def run_experiment(model_params: Dict[str, Any],
X_train: pd.DataFrame,
y: pd.Series,
X_test: Optional[pd.DataFrame] = None,
logging_directory: str = 'output/{time}',
if_exists: str = 'error',
eval_func: Optional[Callable] = None,
algorithm_type: Union[str, Type[BaseEstimator]] = 'lgbm',
fit_params: Optional[Union[Dict[str, Any],
Callable]] = None,
cv: Optional[Union[int, Iterable,
BaseCrossValidator]] = None,
groups: Optional[pd.Series] = None,
categorical_feature: Optional[List[str]] = None,
sample_submission: Optional[pd.DataFrame] = None,
submission_filename: Optional[str] = None,
type_of_target: str = 'auto',
feature_list: Optional[List[Union[int, str]]] = None,
feature_directory: Optional[str] = None,
inherit_experiment: Optional[Experiment] = None,
with_auto_hpo: bool = False,
with_auto_prep: bool = False,
with_mlflow: bool = False):
"""
Evaluate metrics by cross-validation and stores result
(log, oof prediction, test prediction, feature importance plot and submission file)
under the directory specified.
One of the following estimators are used (automatically dispatched by ``type_of_target(y)`` and ``gbdt_type``).
* LGBMClassifier
* LGBMRegressor
* CatBoostClassifier
* CatBoostRegressor
The output files are laid out as follows:
.. code-block:: none
<logging_directory>/
log.txt <== Logging file
importance.png <== Feature importance plot generated by nyaggle.util.plot_importance
oof_prediction.npy <== Out of fold prediction in numpy array format
test_prediction.npy <== Test prediction in numpy array format
submission.csv <== Submission csv file
metrics.json <== Metrics
params.json <== Parameters
models/
fold1 <== The trained model in fold 1
...
Args:
model_params:
Parameters passed to the constructor of the classifier/regressor object (i.e. LGBMRegressor).
X_train:
Training data. Categorical feature should be casted to pandas categorical type or encoded to integer.
y:
Target
X_test:
Test data (Optional). If specified, prediction on the test data is performed using ensemble of models.
logging_directory:
Path to directory where output of experiment is stored.
if_exists:
How to behave if the logging directory already exists.
- error: Raise a ValueError.
- replace: Delete logging directory before logging.
- append: Append to exisitng experiment.
- rename: Rename current directory by adding "_1", "_2"... prefix
fit_params:
Parameters passed to the fit method of the estimator. If dict is passed, the same parameter except
eval_set passed for each fold. If callable is passed,
returning value of ``fit_params(fold_id, train_index, test_index)`` will be used for each fold.
eval_func:
Function used for logging and calculation of returning scores.
This parameter isn't passed to GBDT, so you should set objective and eval_metric separately if needed.
If ``eval_func`` is None, ``roc_auc_score`` or ``mean_squared_error`` is used by default.
gbdt_type:
Type of gradient boosting library used. "lgbm" (lightgbm) or "cat" (catboost)
cv:
int, cross-validation generator or an iterable which determines the cross-validation splitting strategy.
- None, to use the default ``KFold(5, random_state=0, shuffle=True)``,
- integer, to specify the number of folds in a ``(Stratified)KFold``,
- CV splitter (the instance of ``BaseCrossValidator``),
- An iterable yielding (train, test) splits as arrays of indices.
groups:
Group labels for the samples. Only used in conjunction with a “Group” cv instance (e.g., ``GroupKFold``).
sample_submission:
A sample dataframe alined with test data (Usually in Kaggle, it is available as sample_submission.csv).
The submission file will be created with the same schema as this dataframe.
submission_filename:
The name of submission file will be created under logging directory. If ``None``, the basename of the logging
directory will be used as a filename.
categorical_feature:
List of categorical column names. If ``None``, categorical columns are automatically determined by dtype.
type_of_target:
The type of target variable. If ``auto``, type is inferred by ``sklearn.utils.multiclass.type_of_target``.
