quantlab.model.torch_training.drop_extreme#

quantlab.model.torch_training.drop_extreme(y, fraction, column=0)[source]#

Return the rows kept after dropping both tails of one label column.

With k = int(fraction * n), n the number of finite labels in column, the k smallest and the k largest are dropped (ties broken by position), as MASTER does in training. Rows with a NaN label are kept.

Parameters:

ytorch.Tensor: One bar’s labels, [S_t, L].
fractionfloat: Share dropped from each tail, in [0, 0.5).
columnint, default 0: The label column that decides.

Returns:

torch.Tensor: [S_t] booleans, True for the rows kept.

Raises:

ValueError: If fraction is outside [0, 0.5).

Examples

>>> y = torch.tensor([[5.0], [1.0], [float("nan")], [3.0], [2.0], [4.0]])
>>> drop_extreme(y, 0.2)
tensor([False, False,  True,  True,  True,  True])