Implementation:Sktime Pytorch forecasting TimeSeriesDataSet From Dataset

Knowledge Sources	pytorch-forecasting PyTorch Forecasting Docs
Domains	Time_Series, Data_Engineering, Model_Evaluation
Last Updated	2026-02-08 07:00 GMT

Overview

Concrete tool for creating validation or test TimeSeriesDataSets by cloning configuration from an existing training dataset.

Description

The TimeSeriesDataSet.from_dataset class method constructs a new TimeSeriesDataSet using the same variable encoders, scalers, normalizers, and configuration as a source dataset but with different data. It calls get_parameters() on the source dataset to extract all constructor arguments (including fitted encoders and scalers), then passes them along with the new data to the constructor. This ensures consistency between training and evaluation data processing. Optional overrides can be passed via keyword arguments.

Usage

Use this method after constructing a training TimeSeriesDataSet to create a matching validation dataset. Typically called with stop_randomization=True for validation sets and optionally predict=True for single-prediction-per-group evaluation. Used in all four workflows.

Code Reference

Source Location

• Repository: pytorch-forecasting
• File: pytorch_forecasting/data/timeseries/_timeseries.py
• Lines: L1641-1682

Signature

@classmethod
def from_dataset(
    cls: type[TimeSeriesDataType],
    dataset: TimeSeriesDataType,
    data: pd.DataFrame,
    stop_randomization: bool = False,
    predict: bool = False,
    **update_kwargs,
) -> TimeSeriesDataType:
    """
    Construct dataset with different data, same variable encoders, scalers, etc.

    Parameters
    ----------
    dataset : TimeSeriesDataSet
        dataset from which to copy parameters
    data : pd.DataFrame
        data from which new dataset will be generated
    stop_randomization : bool, optional, default=False
        Whether to stop randomizing encoder and decoder lengths
    predict : bool, optional, default=False
        Whether to predict on last entries in time index only
    **update_kwargs
        keyword arguments overrides passed to constructor

    Returns
    -------
    TimeSeriesDataSet
    """

Import

from pytorch_forecasting import TimeSeriesDataSet
# Then call: TimeSeriesDataSet.from_dataset(training, data, ...)

I/O Contract

Inputs

Name	Type	Required	Description
dataset	TimeSeriesDataSet	Yes	Source dataset to copy configuration from
data	pd.DataFrame	Yes	New data for the dataset (e.g., full DataFrame including validation period)
stop_randomization	bool	No	Disable length randomization for validation (default: False)
predict	bool	No	Use only last entries per group for prediction (default: False)
**update_kwargs	dict	No	Override any constructor parameter

Outputs

Name	Type	Description
return	TimeSeriesDataSet	New dataset with same encoders/scalers but different data

Usage Examples

Create Validation Dataset

from pytorch_forecasting import TimeSeriesDataSet

# Assume 'training' is an existing TimeSeriesDataSet and 'data' is the full DataFrame
validation = TimeSeriesDataSet.from_dataset(
    training,
    data,
    stop_randomization=True,
    predict=False,
)

print(f"Training samples: {len(training)}")
print(f"Validation samples: {len(validation)}")

N-BEATS Validation with Min Prediction Index

# For N-BEATS, restrict validation to predictions after training cutoff
validation = TimeSeriesDataSet.from_dataset(
    training,
    data,
    min_prediction_idx=training_cutoff + 1,
)

Related Pages

Implements Principle

• Principle:Sktime_Pytorch_forecasting_Validation_Dataset_Creation

Requires Environment

• Environment:Sktime_Pytorch_forecasting_Core_Python_Dependencies