[WIP] Online beat trackers #314

SebastianPoell · 2017-07-17T21:45:15Z

This PR extends The BeatTracker with online capability based on #292.
Also a simple CCF beat detecting method for experimenting was added.

superbock

Nice work! I am not sure about some defaults (e.g. the 5 frames look back or a threshold of 0.1), these should at least be documented somehow or reconsidered completely -- e.g. the 5 frames could maybe expressed as the minimum interval or something like that.

superbock · 2017-07-19T06:00:10Z

madmom/features/beats.py

+ buffer = self.buffer(activation)
+ # create an online interval histogram
+ histogram = self.tempo_estimator.online_interval_histogram(
+ activation, reset=False)


You should pass the reset argument to the tempo_estimator as well, i.e. reset=reset. Otherwise the beat tracker resets its state, but the tempo estimation still keeps its old state.

Hmm, actually I think it's ok the way it is? The tempo_estimator gets reset inside the reset() method of the BeatTracker and might be called whenever process_online() is called.

If I want to compute a whole song then process_online() gets called only once with all the activations and reset=True. If I would use reset=reset here then while iterating through all activations it would reset the tempo_estimator for each activation.

Ok, I missed the resetting inside reset(). Anyways, I still think passing reset=reset is the cleaner solution than calling self.tempo_estimator.reset(). When iterating over activations frame by frame, reset must be set to false of course -- and this is exactly what is done in online mode.

Ah sorry, I forgot to pass the online=True parameter to process_single() (I'm calling stuff with a script). Nevermind, works now ;)

superbock · 2017-07-19T06:02:01Z

madmom/features/beats.py

+ activation, reset=False)
+ # get the dominant interval
+ interval = self.tempo_estimator.dominant_interval(histogram)
+ # compute current and next beat times


This comment is a bit misleading, especially the part of next beat. If I get it right it is the first position which is possible to be the next beat, right?

superbock · 2017-07-19T06:07:37Z

madmom/features/beats.py

+ # for every frame the tempo may change and therefore the last beat
+ # can easily be missed.
+ if len(detections) > 0 and detections[-1] >= len(buffer) - 5 and \
+ buffer[detections[-1]] > self.threshold:


The indentation is weird and the if-statement is very long. I'd prefer to have multiple logical expressions and then test for them, e.g.:

cond_1 = detections[-1] >= len(buffer) - 5 cond_2 = buffer[detections[-1]] > self.threshold if detections and cond_1 and cond_2:

Of course, more meaningful names would be nice, but you get the idea. What I still don't like is some hard coded values such as -5. Why 5 frames?

I just tested different values here and this just seemed to work best. But I know what you mean... maybe i can find something more rational. For the threshold the intention was that we probably don't want false positive predictions on silent signals. This also seem to improve the dataset results a bit.

I was never questioning the rationale behind using a threshold, e.g.DBNBeatTrackingProcessor uses it for the very same reason 😏. Back then I added the threshold for the DBN stuff and did not back-port it to the other methods. Maybe we should do so now and add it to the other methods (after evaluating it's positive effect of course).

Some more notes regarding the threshold:

What I do is to determine the first/last frame where the activation is above/below the threshold and use only the segment in between. Although this works quite nicely, I am not sure if this is not simply some sort of adaptation towards the annotations...

The same could be done for multiple segments, not just one per song. Although this sounds logical and straight forward, I am not sure if this segmentation helps in general. We must be aware of the many songs which have segments without any perceivable beats where counting continues. For these cases splitting into segments doesn't make sense at all.

Ah ok, cool! Yeah using only "active" segments seems like a good idea. Maybe the distance between those active segments gives a hint whether it should stop tracking or not.
For online it's a bit different anyway because we don't know how long the break will be so we have to decide on the spot (or maybe with a bit of delay).

superbock · 2017-07-19T06:24:10Z

madmom/features/beats.py

+ # display tempo
+ display.append('| %5.1f | ' % float(self.fps * 60 / interval))
+ sys.stderr.write('\r%s' % ''.join(display))
+ sys.stderr.flush()


We should consider refactoring the visualisation stuff as a mixin which provides a visualize() method. BeatTrackingProcessor and DBNBeatTrackingProcessor could then inherit from this mixin and simply call visualize() whenever needed. The mixin should have/keep its own state and counters. Some more notes: we should consider using constants to define how long the beats are displayed (e.g. self.beat_counter = DISPLAY_DURATION) or the inertia of the RNN beat activation strength (e.g. DISPLAY_INERTIA).

I think that would be very nice! Every BeatTracker could show the activation strength, the tempo and whether the current frame is a beat. My only concern is that different trackers may want to show different additional information (like the DBNTracker does with the state space).

superbock · 2017-07-19T06:25:19Z

tests/test_features_beats.py

@@ -79,6 +79,25 @@ def test_process(self):
 self.assertTrue(np.allclose(beats, [0.11, 0.45, 0.79, 1.13, 1.47,
 1.81, 2.15, 2.49]))

+ def test_process_online(self):
+ processor = BeatTrackingProcessor(fps=sample_blstm_act.fps,


minor note: use fps=sample_lstm_act.fps instead.

SebastianPoell · 2017-07-19T13:14:48Z

Thx for all your comments, will work through all of them and then amend ;)

SebastianPoell · 2017-07-20T21:24:04Z

madmom/features/beats.py

@@ -390,6 +390,51 @@ def recursive(position):
 return np.array(positions)


+def detect_beats_with_ccf(activations, interval, look_aside=0.2):


I am actually not sure if you want to keep this method :) It is completely interchangeable with detect_beats() but seems to produce results which are a little bit worse. On the other hand its much faster. Maybe its an option for the BeatDetector (with constant tempo).

