How to save and load best DeepAR model ?

[karndeepsingh]

Hi,
I have been training DeepAR model for timeseries forecasting. I want to know how we can save best weights or best trained model and how it can be loaded back and used for prediction?

Please, provide some help on this.
Thanks

[dai-ichiro]

import pickle

with open('gluonts_model', 'wb') as f:
    pickle.dump(predictor, f)

with open('gluonts_model', 'rb') as f:
    predictor = pickle.load(f)

[karndeepsingh]

How could I save best weights? Does predictor object saves best checkpoints or the latest checkpoints?

[jaheba]

Hi,

you should be able to call model.serialize(<path>):

https://github.com/awslabs/gluon-ts/blob/569fcc779c66a81c843857cca76aa1f07a8ea0fc/src/gluonts/model/predictor.py#L85

To load the model, you just call Predictor.deserialize(<path>).

[vmajor]

I am using the pytorch version of TFT and am doing this to train it. Checkpoints are saved automatically as a function of Lightning:

from gluonts.torch.model.tft import TemporalFusionTransformerEstimator

  # Train a TFT model on the data
   # Toggle comment when enabling Optuna
    estimator = TemporalFusionTransformerEstimator(
        prediction_length=prediction_length,
        freq=freq,
        trainer_kwargs={"max_epochs": 25},
        context_length=8,
        dropout_rate=0.40410747064575153,
        hidden_dim=16,
        num_heads=8,
    )

    #model = estimator.train(train_dataset, val_dataset, ckpt_path='trained_models/tft_model.ckpt')
    model = estimator.train(train_dataset, val_dataset, num_workers=11)

[SummitStudiosDev]

did you ever figure out how to load from a pytorch lightning checkpoint by chance?

[vmajor]

I do not remember, sorry. I did manage to find a way to load from checkpoint but I vaguely remember I used the built in method from gluonts which is in fact the pytorch method. I am sorry I am unable to help more. I abandoned this approach because I found that I do not actually want to load pretrained models in my use case so this solution and related code was not retained.

[SummitStudiosDev]

that's ok! i just ended up retraining it and using the intended method of serializing it

[Somendratripathi]

Using train_model method returns the best model which you could serealize.

https://ts.gluon.ai/stable/_modules/gluonts/torch/model/deep_npts/_estimator.html#DeepNPTSEstimator.train_model