def test_minimal(self):
test_matr = np.zeros([20, 100])
test_matr[0, 10] = 1
lines = _identify_ridge_lines(test_matr, 2*np.ones(20), 1)
assert_(len(lines) == 1)
test_matr = np.zeros([20, 100])
test_matr[0:2, 10] = 1
lines = _identify_ridge_lines(test_matr, 2*np.ones(20), 1)
assert_(len(lines) == 1)
def test_single_pass(self):
distances = [0, 1, 2, 5]
gaps = [0, 1, 2, 0, 1]
test_matr = np.zeros([20, 50]) + 1e-12
length = 12
line = _gen_ridge_line([0, 25], test_matr.shape, length, distances, gaps)
test_matr[line[0], line[1]] = 1
max_distances = max(distances)*np.ones(20)
identified_lines = _identify_ridge_lines(test_matr, max_distances, max(gaps) + 1)
assert_array_equal(identified_lines, [line])
def test_single_bigdist(self):
distances = [0, 1, 2, 5]
gaps = [0, 1, 2, 4]
test_matr = np.zeros([20, 50])
length = 12
line = _gen_ridge_line([0, 25], test_matr.shape, length, distances, gaps)
test_matr[line[0], line[1]] = 1
max_dist = 3
max_distances = max_dist*np.ones(20)
#This should get 2 lines, since the distance is too large
identified_lines = _identify_ridge_lines(test_matr, max_distances, max(gaps) + 1)
assert_(len(identified_lines) == 2)
for iline in identified_lines:
adists = np.diff(iline[1])
np.testing.assert_array_less(np.abs(adists), max_dist)
agaps = np.diff(iline[0])
np.testing.assert_array_less(np.abs(agaps), max(gaps) + 0.1)
def test_single_biggaps(self):
distances = [0]
max_gap = 1
gaps = [3, 6]
test_matr = np.zeros([50, 50])
length = 30
line = _gen_ridge_line([0, 25], test_matr.shape, length, distances, gaps)
test_matr[line[0], line[1]] = 1
max_dist = 1
max_distances = max_dist*np.ones(50)
#This should get 3 lines, since the gaps are too large
identified_lines = _identify_ridge_lines(test_matr, max_distances, max_gap)
assert_(len(identified_lines) == 3)
for iline in identified_lines:
adists = np.diff(iline[1])
np.testing.assert_array_less(np.abs(adists), max_dist)
agaps = np.diff(iline[0])
np.testing.assert_array_less(np.abs(agaps), max(gaps) + 0.1)
def find_peaks(data, widths=[1, 2, 7, 30, 182, 365]):
'''
Finds the peaks using the CWTFindPeaks algorithm. This code is mostly a line
by line port of the scipy.signal.wavelets.find_peaks_cwt. We had to port that
code since the default scipy implementation does not return the widths.
Paramaters
----------
data : array like
the time series to find the peaks
widths : array like
the candidate widths to test
Returns
-------
A list of tripples (peak_volume, peak_width, peak_position). The volume is
the value of data[peak_position]. The width is the estimated width of the
wavelet used to find that peak.
'''
data = np.asanyarray(data)
widths = np.asanyarray(widths)
#These are default values from the scipy port which we based our code on.
gap_thresh = np.ceil(widths[0])
max_distances = widths / 4.0
cwt_dat = cwt(data, ricker, widths)
ridge_lines = _identify_ridge_lines(cwt_dat, max_distances, gap_thresh)
filtered = _filter_ridge_lines(cwt_dat, ridge_lines, \
min_snr=1, noise_perc=1) #noise_perc=1 filters more noise.
