MLflow

The platform hosts an MLflow tracking server at https://api.gwenlake.com/v1/mlflow, authenticated with your Gwenlake API key. A training written with gwenlake.training reports there by itself; any MLflow client can too.

From a training

gwenlake.training.train opens an MLflow run for each training and records:

the parameters of the run (steps, batch size, what the decorator and the command line set);
the metrics passed to run.log({...}), at their step;
the checkpoints written in run.checkpoint() as artifacts, shipped as they are written;
a run id kept next to the checkpoints, so a training that resumes after an eviction reports into the same run and the curve stays whole.

from gwenlake.training import train, Input, Output

@train(training_set=Input("crm.churn-train"), output=Output("crm.churn"),
       steps=30_000, eval_every=1_000, model="churn-v3", experiment="crm")
def fit(training_set, output, run):
    ...
    run.log({"loss": loss, "auc": auc})

model= names the run, so two models trained on the same data stay apart; experiment= picks the experiment (default otherwise). tracking=False on the decorator runs without the server. Tracking is an enhancement, never a dependency: with no key, no network or no mlflow installed (pip install 'mlflow-skinny[extras]'), the training runs and says so. The key is only ever sent to the platform’s own server, never to another tracking URI.

Environment

Variable	Meaning
`MLFLOW_TRACKING_URI`	overrides the platform’s server
`MLFLOW_EXPERIMENT_NAME`	the experiment, when the decorator names none
`MLFLOW_DISABLE`	turns tracking off
`GWENLAKE_API_KEY`	what authenticates the run on the platform’s server (`MLFLOW_TRACKING_TOKEN` is set from it)

A .env file in the working directory is loaded by the training module.

With the MLflow client

Anything that speaks MLflow works against the server, with the API key as the token:

export MLFLOW_TRACKING_URI=https://api.gwenlake.com/v1/mlflow
export MLFLOW_TRACKING_TOKEN=$GWENLAKE_API_KEY

import mlflow

mlflow.set_experiment("crm")
with mlflow.start_run(run_name="churn-v3"):
    mlflow.log_params({"lr": 1e-3})
    mlflow.log_metric("auc", 0.91)
    mlflow.log_artifact("model.pkl")

Experiments, runs, metrics and artifacts are browsed from the Atlas interface, or with the MLflow client (mlflow.search_runs, MlflowClient) pointed at the same server. A model’s card keeps the metrics its training returned; the run keeps the whole curve.

Get started

Concepts

Workspace

Data

Code

Build

Models

Inference

Agents

Client

Changelog

From a training

Environment

With the MLflow client

​From a training

​Environment

​With the MLflow client

From a training

Environment

With the MLflow client