I’m working on end-to-end speech recognition as my university graduation project.
For short, I wonder
Is it possible to initialize RNN weight by other RNN’s weight value, or “repeat” it to inflate it’s size?
Recently my team tried to apply beam search during inference, and had discussion about RNN’s weight.
If we set beam size as 3, each beam will have its own hidden state.
We are looking for the way to compute all beam’s hidden state at once, not using for loop and deal with it one by one.
So we repeated encoder output and other values for 3 times(same as beam size).
But what about weights?
If I have some misunderstandings about beam search process, please let me know.
Thank you very much!