scikit-dag (skdag
) is an open-sourced, MIT-licenced library that provides advanced
workflow management to any machine learning operations that follow
scikit-learn conventions. Installation is simple:
pip install skdag
It works by introducing Directed Acyclic Graphs as a drop-in replacement for traditional
scikit-learn Pipeline
. This gives you a simple interface for a range of use cases
including complex pre-processing, model stacking and benchmarking.
from skdag import DAGBuilder
dag = (
DAGBuilder(infer_dataframe=True)
.add_step("impute", SimpleImputer())
.add_step("vitals", "passthrough", deps={"impute": ["age", "sex", "bmi", "bp"]})
.add_step(
"blood",
PCA(n_components=2, random_state=0),
deps={"impute": ["s1", "s2", "s3", "s4", "s5", "s6"]}
)
.add_step(
"rf",
RandomForestRegressor(max_depth=5, random_state=0),
deps=["blood", "vitals"]
)
.add_step("svm", SVR(C=0.7), deps=["blood", "vitals"])
.add_step(
"knn",
KNeighborsRegressor(n_neighbors=5),
deps=["blood", "vitals"]
)
.add_step("meta", LinearRegression(), deps=["rf", "svm", "knn"])
.make_dag()
)
dag.show(detailed=True)
The above DAG imputes missing values, runs PCA on the columns relating to blood test results and leaves the other columns as they are. Then they get passed to three different regressors before being passed onto a final meta-estimator. Because DAGs (unlike pipelines) allow predictors in the middle or a workflow, you can use them to implement model stacking. We also chose to run the DAG steps in parallel wherever possible.
After building our DAG, we can treat it as any other estimator:
from sklearn import datasets
X, y = datasets.load_diabetes(return_X_y=True, as_frame=True)
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=0
)
dag.fit(X_train, y_train)
dag.predict(X_test)
Just like a pipeline, you can optimise it with a gridsearch, pickle it etc.
Note that this package does not deal with things like delayed dependencies and
distributed architectures - consider an established
solution for such use cases. skdag
is just for building and
executing local ensembles from estimators.
Read on to learn more about skdag
...