We'd like your feedback on MLflow Pipelines

mlflow / mlflow

Open source platform for the machine learning lifecycle

https://mlflow.org

Apache License 2.0

18.23k stars 4.13k forks source link

We'd like your feedback on MLflow Pipelines #6335

Open dbczumar opened 2 years ago

dbczumar commented 2 years ago

MLflow Pipelines provides a standardized framework for creating production-grade ML pipelines that combine modular ML code with software engineering best practices to make model development and deployment fast & scalable. You can read more about MLflow Pipelines here and try out the first pipeline for building regression models here.

We are looking to talk to early users that are willing to share feedback on MLflow Pipelines and / or help us expand its features. If you are interested, please use this form to sign up.

machielg commented 2 years ago

Being able to define the entire ML project into a R&D friendly, declarative and portable format is a huge step forward for the ML/Spark community!

sberryman commented 2 years ago

I'm looking forward to the classification pipeline. I don't have a project for the regression side at the moment. I would be happy to share feedback on the classification pipeline when that is available.

dsgibbons commented 2 years ago

MLflow Pipelines look great, and I am excited to see where it will go. My main feature requests at this point would be:

Better support for unstructured datasets that don't neatly fit in pandas DataFrames (e.g. images, text data, video).
Ability to train on datasets that don't fit into memory (maybe by allowing for pandas to be substituted with dask or Modin).
Explicit function for online data augmentation during training (something better than simply overwriting the model's fit method).

notniknot commented 2 years ago

A declarative way to describe the entire pipeline enables Data Scientists to operationalize their models. That's a great step in the right direction!

But many developers are using Windows to write their code. I find it strange that MLflow Pipelines uses make which is not installed on Windows. Thus, local tests cannot be executed.

I would like the idea to automatically orchestrate the pipeline via Airflow. Maybe one could provide a guide for that or increase the interoperability between other open source orchestrators?

echarso commented 1 year ago

I would be interested to see how i can deploy the pipeline steps in kubernetes and how to assign to each step in the pipeline separate compute resources. I think that will simplify a lot a big number of open source MLOps solution out there :)

sunishsheth2009 commented 1 year ago

I'm looking forward to the classification pipeline. I don't have a project for the regression side at the moment. I would be happy to share feedback on the classification pipeline when that is available.

Hey guys, wanted to follow up on this. We now have released binary classification as part of MLFlow 2.0 release. We have an example repo and a template repo to get started. :) Would love to hear some feedback on the same.

jmahlik commented 1 year ago

It seems limited to only the recipes explicitly pre-defined in mlflow at the moment. Support for defining custom recipes would be awesome. There's some cases where one might need to adjust or extend a pre-defined recipe for a specific use case.

I hackily patched sys.modules and used private classes to play with it at the moment, but official support for defining custom recipes would be great.