c_int64(csr.shape[1]),
                ctypes.c_int(predict_type),
1075
                ctypes.c_int(start_iteration),
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
                ctypes.c_int(num_iteration),
                c_str(self.pred_parameter),
                ctypes.c_int(matrix_type),
                out_shape.ctypes.data_as(ctypes.POINTER(ctypes.c_int64)),
                ctypes.byref(out_ptr_indptr),
                ctypes.byref(out_ptr_indices),
                ctypes.byref(out_ptr_data)))
            matrices = self.__create_sparse_native(csr, out_shape, out_ptr_indptr, out_ptr_indices, out_ptr_data,
                                                   type_ptr_indptr, type_ptr_data, is_csr=True)
            nrow = len(csr.indptr) - 1
            return matrices, nrow

        if predict_type == C_API_PREDICT_CONTRIB:
1089
            return inner_predict_sparse(csr, start_iteration, num_iteration, predict_type)
1090
1091
1092
1093
        nrow = len(csr.indptr) - 1
        if nrow > MAX_INT32:
            sections = [0] + list(np.arange(start=MAX_INT32, stop=nrow, step=MAX_INT32)) + [nrow]
            # __get_num_preds() cannot work with nrow > MAX_INT32, so calculate overall number of predictions piecemeal
1094
            n_preds = [self.__get_num_preds(start_iteration, num_iteration, i, predict_type) for i in np.diff(sections)]
1095
            n_preds_sections = np.array([0] + n_preds, dtype=np.intp).cumsum()
1096
            preds = np.empty(sum(n_preds), dtype=np.float64)
1097
1098
            for (start_idx, end_idx), (start_idx_pred, end_idx_pred) in zip(zip(sections, sections[1:]),
                                                                            zip(n_preds_sections, n_preds_sections[1:])):
1099
                # avoid memory consumption by arrays concatenation operations
1100
                inner_predict(csr[start_idx:end_idx], start_iteration, num_iteration, predict_type, preds[start_idx_pred:end_idx_pred])
1101
1102
            return preds, nrow
        else:
1103
            return inner_predict(csr, start_iteration, num_iteration, predict_type)
Guolin Ke's avatar
Guolin Ke committed
1104

1105
    def __pred_for_csc(self, csc, start_iteration, num_iteration, predict_type):
1106
        """Predict for a CSC data."""
1107
        def inner_predict_sparse(csc, start_iteration, num_iteration, predict_type):
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
            ptr_indptr, type_ptr_indptr, __ = c_int_array(csc.indptr)
            ptr_data, type_ptr_data, _ = c_float_array(csc.data)
            csc_indices = csc.indices.astype(np.int32, copy=False)
            matrix_type = C_API_MATRIX_TYPE_CSC
            if type_ptr_indptr == C_API_DTYPE_INT32:
                out_ptr_indptr = ctypes.POINTER(ctypes.c_int32)()
            else:
                out_ptr_indptr = ctypes.POINTER(ctypes.c_int64)()
            out_ptr_indices = ctypes.POINTER(ctypes.c_int32)()
            if type_ptr_data == C_API_DTYPE_FLOAT32:
                out_ptr_data = ctypes.POINTER(ctypes.c_float)()
            else:
                out_ptr_data = ctypes.POINTER(ctypes.c_double)()
1121
            out_shape = np.empty(2, dtype=np.int64)
1122
1123
1124
            _safe_call(_LIB.LGBM_BoosterPredictSparseOutput(
                self.handle,
                ptr_indptr,
1125
                ctypes.c_int(type_ptr_indptr),
1126
1127
1128
1129
1130
1131
1132
                csc_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
                ptr_data,
                ctypes.c_int(type_ptr_data),
                ctypes.c_int64(len(csc.indptr)),
                ctypes.c_int64(len(csc.data)),
                ctypes.c_int64(csc.shape[0]),
                ctypes.c_int(predict_type),
1133
                ctypes.c_int(start_iteration),
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
                ctypes.c_int(num_iteration),
                c_str(self.pred_parameter),
                ctypes.c_int(matrix_type),
                out_shape.ctypes.data_as(ctypes.POINTER(ctypes.c_int64)),
                ctypes.byref(out_ptr_indptr),
                ctypes.byref(out_ptr_indices),
                ctypes.byref(out_ptr_data)))
            matrices = self.__create_sparse_native(csc, out_shape, out_ptr_indptr, out_ptr_indices, out_ptr_data,
                                                   type_ptr_indptr, type_ptr_data, is_csr=False)
            nrow = csc.shape[0]
            return matrices, nrow

Guolin Ke's avatar
Guolin Ke committed
1146
        nrow = csc.shape[0]
1147
        if nrow > MAX_INT32:
1148
            return self.__pred_for_csr(csc.tocsr(), start_iteration, num_iteration, predict_type)
1149
        if predict_type == C_API_PREDICT_CONTRIB:
1150
1151
            return inner_predict_sparse(csc, start_iteration, num_iteration, predict_type)
        n_preds = self.__get_num_preds(start_iteration, num_iteration, nrow, predict_type)
1152
        preds = np.empty(n_preds, dtype=np.float64)
Guolin Ke's avatar
Guolin Ke committed
1153
1154
        out_num_preds = ctypes.c_int64(0)

1155
1156
        ptr_indptr, type_ptr_indptr, __ = c_int_array(csc.indptr)
        ptr_data, type_ptr_data, _ = c_float_array(csc.data)
Guolin Ke's avatar
Guolin Ke committed
1157

1158
        assert csc.shape[0] <= MAX_INT32
1159
        csc_indices = csc.indices.astype(np.int32, copy=False)
1160

Guolin Ke's avatar
Guolin Ke committed
1161
1162
1163
        _safe_call(_LIB.LGBM_BoosterPredictForCSC(
            self.handle,
            ptr_indptr,
1164
            ctypes.c_int(type_ptr_indptr),
1165
            csc_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
Guolin Ke's avatar
Guolin Ke committed
1166
            ptr_data,
Guolin Ke's avatar
Guolin Ke committed
1167
1168
1169
1170
1171
            ctypes.c_int(type_ptr_data),
            ctypes.c_int64(len(csc.indptr)),
            ctypes.c_int64(len(csc.data)),
            ctypes.c_int64(csc.shape[0]),
            ctypes.c_int(predict_type),
1172
            ctypes.c_int(start_iteration),
Guolin Ke's avatar
Guolin Ke committed
1173
            ctypes.c_int(num_iteration),
1174
            c_str(self.pred_parameter),
Guolin Ke's avatar
Guolin Ke committed
1175
            ctypes.byref(out_num_preds),
wxchan's avatar
wxchan committed
1176
            preds.ctypes.data_as(ctypes.POINTER(ctypes.c_double))))
wxchan's avatar
wxchan committed
1177
        if n_preds != out_num_preds.value:
1178
            raise ValueError("Wrong length for predict results")
wxchan's avatar
wxchan committed
1179
1180
        return preds, nrow

1181
    def current_iteration(self) -> int:
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
        """Get the index of the current iteration.

        Returns
        -------
        cur_iter : int
            The index of the current iteration.
        """
        out_cur_iter = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterGetCurrentIteration(
            self.handle,
            ctypes.byref(out_cur_iter)))
        return out_cur_iter.value

wxchan's avatar
wxchan committed
1195

1196
class Dataset:
wxchan's avatar
wxchan committed
1197
    """Dataset in LightGBM."""
1198

1199
    def __init__(self, data, label=None, reference=None,
1200
                 weight=None, group=None, init_score=None,
1201
                 feature_name='auto', categorical_feature='auto', params=None,
wxchan's avatar
wxchan committed
1202
                 free_raw_data=True):
1203
        """Initialize Dataset.
1204

wxchan's avatar
wxchan committed
1205
1206
        Parameters
        ----------
1207
        data : str, pathlib.Path, numpy array, pandas DataFrame, H2O DataTable's Frame, scipy.sparse, Sequence, list of Sequence or list of numpy array
wxchan's avatar
wxchan committed
1208
            Data source of Dataset.
1209
            If str or pathlib.Path, it represents the path to a text file (CSV, TSV, or LibSVM) or a LightGBM Dataset binary file.
1210
        label : list, numpy 1-D array, pandas Series / one-column DataFrame or None, optional (default=None)
1211
1212
1213
            Label of the data.
        reference : Dataset or None, optional (default=None)
            If this is Dataset for validation, training data should be used as reference.
1214
        weight : list, numpy 1-D array, pandas Series or None, optional (default=None)
1215
            Weight for each instance. Weights should be non-negative.
1216
        group : list, numpy 1-D array, pandas Series or None, optional (default=None)
1217
1218
1219
            Group/query data.
            Only used in the learning-to-rank task.
            sum(group) = n_samples.
1220
1221
            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.
1222
        init_score : list, list of lists (for multi-class task), numpy array, pandas Series, pandas DataFrame (for multi-class task), or None, optional (default=None)
1223
            Init score for Dataset.
1224
        feature_name : list of str, or 'auto', optional (default="auto")
1225
1226
            Feature names.
            If 'auto' and data is pandas DataFrame, data columns names are used.
1227
        categorical_feature : list of str or int, or 'auto', optional (default="auto")
1228
1229
            Categorical features.
            If list of int, interpreted as indices.
1230
            If list of str, interpreted as feature names (need to specify ``feature_name`` as well).
1231
            If 'auto' and data is pandas DataFrame, pandas unordered categorical columns are used.
1232
            All values in categorical features will be cast to int32 and thus should be less than int32 max value (2147483647).
1233
            Large values could be memory consuming. Consider using consecutive integers starting from zero.
1234
            All negative values in categorical features will be treated as missing values.
1235
            The output cannot be monotonically constrained with respect to a categorical feature.
1236
            Floating point numbers in categorical features will be rounded towards 0.
Nikita Titov's avatar
Nikita Titov committed
1237
        params : dict or None, optional (default=None)
1238
            Other parameters for Dataset.
Nikita Titov's avatar
Nikita Titov committed
1239
        free_raw_data : bool, optional (default=True)
1240
            If True, raw data is freed after constructing inner Dataset.
wxchan's avatar
wxchan committed
1241
        """
wxchan's avatar
wxchan committed
1242
1243
1244
1245
1246
1247
        self.handle = None
        self.data = data
        self.label = label
        self.reference = reference
        self.weight = weight
        self.group = group
1248
        self.init_score = init_score
wxchan's avatar
wxchan committed
1249
        self.feature_name = feature_name
1250
        self.categorical_feature = categorical_feature
1251
        self.params = deepcopy(params)
wxchan's avatar
wxchan committed
1252
1253
        self.free_raw_data = free_raw_data
        self.used_indices = None
1254
        self.need_slice = True
wxchan's avatar
wxchan committed
1255
        self._predictor = None
1256
        self.pandas_categorical = None
1257
        self.params_back_up = None
1258
1259
        self.feature_penalty = None
        self.monotone_constraints = None
1260
        self.version = 0
1261
        self._start_row = 0  # Used when pushing rows one by one.
wxchan's avatar
wxchan committed
1262

1263
    def __del__(self) -> None:
1264
1265
1266
1267
        try:
            self._free_handle()
        except AttributeError:
            pass
1268

1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
    def _create_sample_indices(self, total_nrow: int) -> np.ndarray:
        """Get an array of randomly chosen indices from this ``Dataset``.

        Indices are sampled without replacement.

        Parameters
        ----------
        total_nrow : int
            Total number of rows to sample from.
            If this value is greater than the value of parameter ``bin_construct_sample_cnt``, only ``bin_construct_sample_cnt`` indices will be used.
            If Dataset has multiple input data, this should be the sum of rows of every file.

        Returns
        -------
        indices : numpy array
            Indices for sampled data.
        """
        param_str = param_dict_to_str(self.get_params())
        sample_cnt = _get_sample_count(total_nrow, param_str)
        indices = np.empty(sample_cnt, dtype=np.int32)
        ptr_data, _, _ = c_int_array(indices)
        actual_sample_cnt = ctypes.c_int32(0)

        _safe_call(_LIB.LGBM_SampleIndices(
            ctypes.c_int32(total_nrow),
            c_str(param_str),
            ptr_data,
            ctypes.byref(actual_sample_cnt),
        ))
1298
1299
        assert sample_cnt == actual_sample_cnt.value
        return indices
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334

    def _init_from_ref_dataset(self, total_nrow: int, ref_dataset: 'Dataset') -> 'Dataset':
        """Create dataset from a reference dataset.

        Parameters
        ----------
        total_nrow : int
            Number of rows expected to add to dataset.
        ref_dataset : Dataset
            Reference dataset to extract meta from.

        Returns
        -------
        self : Dataset
            Constructed Dataset object.
        """
        self.handle = ctypes.c_void_p()
        _safe_call(_LIB.LGBM_DatasetCreateByReference(
            ref_dataset,
            ctypes.c_int64(total_nrow),
            ctypes.byref(self.handle),
        ))
        return self

    def _init_from_sample(
        self,
        sample_data: List[np.ndarray],
        sample_indices: List[np.ndarray],
        sample_cnt: int,
        total_nrow: int,
    ) -> "Dataset":
        """Create Dataset from sampled data structures.

        Parameters
        ----------
1335
        sample_data : list of numpy array
1336
            Sample data for each column.
1337
        sample_indices : list of numpy array
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
            Sample data row index for each column.
        sample_cnt : int
            Number of samples.
        total_nrow : int
            Total number of rows for all input files.

        Returns
        -------
        self : Dataset
            Constructed Dataset object.
        """
        ncol = len(sample_indices)
        assert len(sample_data) == ncol, "#sample data column != #column indices"

        for i in range(ncol):
            if sample_data[i].dtype != np.double:
                raise ValueError(f"sample_data[{i}] type {sample_data[i].dtype} is not double")
            if sample_indices[i].dtype != np.int32:
                raise ValueError(f"sample_indices[{i}] type {sample_indices[i].dtype} is not int32")

        # c type: double**
        # each double* element points to start of each column of sample data.
        sample_col_ptr = (ctypes.POINTER(ctypes.c_double) * ncol)()
        # c type int**
        # each int* points to start of indices for each column
        indices_col_ptr = (ctypes.POINTER(ctypes.c_int32) * ncol)()
        for i in range(ncol):
            sample_col_ptr[i] = c_float_array(sample_data[i])[0]
            indices_col_ptr[i] = c_int_array(sample_indices[i])[0]

        num_per_col = np.array([len(d) for d in sample_indices], dtype=np.int32)
        num_per_col_ptr, _, _ = c_int_array(num_per_col)

        self.handle = ctypes.c_void_p()
        params_str = param_dict_to_str(self.get_params())
        _safe_call(_LIB.LGBM_DatasetCreateFromSampledColumn(
            ctypes.cast(sample_col_ptr, ctypes.POINTER(ctypes.POINTER(ctypes.c_double))),
            ctypes.cast(indices_col_ptr, ctypes.POINTER(ctypes.POINTER(ctypes.c_int32))),
            ctypes.c_int32(ncol),
            num_per_col_ptr,
            ctypes.c_int32(sample_cnt),
            ctypes.c_int32(total_nrow),
1380
            ctypes.c_int64(total_nrow),
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
            c_str(params_str),
            ctypes.byref(self.handle),
        ))
        return self

    def _push_rows(self, data: np.ndarray) -> 'Dataset':
        """Add rows to Dataset.

        Parameters
        ----------
        data : numpy 1-D array
            New data to add to the Dataset.

        Returns
        -------
        self : Dataset
            Dataset object.
        """
        nrow, ncol = data.shape
        data = data.reshape(data.size)
        data_ptr, data_type, _ = c_float_array(data)

        _safe_call(_LIB.LGBM_DatasetPushRows(
            self.handle,
            data_ptr,
            data_type,
            ctypes.c_int32(nrow),
            ctypes.c_int32(ncol),
            ctypes.c_int32(self._start_row),
        ))
        self._start_row += nrow
        return self

1414
    def get_params(self) -> Dict[str, Any]:
1415
1416
1417
1418
        """Get the used parameters in the Dataset.