#Filtered will be of the form [[peak_widths], [peak_positions]]
candidates = []
for x in filtered:
assert x[0].min() >= 0
assert x[0].max() < widths.shape[0]
peak_pos, peak_width = x[1][0], widths[x[0].max()]
candidates.append((data[peak_pos], peak_width, peak_pos))
return sorted(candidates, reverse=True)
def wavelet_peak_find(s, min_snr=1., assume_sig=4., min_length=8.0,
max_dist=4.0, gap_thresh=2.0):
# this import is here to let scipy 0.9.0 at least
# load this module
import scipy.signal._peak_finding as spf
y, t = s.values, s.index
widths = np.linspace(1, 100, 200)
cwtm = spf.cwt(y, spf.ricker, widths)
ridges = spf._identify_ridge_lines(cwtm, widths / max_dist, gap_thresh)
filt_ridges = spf._filter_ridge_lines(cwtm, ridges, \
min_length=cwtm.shape[0] / min_length, min_snr=min_snr)
### the next code is just to visualize how this works
#import matplotlib.pyplot as plt
#ctr_x, ctr_y = np.meshgrid(t, widths)
#ctr_y *= (t[1] - t[0])
#plt.contourf(ctr_x, ctr_y, cwtm)
##plt.imshow(cwtm) #, extent=(widths[0], widths[-1], times[0], times[-1]))
#for l in ridges:
# plt.plot(t[l[1]], l[0] * 0.5 * (t[1] - t[0]), 'k-')
#for l in filt_ridges:
# plt.plot(t[l[1]], l[0] * 0.5 * (t[1] - t[0]), 'r-')
##plt.plot(peaks_t, peaks_w, 'k*') # not working
#plt.show()
# loop through the ridges and find the point of maximum
# intensity on the ridge and save its characteristics
peak_list = []
for i, l in enumerate(filt_ridges):
pl = np.argmax([cwtm[j, k] for j, k in zip(l[0], l[1])])
peak_w = widths[l[0][pl]] * 0.5 * (t[1] - t[0])
peak_amp = cwtm[l[0][pl], l[1][pl]] / (widths[l[0]][pl] ** 0.5)
peak_t = t[l[1][pl]]
t0, t1 = peak_t - assume_sig * peak_w, peak_t + assume_sig * peak_w
peak_list.append({'t0': t0, 't1': t1, 'x': peak_t, \
'h': peak_amp, 'w': peak_w})
return peak_list
def test_empty(self):
test_matr = np.zeros([20, 100])
lines = _identify_ridge_lines(test_matr, 2*np.ones(20), 1)
assert_(len(lines) == 0)
def test_empty(self):
test_matr = np.zeros([20, 100])
lines = _identify_ridge_lines(test_matr, 2 * np.ones(20), 1)
assert_(len(lines) == 0)
def detect_peaks(x: np.ndarray, y: np.ndarray, widths: np.ndarray,
min_length: int = 5, max_distance: int = 2,
gap_threshold: int = 1, snr: float = 3, min_width: float = 5,
max_width: float = 60,
estimators: Union[str, _estimator_type] = "default"):
r"""
Find peaks in a 1D signal.
Peaks are detected using a modified version of the algorithm described in
[1].
Parameters
----------
x : sorted array
y : array of intensities
widths : array
Array of widths, in x units. Used as scales to build the wavelet
array.
min_length : int
Minimum number of points in a ridge line.
max_distance : float
Maximum x distance between consecutive points in a ridge line, in x
units.
gap_threshold : int
Maximum number of consecutive missing peaks in a ridge line.
snr : positive number
Signal-to-noise- ratio used to filter peaks. Defined as follows:
.. math::
SNR = \frac{peak height - baseline}{noise}
min_width : positive number
Minimum width of the peaks
max_width : positive number
Maximum width of the peaks
estimators : str or dict
How to estimate baseline, noise, peak height, peak width, peak area and
peak location. If `estimators` is 'cwt', parameters are computed as
described in [1]. Check the Notes to see how estimations in 'default'
mode are computed or how custom estimators can be used.
Returns
-------
peaks : List of PeakLocation
params : dict of peak parameters
Notes
-----
Peaks are detected using the CWT algorithm described in [DP06]. The optimum
scale where each peak is detected is the local maximum at the lowest scale
in the ridge line. If no local maximum was found, the scale with the maximum
coefficient is chosen. After finding a peak, the extension of the peak
is found by finding the nearest local minimum at both sides of the peak,
using the wavelet coefficients with the best scale. A peak is represented
then by three indices specifying the peak location, peak start and peak end.
These three values, together with baseline and noise estimations are used
to estimate peak parameters. If the mode used is 'default`, the peak
parameters are defined as follows:
baseline :
A baseline is built using y values where no peak was detected. These
values are interpolated to build the baseline.
noise :
The noise is computed as the standard deviation of the values used
to build the baseline. To obtain a robust estimation, the median
absolute deviation of the baseline is used.
height :
The height of a peak is computed as the difference between the
y value baseline value at the peak location
snr :
The quotient between the height of the peak and the noise.
area :
Area of the peak obtained by integration between the start and
the end of the peak. The area of the baseline is subtracted.
width :
The peak width is computed as the peak extension, that is, the
difference between the end and the start of the peak.
