How to compute std of predcited tensors?

I have an undirected graph data where G = (V, E). I implemented graph convolution to calculate node embedding. For each fold, I have 38 testing samples. My model computes node embedding and transforms them into adjacency matrix format. My question how can I compute the standard deviation of these 38 predicted adjacency matrices in each fold? Should I compute the standard deviation from node embedding? If so, what is the true way to calculate?