        Returns
        -------
1419
        params : dict
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
            The used parameters in this Dataset object.
        """
        if self.params is not None:
            # no min_data, nthreads and verbose in this function
            dataset_params = _ConfigAliases.get("bin_construct_sample_cnt",
                                                "categorical_feature",
                                                "data_random_seed",
                                                "enable_bundle",
                                                "feature_pre_filter",
                                                "forcedbins_filename",
                                                "group_column",
                                                "header",
                                                "ignore_column",
                                                "is_enable_sparse",
                                                "label_column",
1435
                                                "linear_tree",
1436
1437
1438
1439
                                                "max_bin",
                                                "max_bin_by_feature",
                                                "min_data_in_bin",
                                                "pre_partition",
Nikita Titov's avatar
Nikita Titov committed
1440
                                                "precise_float_parser",
1441
1442
1443
1444
1445
                                                "two_round",
                                                "use_missing",
                                                "weight_column",
                                                "zero_as_missing")
            return {k: v for k, v in self.params.items() if k in dataset_params}
1446
1447
        else:
            return {}
1448

1449
    def _free_handle(self) -> "Dataset":
1450
        if self.handle is not None:
1451
            _safe_call(_LIB.LGBM_DatasetFree(self.handle))
1452
            self.handle = None
Guolin Ke's avatar
Guolin Ke committed
1453
1454
1455
        self.need_slice = True
        if self.used_indices is not None:
            self.data = None
Nikita Titov's avatar
Nikita Titov committed
1456
        return self
wxchan's avatar
wxchan committed
1457

Guolin Ke's avatar
Guolin Ke committed
1458
1459
    def _set_init_score_by_predictor(self, predictor, data, used_indices=None):
        data_has_header = False
1460
        if isinstance(data, (str, Path)):
Guolin Ke's avatar
Guolin Ke committed
1461
            # check data has header or not
1462
            data_has_header = any(self.params.get(alias, False) for alias in _ConfigAliases.get("header"))
Guolin Ke's avatar
Guolin Ke committed
1463
        num_data = self.num_data()
1464
1465
1466
        if predictor is not None:
            init_score = predictor.predict(data,
                                           raw_score=True,
1467
1468
                                           data_has_header=data_has_header)
            init_score = init_score.ravel()
1469
1470
            if used_indices is not None:
                assert not self.need_slice
1471
                if isinstance(data, (str, Path)):
1472
                    sub_init_score = np.empty(num_data * predictor.num_class, dtype=np.float64)
1473
                    assert num_data == len(used_indices)
1474
1475
                    for i in range(len(used_indices)):
                        for j in range(predictor.num_class):
1476
1477
1478
1479
                            sub_init_score[i * predictor.num_class + j] = init_score[used_indices[i] * predictor.num_class + j]
                    init_score = sub_init_score
            if predictor.num_class > 1:
                # need to regroup init_score
1480
                new_init_score = np.empty(init_score.size, dtype=np.float64)
1481
1482
                for i in range(num_data):
                    for j in range(predictor.num_class):
1483
1484
1485
                        new_init_score[j * num_data + i] = init_score[i * predictor.num_class + j]
                init_score = new_init_score
        elif self.init_score is not None:
1486
            init_score = np.zeros(self.init_score.shape, dtype=np.float64)
1487
1488
        else:
            return self
Guolin Ke's avatar
Guolin Ke committed
1489
1490
        self.set_init_score(init_score)

1491
    def _lazy_init(self, data, label=None, reference=None,
1492
                   weight=None, group=None, init_score=None, predictor=None,
1493
                   feature_name='auto', categorical_feature='auto', params=None):
wxchan's avatar
wxchan committed
1494
1495
        if data is None:
            self.handle = None
Nikita Titov's avatar
Nikita Titov committed
1496
            return self
Guolin Ke's avatar
Guolin Ke committed
1497
1498
1499
        if reference is not None:
            self.pandas_categorical = reference.pandas_categorical
            categorical_feature = reference.categorical_feature
1500
1501
1502
1503
        data, feature_name, categorical_feature, self.pandas_categorical = _data_from_pandas(data,
                                                                                             feature_name,
                                                                                             categorical_feature,
                                                                                             self.pandas_categorical)
wxchan's avatar
wxchan committed
1504
        label = _label_from_pandas(label)
Guolin Ke's avatar
Guolin Ke committed
1505

1506
        # process for args
wxchan's avatar
wxchan committed
1507
        params = {} if params is None else params
1508
1509
1510
        args_names = (getattr(self.__class__, '_lazy_init')
                      .__code__
                      .co_varnames[:getattr(self.__class__, '_lazy_init').__code__.co_argcount])
1511
        for key in params.keys():
1512
            if key in args_names:
1513
1514
                _log_warning(f'{key} keyword has been found in `params` and will be ignored.\n'
                             f'Please use {key} argument of the Dataset constructor to pass this parameter.')
1515
        # get categorical features
1516
1517
1518
1519
1520
1521
        if categorical_feature is not None:
            categorical_indices = set()
            feature_dict = {}
            if feature_name is not None:
                feature_dict = {name: i for i, name in enumerate(feature_name)}
            for name in categorical_feature:
1522
                if isinstance(name, str) and name in feature_dict:
1523
                    categorical_indices.add(feature_dict[name])
1524
                elif isinstance(name, int):
1525
1526
                    categorical_indices.add(name)
                else:
1527
                    raise TypeError(f"Wrong type({type(name).__name__}) or unknown name({name}) in categorical_feature")
1528
            if categorical_indices:
1529
1530
                for cat_alias in _ConfigAliases.get("categorical_feature"):
                    if cat_alias in params:
1531
                        # If the params[cat_alias] is equal to categorical_indices, do not report the warning.
1532
                        if not (isinstance(params[cat_alias], list) and set(params[cat_alias]) == categorical_indices):
1533
                            _log_warning(f'{cat_alias} in param dict is overridden.')
1534
                        params.pop(cat_alias, None)
1535
                params['categorical_column'] = sorted(categorical_indices)
1536

wxchan's avatar
wxchan committed
1537
        params_str = param_dict_to_str(params)
1538
        self.params = params
1539
        # process for reference dataset
wxchan's avatar
wxchan committed
1540
        ref_dataset = None
wxchan's avatar
wxchan committed
1541
        if isinstance(reference, Dataset):
1542
            ref_dataset = reference.construct().handle
wxchan's avatar
wxchan committed
1543
1544
        elif reference is not None:
            raise TypeError('Reference dataset should be None or dataset instance')
1545
        # start construct data
1546
        if isinstance(data, (str, Path)):
wxchan's avatar
wxchan committed
1547
1548
            self.handle = ctypes.c_void_p()
            _safe_call(_LIB.LGBM_DatasetCreateFromFile(
1549
                c_str(str(data)),
wxchan's avatar
wxchan committed
1550
1551
1552
1553
1554
                c_str(params_str),
                ref_dataset,
                ctypes.byref(self.handle)))
        elif isinstance(data, scipy.sparse.csr_matrix):
            self.__init_from_csr(data, params_str, ref_dataset)
Guolin Ke's avatar
Guolin Ke committed
1555
1556
        elif isinstance(data, scipy.sparse.csc_matrix):
            self.__init_from_csc(data, params_str, ref_dataset)
wxchan's avatar
wxchan committed
1557
1558
        elif isinstance(data, np.ndarray):
            self.__init_from_np2d(data, params_str, ref_dataset)
1559
1560
1561
1562
1563
1564
1565
1566
1567
        elif isinstance(data, list) and len(data) > 0:
            if all(isinstance(x, np.ndarray) for x in data):
                self.__init_from_list_np2d(data, params_str, ref_dataset)
            elif all(isinstance(x, Sequence) for x in data):
                self.__init_from_seqs(data, ref_dataset)
            else:
                raise TypeError('Data list can only be of ndarray or Sequence')
        elif isinstance(data, Sequence):
            self.__init_from_seqs([data], ref_dataset)
1568
        elif isinstance(data, dt_DataTable):
1569
            self.__init_from_np2d(data.to_numpy(), params_str, ref_dataset)
wxchan's avatar
wxchan committed
1570
1571
1572
1573
        else:
            try:
                csr = scipy.sparse.csr_matrix(data)
                self.__init_from_csr(csr, params_str, ref_dataset)
1574
            except BaseException:
1575
                raise TypeError(f'Cannot initialize Dataset from {type(data).__name__}')
wxchan's avatar
wxchan committed
1576
1577
1578
        if label is not None:
            self.set_label(label)
        if self.get_label() is None:
1579
            raise ValueError("Label should not be None")
wxchan's avatar
wxchan committed
1580
1581
1582
1583
        if weight is not None:
            self.set_weight(weight)
        if group is not None:
            self.set_group(group)
1584
1585
        if isinstance(predictor, _InnerPredictor):
            if self._predictor is None and init_score is not None:
1586
                _log_warning("The init_score will be overridden by the prediction of init_model.")
Guolin Ke's avatar
Guolin Ke committed
1587
            self._set_init_score_by_predictor(predictor, data)
1588
1589
        elif init_score is not None:
            self.set_init_score(init_score)
Guolin Ke's avatar
Guolin Ke committed
1590
        elif predictor is not None:
1591
            raise TypeError(f'Wrong predictor type {type(predictor).__name__}')
Guolin Ke's avatar
Guolin Ke committed
1592
        # set feature names
Nikita Titov's avatar
Nikita Titov committed
1593
        return self.set_feature_name(feature_name)
wxchan's avatar
wxchan committed
1594

1595
1596
    @staticmethod
    def _yield_row_from_seqlist(seqs: List[Sequence], indices: Iterable[int]):
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
        offset = 0
        seq_id = 0
        seq = seqs[seq_id]
        for row_id in indices:
            assert row_id >= offset, "sample indices are expected to be monotonic"
            while row_id >= offset + len(seq):
                offset += len(seq)
                seq_id += 1
                seq = seqs[seq_id]
            id_in_seq = row_id - offset
            row = seq[id_in_seq]
            yield row if row.flags['OWNDATA'] else row.copy()

    def __sample(self, seqs: List[Sequence], total_nrow: int) -> Tuple[List[np.ndarray], List[np.ndarray]]:
        """Sample data from seqs.

        Mimics behavior in c_api.cpp:LGBM_DatasetCreateFromMats()

        Returns
        -------
            sampled_rows, sampled_row_indices
        """
        indices = self._create_sample_indices(total_nrow)

        # Select sampled rows, transpose to column order.
1622
        sampled = np.array([row for row in self._yield_row_from_seqlist(seqs, indices)])
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
        sampled = sampled.T

        filtered = []
        filtered_idx = []
        sampled_row_range = np.arange(len(indices), dtype=np.int32)
        for col in sampled:
            col_predicate = (np.abs(col) > ZERO_THRESHOLD) | np.isnan(col)
            filtered_col = col[col_predicate]
            filtered_row_idx = sampled_row_range[col_predicate]

            filtered.append(filtered_col)
            filtered_idx.append(filtered_row_idx)

        return filtered, filtered_idx

    def __init_from_seqs(self, seqs: List[Sequence], ref_dataset: Optional['Dataset'] = None):
        """
        Initialize data from list of Sequence objects.

        Sequence: Generic Data Access Object
            Supports random access and access by batch if properly defined by user

        Data scheme uniformity are trusted, not checked
        """
        total_nrow = sum(len(seq) for seq in seqs)

        # create validation dataset from ref_dataset
        if ref_dataset is not None:
            self._init_from_ref_dataset(total_nrow, ref_dataset)
        else:
            param_str = param_dict_to_str(self.get_params())
            sample_cnt = _get_sample_count(total_nrow, param_str)

            sample_data, col_indices = self.__sample(seqs, total_nrow)
            self._init_from_sample(sample_data, col_indices, sample_cnt, total_nrow)

        for seq in seqs:
            nrow = len(seq)
            batch_size = getattr(seq, 'batch_size', None) or Sequence.batch_size
            for start in range(0, nrow, batch_size):
                end = min(start + batch_size, nrow)
                self._push_rows(seq[start:end])
        return self

wxchan's avatar
wxchan committed
1667
    def __init_from_np2d(self, mat, params_str, ref_dataset):
1668
        """Initialize data from a 2-D numpy matrix."""
wxchan's avatar
wxchan committed
1669
1670
1671
1672
1673
1674
        if len(mat.shape) != 2:
            raise ValueError('Input numpy.ndarray must be 2 dimensional')

        self.handle = ctypes.c_void_p()
        if mat.dtype == np.float32 or mat.dtype == np.float64:
            data = np.array(mat.reshape(mat.size), dtype=mat.dtype, copy=False)
1675
        else:  # change non-float data to float data, need to copy
wxchan's avatar
wxchan committed
1676
1677
            data = np.array(mat.reshape(mat.size), dtype=np.float32)

1678
        ptr_data, type_ptr_data, _ = c_float_array(data)
wxchan's avatar
wxchan committed
1679
1680
        _safe_call(_LIB.LGBM_DatasetCreateFromMat(
            ptr_data,
Guolin Ke's avatar
Guolin Ke committed
1681
            ctypes.c_int(type_ptr_data),
1682
1683
            ctypes.c_int32(mat.shape[0]),
            ctypes.c_int32(mat.shape[1]),
Guolin Ke's avatar
Guolin Ke committed
1684
            ctypes.c_int(C_API_IS_ROW_MAJOR),
wxchan's avatar
wxchan committed
1685
1686
1687
            c_str(params_str),
            ref_dataset,
            ctypes.byref(self.handle)))
Nikita Titov's avatar
Nikita Titov committed
1688
        return self
wxchan's avatar
wxchan committed
1689

1690
    def __init_from_list_np2d(self, mats, params_str, ref_dataset):
1691
        """Initialize data from a list of 2-D numpy matrices."""
1692
        ncol = mats[0].shape[1]
1693
        nrow = np.empty((len(mats),), np.int32)
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
        if mats[0].dtype == np.float64:
            ptr_data = (ctypes.POINTER(ctypes.c_double) * len(mats))()
        else:
            ptr_data = (ctypes.POINTER(ctypes.c_float) * len(mats))()

        holders = []
        type_ptr_data = None

        for i, mat in enumerate(mats):
            if len(mat.shape) != 2:
                raise ValueError('Input numpy.ndarray must be 2 dimensional')

            if mat.shape[1] != ncol:
                raise ValueError('Input arrays must have same number of columns')

            nrow[i] = mat.shape[0]

            if mat.dtype == np.float32 or mat.dtype == np.float64:
                mats[i] = np.array(mat.reshape(mat.size), dtype=mat.dtype, copy=False)
1713
            else:  # change non-float data to float data, need to copy
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
                mats[i] = np.array(mat.reshape(mat.size), dtype=np.float32)

            chunk_ptr_data, chunk_type_ptr_data, holder = c_float_array(mats[i])
            if type_ptr_data is not None and chunk_type_ptr_data != type_ptr_data:
                raise ValueError('Input chunks must have same type')
            ptr_data[i] = chunk_ptr_data
            type_ptr_data = chunk_type_ptr_data
            holders.append(holder)

        self.handle = ctypes.c_void_p()
        _safe_call(_LIB.LGBM_DatasetCreateFromMats(
1725
            ctypes.c_int32(len(mats)),
1726
1727
1728
            ctypes.cast(ptr_data, ctypes.POINTER(ctypes.POINTER(ctypes.c_double))),
            ctypes.c_int(type_ptr_data),
            nrow.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
1729
            ctypes.c_int32(ncol),
1730
1731
1732
1733
            ctypes.c_int(C_API_IS_ROW_MAJOR),
            c_str(params_str),
            ref_dataset,
            ctypes.byref(self.handle)))
Nikita Titov's avatar
Nikita Titov committed
1734
        return self
1735

wxchan's avatar
wxchan committed
1736
    def __init_from_csr(self, csr, params_str, ref_dataset):
1737
        """Initialize data from a CSR matrix."""
wxchan's avatar
wxchan committed
1738
        if len(csr.indices) != len(csr.data):
1739
            raise ValueError(f'Length mismatch: {len(csr.indices)} vs {len(csr.data)}')
wxchan's avatar
wxchan committed
1740
1741
        self.handle = ctypes.c_void_p()

1742
1743
        ptr_indptr, type_ptr_indptr, __ = c_int_array(csr.indptr)
        ptr_data, type_ptr_data, _ = c_float_array(csr.data)
wxchan's avatar
wxchan committed
1744

1745
        assert csr.shape[1] <= MAX_INT32
1746
        csr_indices = csr.indices.astype(np.int32, copy=False)
1747

wxchan's avatar
wxchan committed
1748
1749
        _safe_call(_LIB.LGBM_DatasetCreateFromCSR(
            ptr_indptr,
Guolin Ke's avatar
Guolin Ke committed
1750
            ctypes.c_int(type_ptr_indptr),
1751
            csr_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
wxchan's avatar
wxchan committed
1752
            ptr_data,
Guolin Ke's avatar
Guolin Ke committed
1753
1754
1755
1756
            ctypes.c_int(type_ptr_data),
            ctypes.c_int64(len(csr.indptr)),
            ctypes.c_int64(len(csr.data)),
            ctypes.c_int64(csr.shape[1]),
wxchan's avatar
wxchan committed
1757
1758
1759
            c_str(params_str),
            ref_dataset,
            ctypes.byref(self.handle)))
Nikita Titov's avatar
Nikita Titov committed
1760
        return self
wxchan's avatar
wxchan committed
1761

Guolin Ke's avatar
Guolin Ke committed
1762
    def __init_from_csc(self, csc, params_str, ref_dataset):
1763
        """Initialize data from a CSC matrix."""
Guolin Ke's avatar
Guolin Ke committed
1764
        if len(csc.indices) != len(csc.data):
1765
            raise ValueError(f'Length mismatch: {len(csc.indices)} vs {len(csc.data)}')
Guolin Ke's avatar
Guolin Ke committed
1766
1767
        self.handle = ctypes.c_void_p()

1768
1769
        ptr_indptr, type_ptr_indptr, __ = c_int_array(csc.indptr)
        ptr_data, type_ptr_data, _ = c_float_array(csc.data)
Guolin Ke's avatar
Guolin Ke committed
1770

1771
        assert csc.shape[0] <= MAX_INT32
1772
        csc_indices = csc.indices.astype(np.int32, copy=False)
1773

Guolin Ke's avatar
Guolin Ke committed
1774
1775
        _safe_call(_LIB.LGBM_DatasetCreateFromCSC(
            ptr_indptr,
Guolin Ke's avatar
Guolin Ke committed
1776
            ctypes.c_int(type_ptr_indptr),
1777
            csc_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
Guolin Ke's avatar
Guolin Ke committed
1778
            ptr_data,
Guolin Ke's avatar
Guolin Ke committed
1779
1780
1781
1782
            ctypes.c_int(type_ptr_data),
            ctypes.c_int64(len(csc.indptr)),
            ctypes.c_int64(len(csc.data)),
            ctypes.c_int64(csc.shape[0]),
Guolin Ke's avatar
Guolin Ke committed
1783
1784
1785
            c_str(params_str),
            ref_dataset,
            ctypes.byref(self.handle)))
Nikita Titov's avatar
Nikita Titov committed
1786
        return self
Guolin Ke's avatar
Guolin Ke committed
1787

1788
    @staticmethod
1789
1790
1791
1792
1793
1794
    def _compare_params_for_warning(
        params: Optional[Dict[str, Any]],
        other_params: Optional[Dict[str, Any]],
        ignore_keys: Set[str]
    ) -> bool:
        """Compare two dictionaries with params ignoring some keys.
1795

1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
        It is only for the warning purpose.

        Parameters
        ----------
        params : dict or None
            One dictionary with parameters to compare.
        other_params : dict or None
            Another dictionary with parameters to compare.
        ignore_keys : set
            Keys that should be ignored during comparing two dictionaries.
1806
1807
1808

        Returns
        -------
1809
1810
        compare_result : bool
          Returns whether two dictionaries with params are equal.
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
        """
        if params is None:
            params = {}
        if other_params is None:
            other_params = {}
        for k in other_params:
            if k not in ignore_keys:
                if k not in params or params[k] != other_params[k]:
                    return False
        for k in params:
            if k not in ignore_keys:
                if k not in other_params or params[k] != other_params[k]:
                    return False
        return True

1826
    def construct(self) -> "Dataset":
1827
1828
1829
1830
1831
        """Lazy init.

