from __future__ import annotations
import gc
import logging
import os
import re
import time
import warnings
from types import MappingProxyType
import numpy as np
import pandas as pd
from pandas import DataFrame, Series
from autogluon.common.features.types import R_BOOL, R_CATEGORY, R_FLOAT, R_INT
from autogluon.common.utils.pandas_utils import get_approximate_df_mem_usage
from autogluon.common.utils.resource_utils import ResourceManager
from autogluon.common.utils.try_import import try_import_lightgbm
from autogluon.core.constants import BINARY, MULTICLASS, QUANTILE, REGRESSION, SOFTCLASS
from autogluon.core.models import AbstractModel
from autogluon.core.models._utils import get_early_stopping_rounds
from . import lgb_utils
from .hyperparameters.parameters import DEFAULT_NUM_BOOST_ROUND, get_lgb_objective, get_param_baseline
from .hyperparameters.searchspaces import get_default_searchspace
from .lgb_utils import construct_dataset, train_lgb_model
warnings.filterwarnings("ignore", category=UserWarning, message="Starting from version") # lightGBM brew libomp warning
warnings.filterwarnings("ignore", category=FutureWarning, message="Dask dataframe query") # lightGBM dask-expr warning
logger = logging.getLogger(__name__)
# TODO: Save dataset to binary and reload for HPO. This will avoid the memory spike overhead when training each model and instead it will only occur once upon saving the dataset.
[docs]
class LGBModel(AbstractModel):
"""
LightGBM model: https://lightgbm.readthedocs.io/en/latest/
Hyperparameter options: https://lightgbm.readthedocs.io/en/latest/Parameters.html
Extra hyperparameter options:
ag.early_stop : int, specifies the early stopping rounds. Defaults to an adaptive strategy. Recommended to keep default.
"""
ag_key = "GBM"
ag_name = "LightGBM"
ag_priority = 90
ag_priority_by_problem_type = MappingProxyType({
SOFTCLASS: 100
})
def __init__(self, **kwargs):
super().__init__(**kwargs)
self._features_internal_map = None
self._features_internal_list = None
self._requires_remap = None
def _set_default_params(self):
default_params = get_param_baseline(problem_type=self.problem_type)
for param, val in default_params.items():
self._set_default_param_value(param, val)
def _get_default_searchspace(self):
return get_default_searchspace(problem_type=self.problem_type)
# Use specialized LightGBM metric if available (fast), otherwise use custom func generator
def _get_stopping_metric_internal(self):
stopping_metric = lgb_utils.convert_ag_metric_to_lgbm(ag_metric_name=self.stopping_metric.name, problem_type=self.problem_type)
if stopping_metric is None:
stopping_metric = lgb_utils.func_generator(
metric=self.stopping_metric, is_higher_better=True, needs_pred_proba=not self.stopping_metric.needs_pred, problem_type=self.problem_type
)
stopping_metric_name = self.stopping_metric.name
else:
stopping_metric_name = stopping_metric
return stopping_metric, stopping_metric_name
def _estimate_memory_usage(self, X: pd.DataFrame, **kwargs) -> int:
hyperparameters = self._get_model_params()
return self.estimate_memory_usage_static(X=X, problem_type=self.problem_type, num_classes=self.num_classes, hyperparameters=hyperparameters, **kwargs)
# FIXME: Don't use `hyperparameters.get("max_bins", 255)`, instead get the defaults all at once!
@classmethod
def _estimate_memory_usage_static(
cls,
*,
X: DataFrame,
hyperparameters: dict = None,
num_classes: int = 1,
**kwargs,
) -> int:
"""
Returns the expected peak memory usage in bytes of the LightGBM model during fit.
The memory usage of LightGBM is primarily made up of three sources:
1. The size of the data
2. The size of the histogram cache
Scales roughly by 5100*num_features*num_leaves bytes
For 10000 features and 128 num_leaves, the histogram would be 6.5 GB.
3. The size of the model
Scales linearly with the number of estimators, number of classes, and number of leaves.
Memory usage peaks during model saving, with the peak consuming approximately 2-4x the size of the model in memory.
