when logging the selected slices to Weights and Biases, the pseudo-labels were also logged as true labels
when using the disable_interpolation option of the interpolation strategy together with the prefer_blocks_without_pseudo_labels option, the baseline selected other blocks in some cases since it had no pseudo-labels at all
The latter bug is fixed by adding an return_only_true_labels option in the dataset. This way, the dataset internally stores all the pseudo-labels but only return them on demand. In case that interpolation is disabled, the pseudo-labels are not returned during training but returned when running the block selection strategy. Therefore, at least for randomly sampling exactly the same true labels will be used as with interpolation enabled.
This PR fixes the following bugs:
disable_interpolationoption of the interpolation strategy together with theprefer_blocks_without_pseudo_labelsoption, the baseline selected other blocks in some cases since it had no pseudo-labels at allThe latter bug is fixed by adding an
return_only_true_labelsoption in the dataset. This way, the dataset internally stores all the pseudo-labels but only return them on demand. In case that interpolation is disabled, the pseudo-labels are not returned during training but returned when running the block selection strategy. Therefore, at least for randomly sampling exactly the same true labels will be used as with interpolation enabled.