hi,
the labels (over thumos14 at least) in labels_multihothere do not seem to be multihot. is it normal?
they are class frequencies formatted in something that looks like hot encoding.
can you please check?
i run separately the piece of code that create labels_multihot over thumos14 files:
hi, the labels (over thumos14 at least) in
labels_multihot
here do not seem to be multihot. is it normal? they are class frequencies formatted in something that looks like hot encoding. can you please check?i run separately the piece of code that create
labels_multihot
over thumos14 files:each number is the frequency of the corresponding class.
it seems they stay in this form unchanged even when being used in the loss. https://github.com/XiaojunTang22/ICCV2023-DDGNet/blob/0fe0b1421b432c6931c59c81683d6597d95beced/wsad_dataset.py#L130 https://github.com/XiaojunTang22/ICCV2023-DDGNet/blob/0fe0b1421b432c6931c59c81683d6597d95beced/train.py#L15 https://github.com/XiaojunTang22/ICCV2023-DDGNet/blob/0fe0b1421b432c6931c59c81683d6597d95beced/train.py#L20 https://github.com/XiaojunTang22/ICCV2023-DDGNet/blob/0fe0b1421b432c6931c59c81683d6597d95beced/train.py#L23
inside the loss, they seem to be used directly without change, except adding an extra class for the background as here: https://github.com/XiaojunTang22/ICCV2023-DDGNet/blob/0fe0b1421b432c6931c59c81683d6597d95beced/model.py#L785 as if they are binary.
can you please double-check at your end? thanks