Add fast beam search emformer #68
Conversation
| if decoding_method == "greedy_search": | ||
| decoder_out_list = [] | ||
| hyp_list = [] | ||
| else: |
There was a problem hiding this comment.
Please refactor it in a way that makes it easier to support other decoding methods.
| return parser.parse_args() | ||
|
|
||
|
|
||
| @torch.no_grad() | ||
| def run_model_and_do_greedy_search( | ||
| def run_model_and_do_search( |
There was a problem hiding this comment.
I suggest refactoring it into two functions:
- run_model_and_do_greedy_search
- run_model_and_do_fast_beam_search
It makes the code more readable and more maintainable, I think.
If we want to support another decoding method, we can just create a new function to do that.
It is just a suggestion. You can leave it as it is.
There was a problem hiding this comment.
It make sense, I will finish first the work of conv_emformer at first
|
Could you also update GitHub CI for https://github.com/k2-fsa/sherpa/blob/master/.github/workflows/run-streaming-conv-emformer-test.yaml#L36 to include You can apply a label to this PR with the name |
|
Should be all good @csukuangfj ! |
ezerhouni
changed the title
Thanks! Let us wait for the CI. Will merge after all tests pass. Thanks again! |
This PR adds
fast_beam_search_one_bestto the following models :pruned_stateless_emformer_rnnt2
conv_emformer_transducer_stateless