Daniel Gleaves comments

Repositories
Issues
Comments

Results 1 comments of


                                            Daniel Gleaves

dot_product = s_loss_old - s_loss_new but s_loss_new - s_loss_old?

Where is that from? According to the original code, it is s_loss_new - s_loss_old: [Link](https://github.com/google-research/google-research/blob/8f456f55e95d25ff337ceec787bdbe9bb4ad3496/meta_pseudo_labels/training_utils.py#L472) `dot_product = cross_entropy['s_on_l_new'] - shadow` where shadow is defined as: [Link](https://github.com/google-research/google-research/blob/8f456f55e95d25ff337ceec787bdbe9bb4ad3496/meta_pseudo_labels/training_utils.py#L424-L426) ``` shadow = tf.get_variable(name='cross_entropy_old',...