Question Regarding Virtual Staining with High Proportion of Positive Samples #1694
Description
Hello,
I am currently using your CycleGAN model for virtual staining in my research. I have observed that the proportion of positive samples in my dataset is relatively high. I'm wondering which training parameters I should adjust to improve the model's performance and generation capability under these circumstances.
Specifically, I would like to know:
- Which loss functions or training parameters are more suitable for addressing the issue of an overabundance of positive samples?
- Do you have any recommendations for adjusting the learning rate, loss weights, or other hyperparameters?
Thank you for your contributions, and I would greatly appreciate your insights!
Best regards