MLFlow support in ignite

[Nic-Ma]

🚀 Feature

Hi @vfdev-5 ,

As you may know, PyTorch and MLflow Integration announced last year: https://mlflow.org/news/2020/11/12/pytorch-mlflow-integration/index.html, it works well with PyTorch-Lightning but seems ignite doesn't support it so far? What do you guys think about MLFlow? Is it possible to add support for it? Here is the current API doc of MLFlow based on Lightning: https://www.mlflow.org/docs/latest/python_api/mlflow.pytorch.html

Thanks.

[vfdev-5]

Hi @Nic-Ma , thanks for the question and FR !

Actually, MLflow was one of the first loggers we added after tensorboard and visdom. We have currently a specific logger for mlflow: https://pytorch.org/ignite/contrib/handlers.html#mlflow-logger and previously had few examples like pascal voc using it: https://github.com/pytorch/ignite/blob/v0.4.3/examples/references/segmentation/pascal_voc2012/NOTES_MLflow.md Now we are migrated to ClearML which IMO nicer than MLflow in terms of functionalities, dashboard and its deployment (e.g community application).

I agree that integrating mlflow autologging with ignite would be interesting and would help to increase our visibility. cc @Priyansi

[smurching]

Hey all! On the MLflow side, we'd be happy to see a contribution adding support for ignite autologging to the mlflow.pytorch.autolog API (https://www.mlflow.org/docs/latest/python_api/mlflow.pytorch.html#mlflow.pytorch.autolog), feel free to file a github issue so we can discuss the details. The way MLflow integrates with most ML frameworks is by exposing these sorts of autolog APIs so that users can log model training sessions to MLflow with a single autolog() call.

It seems in practice patching the ignite model training APIs so that they additionally attach an MLflowLogger would work - the challenge I think is inferring what metric/param names to log