GPN 19 - Talk on ML Workflow Tools

https://entropia.de/GPN19:ML_Workflow_Tools_Overview https://pretalx.entropia.de/gpn19/talk/B3HDR3/

Modern ML workflow requires to run experiments fast at a large scale. In order to stay sane and keep an overview of what is going on, there are some tools out there.

We will demonstrate 2 tools in this workshop : ML flow and edflow. These tools focus on different parts of the development workflow : fast model iteration and monitoring (dashboard).

We will provide some basic code and instructions on how to add the tools to basic examples. We will cover a standard classification based problem to demonstrate a simple use-case.

Environment setup

• create new virtual environment, for example using conda

conda create --name ml_talk python=3.6
source activate ml_talk

• then install requirements

pip install -r mlflow_excersize/requirements.txt
pip install -r requirements_edflow.txt

Slides

• https://docs.google.com/presentation/d/10_kPq734hIEYeDczrTDHYUOobqRGwNxZC7qdbSH2VeE/edit#slide=id.g58a2b8ebdc_0_42

Part 1 - EDFlow

https://github.com/pesser/edflow

Solutions for Edflow Part

edflow -t problem1_solution/train.yaml # train model
edflow -t problem1_solution/train.yaml -p project_folder # continue training model


edflow -t problem2_solution/train.yaml -e problem2_solution/validation.yaml # add validation

Part 2 - MLFlow

It consists only of adding some simple logging statements to mlflow_excersize/linear_model.py and mlflow_excersize/linear_model_lasso.py (compare mlflow_excersize/linear_model_mlflow.py) and afterwards running mlflow ui --filestore mlflow_excersize/mlflow.

Part 3 - Project Structure

The template is available under: https://github.com/LeanderK/cookiecutter-ml