def test_invalid_rolling_parameters(prediction_length, unique):
try:
generate_rolling_dataset(
dataset=generate_dataset("constant"),
start_time=pd.Timestamp("2000-01-01-20", freq="1H"),
end_time=pd.Timestamp("2000-01-02-00", freq="1H"),
strategy=StepStrategy(
step_size=prediction_length if unique else 1,
prediction_length=prediction_length,
),
)
# program should have failed at this point
pytest.raises(RuntimeError)
except AssertionError:
pass
def test_dynamic_features(
test_length: int,
prediction_length: int,
target_start: str,
rolling_start: str,
num_dynamic_feat: int,
):
# generate datasets with dynamic values equal to the target values
rolled_ds = generate_rolling_dataset(
dataset=create_dynamic_dataset(target_start, test_length,
num_dynamic_feat),
strategy=StepStrategy(prediction_length=prediction_length),
start_time=pd.Period(rolling_start),
)
for timeseries in rolled_ds:
for feature_name in (
FieldName.FEAT_DYNAMIC_REAL,
FieldName.FEAT_DYNAMIC_CAT,
):
assert len(timeseries[feature_name]) == num_dynamic_feat
for feature in timeseries[feature_name]:
# since target and dynamic features has the same values
# in the dataset we can ensure ordering and length of
# the features by comparing them with the target value
assert (timeseries[FieldName.TARGET] == feature).all()
def test_dynamic_integration(
train_length: int,
test_length: int,
prediction_length: int,
target_start: str,
rolling_start: str,
num_dynamic_feat: int,
):
"""
Trains an estimator on a rolled dataset with dynamic features.
Tests https://github.com/awslabs/gluon-ts/issues/1390
"""
train_ds = create_dynamic_dataset(target_start, train_length,
num_dynamic_feat)
rolled_ds = generate_rolling_dataset(
dataset=create_dynamic_dataset(target_start, test_length,
num_dynamic_feat),
strategy=StepStrategy(prediction_length=prediction_length),
start_time=pd.Timestamp(rolling_start),
)
estimator = DeepAREstimator(
freq="D",
prediction_length=prediction_length,
context_length=2 * prediction_length,
use_feat_dynamic_real=True,
trainer=Trainer(epochs=1),
)
predictor = estimator.train(training_data=train_ds)
forecast_it, ts_it = make_evaluation_predictions(rolled_ds,
predictor=predictor,
num_samples=100)
training_agg_metrics, _ = Evaluator(num_workers=0)(ts_it, forecast_it)
# it should have failed by this point if the dynamic features were wrong
assert training_agg_metrics
def test_step_strategy(ds_name, prediction_length, unique, ignore_end,
ds_expected):
rolled_ds = generate_rolling_dataset(
dataset=generate_dataset(ds_name),
start_time=pd.Period("2000-01-01-20", freq="1H"),
end_time=None if ignore_end else pd.Period("2000-01-02-00", freq="1H"),
strategy=StepStrategy(
step_size=prediction_length if unique else 1,
prediction_length=prediction_length,
),
)
i = 0
for ts in rolled_ds:
assert len(ts["target"]) == len(ds_expected[i])
for rolled_result, expected in zip(ts["target"], ds_expected[i]):
assert rolled_result == expected
i += 1
assert len(ds_expected) == i
StepStrategy,
generate_rolling_dataset,
)
if __name__ == "__main__":
dataset = get_dataset("constant", regenerate=False)
estimator = SimpleFeedForwardEstimator(
prediction_length=dataset.metadata.prediction_length,
freq=dataset.metadata.freq,
trainer=Trainer(epochs=5, num_batches_per_epoch=10),
)
predictor = estimator.train(dataset.train)
# create the rolled dataset to use for forecasting and evaluation
dataset_rolled = generate_rolling_dataset(
dataset=dataset.test,
start_time=pd.Timestamp("2000-01-01-15", freq="1H"),
end_time=pd.Timestamp("2000-01-02-04", freq="1H"),
strategy=StepStrategy(
prediction_length=dataset.metadata.prediction_length, ),
)
forecast_it, ts_it = make_evaluation_predictions(
dataset_rolled, predictor=predictor, num_samples=len(dataset_rolled))
agg_metrics, _ = Evaluator()(ts_it, forecast_it)
pprint(agg_metrics)