Have trouble accessing to the prediction result generated by DeepAREstimator using PandasDataset

[xrl0123]

Hi,

In this PandasDataset tutorial, it shows an example of how to do the forecasting with multiple time series data including static and dynamic features. I followed the code in the example. It finished with no errors. However, when I tried to view the prediction result using list(forecast_it), I got an empty list. Was that the correct way to get the prediction result from the DeepAREstimator model? Thanks!!!

Here is the code:

import numpy as np
import pandas as pd
from gluonts.dataset.pandas import PandasDataset
from gluonts.model.deepar import DeepAREstimator
from gluonts.mx import Trainer
from gluonts.evaluation import make_evaluation_predictions, Evaluator

prediction_length, freq = 24, "1H"
T = 10 * prediction_length
date_range = pd.date_range("2021-01-01", periods=T, freq=freq)

def generate_single_ts(date_range, item_id=None) -> pd.DataFrame:
    """create sum of `n_f` sin/cos curves with random scale and phase."""
    n_f = 2
    period = np.array([24 / (i + 1) for i in range(n_f)]).reshape(1, n_f)
    scale = np.random.normal(1, 0.3, size=(1, n_f))
    phase = 2 * np.pi * np.random.uniform(size=(1, n_f))
    periodic_f = lambda x: scale * np.sin(np.pi * x / period + phase)

    t = np.arange(0, len(date_range)).reshape(-1, 1)
    target = periodic_f(t).sum(axis=1) + np.random.normal(0, 0.1, size=len(t))
    ts = pd.DataFrame({"target": target}, index=date_range)
    if item_id is not None:
        ts["item_id"] = item_id
    return ts

def generate_single_ts_with_features(date_range, item_id) -> pd.DataFrame:
    ts = generate_single_ts(date_range, item_id)
    T = ts.shape[0]
    # static features are constant for each series
    ts["static_cat_1"] = np.random.randint(low=0, high=2)  # high is exclusive
    ts["dynamic_real_1"] = np.random.normal(size=T)
    ts["dynamic_real_2"] = np.random.normal(size=T)
    # ... we can have as many static or dynamic features as we like
    return ts

N = 10
multiple_ts = [
    generate_single_ts_with_features(date_range, item_id=i) for i in range(N)
]

to_pandasdataset = lambda data: PandasDataset(
    data,
    target="target",
    feat_dynamic_real=["dynamic_real_1", "dynamic_real_2"],
    feat_static_cat=["static_cat_1"],
)
train = to_pandasdataset([ts.iloc[:-prediction_length, :] for ts in multiple_ts])
test = to_pandasdataset(multiple_ts)


estimator_with_features = DeepAREstimator(
    freq=freq,
    prediction_length=prediction_length,
    use_feat_dynamic_real=True,
    use_feat_static_cat=True,
    cardinality=[
        2,
    ],
    trainer=Trainer(epochs=1),
)

predictor = estimator_with_features.train(train)
forecast_it, ts_it = make_evaluation_predictions(dataset=test, predictor=predictor)
evaluator = Evaluator(quantiles=(np.arange(20) / 20.0)[1:])
agg_metrics, item_metrics = evaluator(ts_it, forecast_it, num_series=len(test))

list(forecast_it)

Environment:
Python version: 3.7.13
GluonTS version: 0.10.2
mxnet version: 1.7.0.post2

[dai-ichiro]

You have to run this line again before list(forecast_it) because forecast_it is iterator.

forecast_it, ts_it = make_evaluation_predictions(dataset=test, predictor=predictor)

Hope this helps.

[xrl0123]

Hi @dai-ichiro, thanks so much for answering my question. I tried what you suggested and it worked! It seems like running this line

agg_metrics, item_metrics = evaluator(ts_it, forecast_it, num_series=len(test))

erased the value of forecast_it calculated through make_evaluation_predictions?