I dont know if its still relevant but get final embeddings before decodings as vectors for your data and fit t-sne with that. You may need to play with perplexity for good visualisation.