        Returns
        -------
        self : Dataset
Nikita Titov's avatar
Nikita Titov committed
1832
            Constructed Dataset object.
1833
        """
1834
        if self.handle is None:
wxchan's avatar
wxchan committed
1835
            if self.reference is not None:
1836
                reference_params = self.reference.get_params()
1837
1838
                params = self.get_params()
                if params != reference_params:
1839
1840
1841
1842
1843
                    if not self._compare_params_for_warning(
                        params=params,
                        other_params=reference_params,
                        ignore_keys=_ConfigAliases.get("categorical_feature")
                    ):
1844
                        _log_warning('Overriding the parameters from Reference Dataset.')
1845
                    self._update_params(reference_params)
wxchan's avatar
wxchan committed
1846
                if self.used_indices is None:
1847
                    # create valid
1848
                    self._lazy_init(self.data, label=self.label, reference=self.reference,
1849
1850
                                    weight=self.weight, group=self.group,
                                    init_score=self.init_score, predictor=self._predictor,
1851
                                    feature_name=self.feature_name, params=self.params)
wxchan's avatar
wxchan committed
1852
                else:
1853
                    # construct subset
wxchan's avatar
wxchan committed
1854
                    used_indices = list_to_1d_numpy(self.used_indices, np.int32, name='used_indices')
1855
                    assert used_indices.flags.c_contiguous
Guolin Ke's avatar
Guolin Ke committed
1856
                    if self.reference.group is not None:
1857
                        group_info = np.array(self.reference.group).astype(np.int32, copy=False)
1858
                        _, self.group = np.unique(np.repeat(range(len(group_info)), repeats=group_info)[self.used_indices],
1859
                                                  return_counts=True)
1860
                    self.handle = ctypes.c_void_p()
wxchan's avatar
wxchan committed
1861
1862
                    params_str = param_dict_to_str(self.params)
                    _safe_call(_LIB.LGBM_DatasetGetSubset(
1863
                        self.reference.construct().handle,
wxchan's avatar
wxchan committed
1864
                        used_indices.ctypes.data_as(ctypes.POINTER(ctypes.c_int32)),
1865
                        ctypes.c_int32(used_indices.shape[0]),
wxchan's avatar
wxchan committed
1866
1867
                        c_str(params_str),
                        ctypes.byref(self.handle)))
Guolin Ke's avatar
Guolin Ke committed
1868
1869
                    if not self.free_raw_data:
                        self.get_data()
Guolin Ke's avatar
Guolin Ke committed
1870
1871
                    if self.group is not None:
                        self.set_group(self.group)
wxchan's avatar
wxchan committed
1872
1873
                    if self.get_label() is None:
                        raise ValueError("Label should not be None.")
Guolin Ke's avatar
Guolin Ke committed
1874
1875
1876
                    if isinstance(self._predictor, _InnerPredictor) and self._predictor is not self.reference._predictor:
                        self.get_data()
                        self._set_init_score_by_predictor(self._predictor, self.data, used_indices)
wxchan's avatar
wxchan committed
1877
            else:
1878
                # create train
1879
                self._lazy_init(self.data, label=self.label,
1880
1881
                                weight=self.weight, group=self.group,
                                init_score=self.init_score, predictor=self._predictor,
1882
                                feature_name=self.feature_name, categorical_feature=self.categorical_feature, params=self.params)
wxchan's avatar
wxchan committed
1883
1884
            if self.free_raw_data:
                self.data = None
1885
            self.feature_name = self.get_feature_name()
wxchan's avatar
wxchan committed
1886
        return self
wxchan's avatar
wxchan committed
1887

1888
1889
1890
1891
1892
1893
1894
1895
1896
    def create_valid(
        self,
        data,
        label=None,
        weight=None,
        group=None,
        init_score=None,
        params: Optional[Dict[str, Any]] = None
    ) -> "Dataset":
1897
        """Create validation data align with current Dataset.
wxchan's avatar
wxchan committed
1898
1899
1900

        Parameters
        ----------
1901
        data : str, pathlib.Path, numpy array, pandas DataFrame, H2O DataTable's Frame, scipy.sparse, Sequence, list of Sequence or list of numpy array
wxchan's avatar
wxchan committed
1902
            Data source of Dataset.
1903
            If str or pathlib.Path, it represents the path to a text file (CSV, TSV, or LibSVM) or a LightGBM Dataset binary file.
1904
        label : list, numpy 1-D array, pandas Series / one-column DataFrame or None, optional (default=None)
1905
1906
            Label of the data.
        weight : list, numpy 1-D array, pandas Series or None, optional (default=None)
1907
            Weight for each instance. Weights should be non-negative.
1908
        group : list, numpy 1-D array, pandas Series or None, optional (default=None)
1909
1910
1911
            Group/query data.
            Only used in the learning-to-rank task.
            sum(group) = n_samples.
1912
1913
            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.
1914
        init_score : list, list of lists (for multi-class task), numpy array, pandas Series, pandas DataFrame (for multi-class task), or None, optional (default=None)
1915
            Init score for Dataset.
Nikita Titov's avatar
Nikita Titov committed
1916
        params : dict or None, optional (default=None)
1917
            Other parameters for validation Dataset.
1918
1919
1920

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
1921
1922
        valid : Dataset
            Validation Dataset with reference to self.
wxchan's avatar
wxchan committed
1923
        """
1924
        ret = Dataset(data, label=label, reference=self,
1925
                      weight=weight, group=group, init_score=init_score,
1926
                      params=params, free_raw_data=self.free_raw_data)
wxchan's avatar
wxchan committed
1927
        ret._predictor = self._predictor
1928
        ret.pandas_categorical = self.pandas_categorical
wxchan's avatar
wxchan committed
1929
        return ret
wxchan's avatar
wxchan committed
1930

1931
1932
1933
1934
1935
    def subset(
        self,
        used_indices: List[int],
        params: Optional[Dict[str, Any]] = None
    ) -> "Dataset":
1936
        """Get subset of current Dataset.
wxchan's avatar
wxchan committed
1937
1938
1939
1940

        Parameters
        ----------
        used_indices : list of int
1941
            Indices used to create the subset.
Nikita Titov's avatar
Nikita Titov committed
1942
        params : dict or None, optional (default=None)
1943
            These parameters will be passed to Dataset constructor.
1944
1945
1946
1947
1948

        Returns
        -------
        subset : Dataset
            Subset of the current Dataset.
wxchan's avatar
wxchan committed
1949
        """
wxchan's avatar
wxchan committed
1950
1951
        if params is None:
            params = self.params
wxchan's avatar
wxchan committed
1952
        ret = Dataset(None, reference=self, feature_name=self.feature_name,
1953
1954
                      categorical_feature=self.categorical_feature, params=params,
                      free_raw_data=self.free_raw_data)
wxchan's avatar
wxchan committed
1955
        ret._predictor = self._predictor
1956
        ret.pandas_categorical = self.pandas_categorical
1957
        ret.used_indices = sorted(used_indices)
wxchan's avatar
wxchan committed
1958
1959
        return ret

1960
    def save_binary(self, filename: Union[str, Path]) -> "Dataset":
1961
        """Save Dataset to a binary file.
wxchan's avatar
wxchan committed
1962

1963
1964
1965
1966
1967
        .. note::

            Please note that `init_score` is not saved in binary file.
            If you need it, please set it again after loading Dataset.

wxchan's avatar
wxchan committed
1968
1969
        Parameters
        ----------
1970
        filename : str or pathlib.Path
wxchan's avatar
wxchan committed
1971
            Name of the output file.
Nikita Titov's avatar
Nikita Titov committed
1972
1973
1974
1975
1976

        Returns
        -------
        self : Dataset
            Returns self.
wxchan's avatar
wxchan committed
1977
1978
1979
        """
        _safe_call(_LIB.LGBM_DatasetSaveBinary(
            self.construct().handle,
1980
            c_str(str(filename))))
Nikita Titov's avatar
Nikita Titov committed
1981
        return self
wxchan's avatar
wxchan committed
1982

1983
    def _update_params(self, params: Optional[Dict[str, Any]]) -> "Dataset":
1984
1985
        if not params:
            return self
1986
        params = deepcopy(params)
1987
1988
1989
1990
1991

        def update():
            if not self.params:
                self.params = params
            else:
1992
                self.params_back_up = deepcopy(self.params)
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
                self.params.update(params)

        if self.handle is None:
            update()
        elif params is not None:
            ret = _LIB.LGBM_DatasetUpdateParamChecking(
                c_str(param_dict_to_str(self.params)),
                c_str(param_dict_to_str(params)))
            if ret != 0:
                # could be updated if data is not freed
                if self.data is not None:
                    update()
                    self._free_handle()
                else:
2007
                    raise LightGBMError(_LIB.LGBM_GetLastError().decode('utf-8'))
Nikita Titov's avatar
Nikita Titov committed
2008
        return self
wxchan's avatar
wxchan committed
2009

2010
    def _reverse_update_params(self) -> "Dataset":
2011
        if self.handle is None:
2012
            self.params = deepcopy(self.params_back_up)
2013
            self.params_back_up = None
Nikita Titov's avatar
Nikita Titov committed
2014
        return self
2015

2016
2017
2018
2019
2020
    def set_field(
        self,
        field_name: str,
        data
    ) -> "Dataset":
wxchan's avatar
wxchan committed
2021
        """Set property into the Dataset.
wxchan's avatar
wxchan committed
2022
2023
2024

        Parameters
        ----------
2025
        field_name : str
2026
            The field name of the information.
2027
2028
        data : list, list of lists (for multi-class task), numpy array, pandas Series, pandas DataFrame (for multi-class task), or None
            The data to be set.
Nikita Titov's avatar
Nikita Titov committed
2029
2030
2031
2032
2033

        Returns
        -------
        self : Dataset
            Dataset with set property.
wxchan's avatar
wxchan committed
2034
        """
2035
        if self.handle is None:
2036
            raise Exception(f"Cannot set {field_name} before construct dataset")
wxchan's avatar
wxchan committed
2037
        if data is None:
2038
            # set to None
wxchan's avatar
wxchan committed
2039
2040
2041
2042
            _safe_call(_LIB.LGBM_DatasetSetField(
                self.handle,
                c_str(field_name),
                None,
Guolin Ke's avatar
Guolin Ke committed
2043
2044
                ctypes.c_int(0),
                ctypes.c_int(FIELD_TYPE_MAPPER[field_name])))
Nikita Titov's avatar
Nikita Titov committed
2045
            return self
2046
        if field_name == 'init_score':
Guolin Ke's avatar
Guolin Ke committed
2047
            dtype = np.float64
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
            if _is_1d_collection(data):
                data = list_to_1d_numpy(data, dtype, name=field_name)
            elif _is_2d_collection(data):
                data = _data_to_2d_numpy(data, dtype, name=field_name)
                data = data.ravel(order='F')
            else:
                raise TypeError(
                    'init_score must be list, numpy 1-D array or pandas Series.\n'
                    'In multiclass classification init_score can also be a list of lists, numpy 2-D array or pandas DataFrame.'
                )
        else:
            dtype = np.int32 if field_name == 'group' else np.float32
            data = list_to_1d_numpy(data, dtype, name=field_name)

2062
2063
        if data.dtype == np.float32 or data.dtype == np.float64:
            ptr_data, type_data, _ = c_float_array(data)
wxchan's avatar
wxchan committed
2064
        elif data.dtype == np.int32:
2065
            ptr_data, type_data, _ = c_int_array(data)
wxchan's avatar
wxchan committed
2066
        else:
2067
            raise TypeError(f"Expected np.float32/64 or np.int32, met type({data.dtype})")
wxchan's avatar
wxchan committed
2068
        if type_data != FIELD_TYPE_MAPPER[field_name]:
2069
            raise TypeError("Input type error for set_field")
wxchan's avatar
wxchan committed
2070
2071
2072
2073
        _safe_call(_LIB.LGBM_DatasetSetField(
            self.handle,
            c_str(field_name),
            ptr_data,
Guolin Ke's avatar
Guolin Ke committed
2074
2075
            ctypes.c_int(len(data)),
            ctypes.c_int(type_data)))
2076
        self.version += 1
Nikita Titov's avatar
Nikita Titov committed
2077
        return self
wxchan's avatar
wxchan committed
2078

2079
    def get_field(self, field_name: str) -> Optional[np.ndarray]:
wxchan's avatar
wxchan committed
2080
        """Get property from the Dataset.
wxchan's avatar
wxchan committed
2081
2082
2083

        Parameters
        ----------
2084
        field_name : str
2085
            The field name of the information.
wxchan's avatar
wxchan committed
2086
2087
2088

        Returns
        -------
2089
        info : numpy array or None
2090
            A numpy array with information from the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2091
        """
2092
        if self.handle is None:
2093
            raise Exception(f"Cannot get {field_name} before construct Dataset")
2094
2095
        tmp_out_len = ctypes.c_int(0)
        out_type = ctypes.c_int(0)
wxchan's avatar
wxchan committed
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
        ret = ctypes.POINTER(ctypes.c_void_p)()
        _safe_call(_LIB.LGBM_DatasetGetField(
            self.handle,
            c_str(field_name),
            ctypes.byref(tmp_out_len),
            ctypes.byref(ret),
            ctypes.byref(out_type)))
        if out_type.value != FIELD_TYPE_MAPPER[field_name]:
            raise TypeError("Return type error for get_field")
        if tmp_out_len.value == 0:
            return None
        if out_type.value == C_API_DTYPE_INT32:
2108
            arr = cint32_array_to_numpy(ctypes.cast(ret, ctypes.POINTER(ctypes.c_int32)), tmp_out_len.value)
wxchan's avatar
wxchan committed
2109
        elif out_type.value == C_API_DTYPE_FLOAT32:
2110
            arr = cfloat32_array_to_numpy(ctypes.cast(ret, ctypes.POINTER(ctypes.c_float)), tmp_out_len.value)
Guolin Ke's avatar
Guolin Ke committed
2111
        elif out_type.value == C_API_DTYPE_FLOAT64:
2112
            arr = cfloat64_array_to_numpy(ctypes.cast(ret, ctypes.POINTER(ctypes.c_double)), tmp_out_len.value)
2113
        else:
wxchan's avatar
wxchan committed
2114
            raise TypeError("Unknown type")
2115
2116
2117
2118
2119
2120
        if field_name == 'init_score':
            num_data = self.num_data()
            num_classes = arr.size // num_data
            if num_classes > 1:
                arr = arr.reshape((num_data, num_classes), order='F')
        return arr
Guolin Ke's avatar
Guolin Ke committed
2121

2122
2123
2124
2125
    def set_categorical_feature(
        self,
        categorical_feature: Union[List[int], List[str]]
    ) -> "Dataset":
2126
        """Set categorical features.
2127
2128
2129

        Parameters
        ----------
2130
        categorical_feature : list of int or str
2131
            Names or indices of categorical features.
Nikita Titov's avatar
Nikita Titov committed
2132
2133
2134
2135
2136

        Returns
        -------
        self : Dataset
            Dataset with set categorical features.
2137
2138
        """
        if self.categorical_feature == categorical_feature:
Nikita Titov's avatar
Nikita Titov committed
2139
            return self
2140
        if self.data is not None:
2141
2142
            if self.categorical_feature is None:
                self.categorical_feature = categorical_feature
Nikita Titov's avatar
Nikita Titov committed
2143
                return self._free_handle()
2144
            elif categorical_feature == 'auto':
Nikita Titov's avatar
Nikita Titov committed
2145
                return self
2146
            else:
2147
2148
2149
                if self.categorical_feature != 'auto':
                    _log_warning('categorical_feature in Dataset is overridden.\n'
                                 f'New categorical_feature is {sorted(list(categorical_feature))}')
2150
                self.categorical_feature = categorical_feature
Nikita Titov's avatar
Nikita Titov committed
2151
                return self._free_handle()
2152
        else:
2153
2154
            raise LightGBMError("Cannot set categorical feature after freed raw data, "
                                "set free_raw_data=False when construct Dataset to avoid this.")
2155

2156
2157
2158
2159
    def _set_predictor(
        self,
        predictor: Optional[_InnerPredictor]
    ) -> "Dataset":
2160
2161
2162
2163
        """Set predictor for continued training.

