FAQ¶

General¶

What is open-mlpipe?¶

open-mlpipe is a production-level automated ML pipeline that takes a raw dataset and automatically handles the entire machine learning workflow: EDA, cleaning, feature engineering, model comparison, hyperparameter tuning, explainability, and saving a production-ready pipeline.

When should I use open-mlpipe?¶

Use open-mlpipe when you want to quickly benchmark multiple models on a tabular dataset without writing boilerplate code. It's ideal for:

Rapid prototyping and baseline establishment
AutoML for regression and classification tasks
Getting a production-ready model in minutes instead of hours
Comparing 14+ models head-to-head

What tasks does open-mlpipe support?¶

Regression — predicting continuous values (house prices, revenue, etc.)
Classification — predicting categories, both binary and multiclass (fraud detection, churn prediction, etc.)

Does it work with my data format?¶

open-mlpipe supports:

CSV files
Parquet files
Excel files (.xlsx, .xls)

Just pass the file path and open-mlpipe handles the rest.

Installation¶

What Python versions are supported?¶

Python 3.10, 3.11, 3.12, and 3.13.

How do I install with all extras?¶

pip install open-mlpipe[full]

This installs CatBoost, MLflow tracking, FastAPI deployment, and visualization extras.

Do I need GPU support?¶

No. open-mlpipe works on CPU. Some models (like XGBoost) can optionally use GPU if available, but it's not required.

Models and Performance¶

How many models does open-mlpipe compare?¶

14+ models including:

Linear: Ridge, Lasso, ElasticNet, Linear/Logistic Regression
Tree: Decision Tree
Bagging: Random Forest, Extra Trees
Boosting: XGBoost, LightGBM, Gradient Boosting, Hist Gradient Boosting, AdaBoost
Instance-based: KNN, SVM
Probabilistic: Naive Bayes
Ensemble: Stacking, Voting

Which model is usually best?¶

It depends on the dataset, but LightGBM and XGBoost are typically among the top performers for both regression and classification tasks on tabular data.

How does hyperparameter tuning work?¶

open-mlpipe uses Optuna for Bayesian hyperparameter optimization with 18 model-specific search spaces. Each model gets tuned with parameters appropriate for its architecture.

Can I limit which models are tested?¶

Yes, specify candidates in the configuration:

config = PipelineConfig(
    data={"path": "data.csv", "target": "target"},
    model_selection={
        "candidates": ["lightgbm", "xgboost"],
    },
)

Configuration¶

Do I need to write configuration?¶

No. open-mlpipe works with zero configuration:

from open_mlpipe import run

ctx = run("data.csv", target="price")

Configuration is only needed when you want to customize behavior.

How do I configure the pipeline?¶

Three ways:

Python API — pass parameters to PipelineConfig
YAML files — write a config YAML and pass it to run_config()
CLI flags — use command-line options with open-mlpipe run

Can I save and reuse configurations?¶

Yes. Write a YAML config file and reuse it:

# my_config.yaml
project: my-project
data:
  path: data.csv
  target: price
tuning:
  enabled: true
  n_trials: 50

open-mlpipe run --config my_config.yaml

Output and Deployment¶

What does open-mlpipe save?¶

Model file (artifacts/model_v1.joblib) — the full inference pipeline
EDA report — data profiling summary
Metrics JSON — all evaluation metrics
SHAP plots — feature importance visualizations
Model comparison — cross-validation results for all models

How do I load and use the saved model?¶

import joblib
import pandas as pd

model = joblib.load("artifacts/model_v1.joblib")
predictions = model.predict(new_data)

Can I deploy the model as an API?¶

Yes. Install the deploy extras and use the built-in FastAPI support:

pip install open-mlpipe[deploy]
open-mlpipe run --data dataset.csv --deploy

Troubleshooting¶

The pipeline is slow. How can I speed it up?¶

Reduce the number of Optuna trials: tuning={"n_trials": 20}
Use fewer models: model_selection={"candidates": ["lightgbm"]}
Disable feature selection: feature_selection={"enabled": False}
Use Level 1 (zero-touch) with smart defaults

I'm getting "target column not found" error¶

The target column must exist in your dataset. Check:

The column name is spelled correctly
There are no extra spaces in the name
The column exists in the CSV/Parquet file

How do I report a bug or request a feature?¶

Open an issue on GitHub.

Back to Home