How to predict/evaluate a split

[david-waterworth]

I've just started using gluonts and I'm a bit confused about how to use splitters (i.e. https://ts.gluon.ai/stable/tutorials/data_manipulation/dataset_splitting_example.html)

Specifically, it looks like the concept of splitters was added at some stage, most of the documentation simply creates a test and train dataset, then uses make_evaluation_predictions - I cannot find any documentation and how to use splits generated from generate_instance (i.e. generate_instance returns gluonts.dataset.split.TestData) to do the equivalent.

My data has 15m frequency, I generated by splits and train the model as follows:

splitter = OffsetSplitter(offset=-96)
training_dataset, test_template = splitter.split(dataset)

test_pairs = test_template.generate_instances(
    prediction_length=96,
    windows=1,
)

estimator = DeepAREstimator(
    freq=dataset.freq, context_length=7*96, prediction_length=96, num_layers=3, trainer_kwargs={'accelerator': 'gpu', 'max_epochs': 10}
)
predictor = estimator.train(training_dataset, num_workers=2)

But now how do I evaluate the predictions on test_pairs? Each test pair consists of two windows (the "past" and "future") so I need to combine them, ensure the model predicts the "future" only using the "past" etc. I'm assuming there's some helper somewhere - all my attempts result in errors, i.e.

for past, future in test_pairs:
    predictions = list(predictor.predict(past))

[david-waterworth]

I found what I wanted in the source README (https://github.com/awslabs/gluonts/blob/dev/README.md)

In particular forecasts = list(model.predict(test_data.input))