import logging
import math
import os
import time
import numpy as np
import pandas as pd
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_catboost
from autogluon.core.constants import MULTICLASS, PROBLEM_TYPES_CLASSIFICATION, QUANTILE, SOFTCLASS
from autogluon.core.models import AbstractModel
from autogluon.core.models._utils import get_early_stopping_rounds
from autogluon.core.utils.exceptions import TimeLimitExceeded
from .callbacks import EarlyStoppingCallback, MemoryCheckCallback, TimeCheckCallback
from .catboost_utils import get_catboost_metric_from_ag_metric
from .hyperparameters.parameters import get_param_baseline
from .hyperparameters.searchspaces import get_default_searchspace
logger = logging.getLogger(__name__)
# TODO: Consider having CatBoost variant that converts all categoricals to numerical as done in RFModel, was showing improved results in some problems.
[docs]
class CatBoostModel(AbstractModel):
"""
CatBoost model: https://catboost.ai/
Hyperparameter options: https://catboost.ai/en/docs/references/training-parameters
"""
def __init__(self, **kwargs):
super().__init__(**kwargs)
self._category_features = 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)
self._set_default_param_value("random_seed", 0) # Remove randomness for reproducibility
# Set 'allow_writing_files' to True in order to keep log files created by catboost during training (these will be saved in the directory where AutoGluon stores this model)
self._set_default_param_value("allow_writing_files", False) # Disables creation of catboost logging files during training by default
if self.problem_type != SOFTCLASS: # TODO: remove this after catboost 0.24
default_eval_metric = get_catboost_metric_from_ag_metric(self.stopping_metric, self.problem_type, self.quantile_levels)
self._set_default_param_value("eval_metric", default_eval_metric)
def _get_default_searchspace(self):
return get_default_searchspace(self.problem_type, num_classes=self.num_classes)
def _preprocess_nonadaptive(self, X, **kwargs):
X = super()._preprocess_nonadaptive(X, **kwargs)
if self._category_features is None:
self._category_features = list(X.select_dtypes(include="category").columns)
if self._category_features:
X = X.copy()
for category in self._category_features:
current_categories = X[category].cat.categories
if "__NaN__" in current_categories:
X[category] = X[category].fillna("__NaN__")
else:
X[category] = X[category].cat.add_categories("__NaN__").fillna("__NaN__")
return X
def _estimate_memory_usage(self, X: pd.DataFrame, **kwargs) -> float:
"""
Returns the expected peak memory usage in bytes of the CatBoost model during fit.
The memory usage of CatBoost is primarily made up of two sources:
1. The size of the data
2. The size of the histogram cache
Scales roughly by 5080*num_features*2^depth bytes
For 10000 features and 6 depth, the histogram would be 3.2 GB.
"""
num_classes = self.num_classes if self.num_classes else 1 # self.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
params = self._get_model_params(convert_search_spaces_to_default=True)
border_count = params.get("border_count", 254)
depth = params.get("depth", 6)
# Formula based on manual testing, aligns with LightGBM histogram sizes
histogram_mem_usage_bytes = 20 * math.pow(2, depth) * len(X.columns) * border_count
histogram_mem_usage_bytes *= 1.2 # Add a 20% buffer
approx_mem_size_req = data_mem_usage_bytes + histogram_mem_usage_bytes
return approx_mem_size_req
# TODO: Use Pool in preprocess, optimize bagging to do Pool.split() to avoid re-computing pool for each fold! Requires stateful + y
# Pool is much more memory efficient, avoids copying data twice in memory
def _fit(self, X, y, X_val=None, y_val=None, time_limit=None, num_gpus=0, num_cpus=-1, sample_weight=None, sample_weight_val=None, **kwargs):
time_start = time.time()
try_import_catboost()
from catboost import CatBoostClassifier, CatBoostRegressor, Pool
ag_params = self._get_ag_params()
params = self._get_model_params()
params["thread_count"] = num_cpus
if self.problem_type == SOFTCLASS:
# FIXME: This is extremely slow due to unoptimized metric / objective sent to CatBoost
from .catboost_softclass_utils import SoftclassCustomMetric, SoftclassObjective
params["loss_function"] = SoftclassObjective.SoftLogLossObjective()
params["eval_metric"] = SoftclassCustomMetric.SoftLogLossMetric()
elif self.problem_type == QUANTILE:
# FIXME: Unless specified, CatBoost defaults to loss_function='MultiQuantile' and raises an exception
params["loss_function"] = params["eval_metric"]
model_type = CatBoostClassifier if self.problem_type in PROBLEM_TYPES_CLASSIFICATION else CatBoostRegressor
num_rows_train = len(X)
num_cols_train = len(X.columns)
num_classes = self.num_classes if self.num_classes else 1 # self.num_classes could be None after initialization if it's a regression problem
X = self.preprocess(X)
cat_features = list(X.select_dtypes(include="category").columns)
X = Pool(data=X, label=y, cat_features=cat_features, weight=sample_weight)
if X_val is None:
eval_set = None
early_stopping_rounds = None
else:
X_val = self.preprocess(X_val)
X_val = Pool(data=X_val, label=y_val, cat_features=cat_features, weight=sample_weight_val)
eval_set = X_val
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 params.get("allow_writing_files", False):
if "train_dir" not in params:
try:
