Python

ml-pipeline-toolkit

End-to-end machine learning pipeline toolkit for Python: preprocessing, training, evaluation and deployment of scikit-learn models from one YAML file.

0 stars

0 forks

0 watchers

MIT license

pythonmachine-learningmlopsscikit-learnfastapimodel-servingdata-preprocessingml-pipelinedocker

Stack

Python 3.11+scikit-learnpandas and NumPyFastAPI and UvicornuvDockerDocker ComposeGitHub Actions

Features

Configuration

One YAML file per pipeline, validated up front with errors that name the exact key
Command-line overrides such as --set model.params.n_estimators=50
mlpipeline check validates the config, model and data without training

Data preprocessing

Loads CSV, TSV, Parquet, JSON and JSON lines, plus the datasets bundled with scikit-learn
Infers numeric and categorical columns, saves the schema with the model and enforces it on every request
Imputation, scaling, and one-hot or ordinal encoding that handles rare and unseen categories
Seeded, stratified train/test splits

Model training

20 built-in classification and regression models
Your own models through register_model or a package.module:Class import path
Grid and random hyperparameter search, with preprocessing fitted inside each fold

Evaluation

Held-out test metrics compared against a baseline that ignores the features
Cross-validation on the training set, a confusion matrix and a Markdown report
mlpipeline evaluate scores a saved model on new labelled data to catch drift

Deployment

Versioned model bundles with metadata and library versions, written atomically and checked with SHA-256
A FastAPI prediction service with schema validation, size limits and OpenAPI docs
Batch predictions to CSV, Parquet or JSON lines with class probabilities

Operations

Non-root Docker image that trains or serves, and a Docker Compose demo
JSON logs and settings from flags or environment variables
CI with lint, strict type checking and tests on the minimum and latest Python releases

Quick Start

Clone the repository (needs Python 3.11 or later and uv)

git clone https://github.com/tec-smith/ml-pipeline-toolkit.git

2
Enter the project folder
```
cd ml-pipeline-toolkit
```
3
Install the toolkit and its dependencies
```
make install
```
4
Train the example churn model and compare it with a baseline
```
make train
```
5
Serve it, then open localhost:8000/docs to try the API
```
make serve
```

Interested In Contributing?

Reach out to TECSmith about collaborating on this repository, or check out what else is in the works.

Features

Configuration

One YAML file per pipeline, validated up front with errors that name the exact key
Command-line overrides such as --set model.params.n_estimators=50
mlpipeline check validates the config, model and data without training

Data preprocessing

Loads CSV, TSV, Parquet, JSON and JSON lines, plus the datasets bundled with scikit-learn
Infers numeric and categorical columns, saves the schema with the model and enforces it on every request
Imputation, scaling, and one-hot or ordinal encoding that handles rare and unseen categories
Seeded, stratified train/test splits

Model training

20 built-in classification and regression models
Your own models through register_model or a package.module:Class import path
Grid and random hyperparameter search, with preprocessing fitted inside each fold

Evaluation

Held-out test metrics compared against a baseline that ignores the features
Cross-validation on the training set, a confusion matrix and a Markdown report
mlpipeline evaluate scores a saved model on new labelled data to catch drift

Deployment

Versioned model bundles with metadata and library versions, written atomically and checked with SHA-256
A FastAPI prediction service with schema validation, size limits and OpenAPI docs
Batch predictions to CSV, Parquet or JSON lines with class probabilities

Operations

Non-root Docker image that trains or serves, and a Docker Compose demo
JSON logs and settings from flags or environment variables
CI with lint, strict type checking and tests on the minimum and latest Python releases

Quick Start

Clone the repository (needs Python 3.11 or later and uv)

git clone https://github.com/tec-smith/ml-pipeline-toolkit.git

Enter the project folder

cd ml-pipeline-toolkit

Install the toolkit and its dependencies

make install

Train the example churn model and compare it with a baseline

make train

Serve it, then open localhost:8000/docs to try the API

make serve