        It is not recommended for user to call this function.
        Please use init_model argument in engine.train() or engine.cv() instead.
Guolin Ke's avatar
Guolin Ke committed
2164
        """
2165
        if predictor is self._predictor and (predictor is None or predictor.current_iteration() == self._predictor.current_iteration()):
Nikita Titov's avatar
Nikita Titov committed
2166
            return self
2167
        if self.handle is None:
Guolin Ke's avatar
Guolin Ke committed
2168
            self._predictor = predictor
2169
2170
2171
2172
2173
2174
        elif self.data is not None:
            self._predictor = predictor
            self._set_init_score_by_predictor(self._predictor, self.data)
        elif self.used_indices is not None and self.reference is not None and self.reference.data is not None:
            self._predictor = predictor
            self._set_init_score_by_predictor(self._predictor, self.reference.data, self.used_indices)
Guolin Ke's avatar
Guolin Ke committed
2175
        else:
2176
2177
            raise LightGBMError("Cannot set predictor after freed raw data, "
                                "set free_raw_data=False when construct Dataset to avoid this.")
2178
        return self
Guolin Ke's avatar
Guolin Ke committed
2179

2180
    def set_reference(self, reference: "Dataset") -> "Dataset":
2181
        """Set reference Dataset.
Guolin Ke's avatar
Guolin Ke committed
2182
2183
2184
2185

        Parameters
        ----------
        reference : Dataset
2186
            Reference that is used as a template to construct the current Dataset.
Nikita Titov's avatar
Nikita Titov committed
2187
2188
2189
2190
2191

        Returns
        -------
        self : Dataset
            Dataset with set reference.
Guolin Ke's avatar
Guolin Ke committed
2192
        """
2193
2194
2195
        self.set_categorical_feature(reference.categorical_feature) \
            .set_feature_name(reference.feature_name) \
            ._set_predictor(reference._predictor)
2196
        # we're done if self and reference share a common upstream reference
2197
        if self.get_ref_chain().intersection(reference.get_ref_chain()):
Nikita Titov's avatar
Nikita Titov committed
2198
            return self
Guolin Ke's avatar
Guolin Ke committed
2199
2200
        if self.data is not None:
            self.reference = reference
Nikita Titov's avatar
Nikita Titov committed
2201
            return self._free_handle()
Guolin Ke's avatar
Guolin Ke committed
2202
        else:
2203
2204
            raise LightGBMError("Cannot set reference after freed raw data, "
                                "set free_raw_data=False when construct Dataset to avoid this.")
Guolin Ke's avatar
Guolin Ke committed
2205

2206
    def set_feature_name(self, feature_name: List[str]) -> "Dataset":
2207
        """Set feature name.
Guolin Ke's avatar
Guolin Ke committed
2208
2209
2210

        Parameters
        ----------
2211
        feature_name : list of str
2212
            Feature names.
Nikita Titov's avatar
Nikita Titov committed
2213
2214
2215
2216
2217

        Returns
        -------
        self : Dataset
            Dataset with set feature name.
Guolin Ke's avatar
Guolin Ke committed
2218
        """
2219
2220
        if feature_name != 'auto':
            self.feature_name = feature_name
2221
        if self.handle is not None and feature_name is not None and feature_name != 'auto':
wxchan's avatar
wxchan committed
2222
            if len(feature_name) != self.num_feature():
2223
                raise ValueError(f"Length of feature_name({len(feature_name)}) and num_feature({self.num_feature()}) don't match")
2224
            c_feature_name = [c_str(name) for name in feature_name]
wxchan's avatar
wxchan committed
2225
2226
2227
            _safe_call(_LIB.LGBM_DatasetSetFeatureNames(
                self.handle,
                c_array(ctypes.c_char_p, c_feature_name),
Guolin Ke's avatar
Guolin Ke committed
2228
                ctypes.c_int(len(feature_name))))
Nikita Titov's avatar
Nikita Titov committed
2229
        return self
Guolin Ke's avatar
Guolin Ke committed
2230

2231
    def set_label(self, label) -> "Dataset":
2232
        """Set label of Dataset.
Guolin Ke's avatar
Guolin Ke committed
2233
2234
2235

        Parameters
        ----------
2236
        label : list, numpy 1-D array, pandas Series / one-column DataFrame or None
2237
            The label information to be set into Dataset.
Nikita Titov's avatar
Nikita Titov committed
2238
2239
2240
2241
2242

        Returns
        -------
        self : Dataset
            Dataset with set label.
Guolin Ke's avatar
Guolin Ke committed
2243
2244
        """
        self.label = label
2245
        if self.handle is not None:
2246
            label = list_to_1d_numpy(_label_from_pandas(label), name='label')
wxchan's avatar
wxchan committed
2247
            self.set_field('label', label)
2248
            self.label = self.get_field('label')  # original values can be modified at cpp side
Nikita Titov's avatar
Nikita Titov committed
2249
        return self
Guolin Ke's avatar
Guolin Ke committed
2250

2251
    def set_weight(self, weight) -> "Dataset":
2252
        """Set weight of each instance.
Guolin Ke's avatar
Guolin Ke committed
2253
2254
2255

        Parameters
        ----------
2256
        weight : list, numpy 1-D array, pandas Series or None
2257
            Weight to be set for each data point. Weights should be non-negative.
Nikita Titov's avatar
Nikita Titov committed
2258
2259
2260
2261
2262

        Returns
        -------
        self : Dataset
            Dataset with set weight.
Guolin Ke's avatar
Guolin Ke committed
2263
        """
2264
2265
        if weight is not None and np.all(weight == 1):
            weight = None
Guolin Ke's avatar
Guolin Ke committed
2266
        self.weight = weight
2267
        if self.handle is not None and weight is not None:
wxchan's avatar
wxchan committed
2268
2269
            weight = list_to_1d_numpy(weight, name='weight')
            self.set_field('weight', weight)
2270
            self.weight = self.get_field('weight')  # original values can be modified at cpp side
Nikita Titov's avatar
Nikita Titov committed
2271
        return self
Guolin Ke's avatar
Guolin Ke committed
2272

2273
    def set_init_score(self, init_score) -> "Dataset":
2274
        """Set init score of Booster to start from.
Guolin Ke's avatar
Guolin Ke committed
2275
2276
2277

        Parameters
        ----------
2278
        init_score : list, list of lists (for multi-class task), numpy array, pandas Series, pandas DataFrame (for multi-class task), or None
2279
            Init score for Booster.
Nikita Titov's avatar
Nikita Titov committed
2280
2281
2282
2283
2284

        Returns
        -------
        self : Dataset
            Dataset with set init score.
Guolin Ke's avatar
Guolin Ke committed
2285
2286
        """
        self.init_score = init_score
2287
        if self.handle is not None and init_score is not None:
wxchan's avatar
wxchan committed
2288
            self.set_field('init_score', init_score)
2289
            self.init_score = self.get_field('init_score')  # original values can be modified at cpp side
Nikita Titov's avatar
Nikita Titov committed
2290
        return self
Guolin Ke's avatar
Guolin Ke committed
2291

2292
    def set_group(self, group) -> "Dataset":
2293
        """Set group size of Dataset (used for ranking).
Guolin Ke's avatar
Guolin Ke committed
2294
2295
2296

        Parameters
        ----------
2297
        group : list, numpy 1-D array, pandas Series or None
2298
2299
2300
            Group/query data.
            Only used in the learning-to-rank task.
            sum(group) = n_samples.
2301
2302
            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.
Nikita Titov's avatar
Nikita Titov committed
2303
2304
2305
2306
2307

        Returns
        -------
        self : Dataset
            Dataset with set group.
Guolin Ke's avatar
Guolin Ke committed
2308
2309
        """
        self.group = group
2310
        if self.handle is not None and group is not None:
wxchan's avatar
wxchan committed
2311
2312
            group = list_to_1d_numpy(group, np.int32, name='group')
            self.set_field('group', group)
Nikita Titov's avatar
Nikita Titov committed
2313
        return self
Guolin Ke's avatar
Guolin Ke committed
2314

2315
    def get_feature_name(self) -> List[str]:
2316
2317
2318
2319
        """Get the names of columns (features) in the Dataset.

        Returns
        -------
2320
        feature_names : list of str
2321
2322
2323
2324
2325
2326
2327
2328
            The names of columns (features) in the Dataset.
        """
        if self.handle is None:
            raise LightGBMError("Cannot get feature_name before construct dataset")
        num_feature = self.num_feature()
        tmp_out_len = ctypes.c_int(0)
        reserved_string_buffer_size = 255
        required_string_buffer_size = ctypes.c_size_t(0)
2329
        string_buffers = [ctypes.create_string_buffer(reserved_string_buffer_size) for _ in range(num_feature)]
2330
2331
2332
        ptr_string_buffers = (ctypes.c_char_p * num_feature)(*map(ctypes.addressof, string_buffers))
        _safe_call(_LIB.LGBM_DatasetGetFeatureNames(
            self.handle,
2333
            ctypes.c_int(num_feature),
2334
            ctypes.byref(tmp_out_len),
2335
            ctypes.c_size_t(reserved_string_buffer_size),
2336
2337
2338
2339
            ctypes.byref(required_string_buffer_size),
            ptr_string_buffers))
        if num_feature != tmp_out_len.value:
            raise ValueError("Length of feature names doesn't equal with num_feature")
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
        actual_string_buffer_size = required_string_buffer_size.value
        # if buffer length is not long enough, reallocate buffers
        if reserved_string_buffer_size < actual_string_buffer_size:
            string_buffers = [ctypes.create_string_buffer(actual_string_buffer_size) for _ in range(num_feature)]
            ptr_string_buffers = (ctypes.c_char_p * num_feature)(*map(ctypes.addressof, string_buffers))
            _safe_call(_LIB.LGBM_DatasetGetFeatureNames(
                self.handle,
                ctypes.c_int(num_feature),
                ctypes.byref(tmp_out_len),
                ctypes.c_size_t(actual_string_buffer_size),
                ctypes.byref(required_string_buffer_size),
                ptr_string_buffers))
2352
        return [string_buffers[i].value.decode('utf-8') for i in range(num_feature)]
2353

2354
    def get_label(self) -> Optional[np.ndarray]:
2355
        """Get the label of the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2356
2357
2358

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
2359
        label : numpy array or None
2360
            The label information from the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2361
        """
2362
        if self.label is None:
wxchan's avatar
wxchan committed
2363
            self.label = self.get_field('label')
Guolin Ke's avatar
Guolin Ke committed
2364
2365
        return self.label

2366
    def get_weight(self) -> Optional[np.ndarray]:
2367
        """Get the weight of the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2368
2369
2370

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
2371
        weight : numpy array or None
2372
            Weight for each data point from the Dataset. Weights should be non-negative.
Guolin Ke's avatar
Guolin Ke committed
2373
        """
2374
        if self.weight is None:
wxchan's avatar
wxchan committed
2375
            self.weight = self.get_field('weight')
Guolin Ke's avatar
Guolin Ke committed
2376
2377
        return self.weight

2378
    def get_init_score(self) -> Optional[np.ndarray]:
2379
        """Get the initial score of the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2380
2381
2382

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
2383
        init_score : numpy array or None
2384
            Init score of Booster.
Guolin Ke's avatar
Guolin Ke committed
2385
        """
2386
        if self.init_score is None:
wxchan's avatar
wxchan committed
2387
            self.init_score = self.get_field('init_score')
Guolin Ke's avatar
Guolin Ke committed
2388
2389
        return self.init_score

2390
2391
2392
2393
2394
    def get_data(self):
        """Get the raw data of the Dataset.

        Returns
        -------
2395
        data : str, pathlib.Path, numpy array, pandas DataFrame, H2O DataTable's Frame, scipy.sparse, Sequence, list of Sequence or list of numpy array or None
2396
2397
2398
2399
            Raw data used in the Dataset construction.
        """
        if self.handle is None:
            raise Exception("Cannot get data before construct Dataset")
Guolin Ke's avatar
Guolin Ke committed
2400
2401
2402
2403
2404
        if self.need_slice and self.used_indices is not None and self.reference is not None:
            self.data = self.reference.data
            if self.data is not None:
                if isinstance(self.data, np.ndarray) or scipy.sparse.issparse(self.data):
                    self.data = self.data[self.used_indices, :]
2405
                elif isinstance(self.data, pd_DataFrame):
Guolin Ke's avatar
Guolin Ke committed
2406
                    self.data = self.data.iloc[self.used_indices].copy()
2407
                elif isinstance(self.data, dt_DataTable):
Guolin Ke's avatar
Guolin Ke committed
2408
                    self.data = self.data[self.used_indices, :]
2409
2410
2411
2412
                elif isinstance(self.data, Sequence):
                    self.data = self.data[self.used_indices]
                elif isinstance(self.data, list) and len(self.data) > 0 and all(isinstance(x, Sequence) for x in self.data):
                    self.data = np.array([row for row in self._yield_row_from_seqlist(self.data, self.used_indices)])
Guolin Ke's avatar
Guolin Ke committed
2413
                else:
2414
2415
                    _log_warning(f"Cannot subset {type(self.data).__name__} type of raw data.\n"
                                 "Returning original raw data")
2416
            self.need_slice = False
Guolin Ke's avatar
Guolin Ke committed
2417
2418
2419
        if self.data is None:
            raise LightGBMError("Cannot call `get_data` after freed raw data, "
                                "set free_raw_data=False when construct Dataset to avoid this.")
2420
2421
        return self.data

Guolin Ke's avatar
Guolin Ke committed
2422
    def get_group(self):
2423
        """Get the group of the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2424
2425
2426

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
2427
        group : numpy array or None
2428
2429
2430
            Group/query data.
            Only used in the learning-to-rank task.
            sum(group) = n_samples.
2431
2432
            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.
Guolin Ke's avatar
Guolin Ke committed
2433
        """
2434
        if self.group is None:
wxchan's avatar
wxchan committed
2435
            self.group = self.get_field('group')
Guolin Ke's avatar
Guolin Ke committed
2436
2437
            if self.group is not None:
                # group data from LightGBM is boundaries data, need to convert to group size
Nikita Titov's avatar
Nikita Titov committed
2438
                self.group = np.diff(self.group)
Guolin Ke's avatar
Guolin Ke committed
2439
2440
        return self.group

2441
    def num_data(self) -> int:
2442
        """Get the number of rows in the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2443
2444
2445

        Returns
        -------
2446
2447
        number_of_rows : int
            The number of rows in the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2448
        """
2449
        if self.handle is not None:
2450
            ret = ctypes.c_int(0)
wxchan's avatar
wxchan committed
2451
2452
2453
            _safe_call(_LIB.LGBM_DatasetGetNumData(self.handle,
                                                   ctypes.byref(ret)))
            return ret.value
Guolin Ke's avatar
Guolin Ke committed
2454
        else:
2455
            raise LightGBMError("Cannot get num_data before construct dataset")
Guolin Ke's avatar
Guolin Ke committed
2456

2457
    def num_feature(self) -> int:
2458
        """Get the number of columns (features) in the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2459
2460
2461

        Returns
        -------
2462
2463
        number_of_columns : int
            The number of columns (features) in the Dataset.
Guolin Ke's avatar
Guolin Ke committed
2464
        """
2465
        if self.handle is not None:
2466
            ret = ctypes.c_int(0)
wxchan's avatar
wxchan committed
2467
2468
2469
            _safe_call(_LIB.LGBM_DatasetGetNumFeature(self.handle,
                                                      ctypes.byref(ret)))
            return ret.value
Guolin Ke's avatar
Guolin Ke committed
2470
        else:
2471
            raise LightGBMError("Cannot get num_feature before construct dataset")
Guolin Ke's avatar
Guolin Ke committed
2472

2473
    def feature_num_bin(self, feature: Union[int, str]) -> int:
2474
2475
2476
2477
        """Get the number of bins for a feature.

        Parameters
        ----------
2478
2479
        feature : int or str
            Index or name of the feature.
2480
2481
2482
2483
2484
2485
2486

        Returns
        -------
        number_of_bins : int
            The number of constructed bins for the feature in the Dataset.
        """
        if self.handle is not None:
2487
2488
            if isinstance(feature, str):
                feature = self.feature_name.index(feature)
2489
2490
2491
2492
2493
2494
2495
2496
            ret = ctypes.c_int(0)
            _safe_call(_LIB.LGBM_DatasetGetFeatureNumBin(self.handle,
                                                         ctypes.c_int(feature),
                                                         ctypes.byref(ret)))
            return ret.value
        else:
            raise LightGBMError("Cannot get feature_num_bin before construct dataset")

2497
    def get_ref_chain(self, ref_limit: int = 100) -> Set["Dataset"]:
2498
2499
2500
2501
2502
        """Get a chain of Dataset objects.

        Starts with r, then goes to r.reference (if exists),
        then to r.reference.reference, etc.
        until we hit ``ref_limit`` or a reference loop.
2503
2504
2505
2506
2507

        Parameters
        ----------
        ref_limit : int, optional (default=100)
            The limit number of references.
2508
2509
2510

        Returns
        -------
2511
2512
2513
        ref_chain : set of Dataset
            Chain of references of the Datasets.
        """
2514
        head = self
2515
        ref_chain = set()
2516
2517
        while len(ref_chain) < ref_limit:
            if isinstance(head, Dataset):
2518
                ref_chain.add(head)
2519
2520
2521
2522
2523
2524
                if (head.reference is not None) and (head.reference not in ref_chain):
                    head = head.reference
                else:
                    break
            else:
                break
Nikita Titov's avatar
Nikita Titov committed
2525
        return ref_chain
2526

2527
    def add_features_from(self, other: "Dataset") -> "Dataset":
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
        """Add features from other Dataset to the current Dataset.

        Both Datasets must be constructed before calling this method.

        Parameters
        ----------
        other : Dataset
            The Dataset to take features from.

        Returns
        -------
        self : Dataset
            Dataset with the new features added.
        """
        if self.handle is None or other.handle is None:
            raise ValueError('Both source and target Datasets must be constructed before adding features')
        _safe_call(_LIB.LGBM_DatasetAddFeaturesFrom(self.handle, other.handle))
Guolin Ke's avatar
Guolin Ke committed
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
        was_none = self.data is None
        old_self_data_type = type(self.data).__name__
        if other.data is None:
            self.data = None
        elif self.data is not None:
            if isinstance(self.data, np.ndarray):
                if isinstance(other.data, np.ndarray):
                    self.data = np.hstack((self.data, other.data))
                elif scipy.sparse.issparse(other.data):
                    self.data = np.hstack((self.data, other.data.toarray()))
2555
                elif isinstance(other.data, pd_DataFrame):
Guolin Ke's avatar
Guolin Ke committed
2556
                    self.data = np.hstack((self.data, other.data.values))
2557
                elif isinstance(other.data, dt_DataTable):
Guolin Ke's avatar
Guolin Ke committed
2558
2559
2560
2561
2562
2563
2564
                    self.data = np.hstack((self.data, other.data.to_numpy()))
                else:
                    self.data = None
            elif scipy.sparse.issparse(self.data):
                sparse_format = self.data.getformat()
                if isinstance(other.data, np.ndarray) or scipy.sparse.issparse(other.data):
                    self.data = scipy.sparse.hstack((self.data, other.data), format=sparse_format)
2565
                elif isinstance(other.data, pd_DataFrame):
Guolin Ke's avatar
Guolin Ke committed
2566
                    self.data = scipy.sparse.hstack((self.data, other.data.values), format=sparse_format)
2567
                elif isinstance(other.data, dt_DataTable):
Guolin Ke's avatar
Guolin Ke committed
2568
2569
2570
                    self.data = scipy.sparse.hstack((self.data, other.data.to_numpy()), format=sparse_format)
                else:
                    self.data = None
2571
            elif isinstance(self.data, pd_DataFrame):
Guolin Ke's avatar
Guolin Ke committed
2572
2573
                if not PANDAS_INSTALLED:
                    raise LightGBMError("Cannot add features to DataFrame type of raw data "
2574
2575
                                        "without pandas installed. "
                                        "Install pandas and restart your session.")
Guolin Ke's avatar
Guolin Ke committed
2576
                if isinstance(other.data, np.ndarray):
2577
                    self.data = concat((self.data, pd_DataFrame(other.data)),
Guolin Ke's avatar
Guolin Ke committed
2578
2579
                                       axis=1, ignore_index=True)
                elif scipy.sparse.issparse(other.data):
2580
                    self.data = concat((self.data, pd_DataFrame(other.data.toarray())),
Guolin Ke's avatar
Guolin Ke committed
2581
                                       axis=1, ignore_index=True)
2582
                elif isinstance(other.data, pd_DataFrame):
Guolin Ke's avatar
Guolin Ke committed
2583
2584
                    self.data = concat((self.data, other.data),
                                       axis=1, ignore_index=True)
2585
2586
                elif isinstance(other.data, dt_DataTable):
                    self.data = concat((self.data, pd_DataFrame(other.data.to_numpy())),
Guolin Ke's avatar
Guolin Ke committed
2587
2588
2589
                                       axis=1, ignore_index=True)
                else:
                    self.data = None
2590
            elif isinstance(self.data, dt_DataTable):
Guolin Ke's avatar
Guolin Ke committed
2591
                if isinstance(other.data, np.ndarray):
2592
                    self.data = dt_DataTable(np.hstack((self.data.to_numpy(), other.data)))
Guolin Ke's avatar
Guolin Ke committed
2593
                elif scipy.sparse.issparse(other.data):
2594
2595
2596
2597
2598
                    self.data = dt_DataTable(np.hstack((self.data.to_numpy(), other.data.toarray())))
                elif isinstance(other.data, pd_DataFrame):
                    self.data = dt_DataTable(np.hstack((self.data.to_numpy(), other.data.values)))
                elif isinstance(other.data, dt_DataTable):
                    self.data = dt_DataTable(np.hstack((self.data.to_numpy(), other.data.to_numpy())))
Guolin Ke's avatar
Guolin Ke committed
2599
2600
2601
2602
2603
                else:
                    self.data = None
            else:
                self.data = None
        if self.data is None:
2604
2605
            err_msg = (f"Cannot add features from {type(other.data).__name__} type of raw data to "
                       f"{old_self_data_type} type of raw data.\n")
Guolin Ke's avatar
Guolin Ke committed
2606
2607
            err_msg += ("Set free_raw_data=False when construct Dataset to avoid this"
                        if was_none else "Freeing raw data")
2608
            _log_warning(err_msg)
Guolin Ke's avatar
Guolin Ke committed
2609
        self.feature_name = self.get_feature_name()
2610
2611
        _log_warning("Reseting categorical features.\n"
                     "You can set new categorical features via ``set_categorical_feature`` method")
Guolin Ke's avatar
Guolin Ke committed
2612
2613
        self.categorical_feature = "auto"
        self.pandas_categorical = None
2614
2615
        return self

2616
    def _dump_text(self, filename: Union[str, Path]) -> "Dataset":
2617
2618
2619
2620
2621
2622
        """Save Dataset to a text file.