"""
if hyperparameters is None:
hyperparameters = {}
num_classes = num_classes if num_classes else 1 # num_classes could be None after initialization if it's a regression problem
data_mem_usage = get_approximate_df_mem_usage(X).sum()
data_mem_usage_bytes = data_mem_usage * 5 + data_mem_usage / 4 * num_classes # TODO: Extremely crude approximation, can be vastly improved
n_trees_per_estimator = num_classes if num_classes > 2 else 1
max_bins = hyperparameters.get("max_bins", 255)
num_leaves = hyperparameters.get("num_leaves", 31)
# Memory usage of histogram based on https://github.com/microsoft/LightGBM/issues/562#issuecomment-304524592
histogram_mem_usage_bytes = 20 * max_bins * len(X.columns) * num_leaves
histogram_mem_usage_bytes_max = hyperparameters.get("histogram_pool_size", None)
if histogram_mem_usage_bytes_max is not None:
histogram_mem_usage_bytes_max *= 1e6 # Convert megabytes to bytes, `histogram_pool_size` is in MB.
if histogram_mem_usage_bytes > histogram_mem_usage_bytes_max:
histogram_mem_usage_bytes = histogram_mem_usage_bytes_max
histogram_mem_usage_bytes *= 1.2 # Add a 20% buffer
mem_size_per_estimator = n_trees_per_estimator * num_leaves * 100 # very rough estimate
n_estimators = hyperparameters.get("num_boost_round", DEFAULT_NUM_BOOST_ROUND)
n_estimators_min = min(n_estimators, 1000)
mem_size_estimators = n_estimators_min * mem_size_per_estimator # memory estimate after fitting up to 1000 estimators
approx_mem_size_req = data_mem_usage_bytes + histogram_mem_usage_bytes + mem_size_estimators
return approx_mem_size_req
def _fit(self, X, y, X_val=None, y_val=None, time_limit=None, num_gpus=0, num_cpus=0, sample_weight=None, sample_weight_val=None, verbosity=2, **kwargs):
try_import_lightgbm() # raise helpful error message if LightGBM isn't installed
start_time = time.time()
ag_params = self._get_ag_params()
params = self._get_model_params()
generate_curves = ag_params.get("generate_curves", False)
if generate_curves:
X_test = kwargs.get("X_test", None)
y_test = kwargs.get("y_test", None)
else:
X_test = None
y_test = None
if verbosity <= 1:
log_period = False
elif verbosity == 2:
log_period = 1000
elif verbosity == 3:
log_period = 50
else:
log_period = 1
stopping_metric, stopping_metric_name = self._get_stopping_metric_internal()
num_boost_round = params.pop("num_boost_round", DEFAULT_NUM_BOOST_ROUND)
dart_retrain = params.pop("dart_retrain", False) # Whether to retrain the model to get optimal iteration if model is trained in 'dart' mode.
if num_gpus != 0:
if "device" not in params:
# TODO: lightgbm must have a special install to support GPU: https://github.com/Microsoft/LightGBM/tree/master/python-package#build-gpu-version
# Before enabling GPU, we should add code to detect that GPU-enabled version is installed and that a valid GPU exists.
# GPU training heavily alters accuracy, often in a negative manner. We will have to be careful about when to use GPU.
params["device"] = "gpu"
logger.log(20, f"\tTraining {self.name} with GPU, note that this may negatively impact model quality compared to CPU training.")
logger.log(15, f"\tFitting {num_boost_round} rounds... Hyperparameters: {params}")
if "num_threads" not in params:
params["num_threads"] = num_cpus
if "objective" not in params:
params["objective"] = get_lgb_objective(problem_type=self.problem_type)
if self.problem_type in [MULTICLASS, SOFTCLASS] and "num_classes" not in params:
params["num_classes"] = self.num_classes
if "verbose" not in params:
params["verbose"] = -1
num_rows_train = len(X)
dataset_train, dataset_val, dataset_test = self.generate_datasets(
X=X, y=y, params=params, X_val=X_val, y_val=y_val, X_test=X_test, y_test=y_test, sample_weight=sample_weight, sample_weight_val=sample_weight_val
)
gc.collect()
callbacks = []
valid_names = []
valid_sets = []
if dataset_val is not None:
from .callbacks import early_stopping_custom
# TODO: Better solution: Track trend to early stop when score is far worse than best score, or score is trending worse over time
early_stopping_rounds = ag_params.get("early_stop", "adaptive")
if isinstance(early_stopping_rounds, (str, tuple, list)):
early_stopping_rounds = self._get_early_stopping_rounds(num_rows_train=num_rows_train, strategy=early_stopping_rounds)
if early_stopping_rounds is None:
early_stopping_rounds = 999999
reporter = kwargs.get("reporter", None)
train_loss_name = self._get_train_loss_name() if reporter is not None else None
if train_loss_name is not None:
if "metric" not in params or params["metric"] == "":
params["metric"] = train_loss_name
elif train_loss_name not in params["metric"]:
params["metric"] = f'{params["metric"]},{train_loss_name}'
# early stopping callback will be added later by QuantileBooster if problem_type==QUANTILE
early_stopping_callback_kwargs = dict(
stopping_rounds=early_stopping_rounds,
metrics_to_use=[("valid_set", stopping_metric_name)],
max_diff=None,
start_time=start_time,
time_limit=time_limit,
ignore_dart_warning=True,
verbose=False,
manual_stop_file=False,
reporter=reporter,
train_loss_name=train_loss_name,
)
callbacks += [
# Note: Don't use self.params_aux['max_memory_usage_ratio'] here as LightGBM handles memory per iteration optimally. # TODO: Consider using when ratio < 1.
