Refactoring streaming transducer

[ezerhouni]

This PR is about refactoring beam search code for offline and online.
The idea is to create one class per beam search type. When we add new beam search (fast_beam_search_nbest, etc), we can inherit the necessary class and write as little code as possible.

[ezerhouni]

@csukuangfj I started only for streaming rnn-t at the moment. Could you have a quick look and tell me what you think about it. If this refactoring is satisfactory for you, I will do the same for the other models.

[csukuangfj]

@csukuangfj I started only for streaming rnn-t at the moment. Could you have a quick look and tell me what you think about it. If this refactoring is satisfactory for you, I will do the same for the other models.

Thanks! It looks good to me.

[ezerhouni]

@csukuangfj The PR should be ready to be reviewed. I tested all the streaming models but not the offline yet (will do next week).
I didn't touch offline_asr.py as it seems to be the same as offline_server.py, is there any difference ?

[csukuangfj]

offline_asr.py does not involve network communication. It's standalone, taking sound files as inputs and outputs text.

[csukuangfj]

Thanks. I will review it when I come to home(in 30 minutes).

[ezerhouni]

Thanks. I will review it when I come to home(in 30 minutes).

Great ! Thank you ! There is no rush

[csukuangfj]

Thanks!

Left some minor comments.

[csukuangfj]

Sorry, there are conflicts now.

[ezerhouni]

@csukuangfj I will take care of the conflicts, hopefully it won't be too bad.

[csukuangfj]

I just checked it locally and found that most of the conflicts are about style issues:

Hope that it will not cause too much trouble for you.

[ezerhouni]

@csukuangfj Just rebased, it was quite straightforward.

[csukuangfj]

I suggest ignoring errors reported by flake8 in https://github.com/k2-fsa/sherpa/runs/7479612670?check_suite_focus=true

We can edit https://github.com/k2-fsa/sherpa/blob/master/.flake8 to do that. (Either ignore the whole file or ignore
just some specific types of errors.)

You can use
https://github.com/k2-fsa/icefall/blob/master/.flake8
as an example.

[csukuangfj]

The CI complains
https://github.com/k2-fsa/sherpa/runs/7479632765?check_suite_focus=true

2022-07-23 08:00:57,024 INFO [offline_server.py:258] Using device: cpu
Traceback (most recent call last):
  File "sherpa/bin/pruned_transducer_statelessX/offline_server.py", line 675, in <module>
    main()
  File "/opt/hostedtoolcache/Python/3.7.13/x64/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line [28](https://github.com/k2-fsa/sherpa/runs/7479632765?check_suite_focus=true#step:11:29), in decorate_context
    return func(*args, **kwargs)
  File "sherpa/bin/pruned_transducer_statelessX/offline_server.py", line 643, in main
    num_active_paths=num_active_paths,
  File "sherpa/bin/pruned_transducer_statelessX/offline_server.py", line [29](https://github.com/k2-fsa/sherpa/runs/7479632765?check_suite_focus=true#step:11:30)5, in __init__
    self.model,
AttributeError: 'OfflineServer' object has no attribute 'model'

[ezerhouni]

I haven't test this part of the code yet. I need to do some extra testing before moving to merging

[ezerhouni]

@csukuangfj Seems like everything passed

[csukuangfj]

Thanks a lot!

Merging