        This format cannot be loaded back in by LightGBM, but is useful for debugging purposes.

        Parameters
        ----------
2623
        filename : str or pathlib.Path
2624
2625
2626
2627
2628
2629
2630
2631
2632
            Name of the output file.

        Returns
        -------
        self : Dataset
            Returns self.
        """
        _safe_call(_LIB.LGBM_DatasetDumpText(
            self.construct().handle,
2633
            c_str(str(filename))))
2634
2635
        return self

wxchan's avatar
wxchan committed
2636

2637
2638
2639
2640
_LGBM_CustomObjectiveFunction = Callable[
    [np.ndarray, Dataset],
    Tuple[np.ndarray, np.ndarray]
]
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
_LGBM_CustomEvalFunction = Union[
    Callable[
        [np.ndarray, Dataset],
        _LGBM_EvalFunctionResultType
    ],
    Callable[
        [np.ndarray, Dataset],
        List[_LGBM_EvalFunctionResultType]
    ]
]
2651
2652


2653
class Booster:
2654
    """Booster in LightGBM."""
2655

2656
2657
2658
2659
2660
2661
2662
    def __init__(
        self,
        params: Optional[Dict[str, Any]] = None,
        train_set: Optional[Dataset] = None,
        model_file: Optional[Union[str, Path]] = None,
        model_str: Optional[str] = None
    ):
2663
        """Initialize the Booster.
wxchan's avatar
wxchan committed
2664
2665
2666

        Parameters
        ----------
Nikita Titov's avatar
Nikita Titov committed
2667
        params : dict or None, optional (default=None)
2668
2669
2670
            Parameters for Booster.
        train_set : Dataset or None, optional (default=None)
            Training dataset.
2671
        model_file : str, pathlib.Path or None, optional (default=None)
wxchan's avatar
wxchan committed
2672
            Path to the model file.
2673
        model_str : str or None, optional (default=None)
2674
            Model will be loaded from this string.
wxchan's avatar
wxchan committed
2675
        """
2676
        self.handle = None
2677
        self.network = False
wxchan's avatar
wxchan committed
2678
        self.__need_reload_eval_info = True
2679
        self._train_data_name = "training"
2680
        self.__set_objective_to_none = False
wxchan's avatar
wxchan committed
2681
        self.best_iteration = -1
wxchan's avatar
wxchan committed
2682
        self.best_score = {}
2683
        params = {} if params is None else deepcopy(params)
wxchan's avatar
wxchan committed
2684
        if train_set is not None:
2685
            # Training task
wxchan's avatar
wxchan committed
2686
            if not isinstance(train_set, Dataset):
2687
                raise TypeError(f'Training data should be Dataset instance, met {type(train_set).__name__}')
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
            params = _choose_param_value(
                main_param_name="machines",
                params=params,
                default_value=None
            )
            # if "machines" is given, assume user wants to do distributed learning, and set up network
            if params["machines"] is None:
                params.pop("machines", None)
            else:
                machines = params["machines"]
                if isinstance(machines, str):
                    num_machines_from_machine_list = len(machines.split(','))
                elif isinstance(machines, (list, set)):
                    num_machines_from_machine_list = len(machines)
                    machines = ','.join(machines)
                else:
                    raise ValueError("Invalid machines in params.")

                params = _choose_param_value(
                    main_param_name="num_machines",
                    params=params,
                    default_value=num_machines_from_machine_list
                )
                params = _choose_param_value(
                    main_param_name="local_listen_port",
                    params=params,
                    default_value=12400
                )
                self.set_network(
                    machines=machines,
                    local_listen_port=params["local_listen_port"],
                    listen_time_out=params.get("time_out", 120),
                    num_machines=params["num_machines"]
                )
2722
            # construct booster object
2723
2724
2725
2726
            train_set.construct()
            # copy the parameters from train_set
            params.update(train_set.get_params())
            params_str = param_dict_to_str(params)
2727
            self.handle = ctypes.c_void_p()
wxchan's avatar
wxchan committed
2728
            _safe_call(_LIB.LGBM_BoosterCreate(
2729
                train_set.handle,
wxchan's avatar
wxchan committed
2730
2731
                c_str(params_str),
                ctypes.byref(self.handle)))
2732
            # save reference to data
wxchan's avatar
wxchan committed
2733
2734
2735
2736
            self.train_set = train_set
            self.valid_sets = []
            self.name_valid_sets = []
            self.__num_dataset = 1
Guolin Ke's avatar
Guolin Ke committed
2737
2738
            self.__init_predictor = train_set._predictor
            if self.__init_predictor is not None:
wxchan's avatar
wxchan committed
2739
2740
                _safe_call(_LIB.LGBM_BoosterMerge(
                    self.handle,
Guolin Ke's avatar
Guolin Ke committed
2741
                    self.__init_predictor.handle))
Guolin Ke's avatar
Guolin Ke committed
2742
            out_num_class = ctypes.c_int(0)
wxchan's avatar
wxchan committed
2743
2744
2745
2746
            _safe_call(_LIB.LGBM_BoosterGetNumClasses(
                self.handle,
                ctypes.byref(out_num_class)))
            self.__num_class = out_num_class.value
2747
            # buffer for inner predict
wxchan's avatar
wxchan committed
2748
2749
2750
            self.__inner_predict_buffer = [None]
            self.__is_predicted_cur_iter = [False]
            self.__get_eval_info()
2751
            self.pandas_categorical = train_set.pandas_categorical
2752
            self.train_set_version = train_set.version
wxchan's avatar
wxchan committed
2753
        elif model_file is not None:
2754
            # Prediction task
Guolin Ke's avatar
Guolin Ke committed
2755
            out_num_iterations = ctypes.c_int(0)
2756
            self.handle = ctypes.c_void_p()
wxchan's avatar
wxchan committed
2757
            _safe_call(_LIB.LGBM_BoosterCreateFromModelfile(
2758
                c_str(str(model_file)),
wxchan's avatar
wxchan committed
2759
2760
                ctypes.byref(out_num_iterations),
                ctypes.byref(self.handle)))
Guolin Ke's avatar
Guolin Ke committed
2761
            out_num_class = ctypes.c_int(0)
wxchan's avatar
wxchan committed
2762
2763
2764
2765
            _safe_call(_LIB.LGBM_BoosterGetNumClasses(
                self.handle,
                ctypes.byref(out_num_class)))
            self.__num_class = out_num_class.value
2766
            self.pandas_categorical = _load_pandas_categorical(file_name=model_file)
2767
        elif model_str is not None:
2768
            self.model_from_string(model_str)
wxchan's avatar
wxchan committed
2769
        else:
2770
2771
            raise TypeError('Need at least one training dataset or model file or model string '
                            'to create Booster instance')
2772
        self.params = params
wxchan's avatar
wxchan committed
2773

2774
    def __del__(self) -> None:
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
        try:
            if self.network:
                self.free_network()
        except AttributeError:
            pass
        try:
            if self.handle is not None:
                _safe_call(_LIB.LGBM_BoosterFree(self.handle))
        except AttributeError:
            pass
wxchan's avatar
wxchan committed
2785

2786
    def __copy__(self) -> "Booster":
wxchan's avatar
wxchan committed
2787
2788
        return self.__deepcopy__(None)

2789
    def __deepcopy__(self, _) -> "Booster":
2790
        model_str = self.model_to_string(num_iteration=-1)
2791
        booster = Booster(model_str=model_str)
2792
        return booster
wxchan's avatar
wxchan committed
2793

2794
    def __getstate__(self) -> Dict[str, Any]:
wxchan's avatar
wxchan committed
2795
2796
2797
2798
2799
        this = self.__dict__.copy()
        handle = this['handle']
        this.pop('train_set', None)
        this.pop('valid_sets', None)
        if handle is not None:
2800
            this["handle"] = self.model_to_string(num_iteration=-1)
wxchan's avatar
wxchan committed
2801
2802
        return this

2803
    def __setstate__(self, state: Dict[str, Any]) -> None:
2804
2805
        model_str = state.get('handle', None)
        if model_str is not None:
wxchan's avatar
wxchan committed
2806
            handle = ctypes.c_void_p()
Guolin Ke's avatar
Guolin Ke committed
2807
            out_num_iterations = ctypes.c_int(0)
2808
2809
2810
2811
            _safe_call(_LIB.LGBM_BoosterLoadModelFromString(
                c_str(model_str),
                ctypes.byref(out_num_iterations),
                ctypes.byref(handle)))
wxchan's avatar
wxchan committed
2812
2813
2814
            state['handle'] = handle
        self.__dict__.update(state)

2815
    def free_dataset(self) -> "Booster":
Nikita Titov's avatar
Nikita Titov committed
2816
2817
2818
2819
2820
2821
2822
        """Free Booster's Datasets.

        Returns
        -------
        self : Booster
            Booster without Datasets.
        """
wxchan's avatar
wxchan committed
2823
2824
        self.__dict__.pop('train_set', None)
        self.__dict__.pop('valid_sets', None)
2825
        self.__num_dataset = 0
Nikita Titov's avatar
Nikita Titov committed
2826
        return self
wxchan's avatar
wxchan committed
2827

2828
    def _free_buffer(self) -> "Booster":
2829
2830
        self.__inner_predict_buffer = []
        self.__is_predicted_cur_iter = []
Nikita Titov's avatar
Nikita Titov committed
2831
        return self
2832

2833
2834
2835
2836
2837
2838
2839
    def set_network(
        self,
        machines: Union[List[str], Set[str], str],
        local_listen_port: int = 12400,
        listen_time_out: int = 120,
        num_machines: int = 1
    ) -> "Booster":
2840
2841
2842
2843
        """Set the network configuration.

        Parameters
        ----------
2844
        machines : list, set or str
2845
            Names of machines.
Nikita Titov's avatar
Nikita Titov committed
2846
        local_listen_port : int, optional (default=12400)
2847
            TCP listen port for local machines.
Nikita Titov's avatar
Nikita Titov committed
2848
        listen_time_out : int, optional (default=120)
2849
            Socket time-out in minutes.
Nikita Titov's avatar
Nikita Titov committed
2850
        num_machines : int, optional (default=1)
2851
            The number of machines for distributed learning application.
Nikita Titov's avatar
Nikita Titov committed
2852
2853
2854
2855
2856

        Returns
        -------
        self : Booster
            Booster with set network.
2857
        """
2858
2859
        if isinstance(machines, (list, set)):
            machines = ','.join(machines)
2860
2861
2862
2863
2864
        _safe_call(_LIB.LGBM_NetworkInit(c_str(machines),
                                         ctypes.c_int(local_listen_port),
                                         ctypes.c_int(listen_time_out),
                                         ctypes.c_int(num_machines)))
        self.network = True
Nikita Titov's avatar
Nikita Titov committed
2865
        return self
2866

2867
    def free_network(self) -> "Booster":
Nikita Titov's avatar
Nikita Titov committed
2868
2869
2870
2871
2872
2873
2874
        """Free Booster's network.

        Returns
        -------
        self : Booster
            Booster with freed network.
        """
2875
2876
        _safe_call(_LIB.LGBM_NetworkFree())
        self.network = False
Nikita Titov's avatar
Nikita Titov committed
2877
        return self
2878

2879
    def trees_to_dataframe(self) -> pd_DataFrame:
2880
2881
        """Parse the fitted model and return in an easy-to-read pandas DataFrame.

2882
2883
2884
2885
        The returned DataFrame has the following columns.

            - ``tree_index`` : int64, which tree a node belongs to. 0-based, so a value of ``6``, for example, means "this node is in the 7th tree".
            - ``node_depth`` : int64, how far a node is from the root of the tree. The root node has a value of ``1``, its direct children are ``2``, etc.
2886
2887
2888
2889
2890
            - ``node_index`` : str, unique identifier for a node.
            - ``left_child`` : str, ``node_index`` of the child node to the left of a split. ``None`` for leaf nodes.
            - ``right_child`` : str, ``node_index`` of the child node to the right of a split. ``None`` for leaf nodes.
            - ``parent_index`` : str, ``node_index`` of this node's parent. ``None`` for the root node.
            - ``split_feature`` : str, name of the feature used for splitting. ``None`` for leaf nodes.
2891
2892
            - ``split_gain`` : float64, gain from adding this split to the tree. ``NaN`` for leaf nodes.
            - ``threshold`` : float64, value of the feature used to decide which side of the split a record will go down. ``NaN`` for leaf nodes.
2893
            - ``decision_type`` : str, logical operator describing how to compare a value to ``threshold``.
2894
2895
              For example, ``split_feature = "Column_10", threshold = 15, decision_type = "<="`` means that
              records where ``Column_10 <= 15`` follow the left side of the split, otherwise follows the right side of the split. ``None`` for leaf nodes.
2896
2897
            - ``missing_direction`` : str, split direction that missing values should go to. ``None`` for leaf nodes.
            - ``missing_type`` : str, describes what types of values are treated as missing.
2898
            - ``value`` : float64, predicted value for this leaf node, multiplied by the learning rate.
2899
            - ``weight`` : float64 or int64, sum of Hessian (second-order derivative of objective), summed over observations that fall in this node.
2900
2901
            - ``count`` : int64, number of records in the training data that fall into this node.

2902
2903
2904
2905
2906
2907
        Returns
        -------
        result : pandas DataFrame
            Returns a pandas DataFrame of the parsed model.
        """
        if not PANDAS_INSTALLED:
2908
2909
            raise LightGBMError('This method cannot be run without pandas installed. '
                                'You must install pandas and restart your session to use this method.')
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920

        if self.num_trees() == 0:
            raise LightGBMError('There are no trees in this Booster and thus nothing to parse')

        def _is_split_node(tree):
            return 'split_index' in tree.keys()

        def create_node_record(tree, node_depth=1, tree_index=None,
                               feature_names=None, parent_node=None):

            def _get_node_index(tree, tree_index):
2921
                tree_num = f'{tree_index}-' if tree_index is not None else ''
2922
2923
2924
                is_split = _is_split_node(tree)
                node_type = 'S' if is_split else 'L'
                # if a single node tree it won't have `leaf_index` so return 0
2925
2926
                node_num = tree.get('split_index' if is_split else 'leaf_index', 0)
                return f"{tree_num}{node_type}{node_num}"
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938

            def _get_split_feature(tree, feature_names):
                if _is_split_node(tree):
                    if feature_names is not None:
                        feature_name = feature_names[tree['split_feature']]
                    else:
                        feature_name = tree['split_feature']
                else:
                    feature_name = None
                return feature_name

            def _is_single_node_tree(tree):
2939
                return set(tree.keys()) == {'leaf_value'}
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012

            # Create the node record, and populate universal data members
            node = OrderedDict()
            node['tree_index'] = tree_index
            node['node_depth'] = node_depth
            node['node_index'] = _get_node_index(tree, tree_index)
            node['left_child'] = None
            node['right_child'] = None
            node['parent_index'] = parent_node
            node['split_feature'] = _get_split_feature(tree, feature_names)
            node['split_gain'] = None
            node['threshold'] = None
            node['decision_type'] = None
            node['missing_direction'] = None
            node['missing_type'] = None
            node['value'] = None
            node['weight'] = None
            node['count'] = None

            # Update values to reflect node type (leaf or split)
            if _is_split_node(tree):
                node['left_child'] = _get_node_index(tree['left_child'], tree_index)
                node['right_child'] = _get_node_index(tree['right_child'], tree_index)
                node['split_gain'] = tree['split_gain']
                node['threshold'] = tree['threshold']
                node['decision_type'] = tree['decision_type']
                node['missing_direction'] = 'left' if tree['default_left'] else 'right'
                node['missing_type'] = tree['missing_type']
                node['value'] = tree['internal_value']
                node['weight'] = tree['internal_weight']
                node['count'] = tree['internal_count']
            else:
                node['value'] = tree['leaf_value']
                if not _is_single_node_tree(tree):
                    node['weight'] = tree['leaf_weight']
                    node['count'] = tree['leaf_count']

            return node

        def tree_dict_to_node_list(tree, node_depth=1, tree_index=None,
                                   feature_names=None, parent_node=None):

            node = create_node_record(tree,
                                      node_depth=node_depth,
                                      tree_index=tree_index,
                                      feature_names=feature_names,
                                      parent_node=parent_node)

            res = [node]

            if _is_split_node(tree):
                # traverse the next level of the tree
                children = ['left_child', 'right_child']
                for child in children:
                    subtree_list = tree_dict_to_node_list(
                        tree[child],
                        node_depth=node_depth + 1,
                        tree_index=tree_index,
                        feature_names=feature_names,
                        parent_node=node['node_index'])
                    # In tree format, "subtree_list" is a list of node records (dicts),
                    # and we add node to the list.
                    res.extend(subtree_list)
            return res

        model_dict = self.dump_model()
        feature_names = model_dict['feature_names']
        model_list = []
        for tree in model_dict['tree_info']:
            model_list.extend(tree_dict_to_node_list(tree['tree_structure'],
                                                     tree_index=tree['tree_index'],
                                                     feature_names=feature_names))

3013
        return pd_DataFrame(model_list, columns=model_list[0].keys())
3014

3015
    def set_train_data_name(self, name: str) -> "Booster":
3016
3017
3018
3019
        """Set the name to the training Dataset.

