Open auroua opened 7 years ago
@auroua Eh... You are right it's a bug (I forgot it was used in pt calculate :-( ). Thanks very much for reporting such mistakes!
Is this a two-class version or multi-class version?
tf.nn.sigmoid_cross_entropy_with_logits()不是已经包括了负号嘛
https://github.com/CasiaFan/tensorflow_retinanet/blob/39d53cca8bc5bdb10f57614282ed2b568e349910/loss.py#L35
Why you multiply
onehot_labels
again in this line?