Is there an efficient way to do inference with non-embedding target sequence? Thanks!

[EmberNLP]

I'm trying to train a seq2seq model to predict sequence point coordinates (like [0.1,0.3,0.2]), so I do not have embedding matrix in my model. According to your wonderful tutorial, I know that TrainingHelper is helpful when training my model, but I do not know how to do the inference after training. It seems InferenceHelper may help, but I do not know how to use it.

To simply my question, I need the output of step t-1 as input of step t, and output looks like this [0.1,0.3,0.2].

Thanks!