        Parameters
        ----------
3020
        name : str
Nikita Titov's avatar
Nikita Titov committed
3021
3022
3023
3024
3025
3026
            Name for the training Dataset.

        Returns
        -------
        self : Booster
            Booster with set training Dataset name.
3027
        """
3028
        self._train_data_name = name
Nikita Titov's avatar
Nikita Titov committed
3029
        return self
wxchan's avatar
wxchan committed
3030

3031
    def add_valid(self, data: Dataset, name: str) -> "Booster":
3032
        """Add validation data.
wxchan's avatar
wxchan committed
3033
3034
3035
3036

        Parameters
        ----------
        data : Dataset
3037
            Validation data.
3038
        name : str
3039
            Name of validation data.
Nikita Titov's avatar
Nikita Titov committed
3040
3041
3042
3043
3044

        Returns
        -------
        self : Booster
            Booster with set validation data.
wxchan's avatar
wxchan committed
3045
        """
Guolin Ke's avatar
Guolin Ke committed
3046
        if not isinstance(data, Dataset):
3047
            raise TypeError(f'Validation data should be Dataset instance, met {type(data).__name__}')
Guolin Ke's avatar
Guolin Ke committed
3048
        if data._predictor is not self.__init_predictor:
3049
3050
            raise LightGBMError("Add validation data failed, "
                                "you should use same predictor for these data")
wxchan's avatar
wxchan committed
3051
3052
        _safe_call(_LIB.LGBM_BoosterAddValidData(
            self.handle,
wxchan's avatar
wxchan committed
3053
            data.construct().handle))
wxchan's avatar
wxchan committed
3054
3055
3056
3057
3058
        self.valid_sets.append(data)
        self.name_valid_sets.append(name)
        self.__num_dataset += 1
        self.__inner_predict_buffer.append(None)
        self.__is_predicted_cur_iter.append(False)
Nikita Titov's avatar
Nikita Titov committed
3059
        return self
wxchan's avatar
wxchan committed
3060

3061
    def reset_parameter(self, params: Dict[str, Any]) -> "Booster":
3062
        """Reset parameters of Booster.
wxchan's avatar
wxchan committed
3063
3064
3065
3066

        Parameters
        ----------
        params : dict
3067
            New parameters for Booster.
Nikita Titov's avatar
Nikita Titov committed
3068
3069
3070
3071
3072

        Returns
        -------
        self : Booster
            Booster with new parameters.
wxchan's avatar
wxchan committed
3073
3074
3075
3076
3077
3078
        """
        params_str = param_dict_to_str(params)
        if params_str:
            _safe_call(_LIB.LGBM_BoosterResetParameter(
                self.handle,
                c_str(params_str)))
Guolin Ke's avatar
Guolin Ke committed
3079
        self.params.update(params)
Nikita Titov's avatar
Nikita Titov committed
3080
        return self
wxchan's avatar
wxchan committed
3081

3082
3083
3084
3085
3086
    def update(
        self,
        train_set: Optional[Dataset] = None,
        fobj: Optional[_LGBM_CustomObjectiveFunction] = None
    ) -> bool:
Nikita Titov's avatar
Nikita Titov committed
3087
        """Update Booster for one iteration.
3088

wxchan's avatar
wxchan committed
3089
3090
        Parameters
        ----------
3091
3092
3093
3094
        train_set : Dataset or None, optional (default=None)
            Training data.
            If None, last training data is used.
        fobj : callable or None, optional (default=None)
wxchan's avatar
wxchan committed
3095
            Customized objective function.
3096
3097
3098
            Should accept two parameters: preds, train_data,
            and return (grad, hess).

3099
                preds : numpy 1-D array or numpy 2-D array (for multi-class task)
3100
                    The predicted values.
3101
3102
                    Predicted values are returned before any transformation,
                    e.g. they are raw margin instead of probability of positive class for binary task.
3103
3104
                train_data : Dataset
                    The training dataset.
3105
                grad : numpy 1-D array or numpy 2-D array (for multi-class task)
3106
3107
                    The value of the first order derivative (gradient) of the loss
                    with respect to the elements of preds for each sample point.
3108
                hess : numpy 1-D array or numpy 2-D array (for multi-class task)
3109
3110
                    The value of the second order derivative (Hessian) of the loss
                    with respect to the elements of preds for each sample point.
wxchan's avatar
wxchan committed
3111

3112
            For multi-class task, preds are numpy 2-D array of shape = [n_samples, n_classes],
3113
            and grad and hess should be returned in the same format.
3114

wxchan's avatar
wxchan committed
3115
3116
        Returns
        -------
3117
3118
        is_finished : bool
            Whether the update was successfully finished.
wxchan's avatar
wxchan committed
3119
        """
3120
        # need reset training data
3121
3122
3123
3124
3125
3126
        if train_set is None and self.train_set_version != self.train_set.version:
            train_set = self.train_set
            is_the_same_train_set = False
        else:
            is_the_same_train_set = train_set is self.train_set and self.train_set_version == train_set.version
        if train_set is not None and not is_the_same_train_set:
Guolin Ke's avatar
Guolin Ke committed
3127
            if not isinstance(train_set, Dataset):
3128
                raise TypeError(f'Training data should be Dataset instance, met {type(train_set).__name__}')
Guolin Ke's avatar
Guolin Ke committed
3129
            if train_set._predictor is not self.__init_predictor:
3130
3131
                raise LightGBMError("Replace training data failed, "
                                    "you should use same predictor for these data")
wxchan's avatar
wxchan committed
3132
3133
3134
            self.train_set = train_set
            _safe_call(_LIB.LGBM_BoosterResetTrainingData(
                self.handle,
wxchan's avatar
wxchan committed
3135
                self.train_set.construct().handle))
wxchan's avatar
wxchan committed
3136
            self.__inner_predict_buffer[0] = None
3137
            self.train_set_version = self.train_set.version
wxchan's avatar
wxchan committed
3138
3139
        is_finished = ctypes.c_int(0)
        if fobj is None:
3140
            if self.__set_objective_to_none:
3141
                raise LightGBMError('Cannot update due to null objective function.')
wxchan's avatar
wxchan committed
3142
3143
3144
            _safe_call(_LIB.LGBM_BoosterUpdateOneIter(
                self.handle,
                ctypes.byref(is_finished)))
3145
            self.__is_predicted_cur_iter = [False for _ in range(self.__num_dataset)]
wxchan's avatar
wxchan committed
3146
3147
            return is_finished.value == 1
        else:
3148
            if not self.__set_objective_to_none:
Nikita Titov's avatar
Nikita Titov committed
3149
                self.reset_parameter({"objective": "none"}).__set_objective_to_none = True
wxchan's avatar
wxchan committed
3150
3151
3152
            grad, hess = fobj(self.__inner_predict(0), self.train_set)
            return self.__boost(grad, hess)

3153
3154
3155
3156
3157
    def __boost(
        self,
        grad: np.ndarray,
        hess: np.ndarray
    ) -> bool:
3158
        """Boost Booster for one iteration with customized gradient statistics.
Nikita Titov's avatar
Nikita Titov committed
3159

Nikita Titov's avatar
Nikita Titov committed
3160
3161
        .. note::

3162
3163
            Score is returned before any transformation,
            e.g. it is raw margin instead of probability of positive class for binary task.
3164
            For multi-class task, score are numpy 2-D array of shape = [n_samples, n_classes],
3165
            and grad and hess should be returned in the same format.
3166

wxchan's avatar
wxchan committed
3167
3168
        Parameters
        ----------
3169
        grad : numpy 1-D array or numpy 2-D array (for multi-class task)
3170
3171
            The value of the first order derivative (gradient) of the loss
            with respect to the elements of score for each sample point.
3172
        hess : numpy 1-D array or numpy 2-D array (for multi-class task)
3173
3174
            The value of the second order derivative (Hessian) of the loss
            with respect to the elements of score for each sample point.
wxchan's avatar
wxchan committed
3175
3176
3177

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
3178
3179
        is_finished : bool
            Whether the boost was successfully finished.
wxchan's avatar
wxchan committed
3180
        """
3181
3182
3183
        if self.__num_class > 1:
            grad = grad.ravel(order='F')
            hess = hess.ravel(order='F')
3184
3185
        grad = list_to_1d_numpy(grad, name='gradient')
        hess = list_to_1d_numpy(hess, name='hessian')
3186
3187
        assert grad.flags.c_contiguous
        assert hess.flags.c_contiguous
wxchan's avatar
wxchan committed
3188
        if len(grad) != len(hess):
3189
3190
            raise ValueError(f"Lengths of gradient ({len(grad)}) and Hessian ({len(hess)}) don't match")
        num_train_data = self.train_set.num_data()
3191
        if len(grad) != num_train_data * self.__num_class:
3192
3193
3194
            raise ValueError(
                f"Lengths of gradient ({len(grad)}) and Hessian ({len(hess)}) "
                f"don't match training data length ({num_train_data}) * "
3195
                f"number of models per one iteration ({self.__num_class})"
3196
            )
wxchan's avatar
wxchan committed
3197
3198
3199
3200
3201
3202
        is_finished = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterUpdateOneIterCustom(
            self.handle,
            grad.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
            hess.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
            ctypes.byref(is_finished)))
3203
        self.__is_predicted_cur_iter = [False for _ in range(self.__num_dataset)]
wxchan's avatar
wxchan committed
3204
3205
        return is_finished.value == 1

3206
    def rollback_one_iter(self) -> "Booster":
Nikita Titov's avatar
Nikita Titov committed
3207
3208
3209
3210
3211
3212
3213
        """Rollback one iteration.

        Returns
        -------
        self : Booster
            Booster with rolled back one iteration.
        """
wxchan's avatar
wxchan committed
3214
3215
        _safe_call(_LIB.LGBM_BoosterRollbackOneIter(
            self.handle))
3216
        self.__is_predicted_cur_iter = [False for _ in range(self.__num_dataset)]
Nikita Titov's avatar
Nikita Titov committed
3217
        return self
wxchan's avatar
wxchan committed
3218

3219
    def current_iteration(self) -> int:
3220
3221
3222
3223
3224
3225
3226
        """Get the index of the current iteration.

        Returns
        -------
        cur_iter : int
            The index of the current iteration.
        """
Guolin Ke's avatar
Guolin Ke committed
3227
        out_cur_iter = ctypes.c_int(0)
wxchan's avatar
wxchan committed
3228
3229
3230
3231
3232
        _safe_call(_LIB.LGBM_BoosterGetCurrentIteration(
            self.handle,
            ctypes.byref(out_cur_iter)))
        return out_cur_iter.value

3233
    def num_model_per_iteration(self) -> int:
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
        """Get number of models per iteration.

        Returns
        -------
        model_per_iter : int
            The number of models per iteration.
        """
        model_per_iter = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterNumModelPerIteration(
            self.handle,
            ctypes.byref(model_per_iter)))
        return model_per_iter.value

3247
    def num_trees(self) -> int:
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
        """Get number of weak sub-models.

        Returns
        -------
        num_trees : int
            The number of weak sub-models.
        """
        num_trees = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterNumberOfTotalModel(
            self.handle,
            ctypes.byref(num_trees)))
        return num_trees.value

3261
    def upper_bound(self) -> float:
3262
3263
3264
3265
        """Get upper bound value of a model.

        Returns
        -------
3266
        upper_bound : float
3267
3268
3269
3270
3271
3272
3273
3274
            Upper bound value of the model.
        """
        ret = ctypes.c_double(0)
        _safe_call(_LIB.LGBM_BoosterGetUpperBoundValue(
            self.handle,
            ctypes.byref(ret)))
        return ret.value

3275
    def lower_bound(self) -> float:
3276
3277
3278
3279
        """Get lower bound value of a model.

        Returns
        -------
3280
        lower_bound : float
3281
3282
3283
3284
3285
3286
3287
3288
            Lower bound value of the model.
        """
        ret = ctypes.c_double(0)
        _safe_call(_LIB.LGBM_BoosterGetLowerBoundValue(
            self.handle,
            ctypes.byref(ret)))
        return ret.value

3289
3290
3291
3292
3293
3294
    def eval(
        self,
        data: Dataset,
        name: str,
        feval: Optional[Union[_LGBM_CustomEvalFunction, List[_LGBM_CustomEvalFunction]]] = None
    ) -> List[_LGBM_BoosterEvalMethodResultType]:
3295
        """Evaluate for data.
wxchan's avatar
wxchan committed
3296
3297
3298

        Parameters
        ----------
3299
3300
        data : Dataset
            Data for the evaluating.
3301
        name : str
3302
            Name of the data.
3303
        feval : callable, list of callable, or None, optional (default=None)
3304
            Customized evaluation function.
3305
            Each evaluation function should accept two parameters: preds, eval_data,
3306
            and return (eval_name, eval_result, is_higher_better) or list of such tuples.
3307

3308
                preds : numpy 1-D array or numpy 2-D array (for multi-class task)
3309
                    The predicted values.
3310
                    For multi-class task, preds are numpy 2-D array of shape = [n_samples, n_classes].
3311
                    If custom objective function is used, predicted values are returned before any transformation,
3312
                    e.g. they are raw margin instead of probability of positive class for binary task in this case.
3313
                eval_data : Dataset
3314
                    A ``Dataset`` to evaluate.
3315
                eval_name : str
Andrew Ziem's avatar
Andrew Ziem committed
3316
                    The name of evaluation function (without whitespace).
3317
3318
3319
3320
3321
                eval_result : float
                    The eval result.
                is_higher_better : bool
                    Is eval result higher better, e.g. AUC is ``is_higher_better``.

wxchan's avatar
wxchan committed
3322
3323
        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
3324
        result : list
3325
            List with (dataset_name, eval_name, eval_result, is_higher_better) tuples.
wxchan's avatar
wxchan committed
3326
        """
Guolin Ke's avatar
Guolin Ke committed
3327
3328
        if not isinstance(data, Dataset):
            raise TypeError("Can only eval for Dataset instance")
wxchan's avatar
wxchan committed
3329
3330
3331
3332
        data_idx = -1
        if data is self.train_set:
            data_idx = 0
        else:
3333
            for i in range(len(self.valid_sets)):
wxchan's avatar
wxchan committed
3334
3335
3336
                if data is self.valid_sets[i]:
                    data_idx = i + 1
                    break
3337
        # need to push new valid data
wxchan's avatar
wxchan committed
3338
3339
3340
3341
3342
3343
        if data_idx == -1:
            self.add_valid(data, name)
            data_idx = self.__num_dataset - 1

        return self.__inner_eval(name, data_idx, feval)

3344
3345
3346
3347
    def eval_train(
        self,
        feval: Optional[Union[_LGBM_CustomEvalFunction, List[_LGBM_CustomEvalFunction]]] = None
    ) -> List[_LGBM_BoosterEvalMethodResultType]:
3348
        """Evaluate for training data.
wxchan's avatar
wxchan committed
3349
3350
3351

        Parameters
        ----------
3352
        feval : callable, list of callable, or None, optional (default=None)
3353
            Customized evaluation function.
3354
            Each evaluation function should accept two parameters: preds, eval_data,
3355
            and return (eval_name, eval_result, is_higher_better) or list of such tuples.
3356

3357
                preds : numpy 1-D array or numpy 2-D array (for multi-class task)
3358
                    The predicted values.
3359
                    For multi-class task, preds are numpy 2-D array of shape = [n_samples, n_classes].
3360
                    If custom objective function is used, predicted values are returned before any transformation,
3361
                    e.g. they are raw margin instead of probability of positive class for binary task in this case.
Akshita Dixit's avatar
Akshita Dixit committed
3362
                eval_data : Dataset
3363
                    The training dataset.
3364
                eval_name : str
Andrew Ziem's avatar
Andrew Ziem committed
3365
                    The name of evaluation function (without whitespace).
3366
3367
3368
3369
3370
                eval_result : float
                    The eval result.
                is_higher_better : bool
                    Is eval result higher better, e.g. AUC is ``is_higher_better``.

wxchan's avatar
wxchan committed
3371
3372
        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
3373
        result : list
3374
            List with (train_dataset_name, eval_name, eval_result, is_higher_better) tuples.
wxchan's avatar
wxchan committed
3375
        """
3376
        return self.__inner_eval(self._train_data_name, 0, feval)
wxchan's avatar
wxchan committed
3377

3378
3379
3380
3381
    def eval_valid(
        self,
        feval: Optional[Union[_LGBM_CustomEvalFunction, List[_LGBM_CustomEvalFunction]]] = None
    ) -> List[_LGBM_BoosterEvalMethodResultType]:
3382
        """Evaluate for validation data.
wxchan's avatar
wxchan committed
3383
3384
3385

        Parameters
        ----------
3386
        feval : callable, list of callable, or None, optional (default=None)
3387
            Customized evaluation function.
3388
            Each evaluation function should accept two parameters: preds, eval_data,
3389
            and return (eval_name, eval_result, is_higher_better) or list of such tuples.
3390

