About sample on the latent space

[Jay5212]

Dear author
I am a undergraduate from China，my graduation project's purpose is that use VAE model to generate some "rare" signals. The whole process roughly as follows. For example, i have a dataset which has 10000 signals,including 7500 'A'signals、1000 'B'signals、1200 'C'signals and 300 ‘D’signals. I want to put the signals into the trainloader to learn the distribution of these signals in the latent space through encoder. Then i want to sample the specific location (such as D signals' location ) on the latent space and pass it to the decoder to reconstruct the signals. I think my method is feasible in the theory. But i have troubles when i want to sample the speficial location. I use PCA method to visualize the distribution of latent space and get some clusters and use k-means method to calculate the centers. Then use back-calculate to get the really location in the latent space and pass to the decoder. But now i find it is difficult to realize this method. Could you please advise me on the feasibility of this method? If this approach is not feasible, are there alternative methods that can help me achieve my goal?
I would appreciate it if you could answer me.
Thank you for your read.