early_stopping_custom(**early_stopping_callback_kwargs)
]
valid_names = ["valid_set"] + valid_names
valid_sets = [dataset_val] + valid_sets
else:
early_stopping_callback_kwargs = None
from lightgbm.callback import log_evaluation, record_evaluation
if log_period is not None:
callbacks.append(log_evaluation(period=log_period))
seed_val = params.pop("seed_value", 0)
train_params = {
"params": params,
"train_set": dataset_train,
"num_boost_round": num_boost_round,
"valid_names": valid_names,
"valid_sets": valid_sets,
"callbacks": callbacks,
"keep_training_booster": generate_curves,
}
if generate_curves:
scorers = ag_params.get("curve_metrics", [self.eval_metric])
use_curve_metric_error = ag_params.get("use_error_for_curve_metrics", False)
metric_names = [scorer.name for scorer in scorers]
if stopping_metric_name in metric_names:
idx = metric_names.index(stopping_metric_name)
scorers[idx].name = f"_{stopping_metric_name}"
metric_names[idx] = scorers[idx].name
custom_metrics = [
lgb_utils.func_generator(
metric=scorer,
is_higher_better=scorer.greater_is_better_internal,
needs_pred_proba=not scorer.needs_pred,
problem_type=self.problem_type,
error=use_curve_metric_error,
)
for scorer in scorers
]
eval_results = {}
train_params["callbacks"].append(record_evaluation(eval_results))
train_params["feval"] = custom_metrics
if dataset_test is not None:
train_params["valid_names"] = ["train_set", "test_set"] + train_params["valid_names"]
train_params["valid_sets"] = [dataset_train, dataset_test] + train_params["valid_sets"]
else:
train_params["valid_names"] = ["train_set"] + train_params["valid_names"]
train_params["valid_sets"] = [dataset_train] + train_params["valid_sets"]
# NOTE: lgb stops based on first metric if more than one
if not isinstance(stopping_metric, str):
if generate_curves:
train_params["feval"].insert(0, stopping_metric)
else:
train_params["feval"] = stopping_metric
elif isinstance(stopping_metric, str):
if "metric" not in train_params["params"] or train_params["params"]["metric"] == "":
train_params["params"]["metric"] = stopping_metric
elif stopping_metric not in train_params["params"]["metric"]:
train_params["params"]["metric"] = f'{stopping_metric},{train_params["params"]["metric"]}'
if self.problem_type == SOFTCLASS:
train_params["params"]["objective"] = lgb_utils.softclass_lgbobj
train_params["params"]["num_classes"] = self.num_classes
elif self.problem_type == QUANTILE:
train_params["params"]["quantile_levels"] = self.quantile_levels
if seed_val is not None:
train_params["params"]["seed"] = seed_val
# Train LightGBM model:
# Note that self.model contains a <class 'lightgbm.basic.Booster'> not a LightBGMClassifier or LightGBMRegressor object
from lightgbm.basic import LightGBMError
with warnings.catch_warnings():
# Filter harmless warnings introduced in lightgbm 3.0, future versions plan to remove: https://github.com/microsoft/LightGBM/issues/3379
warnings.filterwarnings("ignore", message="Overriding the parameters from Reference Dataset.")
warnings.filterwarnings("ignore", message="categorical_column in param dict is overridden.")
try:
self.model = train_lgb_model(early_stopping_callback_kwargs=early_stopping_callback_kwargs, **train_params)
except LightGBMError:
if train_params["params"].get("device", "cpu") != "gpu":
raise
else:
logger.warning(
"Warning: GPU mode might not be installed for LightGBM, GPU training raised an exception. Falling back to CPU training..."