After computing these parameters, peaks are filtered based on SNR and peak
width. Peak overlap between the filtered peaks is analyzed then. Two
peaks are overlapping if there is superposition in their peak extensions.
Overlapping peaks are flagged, their extension corrected and corrected peak
parameters are computed again.
Custom estimators can be used for noise, baseline, peak height, peak
location, peak width and peak area:
.. code-block:: python
estimators = {"baseline": baseline_func, "noise": noise_func,
"height": height_func, "loc": loc_func,
"width": width_func, "area": area_func}
# x and y are the same array used in the function
# peaks is a list of PeakLocation instances
# peak is a single PeakLocation instance
# baseline must have the same size as x and y
baseline = baseline_func(x, y, peaks)
# noise is a positive number
noise = noise_func(x, y, peaks)
# peak_parameters are all positive numbers
# (area and height can be zero)
height = height_func(x, y, peak, baseline)
area = area_func(x, y, peak, baseline)
width = width_func(x, y, peak, baseline)
loc = loc_func(x, y, peak, baseline)
References
----------
.. [DP06] Pan Du, Warren A. Kibbe, Simon M. Lin, Improved peak detection in
mass spectrum by incorporating continuous wavelet transform-based
pattern matching, Bioinformatics, Volume 22, Issue 17, 1 September 2006,
Pages 2059–2065, https://doi.org/10.1093/bioinformatics/btl355
"""
# Convert to uniform sampling
xu, yu = _resample_data(x, y)
# convert parameters to number of points
widths, max_distance = \
_convert_to_points(xu, widths, max_distance)
# detect peaks in the ridge lines
w = cwt(yu, ricker, widths)
ridge_lines = \
_peak_finding._identify_ridge_lines(w, max_distance, gap_threshold)
# y_peaks are the local maxima of y and are used to validate peaks
# y_peaks = find_peaks(yu)[0]
y_peaks = argrelmax(yu, order=2)[0]
peaks = _process_ridge_lines(w, y_peaks, ridge_lines, min_length,
max_distance)
# baseline and noise estimation
if estimators == "default":
baseline, noise = baseline_noise_estimation(yu)
elif estimators == "cwt":
baseline, noise = None, None
else:
baseline = estimators["baseline"](xu, yu, peaks)
noise = estimators["noise"](xu, yu, peaks)
# peak filtering and parameter estimation
peaks, params = \
_estimate_params(xu, yu, widths, w, peaks, snr, min_width, max_width,
estimators, baseline=baseline, noise=noise)
# sort peaks based on location
sorted_index = sorted(range(len(peaks)), key=lambda s: peaks[s].loc)
peaks = [peaks[k] for k in sorted_index]
params = [params[k] for k in sorted_index]
# find and correct overlap between consecutive peaks:
overlap_index = list()
rm_index = list()
for k in range(len(peaks) - 1):
left, right = peaks[k], peaks[k + 1]
is_same_peak = right.loc == left.loc
merge = (right.loc - left.loc) <= max_distance[0]
has_overlap = left.end > right.start
if is_same_peak:
rm_index.append(k + (left.scale < right.scale))
elif merge:
rm_index.append(k)
right.start = left.start
right.loc = (left.loc + right.loc) // 2
elif has_overlap:
_fix_peak_extension(left, right, yu)
overlap_index.extend([k, k + 1])
# remove invalid peaks after the extension was fixed
if yu[left.loc] < max(yu[left.start], yu[left.end]):
rm_index.append(k)
overlap_peaks = [peaks[x] for x in overlap_index]
# if there are peaks with overlap, then compute again peak parameters after
# correction
if overlap_index:
_, overlap_params = \
_estimate_params(xu, yu, widths, w, overlap_peaks, snr, min_width,
max_width, estimators, baseline=baseline,
noise=noise, append_empty_params=True)
# replace corrected values in params:
for k, param in zip(overlap_index, overlap_params):
if len(param):
params[k] = param
else:
rm_index.append(k)
# remove invalid peaks and back scale peaks
peaks = [p.rescale(xu, x) for (k, p) in enumerate(peaks)
if k not in rm_index]
params = [p for k, p in enumerate(params) if (len(p) and k not in rm_index)]
return peaks, params
