Dataset

Pure PyTorch Dataset for windowed time series from source files.

WindowedDataset

WindowedDataset(sources: list[SourceEntry], inputs, targets, win_sz: int | None = None, stp_sz: int = 1)

Bases: Dataset

Pure PyTorch Dataset for windowed time series from source files.

Parameters:

Name	Type	Description	Default
sources	list[SourceEntry]	list of SourceEntry (path + resampling factor); one file at several rates is several entries, one per factor	required
inputs		single reader or tuple of readers for input signals	required
targets		single reader or tuple of readers for target signals	required
win_sz	int | None	window size in (resampled) samples, None = full-file mode	None
stp_sz	int	step size between windows	1

Source code in tsfast/tsdata/dataset.py

def __init__(
    self,
    sources: list[SourceEntry],
    inputs,
    targets,
    win_sz: int | None = None,
    stp_sz: int = 1,
):
    self.sources = sources
    self._inputs = (inputs,) if not isinstance(inputs, tuple) else inputs
    self._targets = (targets,) if not isinstance(targets, tuple) else targets
    self._single_input = not isinstance(inputs, tuple)
    self._single_target = not isinstance(targets, tuple)
    self.win_sz = win_sz
    self.stp_sz = stp_sz
    self._ref_block = self._find_temporal(*self._inputs, *self._targets)

    if sources:
        for block in (*self._inputs, *self._targets):
            if hasattr(block, "probe"):
                block.probe(sources[0])

    if win_sz is not None:
        ref_block = self._ref_block
        counts = []
        for e in sources:
            eff_len = ref_block.file_len(e)  # already in resampled coords
            n = max(0, (eff_len - win_sz) // stp_sz + 1)
            counts.append(n)
        self._cumsum = np.cumsum(counts)
        self._counts = np.array(counts)