Project Usage Guide

Prepare Environment

1. In case of restricted access: $ pip config set global.trusted-host "files.pythonhosted.org pypi.org pypi.python.org"
2. For Windows systems: $ set MLFLOW_TRACKING_URI=http://<host>:<port>
   For Unix systems: $ export MLFLOW_TRACKING_URI=http://<host>:<port>
3. In case of using remote storage to store artifacts: $ export MLFLOW_S3_ENDPOINT_URL=https://<domain-name>
   For Windows: $ set MLFLOW_S3_ENDPOINT_URL=https://<domain-name>

Run Project

1. Run MLflow server

Save artifacts locally: $ mlflow ui --host <host> --port <port>
Save artifacts into remote storage: $ mlflow ui --host <host> --port <port> --artifacts-destination <remote-uri>
Save artifacts into remote storage w/o proxying access through tracking server: $ mlflow ui --host <host> --port <port> --no-serve-artifacts --default-artifact-root <remote-uri>

2. Run step

NOTE: To see available options run: $ mlflow run --help

To run particular project step: $ mlflow run . --entry-point <entry-point-name> -P <param1>=<value1> -P <param2>=<value2> --experiment-name <Experiment_Name> --run-name <Run_Name> --env-manager <env-manager>

Serve/deploy the model

1. Prepare environment

Install pyenv requirements: $ sudo apt-get install zlib1g-dev libssl-dev libbz2-dev libncursesw5-dev libffi-dev libreadline-dev libsqlite3-dev tk-dev liblzma-dev
Install pyenv for Python version management: $ curl https://pyenv.run | bash
Add pyenv to PATH: $ export PYENV_ROOT="$HOME/.pyenv" [[ -d $PYENV_ROOT/bin ]] && export PATH="$PYENV_ROOT/bin:$PATH" eval "$(pyenv init -)"
Install virtualenv for Python environment management: $ python -m pip install virtualenv

2. Server the model locally

Serve via REST API: $ mlflow models serve -m "models:/<model-name>@<model-alias>" --port <port> (can also use <model-name>/<model-version> instead of alias)

Build docker image: $ mlflow models build-docker -m "models:/<model-name>@<model-alias>" -n "<image-name>" (can also use <model-name>/<model-version> instead of alias)
Serve model via docker container: $ docker run -p <serving-port>:8080 <image-name>

Test model API: $ curl -d '<input-data>' -H 'Content-Type: application/json' -X POST <host>:<port>/invocations

Get model inference w/o serving

$ mlflow models predict -m models:/<model-name>@<model-alias> -i <inputs-path> -o <predictions-path>