def process(self, activations, **kwargs):
"""
Detect the tempi from the (beat) activations.
Parameters
----------
activations : numpy array
Beat activation function.
Returns
-------
tempi : numpy array
Array with the dominant tempi [bpm] (first column) and their
relative strengths (second column).
"""
# smooth the activations
act_smooth = int(round(self.fps * self.act_smooth))
activations = smooth_signal(activations, act_smooth)
# generate a histogram of beat intervals
histogram = self.interval_histogram(activations.astype(np.float))
# smooth the histogram
histogram = smooth_histogram(histogram, self.hist_smooth)
# detect the tempi and return them
return detect_tempo(histogram, self.fps)
def smooth_histogram(histogram, smooth):
"""
Smooth the given histogram.
Parameters
----------
histogram : tuple
Histogram (tuple of 2 numpy arrays, the first giving the strengths of
the bins and the second corresponding delay values).
smooth : int or numpy array
Smoothing kernel (size).
Returns
-------
histogram_bins : numpy array
Bins of the smoothed histogram.
histogram_delays : numpy array
Corresponding delays.
Notes
-----
If `smooth` is an integer, a Hamming window of that length will be used as
a smoothing kernel.
"""
# smooth only the histogram bins, not the corresponding delays
return smooth_signal(histogram[0], smooth), histogram[1]
def process(self, activations):
"""
Detect the beats in the given activation function.
:param activations: beat activation function
:return: detected beat positions [seconds]
"""
# smooth activations
act_smooth = int(self.fps * self.tempo_estimator.act_smooth)
activations = smooth_signal(activations, act_smooth)
# TODO: refactor interval stuff to use TempoEstimation
# if look_ahead is not defined, assume a global tempo
if self.look_ahead is None:
# create a interval histogram
histogram = self.tempo_estimator.interval_histogram(activations)
# get the dominant interval
interval = self.tempo_estimator.dominant_interval(histogram)
# detect beats based on this interval
detections = detect_beats(activations, interval, self.look_aside)
else:
# allow varying tempo
look_ahead_frames = int(self.look_ahead * self.fps)
# detect the beats
detections = []
pos = 0
# TODO: make this _much_ faster!
while pos < len(activations):
# look N frames around the actual position
start = pos - look_ahead_frames
end = pos + look_ahead_frames
if start < 0:
# pad with zeros
act = np.append(np.zeros(-start), activations[0:end])
elif end > len(activations):
# append zeros accordingly
zeros = np.zeros(end - len(activations))
act = np.append(activations[start:], zeros)
else:
act = activations[start:end]
# create a interval histogram
histogram = self.tempo_estimator.interval_histogram(act)
# get the dominant interval
interval = self.tempo_estimator.dominant_interval(histogram)
# add the offset (i.e. the new detected start position)
positions = detect_beats(act, interval, self.look_aside)
# correct the beat positions
positions += start
# search the closest beat to the predicted beat position
pos = positions[(np.abs(positions - pos)).argmin()]
# append to the beats
detections.append(pos)
pos += interval
# convert detected beats to a list of timestamps
detections = np.array(detections) / float(self.fps)
# remove beats with negative times and return them
return detections[np.searchsorted(detections, 0):]
def smooth_histogram(histogram, smooth):
"""
Smooth the given histogram.
:param histogram: histogram
:param smooth: smoothing kernel [numpy array or int]
:return: smoothed histogram
Note: If 'smooth' is an integer, a Hamming window of that length will be
used as a smoothing kernel.
"""
# smooth only the the histogram bins, not the corresponding delays
return smooth_signal(histogram[0], smooth), histogram[1]
def process(self, activations):
"""
Detect the beats in the given activation function.
:param activations: beat activation function
:return: detected beat positions [seconds]
"""
import itertools as it
# estimate the tempo
tempi = self.tempo_estimator.process(activations)
intervals = self.fps * 60. / tempi[:, 0]
# compute possible intervals
if self.use_factors:
# use the dominant interval with different factors
possible_intervals = [int(intervals[0] * f) for f in self.factors]
possible_intervals = [i for i in possible_intervals if
self.tempo_estimator.max_interval >= i >=
self.tempo_estimator.min_interval]
else:
# take the top n intervals from the tempo estimator
possible_intervals = intervals[:self.num_intervals]
# sort and start from the greatest interval
possible_intervals.sort()
possible_intervals = [int(i) for i in possible_intervals[::-1]]
# smooth activations
act_smooth = int(self.fps * self.tempo_estimator.act_smooth)
activations = smooth_signal(activations, act_smooth)
# since the cython code uses memory views, we need to make sure that
# the activations are C-contiguous and of C-type float (np.float32)
contiguous_act = np.ascontiguousarray(activations, dtype=np.float32)
results = self.map(_process_crf,
it.izip(it.repeat(contiguous_act),
possible_intervals,
it.repeat(self.interval_sigma)))
# normalize their probabilities
normalized_seq_probabilities = np.array([r[1] / r[0].shape[0]
for r in results])
# pick the best one
best_seq = results[normalized_seq_probabilities.argmax()][0]
# convert the detected beat positions to seconds and return them
return best_seq.astype(np.float) / self.fps
def process(self, activations):
"""
Detect the tempi from the beat activations.
:param activations: RNN beat activation function
:return: numpy array with the dominant tempi (first column)
and their relative strengths (second column)
"""
# smooth the activations
act_smooth = int(round(self.fps * self.act_smooth))
activations = smooth_signal(activations, act_smooth)
# generate a histogram of beat intervals
histogram = self.interval_histogram(activations.astype(np.float))
# smooth the histogram
histogram = smooth_histogram(histogram, self.hist_smooth)
# detect the tempi and return them
return detect_tempo(histogram, self.fps)
def peak_picking(activations,
threshold,
smooth=None,
pre_avg=0,
post_avg=0,
pre_max=1,
post_max=1):
"""
Perform thresholding and peak-picking on the given activation function.
