Negative labels are silently masked

[nicolaspi]

Hi,
Is there any reason for masking the negative labels? As far as I know there is no mention in the documentation that the labels must be positive, neither in the presentation slides.

Negative labels have a semantic significance in my use case and the metrics and losses I am using are still mathematically valid for negative values.

For now I am bypassing the constraint with a ugly patch:
tfr.utils.is_label_valid = lambda labels: tf.ones_like(labels, dtype=tf.bool)
maybe there is a better solution?