The cluster centroids generated by DEC is actually computed by the
sklean.cluster.KMeans object used to fit the kmeans model on the embeddings.
It can be obtained by running a forward pass for all of the data on the trained encoder, calling
kmeans.fit() on the embeddings for all the training data and calling
kmeans.cluster_centers_ to retreive the centroids, which can then be passed into the decoder in order to visualize the centroids in the original data space.
See https://github.com/apache/incubator-mxnet/blob/master/example/deep-embedded-clustering/dec.py#L119-L120 for an example of how kmeans is performed and the centroids are retrieved during training.