def test_unroll(self):
stream = fstream([(10,12,0.5,'a'), (14,15,1.2,'b')], fields=['start','end','score','name'])
expected = [(0,),(0.5,'a'),(0.5,'a'),(0,),(0,),(1.2,'b'),(0,)]
res = list(unroll(stream,(9,16)))
self.assertListEqual(res, expected)
stream = fstream([(0,1,5),(1,2,9),(2,3,11)], fields=['start','end','score'])
expected = [(5,),(9,),(11,)]
res = list(unroll(stream,(0,3)))
self.assertListEqual(res, expected)
def test_unroll(self):
stream = fstream([(10, 12, 0.5, 'a'), (14, 15, 1.2, 'b')],
fields=['start', 'end', 'score', 'name'])
expected = [(0, ), (0.5, 'a'), (0.5, 'a'), (0, ), (0, ), (1.2, 'b'),
(0, )]
res = list(unroll(stream, (9, 16)))
self.assertListEqual(res, expected)
stream = fstream([(0, 1, 5), (1, 2, 9), (2, 3, 11)],
fields=['start', 'end', 'score'])
expected = [(5, ), (9, ), (11, )]
res = list(unroll(stream, (0, 3)))
self.assertListEqual(res, expected)
def correlation(trackList, regions, limits=(-1000,1000), with_acf=False):
"""
Calculates the cross-correlation between two streams and
returns a vector containing the correlation at each lag in this order
(L/R for resp. limits[0], limits[1]):
[L,L+1,...,R-1,R].
If more than two tracks are given in *trackList*,
returns a list of correlation vectors, one for every distinct pair of tracks.
If `with_acf` is True, self-correlations will also be included in the list.
A negative lag indicates that track 2 is shifted to the right w.r.t track 1,
a positive lag - to the left.
So to get the correlation at lag +4, one has to look at the 4-L th
element of the array;
for lag -4, at the -4-L th element.
Example::
|_____ /^\ _________| lag 0
|______________/^\__|
|_____ /^\ _________| lag -8
-> |______________/^\__|
|_____ /^\ _________| lag +8
|______________/^\__| <-
:param trackList: list of FeatureStream objects
:param regions: a tuple (start,end) or a FeatureStream with the bounds of the regions to consider (see `unroll`).
In the latter case, all regions will be concatenated.
:param limits: (tuple (int,int)) maximum lag to consider. [-1000,1000]
:param with_acf: (bool) include auto-correlations. [False]
:rtype: list of floats, or list of lists of floats.
"""
##### One could profit from numpy to reduce the memory space used for
##### storing these - long - arrays ('dtype' is float64 by default).
if isinstance(regions,FeatureStream):
_reg = list(regions)
x = [array([s[0] for s in unroll(t,FeatureStream(_reg,fields=regions.fields))])
for t in trackList]
_reg = []
else:
x = [array([s[0] for s in unroll(t,regions)]) for t in trackList]
if limits[1]-limits[0] > 2*len(x[0]):
limits = (-len(x[0])+1,len(x[0])-1)
N0 = len(x[0])+limits[1]-limits[0]-1
##### convert to a power of 2, fft gets orders of magnitude faster...
N = 2**int(log(max(1,N0/200.0),2)+8.5) #2**int(log(2+N0,2)+.5)
if N < N0:
x = [vec_reduce(t[:N]) for t in x]
else:
x = [vec_reduce(t) for t in x]
def _corr(x1,x2,N):
corr = ifft(conjugate(fft(x1,N))*fft(x2,N))/len(x1)
corr = ncat((corr[N+limits[0]:], corr[:limits[1]+1]))
return real(corr)
if with_acf:
return [[_corr(x1,x2,N) for x2 in x[n:]] for n,x1 in enumerate(x)]
elif len(trackList) == 2:
return _corr(x[0],x[1],N)
else:
return [[_corr(x1,x2,N) for x2 in x[n+1:]] for n,x1 in enumerate(x[:-1])]
def correlation(trackList, regions, limits=(-1000, 1000), with_acf=False):
"""
Calculates the cross-correlation between two streams and
returns a vector containing the correlation at each lag in this order
(L/R for resp. limits[0], limits[1]):
[L,L+1,...,R-1,R].
If more than two tracks are given in *trackList*,
returns a list of correlation vectors, one for every distinct pair of tracks.
If `with_acf` is True, self-correlations will also be included in the list.
A negative lag indicates that track 2 is shifted to the right w.r.t track 1,
a positive lag - to the left.
So to get the correlation at lag +4, one has to look at the 4-L th
element of the array;
for lag -4, at the -4-L th element.
Example::
|_____ /^\ _________| lag 0
|______________/^\__|
|_____ /^\ _________| lag -8
-> |______________/^\__|
|_____ /^\ _________| lag +8
|______________/^\__| <-
:param trackList: list of FeatureStream objects
:param regions: a tuple (start,end) or a FeatureStream with the bounds of the regions to consider (see `unroll`).
In the latter case, all regions will be concatenated.
:param limits: (tuple (int,int)) maximum lag to consider. [-1000,1000]
:param with_acf: (bool) include auto-correlations. [False]
:rtype: list of floats, or list of lists of floats.
"""
##### One could profit from numpy to reduce the memory space used for
##### storing these - long - arrays ('dtype' is float64 by default).
if isinstance(regions, FeatureStream):
_reg = list(regions)
x = [
array([
s[0]
for s in unroll(t, FeatureStream(_reg, fields=regions.fields))
]) for t in trackList
]
_reg = []
else:
x = [array([s[0] for s in unroll(t, regions)]) for t in trackList]
if limits[1] - limits[0] > 2 * len(x[0]):
limits = (-len(x[0]) + 1, len(x[0]) - 1)
N0 = len(x[0]) + limits[1] - limits[0] - 1
##### convert to a power of 2, fft gets orders of magnitude faster...
N = 2**int(log(max(1, N0 / 200.0), 2) + 8.5) #2**int(log(2+N0,2)+.5)
if N < N0:
x = [vec_reduce(t[:N]) for t in x]
else:
x = [vec_reduce(t) for t in x]
def _corr(x1, x2, N):
corr = ifft(conjugate(fft(x1, N)) * fft(x2, N)) / len(x1)
corr = ncat((corr[N + limits[0]:], corr[:limits[1] + 1]))
return real(corr)
if with_acf:
return [[_corr(x1, x2, N) for x2 in x[n:]] for n, x1 in enumerate(x)]
elif len(trackList) == 2:
return _corr(x[0], x[1], N)
else:
return [[_corr(x1, x2, N) for x2 in x[n + 1:]]
for n, x1 in enumerate(x[:-1])]