def test_inference_quantile_prediction(quantiles, inference_quantiles):
hps = {
"seed": 42,
"freq": "D",
"prediction_length": 3,
"quantiles": quantiles,
"epochs": 3,
"num_batches_per_epoch": 3,
"is_iqf": True,
}
dataset_train, dataset_test = make_dummy_datasets_with_features(
cardinality=[3, 10],
num_feat_dynamic_real=2,
freq=hps["freq"],
prediction_length=hps["prediction_length"],
)
estimator = MQCNNEstimator.from_inputs(dataset_train, **hps)
predictor = estimator.train(dataset_train, num_workers=None)
forecasts = list(predictor.predict(dataset_test))
assert len(forecasts) == len(dataset_test)
item_id = 0
for inference_quantile in inference_quantiles:
assert (len(forecasts[item_id].quantile(inference_quantile)) ==
hps["prediction_length"])
inference_quantile = sorted(inference_quantiles)
previous_quantile_prediction = forecasts[item_id].quantile(
inference_quantile[0])
for inference_quantile in inference_quantiles[1:]:
assert all(previous_quantile_prediction[i] <= pred
for (i, pred) in enumerate(forecasts[item_id].quantile(
inference_quantile))), f"quantile-crossing occurred"
def test_mqcnn_covariate_smoke_test(
use_feat_dynamic_real,
add_time_feature,
add_age_feature,
enable_decoder_dynamic_feature,
hybridize,
):
hps = {
"seed": 42,
"freq": "D",
"prediction_length": 3,
"quantiles": [0.5, 0.1],
"epochs": 3,
"num_batches_per_epoch": 3,
"use_feat_dynamic_real": use_feat_dynamic_real,
"add_time_feature": add_time_feature,
"add_age_feature": add_age_feature,
"enable_decoder_dynamic_feature": enable_decoder_dynamic_feature,
"hybridize": hybridize,
}
dataset_train, dataset_test = make_dummy_datasets_with_features(
cardinality=[3, 10],
num_feat_dynamic_real=2,
freq=hps["freq"],
prediction_length=hps["prediction_length"],
)
estimator = MQCNNEstimator.from_hyperparameters(**hps)
predictor = estimator.train(dataset_train, num_workers=0)
forecasts = list(predictor.predict(dataset_test))
assert len(forecasts) == len(dataset_test)
def test_mqcnn_covariate_smoke_test(
use_past_feat_dynamic_real,
use_feat_dynamic_real,
add_time_feature,
add_age_feature,
enable_encoder_dynamic_feature,
enable_decoder_dynamic_feature,
hybridize,
quantiles,
distr_output,
is_iqf,
):
hps = {
"seed": 42,
"freq": "Y",
"context_length": 5,
"prediction_length": 3,
"quantiles": quantiles,
"distr_output": distr_output,
"epochs": 3,
"num_batches_per_epoch": 3,
"use_past_feat_dynamic_real": use_past_feat_dynamic_real,
"use_feat_dynamic_real": use_feat_dynamic_real,
"add_time_feature": add_time_feature,
"add_age_feature": add_age_feature,
"enable_encoder_dynamic_feature": enable_encoder_dynamic_feature,
"enable_decoder_dynamic_feature": enable_decoder_dynamic_feature,
"hybridize": hybridize,
"is_iqf": is_iqf,
}
dataset_train, dataset_test = make_dummy_datasets_with_features(
cardinality=[3, 10],
num_feat_dynamic_real=2,
num_past_feat_dynamic_real=4,
freq=hps["freq"],
prediction_length=hps["prediction_length"],
)
estimator = MQCNNEstimator.from_hyperparameters(**hps)
predictor = estimator.train(dataset_train, num_workers=None)
forecasts = list(predictor.predict(dataset_test))
assert len(forecasts) == len(dataset_test)
def test_backwards_compatibility():
hps = {
"freq": "D",
"context_length": 5,
"num_forking": 4,
"prediction_length": 3,
"quantiles": [0.5, 0.1],
"epochs": 3,
"num_batches_per_epoch": 3,
"use_feat_dynamic_real": True,
"use_past_feat_dynamic_real": True,
"enable_encoder_dynamic_feature": True,
"enable_decoder_dynamic_feature": True,
"scaling": True,
"scaling_decoder_dynamic_feature": True,
}
dataset_train, dataset_test = make_dummy_datasets_with_features(
cardinality=[3, 10],
num_feat_dynamic_real=2,
num_past_feat_dynamic_real=4,
freq=hps["freq"],
prediction_length=hps["prediction_length"],
)
for i in range(len(dataset_train)):
dataset_train.list_data[i]["dynamic_feat"] = dataset_train.list_data[
i
]["feat_dynamic_real"]
del dataset_train.list_data[i]["feat_dynamic_real"]
for i in range(len(dataset_test)):
dataset_test.list_data[i]["dynamic_feat"] = dataset_test.list_data[i][
"feat_dynamic_real"
]
del dataset_test.list_data[i]["feat_dynamic_real"]
estimator = MQCNNEstimator.from_inputs(dataset_train, **hps)
predictor = estimator.train(dataset_train, num_workers=None)
forecasts = list(predictor.predict(dataset_test))
assert len(forecasts) == len(dataset_test)
def test_feat_static_cat_smoke_test(use_feat_static_cat, cardinality):
hps = {
"seed": 42,
"freq": "D",
"prediction_length": 3,
"quantiles": [0.5, 0.1],
"epochs": 3,
"num_batches_per_epoch": 3,
"use_feat_static_cat": use_feat_static_cat,
}
dataset_train, dataset_test = make_dummy_datasets_with_features(
cardinality=cardinality,
num_feat_dynamic_real=2,
freq=hps["freq"],
prediction_length=hps["prediction_length"],
)
estimator = MQCNNEstimator.from_inputs(dataset_train, **hps)
predictor = estimator.train(dataset_train, num_workers=None)
forecasts = list(predictor.predict(dataset_test))
assert len(forecasts) == len(dataset_test)
def test_mqcnn_scaling_smoke_test(scaling):
hps = {
"seed": 42,
"freq": "D",
"prediction_length": 3,
"quantiles": [0.5, 0.1],
"epochs": 3,
"num_batches_per_epoch": 3,
"scaling": scaling,
}
dataset_train, dataset_test = make_dummy_datasets_with_features(
cardinality=[3, 10],
num_feat_dynamic_real=2,
freq=hps["freq"],
prediction_length=hps["prediction_length"],
)
estimator = MQCNNEstimator.from_inputs(dataset_train, **hps)
predictor = estimator.train(dataset_train, num_workers=0)
forecasts = list(predictor.predict(dataset_test))
assert len(forecasts) == len(dataset_test)
from gluonts.testutil.dummy_datasets import make_dummy_datasets_with_features
common_estimator_hps = dict(
freq="D",
prediction_length=3,
trainer=Trainer(epochs=3, num_batches_per_epoch=2, batch_size=4),
)
@pytest.mark.parametrize(
"estimator, datasets",
[
# No features
(
partial(DeepAREstimator, **common_estimator_hps),
make_dummy_datasets_with_features(),
),
# Single static categorical feature
(
partial(
DeepAREstimator,
**common_estimator_hps,
use_feat_static_cat=True,
cardinality=[5],
),
make_dummy_datasets_with_features(cardinality=[5]),
),
# Multiple static categorical features
(
partial(
DeepAREstimator,
