selection of encodings for pose and latent interpolation

In vis.py I have noticed that we have changed the way we select latents for specific class 

https://github.com/alan-turing-institute/affinity-vae/blame/17e65f90445e69b12fd81d50736f75783a120968/avae/vis.py#L1144

we seem to be selecting multiple encodings and then take the mean of them. Is it not better to just take one encoding for the purpose of interpolation and visualisation ?