Skip to content

Loading and/or Deploying a trained model

  • Once the AutoML tool is done, it will point you to a folder with the results/or you can find the results.zip from AutoDW

Tabular model - Loading

  • Once you have this folder (extract the zip if it is one), you can simply do 
    1
2
    from autogluon.tabular import TabularPredictor
predictor = TabularPredictor.load(predictor_path)
  • Now to predict something you can do, predictor.predict(test_data) or predictor.predict(test_data, model = 'X') for a specific model
  • For more instructions on how to use this, please refer to the AutoGluon documentation page

Tabular model - Deployment

Deployment Considerations

When deploying an AutoGluon Tabular model to production, the saved predictor directory should be treated as an immutable artifact. The predictor already includes all preprocessing, feature transformations, and ensemble logic, so no separate pipeline code is required.

  • Package the extracted predictor directory together with:

  • Your inference service code (e.g. FastAPI, Flask)

  • A pinned Python environment (e.g. requirements.txt or Docker image)
  • Load the predictor once at service startup, not per request
  • Reuse the in-memory predictor object for all inference calls

Example (FastAPI-style initialization):

1
2
3
from autogluon.tabular import TabularPredictor

predictor = TabularPredictor.load("/models/automl_predictor")

Input Validation

At inference time:

  • Inputs must be provided as a pandas DataFrame
  • Feature columns must exactly match those used during training
  • The target column must not be included

It is strongly recommended to:

  • Validate column names and data types before prediction
  • Reject or log requests with missing or extra columns

Performance and Scaling

  • AutoGluon predictors are CPU-optimized by default

  • For high-throughput use cases:

  • Run multiple workers (e.g. uvicorn --workers N)

  • Consider batching predictions where possible
  • For low-latency scenarios, avoid model reloading and disk access during requests

Model Versioning and Updates

  • Each trained predictor directory should be versioned (e.g. by dataset ID, timestamp, or hash)
  • Deployments should reference models by versioned paths
  • Rolling updates can be achieved by loading a new predictor and switching traffic

Serialization and Portability

  • The predictor directory is not framework-agnostic; it must be used with AutoGluon
  • Python and AutoGluon versions should match (or be compatible) between training and deployment environments
  • Containerized deployment (Docker) is recommended for reproducibility

Vision Dataset

  • This is a WIP