3391
                preds : numpy 1-D array or numpy 2-D array (for multi-class task)
3392
                    The predicted values.
3393
                    For multi-class task, preds are numpy 2-D array of shape = [n_samples, n_classes].
3394
                    If custom objective function is used, predicted values are returned before any transformation,
3395
                    e.g. they are raw margin instead of probability of positive class for binary task in this case.
Akshita Dixit's avatar
Akshita Dixit committed
3396
                eval_data : Dataset
3397
                    The validation dataset.
3398
                eval_name : str
Andrew Ziem's avatar
Andrew Ziem committed
3399
                    The name of evaluation function (without whitespace).
3400
3401
3402
3403
3404
                eval_result : float
                    The eval result.
                is_higher_better : bool
                    Is eval result higher better, e.g. AUC is ``is_higher_better``.

wxchan's avatar
wxchan committed
3405
3406
        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
3407
        result : list
3408
            List with (validation_dataset_name, eval_name, eval_result, is_higher_better) tuples.
wxchan's avatar
wxchan committed
3409
        """
3410
        return [item for i in range(1, self.__num_dataset)
wxchan's avatar
wxchan committed
3411
                for item in self.__inner_eval(self.name_valid_sets[i - 1], i, feval)]
wxchan's avatar
wxchan committed
3412

3413
3414
3415
3416
3417
3418
3419
    def save_model(
        self,
        filename: Union[str, Path],
        num_iteration: Optional[int] = None,
        start_iteration: int = 0,
        importance_type: str = 'split'
    ) -> "Booster":
3420
        """Save Booster to file.
wxchan's avatar
wxchan committed
3421
3422
3423

        Parameters
        ----------
3424
        filename : str or pathlib.Path
3425
            Filename to save Booster.
3426
3427
3428
3429
        num_iteration : int or None, optional (default=None)
            Index of the iteration that should be saved.
            If None, if the best iteration exists, it is saved; otherwise, all iterations are saved.
            If <= 0, all iterations are saved.
Nikita Titov's avatar
Nikita Titov committed
3430
        start_iteration : int, optional (default=0)
3431
            Start index of the iteration that should be saved.
3432
        importance_type : str, optional (default="split")
3433
3434
3435
            What type of feature importance should be saved.
            If "split", result contains numbers of times the feature is used in a model.
            If "gain", result contains total gains of splits which use the feature.
Nikita Titov's avatar
Nikita Titov committed
3436
3437
3438
3439
3440

        Returns
        -------
        self : Booster
            Returns self.
wxchan's avatar
wxchan committed
3441
        """
3442
        if num_iteration is None:
3443
            num_iteration = self.best_iteration
3444
        importance_type_int = FEATURE_IMPORTANCE_TYPE_MAPPER[importance_type]
wxchan's avatar
wxchan committed
3445
3446
        _safe_call(_LIB.LGBM_BoosterSaveModel(
            self.handle,
3447
            ctypes.c_int(start_iteration),
Guolin Ke's avatar
Guolin Ke committed
3448
            ctypes.c_int(num_iteration),
3449
            ctypes.c_int(importance_type_int),
3450
            c_str(str(filename))))
3451
        _dump_pandas_categorical(self.pandas_categorical, filename)
Nikita Titov's avatar
Nikita Titov committed
3452
        return self
wxchan's avatar
wxchan committed
3453

3454
3455
3456
3457
3458
    def shuffle_models(
        self,
        start_iteration: int = 0,
        end_iteration: int = -1
    ) -> "Booster":
3459
        """Shuffle models.
Nikita Titov's avatar
Nikita Titov committed
3460

3461
3462
3463
        Parameters
        ----------
        start_iteration : int, optional (default=0)
3464
            The first iteration that will be shuffled.
3465
3466
        end_iteration : int, optional (default=-1)
            The last iteration that will be shuffled.
3467
            If <= 0, means the last available iteration.
3468

Nikita Titov's avatar
Nikita Titov committed
3469
3470
3471
3472
        Returns
        -------
        self : Booster
            Booster with shuffled models.
3473
        """
3474
3475
        _safe_call(_LIB.LGBM_BoosterShuffleModels(
            self.handle,
Guolin Ke's avatar
Guolin Ke committed
3476
3477
            ctypes.c_int(start_iteration),
            ctypes.c_int(end_iteration)))
Nikita Titov's avatar
Nikita Titov committed
3478
        return self
3479

3480
    def model_from_string(self, model_str: str) -> "Booster":
3481
3482
3483
3484
        """Load Booster from a string.

        Parameters
        ----------
3485
        model_str : str
3486
3487
3488
3489
            Model will be loaded from this string.

        Returns
        -------
Nikita Titov's avatar
Nikita Titov committed
3490
        self : Booster
3491
3492
            Loaded Booster object.
        """
3493
3494
3495
3496
        if self.handle is not None:
            _safe_call(_LIB.LGBM_BoosterFree(self.handle))
        self._free_buffer()
        self.handle = ctypes.c_void_p()
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
        out_num_iterations = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterLoadModelFromString(
            c_str(model_str),
            ctypes.byref(out_num_iterations),
            ctypes.byref(self.handle)))
        out_num_class = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterGetNumClasses(
            self.handle,
            ctypes.byref(out_num_class)))
        self.__num_class = out_num_class.value
3507
        self.pandas_categorical = _load_pandas_categorical(model_str=model_str)
3508
3509
        return self

3510
3511
3512
3513
3514
3515
    def model_to_string(
        self,
        num_iteration: Optional[int] = None,
        start_iteration: int = 0,
        importance_type: str = 'split'
    ) -> str:
3516
        """Save Booster to string.
3517

3518
3519
3520
3521
3522
3523
        Parameters
        ----------
        num_iteration : int or None, optional (default=None)
            Index of the iteration that should be saved.
            If None, if the best iteration exists, it is saved; otherwise, all iterations are saved.
            If <= 0, all iterations are saved.
Nikita Titov's avatar
Nikita Titov committed
3524
        start_iteration : int, optional (default=0)
3525
            Start index of the iteration that should be saved.
3526
        importance_type : str, optional (default="split")
3527
3528
3529
            What type of feature importance should be saved.
            If "split", result contains numbers of times the feature is used in a model.
            If "gain", result contains total gains of splits which use the feature.
3530
3531
3532

        Returns
        -------
3533
        str_repr : str
3534
3535
            String representation of Booster.
        """
3536
        if num_iteration is None:
3537
            num_iteration = self.best_iteration
3538
        importance_type_int = FEATURE_IMPORTANCE_TYPE_MAPPER[importance_type]
3539
        buffer_len = 1 << 20
3540
        tmp_out_len = ctypes.c_int64(0)
3541
3542
3543
3544
        string_buffer = ctypes.create_string_buffer(buffer_len)
        ptr_string_buffer = ctypes.c_char_p(*[ctypes.addressof(string_buffer)])
        _safe_call(_LIB.LGBM_BoosterSaveModelToString(
            self.handle,
3545
            ctypes.c_int(start_iteration),
3546
            ctypes.c_int(num_iteration),
3547
            ctypes.c_int(importance_type_int),
3548
            ctypes.c_int64(buffer_len),
3549
3550
3551
            ctypes.byref(tmp_out_len),
            ptr_string_buffer))
        actual_len = tmp_out_len.value
3552
        # if buffer length is not long enough, re-allocate a buffer
3553
3554
3555
3556
3557
        if actual_len > buffer_len:
            string_buffer = ctypes.create_string_buffer(actual_len)
            ptr_string_buffer = ctypes.c_char_p(*[ctypes.addressof(string_buffer)])
            _safe_call(_LIB.LGBM_BoosterSaveModelToString(
                self.handle,
3558
                ctypes.c_int(start_iteration),
3559
                ctypes.c_int(num_iteration),
3560
                ctypes.c_int(importance_type_int),
3561
                ctypes.c_int64(actual_len),
3562
3563
                ctypes.byref(tmp_out_len),
                ptr_string_buffer))
3564
        ret = string_buffer.value.decode('utf-8')
3565
3566
        ret += _dump_pandas_categorical(self.pandas_categorical)
        return ret
3567

3568
3569
3570
3571
3572
3573
3574
    def dump_model(
        self,
        num_iteration: Optional[int] = None,
        start_iteration: int = 0,
        importance_type: str = 'split',
        object_hook: Optional[Callable[[Dict[str, Any]], Dict[str, Any]]] = None
    ) -> Dict[str, Any]:
Nikita Titov's avatar
Nikita Titov committed
3575
        """Dump Booster to JSON format.
wxchan's avatar
wxchan committed
3576

3577
3578
        Parameters
        ----------
3579
3580
3581
3582
        num_iteration : int or None, optional (default=None)
            Index of the iteration that should be dumped.
            If None, if the best iteration exists, it is dumped; otherwise, all iterations are dumped.
            If <= 0, all iterations are dumped.
Nikita Titov's avatar
Nikita Titov committed
3583
        start_iteration : int, optional (default=0)
3584
            Start index of the iteration that should be dumped.
3585
        importance_type : str, optional (default="split")
3586
3587
3588
            What type of feature importance should be dumped.
            If "split", result contains numbers of times the feature is used in a model.
            If "gain", result contains total gains of splits which use the feature.
3589
3590
3591
3592
3593
3594
3595
3596
3597
        object_hook : callable or None, optional (default=None)
            If not None, ``object_hook`` is a function called while parsing the json
            string returned by the C API. It may be used to alter the json, to store
            specific values while building the json structure. It avoids
            walking through the structure again. It saves a significant amount
            of time if the number of trees is huge.
            Signature is ``def object_hook(node: dict) -> dict``.
            None is equivalent to ``lambda node: node``.
            See documentation of ``json.loads()`` for further details.
3598

wxchan's avatar
wxchan committed
3599
3600
        Returns
        -------
3601
        json_repr : dict
Nikita Titov's avatar
Nikita Titov committed
3602
            JSON format of Booster.
wxchan's avatar
wxchan committed
3603
        """
3604
        if num_iteration is None:
3605
            num_iteration = self.best_iteration
3606
        importance_type_int = FEATURE_IMPORTANCE_TYPE_MAPPER[importance_type]
wxchan's avatar
wxchan committed
3607
        buffer_len = 1 << 20
3608
        tmp_out_len = ctypes.c_int64(0)
wxchan's avatar
wxchan committed
3609
3610
3611
3612
        string_buffer = ctypes.create_string_buffer(buffer_len)
        ptr_string_buffer = ctypes.c_char_p(*[ctypes.addressof(string_buffer)])
        _safe_call(_LIB.LGBM_BoosterDumpModel(
            self.handle,
3613
            ctypes.c_int(start_iteration),
Guolin Ke's avatar
Guolin Ke committed
3614
            ctypes.c_int(num_iteration),
3615
            ctypes.c_int(importance_type_int),
3616
            ctypes.c_int64(buffer_len),
wxchan's avatar
wxchan committed
3617
            ctypes.byref(tmp_out_len),
Guolin Ke's avatar
Guolin Ke committed
3618
            ptr_string_buffer))
wxchan's avatar
wxchan committed
3619
        actual_len = tmp_out_len.value
3620
        # if buffer length is not long enough, reallocate a buffer
wxchan's avatar
wxchan committed
3621
3622
3623
3624
3625
        if actual_len > buffer_len:
            string_buffer = ctypes.create_string_buffer(actual_len)
            ptr_string_buffer = ctypes.c_char_p(*[ctypes.addressof(string_buffer)])
            _safe_call(_LIB.LGBM_BoosterDumpModel(
                self.handle,
3626
                ctypes.c_int(start_iteration),
Guolin Ke's avatar
Guolin Ke committed
3627
                ctypes.c_int(num_iteration),
3628
                ctypes.c_int(importance_type_int),
3629
                ctypes.c_int64(actual_len),
wxchan's avatar
wxchan committed
3630
                ctypes.byref(tmp_out_len),
Guolin Ke's avatar
Guolin Ke committed
3631
                ptr_string_buffer))
3632
        ret = json.loads(string_buffer.value.decode('utf-8'), object_hook=object_hook)
3633
3634
3635
        ret['pandas_categorical'] = json.loads(json.dumps(self.pandas_categorical,
                                                          default=json_default_with_numpy))
        return ret
wxchan's avatar
wxchan committed
3636

3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
    def predict(
        self,
        data,
        start_iteration: int = 0,
        num_iteration: Optional[int] = None,
        raw_score: bool = False,
        pred_leaf: bool = False,
        pred_contrib: bool = False,
        data_has_header: bool = False,
        validate_features: bool = False,
        **kwargs: Any
    ):
3649
        """Make a prediction.
wxchan's avatar
wxchan committed
3650
3651
3652

        Parameters
        ----------
3653
        data : str, pathlib.Path, numpy array, pandas DataFrame, H2O DataTable's Frame or scipy.sparse
3654
            Data source for prediction.
3655
            If str or pathlib.Path, it represents the path to a text file (CSV, TSV, or LibSVM).
3656
        start_iteration : int, optional (default=0)
3657
            Start index of the iteration to predict.
3658
            If <= 0, starts from the first iteration.
3659
        num_iteration : int or None, optional (default=None)
3660
3661
3662
3663
            Total number of iterations used in the prediction.
            If None, if the best iteration exists and start_iteration <= 0, the best iteration is used;
            otherwise, all iterations from ``start_iteration`` are used (no limits).
            If <= 0, all iterations from ``start_iteration`` are used (no limits).
3664
3665
3666
3667
        raw_score : bool, optional (default=False)
            Whether to predict raw scores.
        pred_leaf : bool, optional (default=False)
            Whether to predict leaf index.
3668
3669
        pred_contrib : bool, optional (default=False)
            Whether to predict feature contributions.
3670

Nikita Titov's avatar
Nikita Titov committed
3671
3672
3673
3674
3675
3676
3677
            .. note::

                If you want to get more explanations for your model's predictions using SHAP values,
                like SHAP interaction values,
                you can install the shap package (https://github.com/slundberg/shap).
                Note that unlike the shap package, with ``pred_contrib`` we return a matrix with an extra
                column, where the last column is the expected value.
3678

3679
3680
        data_has_header : bool, optional (default=False)
            Whether the data has header.
3681
            Used only if data is str.
3682
3683
3684
        validate_features : bool, optional (default=False)
            If True, ensure that the features used to predict match the ones used to train.
            Used only if data is pandas DataFrame.
3685
3686
        **kwargs
            Other parameters for the prediction.
wxchan's avatar
wxchan committed
3687
3688
3689

        Returns
        -------
3690
        result : numpy array, scipy.sparse or list of scipy.sparse
3691
            Prediction result.
3692
            Can be sparse or a list of sparse objects (each element represents predictions for one class) for feature contributions (when ``pred_contrib=True``).
wxchan's avatar
wxchan committed
3693
        """
3694
        predictor = self._to_predictor(deepcopy(kwargs))
3695
        if num_iteration is None:
3696
            if start_iteration <= 0:
3697
3698
3699
3700
                num_iteration = self.best_iteration
            else:
                num_iteration = -1
        return predictor.predict(data, start_iteration, num_iteration,
3701
                                 raw_score, pred_leaf, pred_contrib,
3702
                                 data_has_header, validate_features)
wxchan's avatar
wxchan committed
3703

3704
3705
3706
3707
    def refit(
        self,
        data,
        label,
3708
3709
        decay_rate: float = 0.9,
        reference: Optional[Dataset] = None,
3710
3711
3712
        weight=None,
        group=None,
        init_score=None,
3713
3714
3715
3716
3717
        feature_name: Union[str, List[str]] = 'auto',
        categorical_feature: Union[str, List[str], List[int]] = 'auto',
        dataset_params: Optional[Dict[str, Any]] = None,
        free_raw_data: bool = True,
        validate_features: bool = False,
3718
3719
        **kwargs
    ):
Guolin Ke's avatar
Guolin Ke committed
3720
3721
3722
3723
        """Refit the existing Booster by new data.

