MLFlow wrapper with controllable dependencies and support of WebHooks Can be used for people who want to experiment locally with mlflow

For basic light version use:

pip install mlflow-client

Else if you want to use the mlflow packaged version:

pip install mlflow-client[mlflow]

Just create an MLClient

from mlflow_client import MLClient
client = MLClient()

MLClient accepts the following optional parameters:

• backend: the type of backend that you want to use to store your runs (With or without mlflow). Not specifying the backend will make the Client to provide automatically the backend.

from mlflow_client import Backend
client = MLClient(backend=Backend.MLFLOW) # or MLClient(backend="mlflow")
client = MLClient(backend=Backend.LOCAL) # or MLClient(backend="local")

• backend_uri: specify where you want to store the results, in case of local storage, backend_uri must be a path to the directory you want to save your results in. In case of an mlflow backend, backend_uri must be an url to a tracking server.
• experiment: the name of your experiment
• hooks_uri: this will be detailed in the WebHooks section

Now that your Client is set, you can start a run and with the same syntax as mlflow. client.start_run() will return a run object that allows you to log parameters, metrics, files and the resulting model.

from sklearn.dummy import  DummyClassifier
from mlflow_client import MLFramework, MLClient, Backend

client = MLClient(backend=Backend.LOCAL)

with client.start_run() as run:
    model = DummyClassifier()
    ... train the model
    run.log_parameter('alpha', 0.05)
    run.log_metric('loss', 3.30)
    run.log_artifact('/path/to/some/plot.png')
    # specify the type of the model that you want to persist
    run.log_model(model=model, output_dir='dummy_model', library=MLFramework.SCIKIT_LEARN)

If your model is not pure scikit or not pure (any other ml library), you can persist it as PyFunc: You just need to specify a path to your persisted model or a path to a directory containing these file(s) and an entry_point allowing to load the files in memory before making a prediction

For example if you created a class MyClass that's doing some preprocessing before fitting, you can initiate this class, do whatever you want, and then persist files as pkl in a directory 'local_path' You then just need to specify the load_entry_point which is a function / method that takes a path containing the pkl as parameter and returns any object exposing a predict method.

In the below example this method is the constructor of MyClass

... 
with client.start_run() as run:
    ...
    model = MyClass()
    model.preprocess()
    model.fit()
    model.save_models('local_path')
    # note: library can be omitted since default value is already PYFUNC
    run.log_model(model='local_path', load_entry_point='my_module.core.MyClass', library=MLFrameworks.PYFUNC)

You can have a look at all the possibilities here.

MLFlow Client allows you to configure hooks that will be triggered during the run. The currently supported format is the following:

{
  "event_name" (run_started | run_ended): [
      {
        "url": POST URL TO SEND THE HOOK TO
        "name": HOOK NAME (OPTIONAL)
      }, 
      ...
  ],
  ...
}

There are several ways to configure your run to use hooks:

• fill the "hooks_uri" parameter of the MLClient. This uri can be:
  • a GET URL that will return the hooks as JSON
  • a path to a JSON file
  • a path to a directory containing an MLproject, this MLproject needs to have a field hooks (example here)
• set the MLFLOW_HOOKS_URI env variable that can be either of the 3 uri above
• use client.add_hook(event=, name=, url=) on an MLClient instance before starting the run

When the event occurs the url specified will receive the following payload:

{
    "event": EVENT NAME,
    "status": (success | failed),
    'timestamp': TRIGGER TIME IN FORMAT YYYY-MM-DDTHH:MM:SS,
    'payload': {
        'experiment': EXPERIMENT NAME,
        'run': RUN UUID,
        'message': ADDITIONNAL INFORMATION (CONTAINS THE ERROR LOG IF STATUS FAILED)
    }
}