Hyperparameter candidates

[K0ki116]

Dear all,

Reading Surreal-GAN paper on aging (Yang et al., Nat Med, 2024), I'm now trying to apply Surreal-GAN to our data.

As a hyperparameter, candidates on lamda were set to 0.1, 0.2, 0.4, 0.6, 0.8 in the original Surreal-GAN paper.
However, I haven't found detailed methods on how to tune hyperparameter gamma.

What should the candidate values/range for gamma be?

I appreciate your advice,

Koki

[zhijian-yang]

Hi Koki,

Thank you for trying our methods. Gamma can be set to values between 0.1 and 8. Some more details about other hyper-parameters can be found in section "Important Hyper-parameters", in the readme file. For lam and gamma, I suggest that, if you see no more improvements in model reproducibility by increasing either lam or gamma, there is no need to try even larger values.

Hope that is helpful!

Thanks,
Zhijian

[K0ki116]

Dear Zhjijan,

Thank you for your reply!

I made a sample hyperparameter grid. What do you think about this?

M: 2,3,4,5,6
lamda: 0.1, 0.2, 0.4, 0.6, 0.8
gamma: 0.1, 0.5, 1, 3, 5, 8

[zhijian-yang]

That looks good to me now!