Parameters
----------
activations : numpy array
Activation function.
threshold : float
Threshold for peak-picking
smooth : int or numpy array
Smooth the activation function with the kernel (size).
pre_avg : int, optional
Use `pre_avg` frames past information for moving average.
post_avg : int, optional
Use `post_avg` frames future information for moving average.
pre_max : int, optional
Use `pre_max` frames past information for moving maximum.
post_max : int, optional
Use `post_max` frames future information for moving maximum.
Returns
-------
peak_idx : numpy array
Indices of the detected peaks.
See Also
--------
:func:`smooth`
Notes
-----
If no moving average is needed (e.g. the activations are independent of
the signal's level as for neural network activations), set `pre_avg` and
`post_avg` to 0.
For peak picking of local maxima, set `pre_max` and `post_max` to 1.
For online peak picking, set all `post_` parameters to 0.
References
----------
.. [1] Sebastian Böck, Florian Krebs and Markus Schedl,
"Evaluating the Online Capabilities of Onset Detection Methods",
Proceedings of the 13th International Society for Music Information
Retrieval Conference (ISMIR), 2012.
"""
# smooth activations
if smooth not in (None, 0):
activations = smooth_signal(activations, smooth)
# compute a moving average
avg_length = pre_avg + post_avg + 1
if avg_length > 1:
# TODO: make the averaging function exchangeable (mean/median/etc.)
avg_origin = int(np.floor((pre_avg - post_avg) / 2))
if activations.ndim == 1:
filter_size = avg_length
elif activations.ndim == 2:
filter_size = [avg_length, 1]
else:
raise ValueError('`activations` must be either 1D or 2D')
mov_avg = uniform_filter(activations,
filter_size,
mode='constant',
origin=avg_origin)
else:
# do not use a moving average
mov_avg = 0
# detections are those activations above the moving average + the threshold
detections = activations * (activations >= mov_avg + threshold)
# peak-picking
max_length = pre_max + post_max + 1
if max_length > 1:
# compute a moving maximum
max_origin = int(np.floor((pre_max - post_max) / 2))
if activations.ndim == 1:
filter_size = max_length
elif activations.ndim == 2:
filter_size = [max_length, 1]
else:
raise ValueError('`activations` must be either 1D or 2D')
mov_max = maximum_filter(detections,
filter_size,
mode='constant',
origin=max_origin)
# detections are peak positions
detections *= (detections == mov_max)
# return indices
if activations.ndim == 1:
return np.nonzero(detections)[0]
elif activations.ndim == 2:
return np.nonzero(detections)
else:
raise ValueError('`activations` must be either 1D or 2D')
def peak_picking(activations, threshold, smooth=None, pre_avg=0, post_avg=0,
pre_max=1, post_max=1):
"""
Perform thresholding and peak-picking on the given activation function.
Parameters
----------
activations : numpy array
Activation function.
threshold : float
Threshold for peak-picking
smooth : int or numpy array
Smooth the activation function with the kernel (size).
pre_avg : int, optional
Use `pre_avg` frames past information for moving average.
post_avg : int, optional
Use `post_avg` frames future information for moving average.
pre_max : int, optional
Use `pre_max` frames past information for moving maximum.
post_max : int, optional
Use `post_max` frames future information for moving maximum.
Returns
-------
peak_idx : numpy array
Indices of the detected peaks.
See Also
--------
:func:`smooth`
Notes
-----
If no moving average is needed (e.g. the activations are independent of
the signal's level as for neural network activations), set `pre_avg` and
`post_avg` to 0.
For peak picking of local maxima, set `pre_max` and `post_max` to 1.
For online peak picking, set all `post_` parameters to 0.
References
----------
.. [1] Sebastian Böck, Florian Krebs and Markus Schedl,
"Evaluating the Online Capabilities of Onset Detection Methods",
Proceedings of the 13th International Society for Music Information
Retrieval Conference (ISMIR), 2012.
"""
# smooth activations
if smooth not in (None, 0):
activations = smooth_signal(activations, smooth)
# compute a moving average
avg_length = pre_avg + post_avg + 1
if avg_length > 1:
# TODO: make the averaging function exchangeable (mean/median/etc.)
avg_origin = int(np.floor((pre_avg - post_avg) / 2))
if activations.ndim == 1:
filter_size = avg_length
elif activations.ndim == 2:
filter_size = [avg_length, 1]
else:
raise ValueError('`activations` must be either 1D or 2D')
mov_avg = uniform_filter(activations, filter_size, mode='constant',
origin=avg_origin)
else:
# do not use a moving average
mov_avg = 0
# detections are those activations above the moving average + the threshold
detections = activations * (activations >= mov_avg + threshold)
# peak-picking
max_length = pre_max + post_max + 1
if max_length > 1:
# compute a moving maximum
max_origin = int(np.floor((pre_max - post_max) / 2))
if activations.ndim == 1:
filter_size = max_length
elif activations.ndim == 2:
filter_size = [max_length, 1]
else:
raise ValueError('`activations` must be either 1D or 2D')
mov_max = maximum_filter(detections, filter_size, mode='constant',
origin=max_origin)
# detections are peak positions
detections *= (detections == mov_max)
# return indices
if activations.ndim == 1:
return np.nonzero(detections)[0]
elif activations.ndim == 2:
return np.nonzero(detections)
else:
raise ValueError('`activations` must be either 1D or 2D')
