Experiment tracking with LightGBM#

In this tutorial you will train a model using the LightGBM framework and use the experiment tracking capabilities of Dataiku to log training runs (parameters, performance).


Performing the experiment#

The following code snippet provides a reusable example to train a simple gradient boosting model with these main steps:

(1) Specify the categorical and numeric features and the target variable.

(2) Using the categorical and continuous variables spcecified, set up a preprocessing Pipeline with two transformation steps. First, define a transformer to one-hot-encode categorical variables and then impute any missing values in continuous variables and rescale them in the other.

(3) Define a dictionary containing the search space for hyperparameter tuning. Then lay out a cross-validation strategy to train a classifier and evaluate the model. Log the parameters and resulting metrics as well as the models using Experiment Tracking feature, while looping over combinations of hyperparameters. The model artifact logged for the run is also a Pipeline called clf_pipeline that encapsulates the preprocessing and the model itself.

import dataiku
from lightgbm import LGBMClassifier
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder 
from sklearn.model_selection import cross_validate, ParameterGrid, StratifiedKFold

# !! - Replace these values by your own - !!

client = dataiku.api_client()
project = client.get_project(USER_PROJECT_KEY)

ds = dataiku.Dataset(USER_TRAINING_DATASET)
df = ds.get_dataframe()

# (1)
cat_features = ["job", "marital", "education", "default",
    "housing","loan", "month", "contact", "poutcome"]

num_features = ["age", "balance", "day", "duration", 
    "campaign", "pdays", "previous"]

target ="y"

X = df.drop(target, axis=1)
y = df[target]

# (2)
num_pipeline = Pipeline([
         ("impute", SimpleImputer(strategy="median")),
         ("scale", StandardScaler())
cat_transformer = OneHotEncoder(handle_unknown="ignore")

preprocessor = ColumnTransformer(
         ("num", num_pipeline, num_features),
         ("cat", cat_transformer, cat_features),

# (3)
hparams_dict = {"learning_rate": [1e-3, 1e-4],
    "n_estimators": [250, 500, 1000],
    "seed": [47]

n_folds = 5
param_grid = ParameterGrid(hparams_dict)
cv = StratifiedKFold(n_splits=n_folds)

mf = project.get_managed_folder(USER_XPTRACKING_FOLDER_ID)
mlflow_extension = project.get_mlflow_extension()

with project.setup_mlflow(mf) as mlflow:
    for hparams in param_grid:
        with mlflow.start_run() as run:
            run_id = run.info.run_id
            clf_pipeline = Pipeline(steps=
                    [("preprocessor", preprocessor), 
                     ("classifier", LGBMClassifier(**hparams))
            scores = cross_validate(clf_pipeline, X, y, cv=cv, scoring='roc_auc')

            run_metric_mean = scores['test_score'].mean()
            mlflow.log_metric('train_mean_auc', run_metric_mean)

            for k,v in hparams.items():

            mlflow.sklearn.log_model(sk_model=clf_pipeline, artifact_path='model')
                classes=['no', 'yes'],