export_model

TimeSeriesPredictor.export_model(path: str | Path, model: str) → str[source]

Export a trained model to a standalone checkpoint that can be loaded without AutoGluon.

This is useful if you fine-tuned a pretrained model with AutoGluon and want to use the resulting model outside of AutoGluon, e.g., for deployment or for sharing it with others.

The exported checkpoint is written in the format defined by the library that implements the model. For example, Chronos and Chronos2 models are exported as Hugging Face checkpoints:

predictor.export_model("./my_chronos2_model", model="Chronos2")

# the exported checkpoint can be loaded without AutoGluon
from chronos import BaseChronosPipeline

pipeline = BaseChronosPipeline.from_pretrained("./my_chronos2_model")

The exported checkpoint can also be used inside AutoGluon by passing it as model_path:

predictor.fit(train_data, hyperparameters={"Chronos2": {"model_path": "./my_chronos2_model"}})

Note

The exported checkpoint only contains the model itself. Data transformations that AutoGluon applies around the model, such as target_scaler or covariate_regressor, are not included. Models that use such transformations cannot be exported, since the exported model would produce different forecasts than predict().

Parameters:

  • path (str | Path) – Directory where the exported model will be saved.

  • model (str) – Name of the model to export. Available models can be listed with model_names().

Returns:

path – Directory containing the exported model.

Return type:

str

Raises:

  • NotImplementedError – If the selected model does not support exporting. Only some pretrained models can be exported; the error message lists the model types that support this operation.

  • ValueError – If the selected model uses a target_scaler, covariate_scaler or covariate_regressor.