openml/flow-visualization

Tool to convert openml flows to ONNX and visualize them via Netron

Visualizing openML flows

This repository provides an initial attempt at visualizing flows. Generally, this should work in three steps:

1. Retrieve the pipeline/model from the openML flow.
2. Convert the model to onnx format.
3. Load the onnx model through netron.

The current trial version is restricted to visualizing scikit-learn flows.

Usage

1. Setting up your environment

   cd .
   pip install -r requirements.txt
   

2. implementation.csv provides an overview of all distinct configurations found among openML flows. Run retrieve_sklearn_version.py in order to generate a docker for every distinct sklearn configuration found. Dockerfiles are stored in ./dockers (59 distinct configurations found).
3. Running a docker will call sk_onnx.py or sk_onnx_old.py, depending on the sklearn version used. This script will convert all flows that are built using that specific sklearn version to onnx files. Files will be stored in ./onnx_models.
4. Once you have the onnx file stored, you can visualize it using netron. A function to do so is provided in sk_onnx.visualize_onnx.

Keep in mind that there are currently multiple bugs in the code that need fixing. There are also multiple to-do's (such as integrating netron in the openML website). Find elaborate documentation on limitations and future work in Documentation.pdf.

App

A small Flask app is provided to show how netron can be used to visualize onnx models. This can be done by either loading a stored file or directly via url. Running the Flask app will call sk_onnx_test.py which allows you to visualize flow 19433:

cd .
flask run

Contact

Feel free to reach out to me ([email protected]) in case of questions!