mlflow-dl 0.2.2

Package for interaction with MLFlow Registry

MLFlow model downloader

Allows to download models from MLFlow Model Registry using model's status (Staging or Production) or model's Version.

NOTES:

• supports MLflow's version <1.20.0
• tested only with Tensorflow/Keras models
• tested on S3 and SFTP storages

Installation

Install from PyPi

pip install mlflow-dl

Install from GitHub

git clone [e-mail]:dem-artem/mlflow_dl.git
cd mlflow_dl

• for production:

pip install .

• for development:

cd mlflow_dl
pip install -e ".[dev]"

Environment Variables Configuration

In the most of the cases you need to configure a proper credentials to mlflow and related storage. Built in variables:

ENV Name	Description
MLFLOWDL_TARGET_FOLDER_LOCAL	The folder where downloaded results placed

Also, you may need to configure variables for access to MLFlow and AWS S3 bucket. Some of them:

ENV Name	Description
AWS_ACCESS_KEY_ID
AWS_SECRET_ACCESS_KEY
AWS_DEFAULT_REGION
AWS_BUCKET_NAME
MLFLOW_TRACKING_USERNAME	Username for HTTP Base Auth for MLFlow (if used)
MLFLOW_TRACKING_PASSWORD	Password for HTTP Base Auth for MLFlow (if used)
MLFLOW_TRACKING_URI	URL to MLFlow UI (required)

Usage from Command Line

• Get list of available commands:

python -m mlflow_dl

or

mlflow_dl

• Download models in "Production" stage:

mlflow_dl download-production -m model_name_1,model_name_2

• Download models in "Staging" stage:

mlflow_dl download-staging -m model_name_1,model_name_2

• Download model by name and version number:

mlflow_dl download-model-by-version -m model_name -v 1

• Download specific folder from the root of model's experiment:

mlflow_dl download-folder-by-model-version -m model_name -v 1 -f custom_folder

• Download specific folder for Tensorflow Serving compatible structure:

mlflow_dl download-folder-by-model-version -m model_name -v 1 -f custom_folder --no-subfolder

Usage from Code

from mlflow_dl import MlflowDl

mlflowdl = MlflowDl(target_folder="tmp_folder")
mlflowdl.download_model_by_version("model_name", "1")

Development

Testing

python -m unittest

Coverage

python -m coverage run -m unittest
python -m coverage html -i

Code Style

python -m flake8

Versioning

Commit your changes and run with a proper label (major | minor | patch):

bumpversion patch
git push origin
git push origin --tags

Build package

rm -rf ./dist
python -m build

Upload to custom PyPI

In case of private repo you need to make sure you configured credentials in ~/.pypirc, than run:

twine upload --repository {repo_name} dist/*