Otherwise, ``binary``, ``continuous``, or ``multiclass`` are supported.
feature_list:
The list of feature ids saved through nyaggle.feature_store module.
feature_directory:
The location of features stored. Only used if feature_list is not empty.
inherit_experiment:
An experiment object which is used to log results. if not ``None``, all logs in this function are treated
as a part of this experiment.
with_auto_prep:
If True, the input datasets will be copied and automatic preprocessing will be performed on them.
For example, if ``gbdt_type = 'cat'``, all missing values in categorical features will be filled.
with_auto_hpo:
If True, model parameters will be automatically updated using optuna (only available in lightgbm).
with_mlflow:
If True, `mlflow tracking <https://www.mlflow.org/docs/latest/tracking.html>`_ is used.
One instance of ``nyaggle.experiment.Experiment`` corresponds to one run in mlflow.
Note that all output
mlflow's directory (``mlruns`` by default).
:return:
Namedtuple with following members
* oof_prediction:
numpy array, shape (len(X_train),) Predicted value on Out-of-Fold validation data.
* test_prediction:
numpy array, shape (len(X_test),) Predicted value on test data. ``None`` if X_test is ``None``
* metrics:
list of float, shape(nfolds+1) ``scores[i]`` denotes validation score in i-th fold.
``scores[-1]`` is overall score.
* models:
list of objects, shape(nfolds) Trained models for each folds.
* importance:
list of pd.DataFrame, feature importance for each fold (type="gain").
* time:
Training time in seconds.
* submit_df:
The dataframe saved as submission.csv
"""
start_time = time.time()
cv = check_cv(cv, y)
if feature_list:
X = pd.concat([X_train, X_test]) if X_test is not None else X_train
X.reset_index(drop=True, inplace=True)
X = load_features(X, feature_list, directory=feature_directory)
ntrain = len(X_train)
X_train, X_test = X.iloc[:ntrain, :], X.iloc[ntrain:, :].reset_index(
drop=True)
_check_input(X_train, y, X_test)
if categorical_feature is None:
categorical_feature = [
c for c in X_train.columns
if X_train[c].dtype.name in ['object', 'category']
]
if type_of_target == 'auto':
type_of_target = multiclass.type_of_target(y)
model_type, eval_func, cat_param_name = _dispatch_models(
algorithm_type, type_of_target, eval_func)
if with_auto_prep:
assert algorithm_type in (
'cat', 'xgb', 'lgbm'), "with_auto_prep is only supported for gbdt"
X_train, X_test = autoprep_gbdt(algorithm_type, X_train, X_test,
categorical_feature)
logging_directory = logging_directory.format(
time=datetime.now().strftime('%Y%m%d_%H%M%S'))
if inherit_experiment is not None:
experiment = ExpeimentProxy(inherit_experiment)
else:
experiment = Experiment(logging_directory,
if_exists=if_exists,
with_mlflow=with_mlflow)
with experiment as exp:
exp.log('Algorithm: {}'.format(algorithm_type))
exp.log('Experiment: {}'.format(exp.logging_directory))
exp.log('Params: {}'.format(model_params))
exp.log('Features: {}'.format(list(X_train.columns)))
exp.log_param('algorithm_type', algorithm_type)
exp.log_param('num_features', X_train.shape[1])
if callable(fit_params):
exp.log_param('fit_params', str(fit_params))
else:
exp.log_dict('fit_params', fit_params)
exp.log_dict('model_params', model_params)
if feature_list is not None:
exp.log_param('features', feature_list)
if with_auto_hpo:
assert algorithm_type == 'lgbm', 'auto-tuning is only supported for LightGBM'
model_params = find_best_lgbm_parameter(
model_params,
X_train,
y,
cv=cv,
groups=groups,
type_of_target=type_of_target)
exp.log_param('model_params_tuned', model_params)
exp.log('Categorical: {}'.format(categorical_feature))
models = [model_type(**model_params) for _ in range(cv.get_n_splits())]
