I have a hierarchical time series model that works on multiple time series and uses latent GPs. I can use minibatch for the series dimension (batch of series) but not for the time points dimension. This is due to the fact that, from what I see, minibatch only samples with replacement, which is not applicable to the GP case. Is there a quick way to sample without replacement using minibatches? Can you see other problems arising from using minibatches and GPs?