Hi, @blackfeather-wang, I would like to ask if there is any discrepancy between your code implementation and the description of the paper. In your paper, you say that we are going to minimize $\mathcal{L}{contrast}$, and the form of $\mathcal{L}{contrast}$ is as follows.

However, after I read your code, I find that your code seems to minimize $-\mathcal{L}_{contrast}$, and the consequence is that you're not maximizing the lower bound of $I(h,y)$, instead you're minimizing the lower bound of $I(h,y)$, can you explain that?

Hi. Thanks for the great idea and sharing the code.
I have a question. You changed the position of runner.optimizer.zero_grad() from after_train_iter to before_train_iter, refererence to

I was wondering what this change will do and how does it affect the training and model performance.
I adopted your idea to UNet in medical segmentation but missed the aforementioned change. Yet the training converges correctly as expected. Now I realized the change and like to know its supposed effect.
Thank you again.

Thanks for your awesome work! I'm very interested in mutual information estimation used in your paper.
According to Appendix-G and your finding (figure-6), you train an auxiliary classifier to estimate I(h, y) and end2end supervised training retains all task-relevant information.
From my perspective, it shows that we don't need to build a classifier upon the final feature map and deploying the classifier (trained on feature maps from many layers) to the first feature map is enough.
I'm not sure I understand this correctly. Would you help me clarify this?

Hi. This work is really a fantastic job! I want to do some further research based on it in order to reduce the required resources. I noticed that you've provided memory usage in your paper. May I ask how did you measure the GPU footprint?
Thanks very much. Looking forward to your reply.