# TODO: What if path is in S3?
os.makedirs(os.path.dirname(self.path), exist_ok=True)
except:
pass
else:
params["train_dir"] = os.path.join(self.path, "catboost_info")
# TODO: Add more control over these params (specifically early_stopping_rounds)
verbosity = kwargs.get("verbosity", 2)
if verbosity <= 1:
verbose = False
elif verbosity == 2:
verbose = False
elif verbosity == 3:
verbose = 20
else:
verbose = True
num_features = len(self._features)
if num_gpus != 0:
if "task_type" not in params:
params["task_type"] = "GPU"
logger.log(20, f"\tTraining {self.name} with GPU, note that this may negatively impact model quality compared to CPU training.")
# TODO: Confirm if GPU is used in HPO (Probably not)
# TODO: Adjust max_bins to 254?
if params.get("task_type", None) == "GPU":
if "colsample_bylevel" in params:
params.pop("colsample_bylevel")
logger.log(30, f"\t'colsample_bylevel' is not supported on GPU, using default value (Default = 1).")
if "rsm" in params:
params.pop("rsm")
logger.log(30, f"\t'rsm' is not supported on GPU, using default value (Default = 1).")
if self.problem_type == MULTICLASS and "rsm" not in params and "colsample_bylevel" not in params and num_features > 1000:
# Subsample columns to speed up training
if params.get("task_type", None) != "GPU": # RSM does not work on GPU
params["colsample_bylevel"] = max(min(1.0, 1000 / num_features), 0.05)
logger.log(
30,
f'\tMany features detected ({num_features}), dynamically setting \'colsample_bylevel\' to {params["colsample_bylevel"]} to speed up training (Default = 1).',
)
logger.log(30, f"\tTo disable this functionality, explicitly specify 'colsample_bylevel' in the model hyperparameters.")
else:
params["colsample_bylevel"] = 1.0
logger.log(30, f"\t'colsample_bylevel' is not supported on GPU, using default value (Default = 1).")
logger.log(15, f"\tCatboost model hyperparameters: {params}")
extra_fit_kwargs = dict()
if params.get("task_type", None) != "GPU":
callbacks = []
if early_stopping_rounds is not None:
callbacks.append(EarlyStoppingCallback(stopping_rounds=early_stopping_rounds, eval_metric=params["eval_metric"]))
if num_rows_train * num_cols_train * num_classes > 5_000_000:
# The data is large enough to potentially cause memory issues during training, so monitor memory usage via callback.
callbacks.append(MemoryCheckCallback())
if time_limit is not None:
time_cur = time.time()
time_left = time_limit - (time_cur - time_start)
if time_left <= time_limit * 0.4: # if 60% of time was spent preprocessing, likely not enough time to train model
raise TimeLimitExceeded
callbacks.append(TimeCheckCallback(time_start=time_cur, time_limit=time_left))
extra_fit_kwargs["callbacks"] = callbacks
else:
logger.log(30, f"\tWarning: CatBoost on GPU is experimental. If you encounter issues, use CPU for training CatBoost instead.")