        Parameters
        ----------
3724
        data : str, pathlib.Path, numpy array, pandas DataFrame, H2O DataTable's Frame or scipy.sparse
Guolin Ke's avatar
Guolin Ke committed
3725
            Data source for refit.
3726
            If str or pathlib.Path, it represents the path to a text file (CSV, TSV, or LibSVM).
3727
        label : list, numpy 1-D array or pandas Series / one-column DataFrame
Guolin Ke's avatar
Guolin Ke committed
3728
3729
            Label for refit.
        decay_rate : float, optional (default=0.9)
3730
3731
            Decay rate of refit,
            will use ``leaf_output = decay_rate * old_leaf_output + (1.0 - decay_rate) * new_leaf_output`` to refit trees.
3732
3733
3734
        reference : Dataset or None, optional (default=None)
            Reference for ``data``.
        weight : list, numpy 1-D array, pandas Series or None, optional (default=None)
3735
            Weight for each ``data`` instance. Weights should be non-negative.
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
        group : list, numpy 1-D array, pandas Series or None, optional (default=None)
            Group/query size for ``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.
        init_score : list, list of lists (for multi-class task), numpy array, pandas Series, pandas DataFrame (for multi-class task), or None, optional (default=None)
            Init score for ``data``.
        feature_name : list of str, or 'auto', optional (default="auto")
            Feature names for ``data``.
            If 'auto' and data is pandas DataFrame, data columns names are used.
        categorical_feature : list of str or int, or 'auto', optional (default="auto")
            Categorical features for ``data``.
            If list of int, interpreted as indices.
            If list of str, interpreted as feature names (need to specify ``feature_name`` as well).
            If 'auto' and data is pandas DataFrame, pandas unordered categorical columns are used.
3752
            All values in categorical features will be cast to int32 and thus should be less than int32 max value (2147483647).
3753
3754
3755
            Large values could be memory consuming. Consider using consecutive integers starting from zero.
            All negative values in categorical features will be treated as missing values.
            The output cannot be monotonically constrained with respect to a categorical feature.
3756
            Floating point numbers in categorical features will be rounded towards 0.
3757
3758
3759
3760
        dataset_params : dict or None, optional (default=None)
            Other parameters for Dataset ``data``.
        free_raw_data : bool, optional (default=True)
            If True, raw data is freed after constructing inner Dataset for ``data``.
3761
3762
3763
        validate_features : bool, optional (default=False)
            If True, ensure that the features used to refit the model match the original ones.
            Used only if data is pandas DataFrame.
3764
3765
        **kwargs
            Other parameters for refit.
3766
            These parameters will be passed to ``predict`` method.
Guolin Ke's avatar
Guolin Ke committed
3767
3768
3769
3770
3771
3772

        Returns
        -------
        result : Booster
            Refitted Booster.
        """
3773
3774
        if self.__set_objective_to_none:
            raise LightGBMError('Cannot refit due to null objective function.')
3775
3776
        if dataset_params is None:
            dataset_params = {}
3777
        predictor = self._to_predictor(deepcopy(kwargs))
3778
        leaf_preds = predictor.predict(data, -1, pred_leaf=True, validate_features=validate_features)
3779
        nrow, ncol = leaf_preds.shape
3780
        out_is_linear = ctypes.c_int(0)
3781
3782
3783
        _safe_call(_LIB.LGBM_BoosterGetLinear(
            self.handle,
            ctypes.byref(out_is_linear)))
Nikita Titov's avatar
Nikita Titov committed
3784
3785
3786
3787
3788
        new_params = _choose_param_value(
            main_param_name="linear_tree",
            params=self.params,
            default_value=None
        )
3789
        new_params["linear_tree"] = bool(out_is_linear.value)
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
        new_params.update(dataset_params)
        train_set = Dataset(
            data=data,
            label=label,
            reference=reference,
            weight=weight,
            group=group,
            init_score=init_score,
            feature_name=feature_name,
            categorical_feature=categorical_feature,
            params=new_params,
            free_raw_data=free_raw_data,
        )
3803
        new_params['refit_decay_rate'] = decay_rate
3804
        new_booster = Booster(new_params, train_set)
Guolin Ke's avatar
Guolin Ke committed
3805
3806
3807
3808
3809
        # Copy models
        _safe_call(_LIB.LGBM_BoosterMerge(
            new_booster.handle,
            predictor.handle))
        leaf_preds = leaf_preds.reshape(-1)
3810
        ptr_data, _, _ = c_int_array(leaf_preds)
Guolin Ke's avatar
Guolin Ke committed
3811
3812
3813
        _safe_call(_LIB.LGBM_BoosterRefit(
            new_booster.handle,
            ptr_data,
3814
3815
            ctypes.c_int32(nrow),
            ctypes.c_int32(ncol)))
3816
        new_booster.network = self.network
Guolin Ke's avatar
Guolin Ke committed
3817
3818
        return new_booster

3819
    def get_leaf_output(self, tree_id: int, leaf_id: int) -> float:
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
        """Get the output of a leaf.

        Parameters
        ----------
        tree_id : int
            The index of the tree.
        leaf_id : int
            The index of the leaf in the tree.

        Returns
        -------
        result : float
            The output of the leaf.
        """
3834
3835
3836
3837
3838
3839
3840
3841
        ret = ctypes.c_double(0)
        _safe_call(_LIB.LGBM_BoosterGetLeafValue(
            self.handle,
            ctypes.c_int(tree_id),
            ctypes.c_int(leaf_id),
            ctypes.byref(ret)))
        return ret.value

3842
3843
3844
3845
    def _to_predictor(
        self,
        pred_parameter: Optional[Dict[str, Any]] = None
    ) -> _InnerPredictor:
3846
        """Convert to predictor."""
3847
        predictor = _InnerPredictor(booster_handle=self.handle, pred_parameter=pred_parameter)
3848
        predictor.pandas_categorical = self.pandas_categorical
wxchan's avatar
wxchan committed
3849
3850
        return predictor

3851
    def num_feature(self) -> int:
3852
3853
3854
3855
3856
3857
3858
        """Get number of features.

        Returns
        -------
        num_feature : int
            The number of features.
        """
3859
3860
3861
3862
3863
3864
        out_num_feature = ctypes.c_int(0)
        _safe_call(_LIB.LGBM_BoosterGetNumFeature(
            self.handle,
            ctypes.byref(out_num_feature)))
        return out_num_feature.value

3865
    def feature_name(self) -> List[str]:
3866
        """Get names of features.
wxchan's avatar
wxchan committed
3867
3868
3869

        Returns
        -------
3870
        result : list of str
3871
            List with names of features.
wxchan's avatar
wxchan committed
3872
        """
3873
        num_feature = self.num_feature()
3874
        # Get name of features
wxchan's avatar
wxchan committed
3875
        tmp_out_len = ctypes.c_int(0)
3876
3877
        reserved_string_buffer_size = 255
        required_string_buffer_size = ctypes.c_size_t(0)
3878
        string_buffers = [ctypes.create_string_buffer(reserved_string_buffer_size) for _ in range(num_feature)]
wxchan's avatar
wxchan committed
3879
3880
3881
        ptr_string_buffers = (ctypes.c_char_p * num_feature)(*map(ctypes.addressof, string_buffers))
        _safe_call(_LIB.LGBM_BoosterGetFeatureNames(
            self.handle,
3882
            ctypes.c_int(num_feature),
wxchan's avatar
wxchan committed
3883
            ctypes.byref(tmp_out_len),
3884
            ctypes.c_size_t(reserved_string_buffer_size),
3885
            ctypes.byref(required_string_buffer_size),
wxchan's avatar
wxchan committed
3886
3887
3888
            ptr_string_buffers))
        if num_feature != tmp_out_len.value:
            raise ValueError("Length of feature names doesn't equal with num_feature")
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
        actual_string_buffer_size = required_string_buffer_size.value
        # if buffer length is not long enough, reallocate buffers
        if reserved_string_buffer_size < actual_string_buffer_size:
            string_buffers = [ctypes.create_string_buffer(actual_string_buffer_size) for _ in range(num_feature)]
            ptr_string_buffers = (ctypes.c_char_p * num_feature)(*map(ctypes.addressof, string_buffers))
            _safe_call(_LIB.LGBM_BoosterGetFeatureNames(
                self.handle,
                ctypes.c_int(num_feature),
                ctypes.byref(tmp_out_len),
                ctypes.c_size_t(actual_string_buffer_size),
                ctypes.byref(required_string_buffer_size),
                ptr_string_buffers))
3901
        return [string_buffers[i].value.decode('utf-8') for i in range(num_feature)]
wxchan's avatar
wxchan committed
3902

3903
3904
3905
3906
3907
    def feature_importance(
        self,
        importance_type: str = 'split',
        iteration: Optional[int] = None
    ) -> np.ndarray:
3908
        """Get feature importances.
3909

3910
3911
        Parameters
        ----------
3912
        importance_type : str, optional (default="split")
3913
3914
3915
            How the importance is calculated.
            If "split", result contains numbers of times the feature is used in a model.
            If "gain", result contains total gains of splits which use the feature.
3916
3917
3918
3919
        iteration : int or None, optional (default=None)
            Limit number of iterations in the feature importance calculation.
            If None, if the best iteration exists, it is used; otherwise, all trees are used.
            If <= 0, all trees are used (no limits).
3920

3921
3922
        Returns
        -------
3923
3924
        result : numpy array
            Array with feature importances.
3925
        """
3926
3927
        if iteration is None:
            iteration = self.best_iteration
3928
        importance_type_int = FEATURE_IMPORTANCE_TYPE_MAPPER[importance_type]
3929
        result = np.empty(self.num_feature(), dtype=np.float64)
3930
3931
3932
3933
3934
        _safe_call(_LIB.LGBM_BoosterFeatureImportance(
            self.handle,
            ctypes.c_int(iteration),
            ctypes.c_int(importance_type_int),
            result.ctypes.data_as(ctypes.POINTER(ctypes.c_double))))
3935
        if importance_type_int == C_API_FEATURE_IMPORTANCE_SPLIT:
3936
            return result.astype(np.int32)
3937
3938
        else:
            return result
3939

3940
3941
3942
3943
3944
3945
    def get_split_value_histogram(
        self,
        feature: Union[int, str],
        bins: Optional[Union[int, str]] = None,
        xgboost_style: bool = False
    ) -> Union[Tuple[np.ndarray, np.ndarray], np.ndarray, pd_DataFrame]:
3946
3947
3948
3949
        """Get split value histogram for the specified feature.

        Parameters
        ----------
3950
        feature : int or str
3951
3952
            The feature name or index the histogram is calculated for.
            If int, interpreted as index.
3953
            If str, interpreted as name.
3954

Nikita Titov's avatar
Nikita Titov committed
3955
3956
3957
            .. warning::

                Categorical features are not supported.
3958

3959
        bins : int, str or None, optional (default=None)
3960
3961
3962
            The maximum number of bins.
            If None, or int and > number of unique split values and ``xgboost_style=True``,
            the number of bins equals number of unique split values.
3963
            If str, it should be one from the list of the supported values by ``numpy.histogram()`` function.
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
        xgboost_style : bool, optional (default=False)
            Whether the returned result should be in the same form as it is in XGBoost.
            If False, the returned value is tuple of 2 numpy arrays as it is in ``numpy.histogram()`` function.
            If True, the returned value is matrix, in which the first column is the right edges of non-empty bins
            and the second one is the histogram values.

        Returns
        -------
        result_tuple : tuple of 2 numpy arrays
            If ``xgboost_style=False``, the values of the histogram of used splitting values for the specified feature
            and the bin edges.
        result_array_like : numpy array or pandas DataFrame (if pandas is installed)
            If ``xgboost_style=True``, the histogram of used splitting values for the specified feature.
        """
        def add(root):
            """Recursively add thresholds."""
            if 'split_index' in root:  # non-leaf
3981
                if feature_names is not None and isinstance(feature, str):
3982
3983
3984
3985
                    split_feature = feature_names[root['split_feature']]
                else:
                    split_feature = root['split_feature']
                if split_feature == feature:
3986
                    if isinstance(root['threshold'], str):
3987
3988
3989
                        raise LightGBMError('Cannot compute split value histogram for the categorical feature')
                    else:
                        values.append(root['threshold'])
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
                add(root['left_child'])
                add(root['right_child'])

        model = self.dump_model()
        feature_names = model.get('feature_names')
        tree_infos = model['tree_info']
        values = []
        for tree_info in tree_infos:
            add(tree_info['tree_structure'])

4000
        if bins is None or isinstance(bins, int) and xgboost_style:
4001
4002
4003
4004
4005
4006
4007
            n_unique = len(np.unique(values))
            bins = max(min(n_unique, bins) if bins is not None else n_unique, 1)
        hist, bin_edges = np.histogram(values, bins=bins)
        if xgboost_style:
            ret = np.column_stack((bin_edges[1:], hist))
            ret = ret[ret[:, 1] > 0]
            if PANDAS_INSTALLED:
4008
                return pd_DataFrame(ret, columns=['SplitValue', 'Count'])
4009
4010
4011
4012
4013
            else:
                return ret
        else:
            return hist, bin_edges

4014
4015
4016
4017
4018
4019
    def __inner_eval(
        self,
        data_name: str,
        data_idx: int,
        feval: Optional[Union[_LGBM_CustomEvalFunction, List[_LGBM_CustomEvalFunction]]] = None
    ) -> List[_LGBM_BoosterEvalMethodResultType]:
4020
        """Evaluate training or validation data."""
wxchan's avatar
wxchan committed
4021
        if data_idx >= self.__num_dataset:
4022
            raise ValueError("Data_idx should be smaller than number of dataset")
wxchan's avatar
wxchan committed
4023
4024
4025
        self.__get_eval_info()
        ret = []
        if self.__num_inner_eval > 0:
4026
            result = np.empty(self.__num_inner_eval, dtype=np.float64)
Guolin Ke's avatar
Guolin Ke committed
4027
            tmp_out_len = ctypes.c_int(0)
wxchan's avatar
wxchan committed
4028
4029
            _safe_call(_LIB.LGBM_BoosterGetEval(
                self.handle,
Guolin Ke's avatar
Guolin Ke committed
4030
                ctypes.c_int(data_idx),
wxchan's avatar
wxchan committed
4031
                ctypes.byref(tmp_out_len),
Guolin Ke's avatar
Guolin Ke committed
4032
                result.ctypes.data_as(ctypes.POINTER(ctypes.c_double))))
wxchan's avatar
wxchan committed
4033
            if tmp_out_len.value != self.__num_inner_eval:
4034
                raise ValueError("Wrong length of eval results")
4035
            for i in range(self.__num_inner_eval):
4036
4037
                ret.append((data_name, self.__name_inner_eval[i],
                            result[i], self.__higher_better_inner_eval[i]))
4038
4039
        if callable(feval):
            feval = [feval]
wxchan's avatar
wxchan committed
4040
4041
4042
4043
4044
        if feval is not None:
            if data_idx == 0:
                cur_data = self.train_set
            else:
                cur_data = self.valid_sets[data_idx - 1]
4045
4046
4047
4048
4049
4050
4051
4052
4053
            for eval_function in feval:
                if eval_function is None:
                    continue
                feval_ret = eval_function(self.__inner_predict(data_idx), cur_data)
                if isinstance(feval_ret, list):
                    for eval_name, val, is_higher_better in feval_ret:
                        ret.append((data_name, eval_name, val, is_higher_better))
                else:
                    eval_name, val, is_higher_better = feval_ret
wxchan's avatar
wxchan committed
4054
4055
4056
                    ret.append((data_name, eval_name, val, is_higher_better))
        return ret

4057
    def __inner_predict(self, data_idx: int):
4058
        """Predict for training and validation dataset."""
wxchan's avatar
wxchan committed
4059
        if data_idx >= self.__num_dataset:
4060
            raise ValueError("Data_idx should be smaller than number of dataset")
wxchan's avatar
wxchan committed
4061
4062
4063
4064
4065
        if self.__inner_predict_buffer[data_idx] is None:
            if data_idx == 0:
                n_preds = self.train_set.num_data() * self.__num_class
            else:
                n_preds = self.valid_sets[data_idx - 1].num_data() * self.__num_class
4066
            self.__inner_predict_buffer[data_idx] = np.empty(n_preds, dtype=np.float64)
4067
        # avoid to predict many time in one iteration
wxchan's avatar
wxchan committed
4068
4069
        if not self.__is_predicted_cur_iter[data_idx]:
            tmp_out_len = ctypes.c_int64(0)
Guolin Ke's avatar
Guolin Ke committed
4070
            data_ptr = self.__inner_predict_buffer[data_idx].ctypes.data_as(ctypes.POINTER(ctypes.c_double))
wxchan's avatar
wxchan committed
4071
4072
            _safe_call(_LIB.LGBM_BoosterGetPredict(
                self.handle,
Guolin Ke's avatar
Guolin Ke committed
4073
                ctypes.c_int(data_idx),
wxchan's avatar
wxchan committed
4074
4075
4076
                ctypes.byref(tmp_out_len),
                data_ptr))
            if tmp_out_len.value != len(self.__inner_predict_buffer[data_idx]):
4077
                raise ValueError(f"Wrong length of predict results for data {data_idx}")
wxchan's avatar
wxchan committed
4078
            self.__is_predicted_cur_iter[data_idx] = True
4079
4080
4081
4082
4083
        result = self.__inner_predict_buffer[data_idx]
        if self.__num_class > 1:
            num_data = result.size // self.__num_class
            result = result.reshape(num_data, self.__num_class, order='F')
        return result
wxchan's avatar
wxchan committed
4084

4085
    def __get_eval_info(self) -> None:
4086
        """Get inner evaluation count and names."""
wxchan's avatar
wxchan committed
4087
4088
        if self.__need_reload_eval_info:
            self.__need_reload_eval_info = False
Guolin Ke's avatar
Guolin Ke committed
4089
            out_num_eval = ctypes.c_int(0)
4090
            # Get num of inner evals
wxchan's avatar
wxchan committed
4091
4092
4093
4094
4095
            _safe_call(_LIB.LGBM_BoosterGetEvalCounts(
                self.handle,
                ctypes.byref(out_num_eval)))
            self.__num_inner_eval = out_num_eval.value
            if self.__num_inner_eval > 0:
4096
                # Get name of eval metrics
Guolin Ke's avatar
Guolin Ke committed
4097
                tmp_out_len = ctypes.c_int(0)
4098
4099
4100
                reserved_string_buffer_size = 255
                required_string_buffer_size = ctypes.c_size_t(0)
                string_buffers = [
4101
                    ctypes.create_string_buffer(reserved_string_buffer_size) for _ in range(self.__num_inner_eval)
4102
                ]
wxchan's avatar
wxchan committed
4103
                ptr_string_buffers = (ctypes.c_char_p * self.__num_inner_eval)(*map(ctypes.addressof, string_buffers))
wxchan's avatar
wxchan committed
4104
4105
                _safe_call(_LIB.LGBM_BoosterGetEvalNames(
                    self.handle,
4106
                    ctypes.c_int(self.__num_inner_eval),
wxchan's avatar
wxchan committed
4107
                    ctypes.byref(tmp_out_len),
4108
                    ctypes.c_size_t(reserved_string_buffer_size),
4109
                    ctypes.byref(required_string_buffer_size),
wxchan's avatar
wxchan committed
4110
4111
                    ptr_string_buffers))
                if self.__num_inner_eval != tmp_out_len.value:
4112
                    raise ValueError("Length of eval names doesn't equal with num_evals")
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
                actual_string_buffer_size = required_string_buffer_size.value
                # if buffer length is not long enough, reallocate buffers
                if reserved_string_buffer_size < actual_string_buffer_size:
                    string_buffers = [
                        ctypes.create_string_buffer(actual_string_buffer_size) for _ in range(self.__num_inner_eval)
                    ]
                    ptr_string_buffers = (ctypes.c_char_p * self.__num_inner_eval)(*map(ctypes.addressof, string_buffers))
                    _safe_call(_LIB.LGBM_BoosterGetEvalNames(
                        self.handle,
                        ctypes.c_int(self.__num_inner_eval),
                        ctypes.byref(tmp_out_len),
                        ctypes.c_size_t(actual_string_buffer_size),
                        ctypes.byref(required_string_buffer_size),
                        ptr_string_buffers))
                self.__name_inner_eval = [
                    string_buffers[i].value.decode('utf-8') for i in range(self.__num_inner_eval)
                ]
                self.__higher_better_inner_eval = [
                    name.startswith(('auc', 'ndcg@', 'map@', 'average_precision')) for name in self.__name_inner_eval
                ]