Hello,
I am a little bit confused about the dimensions of mu and sigma in the latent space of scPhere when choosing VMF or WN distribution. As far as I understood, when choosing a latent dimension of 50 in the WN scenario, after the polar project, the dimension of the mu vector will be 51 and the sigma will remain with 50 elements for one datapoint. Similar story happens for the VMF scenario as well, and 51 "z" samples are passed to the decoder.
I am quite confused about how the wrapped normal distribution handles this. One one hand when using the sample function, a 51 dimension vector is sampled. On the other hand when calculating the log probability, the std vector is used which has 50 dimensions. I was wondering if you can clear up how does this work.
Hello,
I am a little bit confused about the dimensions of mu and sigma in the latent space of scPhere when choosing VMF or WN distribution. As far as I understood, when choosing a latent dimension of 50 in the WN scenario, after the polar project, the dimension of the mu vector will be 51 and the sigma will remain with 50 elements for one datapoint. Similar story happens for the VMF scenario as well, and 51 "z" samples are passed to the decoder.
I am quite confused about how the wrapped normal distribution handles this. One one hand when using the sample function, a 51 dimension vector is sampled. On the other hand when calculating the log probability, the std vector is used which has 50 dimensions. I was wondering if you can clear up how does this work.