if time_limit is not None:
params["iterations"] = self._estimate_iter_in_time_gpu(
X=X,
eval_set=eval_set,
time_limit=time_limit,
verbose=verbose,
params=params,
num_rows_train=num_rows_train,
time_start=time_start,
model_type=model_type,
)
if early_stopping_rounds is not None:
if isinstance(early_stopping_rounds, int):
extra_fit_kwargs["early_stopping_rounds"] = early_stopping_rounds
elif isinstance(early_stopping_rounds, tuple):
extra_fit_kwargs["early_stopping_rounds"] = 50
self.model = model_type(**params)
# TODO: Custom metrics don't seem to work anymore
# TODO: Custom metrics not supported in GPU mode
# TODO: Callbacks not supported in GPU mode
fit_final_kwargs = dict(
eval_set=eval_set,
verbose=verbose,
**extra_fit_kwargs,
)
if eval_set is not None:
fit_final_kwargs["use_best_model"] = True
self.model.fit(X, **fit_final_kwargs)
self.params_trained["iterations"] = self.model.tree_count_
# FIXME: This logic is a hack made to maintain compatibility with GPU CatBoost.
# GPU CatBoost does not support callbacks or custom metrics.
# Since we use callbacks to check memory and training time in CPU mode, we need a way to estimate these things prior to training for GPU mode.
# This method will train a model on a toy number of iterations to estimate memory and training time.
# It will return an updated iterations to train on that will avoid running OOM and running over time limit.
# Remove this logic once CatBoost fixes GPU support for callbacks and custom metrics.
def _estimate_iter_in_time_gpu(self, *, X, eval_set, time_limit, verbose, params, num_rows_train, time_start, model_type):
import math
import pickle
import sys
modifier = min(1.0, 10000 / num_rows_train)
num_sample_iter_max = max(round(modifier * 50), 2)
time_left_start = time_limit - (time.time() - time_start)
if time_left_start <= time_limit * 0.4: # if 60% of time was spent preprocessing, likely not enough time to train model
raise TimeLimitExceeded
default_iters = params["iterations"]
params_init = params.copy()
num_sample_iter = min(num_sample_iter_max, params_init["iterations"])
params_init["iterations"] = num_sample_iter
sample_model = model_type(
**params_init,
)
sample_model.fit(
X,
eval_set=eval_set,
use_best_model=True,
verbose=verbose,
)
time_left_end = time_limit - (time.time() - time_start)
time_taken_per_iter = (time_left_start - time_left_end) / num_sample_iter
estimated_iters_in_time = round(time_left_end / time_taken_per_iter)
available_mem = ResourceManager.get_available_virtual_mem()
if self.problem_type == SOFTCLASS:
model_size_bytes = 1 # skip memory check
else:
model_size_bytes = sys.getsizeof(pickle.dumps(sample_model))
max_memory_proportion = 0.3
mem_usage_per_iter = model_size_bytes / num_sample_iter
max_memory_iters = math.floor(available_mem * max_memory_proportion / mem_usage_per_iter)
final_iters = min(default_iters, min(max_memory_iters, estimated_iters_in_time))
return final_iters
def _predict_proba(self, X, **kwargs):
if self.problem_type != SOFTCLASS:
return super()._predict_proba(X, **kwargs)
# For SOFTCLASS problems, manually transform predictions into probabilities via softmax
X = self.preprocess(X, **kwargs)
y_pred_proba = self.model.predict(X, prediction_type="RawFormulaVal")
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])
if y_pred_proba.shape[1] == 2:
y_pred_proba = y_pred_proba[:, 1]
return y_pred_proba
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 _get_early_stopping_rounds(self, num_rows_train, strategy="auto"):
return get_early_stopping_rounds(num_rows_train=num_rows_train, strategy=strategy)
def _ag_params(self) -> set:
return {"early_stop"}
def _validate_fit_memory_usage(self, mem_error_threshold: float = 1, mem_warning_threshold: float = 0.75, mem_size_threshold: int = 1e9, **kwargs):
return super()._validate_fit_memory_usage(
mem_error_threshold=mem_error_threshold, mem_warning_threshold=mem_warning_threshold, mem_size_threshold=mem_size_threshold, **kwargs
)
def get_minimum_resources(self, is_gpu_available=False):
minimum_resources = {
"num_cpus": 1,
}
if is_gpu_available:
# Our custom implementation does not support partial GPU. No gpu usage according to nvidia-smi when the `num_gpus` passed to fit is fractional`
minimum_resources["num_gpus"] = 0.5
return minimum_resources
def _get_default_resources(self):
# logical=False is faster in training
num_cpus = ResourceManager.get_cpu_count_psutil(logical=False)
num_gpus = 0
return num_cpus, num_gpus
def _more_tags(self):
# `can_refit_full=True` because iterations is communicated at end of `_fit`
return {"can_refit_full": True}