Specifically, I'd like to integrate the SVDKL model in place of the VAE-GP combination. Additionally, I aim to train the DKL model dynamically during the Bayesian Optimization (BO) step, instead of relying on a static latent space from a pretrained model, as is the case with the VAE.
Could you provide guidance on how this could be implemented or if there are any potential challenges or limitations with such an approach?
Hi,
I have been following the tutorial VAE for MNIST in BoTorch and was wondering if it's possible to replace the combination of the VAE and single-task GP model with the SVDKL (Stochastic Variational Deep Kernel Learning) model from the GPyTorch tutorial.
Specifically, I'd like to integrate the SVDKL model in place of the VAE-GP combination. Additionally, I aim to train the DKL model dynamically during the Bayesian Optimization (BO) step, instead of relying on a static latent space from a pretrained model, as is the case with the VAE.
Could you provide guidance on how this could be implemented or if there are any potential challenges or limitations with such an approach?
Thank you!