Skip to main content

A set of tools for working with mlflow (see https://mlflow.org)

Project description

mlflowhelper

A set of tools for working with mlflow (see https://mlflow.org)

Features

  • managed artifact logging and loading
    • automatic artifact logging and cleanup
    • no overwriting files when running scripts in parallel
    • loading artifact
    • central configuration of logging and loading behavior
  • log all function parameters and locals with a simple call to mlflowhelper.log_vars()

Documentation

pip install mlflowhelper

Managed artifact logging and loading

General functionality

from matplotlib import pyplot as plt
import mlflowhelper

with mlflowhelper.start_run():
    with mlflowhelper.managed_artifact("plot.png") as artifact:
        fig = plt.figure()
        plt.plot([1,2,3], [1,2,3])
        fig.savefig(artifact.get_path())

This code snippet automatically logs the created artifact (plot.png). At the same time if will create the artifact in a temporary folder so that you don't have to worry about overwriting it when running your scripts in parallel. By default, this also cleans up the artifact and the temporary folder after logging.

You can also manage artifacts on a directory level:

from matplotlib import pyplot as plt
import mlflowhelper

with mlflowhelper.start_run():
    with mlflowhelper.managed_artifact_dir("plots") as artifact_dir:

        # plot 1
        fig = plt.figure()
        plt.plot([1,2,3], [1,2,3])
        fig.savefig(artifact_dir.get_path("plot1.png"))

        # plot 2
        fig = plt.figure()
        plt.plot([1,2,3], [1,2,3])
        fig.savefig(artifact_dir.get_path("plot2.png"))

Artifact loading

You may want to run experiments but reuse some precomputed artifact from a different run (such as preprocessed data, trained models, etc.). This can be done as follows:

import mlflowhelper
import pandas as pd

with mlflowhelper.start_run():
    mlflowhelper.set_load(run_id="e1363f760b1e4ab3a9e93f856f2e9341", stages=["load_data"]) # activate loading from previous run
    with mlflowhelper.managed_artifact_dir("data.csv", stage="load_data") as artifact:
        if artifact.loaded:
            # load artifact
            data = pd.read_csv(artifact.get_path())
        else:
            # create and save artifact
            data = pd.read_csv("/shared/dir/data.csv").sample(frac=1)
            data.to_csv(artifact.get_path())

Similarly, this works for directories of course:

import mlflowhelper
import pandas as pd

mlflowhelper.set_load(run_id="e1363f760b1e4ab3a9e93f856f2e9341", stages=["load_data"]) # activate loading from previous run
with mlflowhelper.start_run():
    with mlflowhelper.managed_artifact_dir("data", stage="load_data") as artifact_dir:
        train_path = artifact_dir.get_path("test.csv")
        test_path = artifact_dir.get_path("train.csv")
        if artifact_dir.loaded:
            # load artifacts
            train = pd.read_csv(train_path)
            test = pd.read_csv(test_path)
        else:
            data = pd.read_csv("/shared/dir/data.csv").sample(frac=1)
            train = data.iloc[:100,:]
            test = data.iloc[100:,:]
            # save artifacts
            train.to_csv(train_path)
            test.to_csv(test_path)

Note: The stage parameter must be set in mlflowhelper.managed_artifact(_dir) to enable loading.

Central logging and loading behavior management

Logging and loading behavior can be managed in a central way:

import mlflowhelper
import pandas as pd

with mlflowhelper.start_run():

    # activate loading the stage `load_data` from previous run `e1363f760b1e4ab3a9e93f856f2e9341`
    mlflowhelper.set_load(run_id="e1363f760b1e4ab3a9e93f856f2e9341", stages=["load_data"])

    # deactivate logging the stage `load_data`, in this case for example because it was loaded from a previous run
    mlflowhelper.set_skip_log(stages=["load_data"])

    with mlflowhelper.managed_artifact_dir("data", stage="load_data") as artifact_dir:
        train_path = artifact_dir.get_path("test.csv")
        test_path = artifact_dir.get_path("train.csv")
        if artifact_dir.loaded:
            # load artifacts
            train = pd.read_csv(train_path)
            test = pd.read_csv(test_path)
        else:
            data = pd.read_csv("/shared/dir/data.csv").sample(frac=1)
            train = data.iloc[:100,:]
            test = data.iloc[100:,:]
            # save artifacts
            train.to_csv(train_path)
            test.to_csv(test_path)

Note: For central managing the stage parameter must be set in mlflowhelper.managed_artifact(_dir).

Easy parameter logging

mlflowhelper helps you to never forget logging parameters again by making it easy to log all existing variables using mlflowhelper.log_vars.

import mlflowhelper

def main(param1, param2, param3="defaultvalue", verbose=0, *args, **kwargs):
    some_variable = "x"
    with mlflowhelper.start_run(): # mlflow.start_run() is also OK here
        mlflowhelper.log_vars(exclude=["verbose"])

if __name__ == '__main__':
    main("a", "b", something_else=6)

This will log:

{
  "param1": "a",
  "param2": "b",
  "param3": "defaultvalue",
  "something_else": 6
}

Persistent dictionary

mlflowhelper provides dictionary-like implementation that persistens elements to MLFlow.

from mlflowhelper.tracking.collections import MlflowDict
d = MlflowDict() # you can also provide tracking URI or an MlflowClient
d["a"] = 5
del d
d = MlflowDict() # you can also provide tracking URI or an MlflowClient
print(d["a"]) # will give you 5

Other

There are a few more convenience functions included in mlflowhelper:

TODOs / Ideas

  • check if loading works across experiments
  • purge local artifacts (check via API which runs are marked as deleted and delete their artifacts)
  • support nested runs by creating subdirectories based on experiment and run
  • support loading from central cache instead of from runs
  • automatically log from where and what has been loaded
  • set tags for logged stages (to check for artifacts before loading them)
  • consider loading extensions:
    • does nested loading make sense (different loads for certain nested runs)?
    • does mixed loading make sense (loading artifacts from different runs for different stages)?

Note

This project has been set up using PyScaffold 3.2.1. For details and usage information on PyScaffold see https://pyscaffold.org/.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

mlflowhelper-1.1.0.tar.gz (27.1 kB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page