"Refer to LightGBM GPU documentation: https://github.com/Microsoft/LightGBM/tree/master/python-package#build-gpu-version"
"One possible method is:"
"\tpip uninstall lightgbm -y"
"\tpip install lightgbm --install-option=--gpu"
)
train_params["params"]["device"] = "cpu"
self.model = train_lgb_model(early_stopping_callback_kwargs=early_stopping_callback_kwargs, **train_params)
retrain = False
if train_params["params"].get("boosting_type", "") == "dart":
if dataset_val is not None and dart_retrain and (self.model.best_iteration != num_boost_round):
retrain = True
if time_limit is not None:
time_left = time_limit + start_time - time.time()
if time_left < 0.5 * time_limit:
retrain = False
if retrain:
logger.log(15, f"Retraining LGB model to optimal iterations ('dart' mode).")
train_params.pop("callbacks", None)
train_params.pop("valid_sets", None)
train_params.pop("valid_names", None)
train_params["num_boost_round"] = self.model.best_iteration
self.model = train_lgb_model(**train_params)
else:
logger.log(15, f"Not enough time to retrain LGB model ('dart' mode)...")
if generate_curves:
def og_name(key):
if key == f"_{stopping_metric_name}":
return stopping_metric_name
return key
def filter(d, keys):
return {og_name(key): d[key] for key in keys if key in d}
curves = {"train": filter(eval_results["train_set"], metric_names)}
if X_val is not None:
curves["val"] = filter(eval_results["valid_set"], metric_names)
if X_test is not None:
curves["test"] = filter(eval_results["test_set"], metric_names)
if f"_{stopping_metric_name}" in metric_names:
idx = metric_names.index(f"_{stopping_metric_name}")
metric_names[idx] = stopping_metric_name
self.save_learning_curves(metrics=metric_names, curves=curves)
if dataset_val is not None and not retrain:
self.params_trained["num_boost_round"] = self.model.best_iteration
else:
self.params_trained["num_boost_round"] = self.model.current_iteration()
def _predict_proba(self, X, num_cpus=0, **kwargs) -> np.ndarray:
X = self.preprocess(X, **kwargs)
y_pred_proba = self.model.predict(X, num_threads=num_cpus)
if self.problem_type == QUANTILE:
# y_pred_proba is a pd.DataFrame, need to convert
y_pred_proba = y_pred_proba.to_numpy()
if self.problem_type in [REGRESSION, QUANTILE, MULTICLASS]:
return y_pred_proba
elif self.problem_type == BINARY:
if len(y_pred_proba.shape) == 1:
return y_pred_proba
elif y_pred_proba.shape[1] > 1:
return y_pred_proba[:, 1]
else:
return y_pred_proba
elif self.problem_type == SOFTCLASS: # apply softmax
y_pred_proba = np.exp(y_pred_proba)
y_pred_proba = np.multiply(y_pred_proba, 1 / np.sum(y_pred_proba, axis=1)[:, np.newaxis])
return y_pred_proba
else:
if len(y_pred_proba.shape) == 1:
return y_pred_proba
elif y_pred_proba.shape[1] > 2: # Should this ever happen?
return y_pred_proba
else: # Should this ever happen?
return y_pred_proba[:, 1]
def _preprocess_nonadaptive(self, X, is_train=False, **kwargs):
X = super()._preprocess_nonadaptive(X=X, **kwargs)
if is_train:
self._requires_remap = False
for column in X.columns:
if isinstance(column, str):
new_column = re.sub(r'[",:{}[\]]', "", column)
if new_column != column:
self._features_internal_map = {feature: i for i, feature in enumerate(list(X.columns))}
self._requires_remap = True
break
if self._requires_remap:
self._features_internal_list = np.array([self._features_internal_map[feature] for feature in list(X.columns)])
else:
self._features_internal_list = self._features_internal
if self._requires_remap:
X_new = X.copy(deep=False)
X_new.columns = self._features_internal_list
return X_new
else:
return X
def generate_datasets(
self,
X: DataFrame,
y: Series,
params,
X_val=None,
y_val=None,
X_test=None,
y_test=None,
sample_weight=None,
sample_weight_val=None,
sample_weight_test=None,
save=False,
):
lgb_dataset_params_keys = ["two_round"] # Keys that are specific to lightGBM Dataset object construction.
