The general idea for generating from mixture component is you generate from each component n data point (n being the number of data point you want), resulting a (k, n, dim) data tensor, then for each of the n data slice you select 1 from k components.
You can also have a look at examples in scikit-learn.