Closed MailSuesarn closed 1 year ago
It is definitely possible, however, implementations may be complicated in the case of TensorFlow. First, you need to create three new learnable variables, i.e., a
, b
, alpha
in the loss function based on our AUCMLoss
loss. Second, you need to modify these variables in TF optimizer during updating parameters, i.e., updating a
, b
by gradient descent and updating alpha
by gradient ascent. Additionally, you need to maintain a list of variables (referred as reference variable v
in the paper) to compute accumulated average of model variables for regularization purpose. For more details, you can take a look at the original paper.
We are planning on releasing some code in TensorFlow in the coming months. Feel free to star this repository to stay updated on the latest developments and news. Thanks!
thank you very much I will wait for your new version release. It's a great job.
Is it possible that I will use AUCMLoss in the version of Tensorflow 2? Can you suggest to me how to implement AUCMLoss with Tensorflow 2