data_params = {key: params[key] for key in lgb_dataset_params_keys if key in params}.copy()
X = self.preprocess(X, is_train=True)
if X_val is not None:
X_val = self.preprocess(X_val)
if X_test is not None:
X_test = self.preprocess(X_test)
# TODO: Try creating multiple Datasets for subsets of features, then combining with Dataset.add_features_from(), this might avoid memory spike
y_og = None
y_val_og = None
y_test_og = None
if self.problem_type == SOFTCLASS:
y_og = np.array(y)
y = None
if X_val is not None:
y_val_og = np.array(y_val)
y_val = None
if X_test is not None:
y_test_og = np.array(y_test)
y_test = None
# X, W_train = self.convert_to_weight(X=X)
dataset_train = construct_dataset(
x=X, y=y, location=os.path.join("self.path", "datasets", "train"), params=data_params, save=save, weight=sample_weight
)
# dataset_train = construct_dataset_lowest_memory(X=X, y=y, location=self.path + 'datasets/train', params=data_params)
if X_val is not None:
# X_val, W_val = self.convert_to_weight(X=X_val)
dataset_val = construct_dataset(
x=X_val,
y=y_val,
location=os.path.join(self.path, "datasets", "val"),
reference=dataset_train,
params=data_params,
save=save,
weight=sample_weight_val,
)
# dataset_val = construct_dataset_lowest_memory(X=X_val, y=y_val, location=self.path + 'datasets/val', reference=dataset_train, params=data_params)
else:
dataset_val = None
if X_test is not None:
dataset_test = construct_dataset(
x=X_test,
y=y_test,
location=os.path.join(self.path, "datasets", "test"),
reference=dataset_train,
params=data_params,
save=save,
weight=sample_weight_test,
)
else:
dataset_test = None
if self.problem_type == SOFTCLASS:
if y_og is not None:
dataset_train.softlabels = y_og
if y_val_og is not None:
dataset_val.softlabels = y_val_og
if y_test_og is not None:
dataset_test.softlabels = y_test_og
return dataset_train, dataset_val, dataset_test
def _get_train_loss_name(self):
if self.problem_type == BINARY:
train_loss_name = "binary_logloss"
elif self.problem_type == MULTICLASS:
train_loss_name = "multi_logloss"
elif self.problem_type == REGRESSION:
train_loss_name = "l2"
else:
raise ValueError(f"unknown problem_type for LGBModel: {self.problem_type}")
return train_loss_name
def _get_early_stopping_rounds(self, num_rows_train, strategy="auto"):
return get_early_stopping_rounds(num_rows_train=num_rows_train, strategy=strategy)
def _get_default_auxiliary_params(self) -> dict:
default_auxiliary_params = super()._get_default_auxiliary_params()
extra_auxiliary_params = dict(
valid_raw_types=[R_BOOL, R_INT, R_FLOAT, R_CATEGORY],
)
default_auxiliary_params.update(extra_auxiliary_params)
return default_auxiliary_params
def _is_gpu_lgbm_installed(self):
# Taken from https://github.com/microsoft/LightGBM/issues/3939
try_import_lightgbm()
import lightgbm
try:
data = np.random.rand(50, 2)
label = np.random.randint(2, size=50)
train_data = lightgbm.Dataset(data, label=label)
params = {"device": "gpu"}
gbm = lightgbm.train(params, train_set=train_data, verbose=-1)
return True
except Exception as e:
return False
def get_minimum_resources(self, is_gpu_available=False):
minimum_resources = {
"num_cpus": 1,
}
if is_gpu_available and self._is_gpu_lgbm_installed():
minimum_resources["num_gpus"] = 0.5
return minimum_resources
def _get_default_resources(self):
# only_physical_cores=True is faster in training
num_cpus = ResourceManager.get_cpu_count(only_physical_cores=True)
num_gpus = 0
return num_cpus, num_gpus
@classmethod
def supported_problem_types(cls) -> list[str] | None:
return ["binary", "multiclass", "regression", "quantile", "softclass"]
@property
def _features(self):
return self._features_internal_list
def _ag_params(self) -> set:
return {"early_stop", "generate_curves", "curve_metrics", "use_error_for_curve_metrics"}
@classmethod
def _class_tags(cls):
return {
"can_estimate_memory_usage_static": True,
"supports_learning_curves": True,
}
def _more_tags(self):
# `can_refit_full=True` because num_boost_round is communicated at end of `_fit`
return {"can_refit_full": True}