I tried running the example tutorial but I got the following error. ''' AssertionError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx '''

Hi Zhuoning,

This is an interesting work! I am wondering if the DAM method can be extended to a multi-class classification task with long-tailed imbalanced data. Intuitively, this should be possible as the famous sklearn tool provides auc score for multi-class setting by using one-versus-rest or one-versus-one technique.

Besides, it seems that optimizer.update_regularizer() is called only when the learning rate is reduced, thus it would be more elegant to incorporate this functional call into a pytorch lr scheduler. E.g.,

scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer)
scheduler.step()    # override the step to fulfill: optimizer.update_regularizer()

For current libauc version, the PESG optimizer is not compatible with schedulers in torch.optim.lr_scheduler . It would be great if this feature can be supported in the future.

Thanks for your work!

Hi,

When to use retain_graph=True in the loss backward function?

In 2 examples (2 and 4), it is True. But not in the other examples.

I appreciate your time.

I'm not very familiar with the maths in the paper so please forgive me if i'm asking something obvious.

The AUCMLoss uses the "imbalance ratio" between positive and negative samples. The ratio is defined as

the ratio of # of positive examples to the # of negative examples

Or imratio=#pos/#neg

When #pos<#neg, imratio is some value between 0 and 1. But when #pos>#neg, imratio>1

Will this break the loss calculations? I have a feeling it would invalidate the many 1-self.p calculations in the LibAUC implementation, but as i'm not familiar with the maths I can't say for sure.

Also, is there a problem (mathematically speaking) with calculating imratio=#pos/#total_samples, to avoid the problem above? When #pos<<#neg, #neg approximates #total_samples.

I noticed in the AUCMLoss class that the margin argument is not used. Following the formulation in the paper, the forward function should be changed in line 20 from 2*self.alpha*(self.p*(1-self.p) + \ to 2*self.alpha*(self.p*(1-self.p)*self.margin + \

I have started using this library and I've read your paper Robust Deep AUC Maximization: A New Surrogate Loss and Empirical Studies on Medical Image Classification, and I'm still not sure how to train a multi-label classification (MLC) model.

Specifically, how did you fine-tune for the Chexpert multi-label classification task? (i.e. classify 5 diseases, where each image may have presence of 0, 1 or more diseases)

• The first step pre-training with Cross-entropy loss seems clear to me
• You mention: "In the second step of AUC maximization, we replace the last classifier layer trained in the first step by random weights and use our DAM method to optimize the last classifier layer and all previous layers.". The new classifier layer is a single or multi-label classifier?
• In the Appendix I, figure 7 shows only one score as output for Deep AUC maximization (i.e. only one disease)
• In the code, both AUCMLoss() and APLoss_SH() receive single-label outputs, not multi-label outputs, apparently

How do you train for the 5 diseases? Train sequentially Cardiomegaly, then Edema, and so on? or with 5 losses added up? or something else?

Thank you for the great library. Does it currently support tensorflow? If so, could you provide an example of how it can be used with tensorflow? Thank you very much