AutoML
AutoML trains models on your data without code. You pick a dataset and the column to predict. AutoML then trains and tunes many models, ranks them on rows held back for validation, and keeps the best one, which is often an ensemble of several. Publishing a job's model puts it in the Model Registry. There it is deployed, tested, monitored for drift and compared in A/B tests like any other registry model.
AutoML runs on AutoGluon, on the platform's managed training runtime. You choose how much compute a job gets, and nothing else needs setting up.
Problem types
| Problem type | For | Predictor | Publishable to the registry |
|---|---|---|---|
| Tabular Data | Classification and regression on rows of numbers and categories | Binary Classifier, Multiclass Classifier or Regressor | Yes |
| Multimodal | Classification and regression on rows that combine free text, images and table columns | Binary Classifier, Multiclass Classifier or Regressor | Yes |
| Time Series | Forecasting future values of one or more series | Forecaster (point and quantile forecasts) | No: the model stays in the job's output volume |
The AutoML page
Open MLOps > AutoML. The cards at the top count:
- Total Jobs, all your jobs;
- Active Jobs, the ones still preparing, training or evaluating;
- Completed jobs;
- Failed jobs.
The list shows each job's Job Name, Type (problem type), Status (with a progress bar while it runs), Dataset, Performance (the validation score, its metric and the best model), Environment (its CPUs, memory and GPUs) and when it was Created. From the list you can open a job (View Details), stop a running one (Stop Job) or delete one (Delete Job). New AutoML Job starts the five-step create page.
A job moves through these statuses:
| Status | Meaning |
|---|---|
| Pending | Waiting to start |
| Preparing | Starting its environment and reading the data |
| Running / Training | Training models |
| Evaluating | Scoring the trained models and saving the results |
| Completed | Finished; its results and model are ready |
| Failed | Stopped by an error; the Logs tab says why |
| Cancelled | Stopped from the page or the list |
In this section
- Create a job: the dataset, its columns, the problem type, splits, metric, preset, time limit and hardware.
- Advanced settings: model families, ensembling, data handling, training behavior, hyperparameter tuning, per-family hyperparameters, time series and multimodal options, and custom JSON.
- Results and publishing: a job's page, its leaderboard, feature importance, configuration and logs, and publishing its model to the registry.
Billing and governance
A job runs as you, and its compute is charged to you. Before a job starts, it is checked against your organization's go-live gates and your budget. When a governance solution covers ML models, creating a job or publishing its model can need approval (see Governance).
REST API and Python SDK
Everything in this section is also available from the AutoML REST API and the Python SDK's client.automl. You can create a job, follow it, read its results, logs and metrics, stop or delete it, and publish its model (deploy_best_model).
from strongly import Strongly
client = Strongly()
for job in client.automl.list():
print(job.name, job.status, job.problem_type)