Online beat tracking

[superbock]

Rebased everything on top of #185.

Before this PR can be merged, a couple of things need to be addressed:

• add uni-directional RNN beat tracking models
• make a decision whether to keep the current behaviour of the process() method

[winchell]

Looking forward to using this! I'm testing it out but here's what happens:

Mac-Pro:~ $ DBNBeatTracker online test.wav 
/Library/Python/2.7/site-packages/madmom/processors.py:750: UserWarning: make sure that the `frame_size` (2048) is equal to the maximum value used by any `FramedSignalProcessor`.
  frame_size)
Traceback (most recent call last):
  File "/usr/local/bin/DBNBeatTracker", line 112, in <module>
    main()
  File "/usr/local/bin/DBNBeatTracker", line 108, in main
    args.func(processor, **vars(args))
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 762, in process_online
    _process((processor, frame, outfile, process_args))
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 200, in _process
    return process_tuple[0](*process_tuple[1:-1], **process_tuple[-1])
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 128, in __call__
    return self.process(*args, **kwargs)
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 491, in process
    data = _process((self.in_processor, data, kwargs))
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 200, in _process
    return process_tuple[0](*process_tuple[1:-1], **process_tuple[-1])
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 128, in __call__
    return self.process(*args, **kwargs)
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 336, in process
    data = _process((processor, data, kwargs))
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 200, in _process
    return process_tuple[0](*process_tuple[1:-1], **process_tuple[-1])
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 128, in __call__
    return self.process(*args, **kwargs)
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 336, in process
    data = _process((processor, data, kwargs))
  File "/Library/Python/2.7/site-packages/madmom/processors.py", line 203, in _process
    return process_tuple[0](*process_tuple[1:-1])
  File "/Library/Python/2.7/site-packages/madmom/ml/nn/__init__.py", line 40, in average_predictions
    predictions = predictions[0]
IndexError: list index out of range

[superbock]

Did you update the models? Please run git submodule update.

[winchell]

I did try that (but it didn't work):

Mac-Pro:madmom $ git submodule update --force --remote
Submodule path 'madmom/models': checked out '9b7b11dde6f586e2b23e9432d42183c9b2fd1758'

[superbock]

Ah, I forgot to include the models in setup.py to be installed as well. Try changing line 40 in setup.py to 'models/beats/201[56]/*'.

[winchell]

Awesome, thanks! It's working now.

[superbock]

Thanks for testing :) Will fix it and (hopefully) release v0.15 very soon.

[fdlm]

cool beans! 👍
Just some minor comments....

[fdlm]

typo :)

[fdlm]

• Is there any use case for the nn_files parameter?
• Does it make sense to have the nn_files parameter if there is no frame_sizes parameter? As I see it, only neural networks trained on the given frame sizes (either for online or offline mode) would work.

[superbock]

Yes. Although you're right about which model can be used (only trained specifically for/with this settings), it can be quite useful to select only one or two of the trained models if processing power is limited.

[fdlm]

I understand. But then this use case should be explained in the docs, I think it's not obvious from the beginning. Now that I've checked, the online parameter is also missing from the docstring of the class.

[superbock]

Added docstrings and doctests.

[fdlm]

Just out of curiosity, is this better than np.atleast_2d?

[superbock]

Yes, it is way faster on all machines I tested. I only spotted this when profiling the whole online beat tracking stuff, and it had a quite remarkable impact.