Discrete sampling in Neural network

I am working with unsupervised object recognition using variational autoencoders. At one step, I have an array of intermediate representations and I have to choose a random sample from them. So, step by step, my problem can be explained as:

1. Z_what is an intermediate representation or latent vector. I have an array of Z_what, like [Z_what_1,Z_what_2, ... , Z_what_N].
2. I have to choose one Z_what_i from this array uniformly at random.

The problems are:

1. This is not a continuous sample. So I cannot use the reparameterization trick.
2. I have to obtain a Z_what_i , not just a softmax distribution on the array. So I think I cannot use Gumbel-softmax trick either.

Is there any way to do this? Thanks in advance.