The MLFlow integration is currently in beta and is not a part of the official wandb python package. To try this integration you can install wandb from our git branch by running:
pip install --upgrade git+git://github.com/wandb/client.git@feature/mlflow#egg=wandbIf you're already using MLflow to track your experiments it's easy to visualize them with W&B. Simply by calling import wandb in your mlflow scripts we'll mirror all metrics, params, and artifacts to W&B. We do this by patching the mlflow python library. Our current integration is write only. All data will also be written to the backend you've configured for mlflow.
When mirroring data to both a wandb and mlflow tracking backend, the following concepts are mapped to each-other.
| MLflow | W&B |
|---|---|
| Experiment | Project |
| mlflow.start_run | wandb.init |
| mlflow.log_params | wandb.config |
| mlflow.log_metrics | wandb.log |
| mlflow.log_artifacts | wandb.save |
| mlflow.start_run(nested=True) | Grouping |
If you want to log rich media like Images, Video, or Plots you can call wandb.log in your code as well. Be sure to pass a step argument to your calls to log so they can be aligned with the metrics you're logging with mlflow.
By default wandb only logs metrics, params and artifacts. If you don't want to store artifacts with wandb, you can set WANDB_SYNC_MLFLOW=metrics,params . If you want to disable mirroring of all data to wandb you can set the WANDB_SYNC_MLFLOW=false.