The problem you are facing here is the fact that many musical pieces simply don't have a constant tempo.

Although you're allowing some deviation from a constant tempo, this is only considered for each beat individually, i.e. the tempo remains stable and only some jitter is allowed. If one beat is a bit earlier (faster), the position of the next is not based on the current beat, but rather on the previously pre-computed positions.

detect_beats on the other hand allows also some deviation from the main tempo, but the position of the next beat is based on this position rather than the pre-computed ones. This makes this method more flexible, but also slower.

If detect_beats_with_ccf performs inferior to detect_beats on most datasets I don't see a point for including it, even if it is faster. I'd rather speed up detect_beats if speed is really an issue.

Yes, i think the same! At least i can use it for my thesis and evaluate it :)

I'm very curious now about the MultiAgent approach because, similar to detect_beats it always jumps from beat to beat but considers multiple tempo hypothesis and should be a very good fit for online mode 😋

superbock · 2017-10-16T08:51:48Z

I merged #292, please rebase your branch on master.

SebastianPoell · 2017-10-17T23:32:40Z

Ok, rebased this branch and applied your stuck-appveyor fixes to the tests. I think the -0.1% Coverage is due to the CCF-Commit (which can be omitted). Anyway, would be awesome if you could take it from here \m/

superbock reviewed Jul 19, 2017

View reviewed changes

SebastianPoell force-pushed the online_beat_tracker branch 5 times, most recently from fb6c54f to cb466ac Compare July 20, 2017 19:57

SebastianPoell commented Jul 20, 2017

View reviewed changes

SebastianPoell force-pushed the online_beat_tracker branch 2 times, most recently from cb466ac to 9d85686 Compare July 21, 2017 01:38

superbock force-pushed the refactor_tempo branch 5 times, most recently from 48d0cf8 to 886937c Compare July 24, 2017 20:17

superbock force-pushed the refactor_tempo branch 7 times, most recently from 37e83bb to cbdbe8c Compare August 18, 2017 18:18

superbock force-pushed the refactor_tempo branch 5 times, most recently from ce27bc1 to 67fb095 Compare August 30, 2017 21:22

SebastianPoell force-pushed the online_beat_tracker branch 3 times, most recently from b169c9b to 31c4d33 Compare August 31, 2017 06:27

SebastianPoell force-pushed the online_beat_tracker branch 4 times, most recently from 2aa06d6 to 4fd7ab9 Compare September 1, 2017 14:03

superbock force-pushed the refactor_tempo branch 3 times, most recently from 8777d30 to ceb5c74 Compare September 3, 2017 10:33

SebastianPoell force-pushed the online_beat_tracker branch 10 times, most recently from 56eb679 to 55d7f1f Compare September 18, 2017 22:03

superbock force-pushed the refactor_tempo branch 4 times, most recently from dcdbb8b to 1c640e5 Compare September 19, 2017 13:19

superbock force-pushed the refactor_tempo branch from 1c640e5 to be97b36 Compare October 16, 2017 08:14

SebastianPoell force-pushed the online_beat_tracker branch from 55d7f1f to a192ef0 Compare October 17, 2017 22:30

SebastianPoell changed the base branch from refactor_tempo to master October 17, 2017 22:41

SebastianPoell added 3 commits October 18, 2017 00:58

Basic online BeatTracker

a4e8a39

Adapted BeatDetector/BeatTracker to TempoDetector defaults

ae70766

Added CCF method

5ba9021

SebastianPoell force-pushed the online_beat_tracker branch from a192ef0 to 5ba9021 Compare October 17, 2017 22:58

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] Online beat trackers #314

[WIP] Online beat trackers #314

SebastianPoell commented Jul 17, 2017 •

edited

superbock left a comment

superbock Jul 19, 2017

SebastianPoell Jul 20, 2017 •

edited

superbock Jul 20, 2017

SebastianPoell Jul 21, 2017

superbock Jul 19, 2017

superbock Jul 19, 2017

SebastianPoell Jul 19, 2017 •

edited

superbock Jul 21, 2017

superbock Jul 21, 2017

SebastianPoell Jul 21, 2017

superbock Jul 19, 2017 •

edited

SebastianPoell Jul 19, 2017

superbock Jul 19, 2017

SebastianPoell commented Jul 19, 2017

SebastianPoell Jul 20, 2017 •

edited

superbock Jul 21, 2017

SebastianPoell Jul 21, 2017

superbock commented Oct 16, 2017

SebastianPoell commented Oct 17, 2017

		@@ -390,6 +390,51 @@ def recursive(position):
		return np.array(positions)


		def detect_beats_with_ccf(activations, interval, look_aside=0.2):

[WIP] Online beat trackers #314

Are you sure you want to change the base?

[WIP] Online beat trackers #314

Conversation

SebastianPoell commented Jul 17, 2017 • edited

superbock left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SebastianPoell Jul 20, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SebastianPoell Jul 19, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

superbock Jul 19, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SebastianPoell commented Jul 19, 2017

SebastianPoell Jul 20, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

superbock commented Oct 16, 2017

SebastianPoell commented Oct 17, 2017

SebastianPoell commented Jul 17, 2017 •

edited

SebastianPoell Jul 20, 2017 •

edited

SebastianPoell Jul 19, 2017 •

edited

superbock Jul 19, 2017 •

edited

SebastianPoell Jul 20, 2017 •

edited