if fit_params is None:
fit_params = {}
if cat_param_name is not None and not callable(
fit_params) and cat_param_name not in fit_params:
fit_params[cat_param_name] = categorical_feature
if isinstance(fit_params, Dict):
exp.log_params(fit_params)
result = cross_validate(models,
X_train=X_train,
y=y,
X_test=X_test,
cv=cv,
groups=groups,
logger=exp.get_logger(),
eval_func=eval_func,
fit_params=fit_params,
type_of_target=type_of_target)
# save oof
exp.log_numpy('oof_prediction', result.oof_prediction)
exp.log_numpy('test_prediction', result.test_prediction)
for i in range(cv.get_n_splits()):
exp.log_metric('Fold {}'.format(i + 1), result.scores[i])
exp.log_metric('Overall', result.scores[-1])
# save importance plot
if result.importance:
importance = pd.concat(result.importance)
plot_file_path = os.path.join(exp.logging_directory,
'importance.png')
plot_importance(importance, plot_file_path)
exp.log_artifact(plot_file_path)
# save trained model
for i, model in enumerate(models):
_save_model(model, exp.logging_directory, i + 1, exp)
# save submission.csv
submit_df = None
if X_test is not None:
submit_df = make_submission_df(result.test_prediction,
sample_submission, y)
exp.log_dataframe(
submission_filename or os.path.basename(exp.logging_directory),
submit_df, 'csv')
elapsed_time = time.time() - start_time
return ExperimentResult(result.oof_prediction, result.test_prediction,
result.scores, models, result.importance,
elapsed_time, submit_df)
def adversarial_validate(X_train: pd.DataFrame,
X_test: pd.DataFrame,
importance_type: str = 'gain',
estimator: Optional[BaseEstimator] = None,
cat_cols = None,
cv = None) -> ADVResult:
"""
Perform adversarial validation between X_train and X_test.
Args:
X_train:
Training data
X_test:
Test data
importance_type:
The type of feature importance calculated.
estimator:
The custom estimator. If None, LGBMClassifier is automatically used.
cv:
Cross validation split. If ``None``, the first fold out of 5 fold is used as validation.
Returns:
Namedtuple with following members
* auc:
float, ROC AUC score of adversarial validation.
* importance:
pandas DataFrame, feature importance of adversarial model (order by importance)
Example:
>>> from sklearn.model_selection import train_test_split
>>> from nyaggle.testing import make_regression_df
>>> from nyaggle.validation import adversarial_validate
>>> X, y = make_regression_df(n_samples=8)
>>> X_train, X_test, y_train, y_test = train_test_split(X, y)
>>> auc, importance = cross_validate(X_train, X_test)
>>>
>>> print(auc)
0.51078231
>>> importance.head()
feature importance
col_1 231.5827204
col_5 207.1837266
col_7 188.6920685
col_4 174.5668498
col_9 170.6438643
"""
concat = pd.concat([X_train, X_test]).copy().reset_index(drop=True)
y = np.array([1]*len(X_train) + [0]*len(X_test))
if estimator is None:
requires_lightgbm()
from lightgbm import LGBMClassifier
estimator = LGBMClassifier(n_estimators=10000, objective='binary', importance_type=importance_type,
random_state=0)
else:
assert is_instance(estimator, ('lightgbm.sklearn.LGBMModel', 'catboost.core.CatBoost')), \
'Only CatBoostClassifier or LGBMClassifier is allowed'
if cv is None:
cv = Take(1, KFold(5, shuffle=True, random_state=0))
fit_params = {'verbose': -1}
if cat_cols:
fit_params['categorical_feature'] = cat_cols
result = cross_validate(estimator, concat, y, None, cv=cv,
eval_func=roc_auc_score, fit_params=fit_params, importance_type=importance_type)
importance = pd.concat(result.importance)
importance = importance.groupby('feature')['importance'].mean().reset_index()
importance.sort_values(by='importance', ascending=False, inplace=True)
importance.reset_index(drop=True, inplace=True)
return ADVResult(result.scores[-1], importance)