Mercurial > hg > segmentation
changeset 1:c11ea9e0357f
adding funcs
| author | mitian |
|---|---|
| date | Thu, 02 Apr 2015 22:16:38 +0100 |
| parents | 26838b1f560f |
| children | ef1fd8b0f3c4 |
| files | SegEval.py cnmf.py foote.py novelty.py sf.py utils/SegProperties.py utils/SegUtil.py |
| diffstat | 7 files changed, 84 insertions(+), 42 deletions(-) [+] |
line wrap: on
line diff
--- a/SegEval.py Thu Apr 02 18:09:27 2015 +0100 +++ b/SegEval.py Thu Apr 02 22:16:38 2015 +0100 @@ -43,6 +43,21 @@ import foote as foote_S import sf as sf_S import fmc2d as fmc2d_S +import novelty as novelty_S + +# Algorithm params +h = 8 # Size of median filter for features in C-NMF +R = 15 # Size of the median filter for the activation matrix C-NMF +rank = 4 # Rank of decomposition for the boundaries +rank_labels = 6 # Rank of decomposition for the labels +R_labels = 6 # Size of the median filter for the labels +# Foote +M = 2 # Median filter for the audio features (in beats) +Mg = 32 # Gaussian kernel size +L = 16 # Size of the median filter for the adaptive threshold +# 2D-FMC +N = 8 # Size of the fixed length segments (for 2D-FMC) + # Define arg parser def parse_args(): @@ -56,7 +71,8 @@ op.add_option('-o', '--ouput', action="store", dest="OUTPUT", default='/Volumes/c4dm-03/people/mit/segmentation/gammatone/qupujicheng', type="str", help="Write segmentation results to ") # boundary retrieval options - op.add_option('-b', '--bounrary-method', action="store", dest="BOUNDARY", default=['novelty', 'cnmf', 'sf', 'fmc2d'], help="Choose boundary retrieval algorithm ('novelty', 'cnmf', 'sf', 'fmc2d')." ) + op.add_option('-b', '--bounrary-method', action="store", dest="BOUNDARY", type='choice', choices=['novelty', 'cnmf', 'foote', 'sf'], default='novelty', help="Choose boundary retrieval algorithm ('novelty', 'cnmf', 'sf', 'fmc2d')." ) + op.add_option('-l', '--labeling-method', action="store", dest="LABEL", type='choice', choices=['cnmf', 'fmc2d'], default='cnmf', help="Choose boundary labeling algorithm ('cnmf', 'fmc2d')." ) # Plot/print/mode options op.add_option('-p', '--plot', action="store_true", dest="PLOT", default=False, help="Save plots") @@ -305,38 +321,55 @@ print 'Segmenting using %s method' %options.BOUNDARY for i,ao in enumerate(audio_list): print 'processing: %s' %ao.name - - - + + # Experiment 1: segmentation using individual features. + if options.BOUNDARY == 'novelty': + # Peak picking from the novelty curve + gammatone_novelty, smoothed_gammatone_novelty, gammatone_bound_idxs = novelty_S.process(ao.gammatone_ssm, self.kernel_size, peak_picker) + timbre_novelty, smoothed_timbre_novelty, timbre_bound_idxs = novelty_S.process(ao.timbre_ssm, self.kernel_size, peak_picker) + tempo_novelty, smoothed_harmonic_novelty, tempo_bound_idxs = novelty_S.process(ao.tempo_ssm, self.kernel_size, peak_picker) + harmonic_novelty, smoothed_tempo_novelty, harmonic_bound_idxs = novelty_S.process(ao.harmonic_ssm, self.kernel_size, peak_picker) + + if options.BOUNDARY == 'cnmf': + gammatone_bound_idxs = cnmf_S.segmentation(ao.gammatone_features, rank=rank, R=R, h=8, niter=300) + timbre_bound_idxs = cnmf_S.segmentation(ao.timbre_features, rank=rank, R=R, h=h, niter=300) + tempo_bound_idxs = cnmf_S.segmentation(ao.tempo_features, rank=rank, R=R, h=h, niter=300) + harmonic_bound_idxs = cnmf_S.segmentation(ao.harmonic_features, rank=rank, R=R, h=h, niter=300) + + if options.BOUNDARY == 'foote': + gammatone_bound_idxs = foote_S.segmentation(ao.gammatone_features, M=M, Mg=Mg, L=L) + timbre_bound_idxs = foote_S.segmentation(ao.timbre_features, M=M, Mg=Mg, L=L) + tempo_bound_idxs = foote_S.segmentation(ao.tempo_features, M=M, Mg=Mg, L=L) + harmonic_bound_idxs = foote_S.segmentation(ao.harmonic_features, M=M, Mg=Mg, L=L) + if options.BOUNDARY == 'sf': + gammatone_bound_idxs = sf_S.segmentation(ao.gammatone_features) + timbre_bound_idxs = sf_S.segmentation(ao.timbre_features) + tempo_bound_idxs = sf_S.segmentation(ao.tempo_features) + harmonic_bound_idxs = sf_S.segmentation(ao.harmonic_features) + + if options.LABEL == 'fmc2d': + gammatone_bound_labels = fmc2d_S.compute_similarity(gammatone_bound_idxs, xmeans=True, N=N) + timbre_bound_labels = fmc2d_S.compute_similarity(timbre_bound_idxs, xmeans=True, N=N) + tempo_bound_labels = fmc2d_S.compute_similarity(tempo_bound_idxs, xmeans=True, N=N) + harmonic_bound_labels = fmc2d_S.compute_similarity(harmonic_bound_idxs, xmeans=True, N=N) + + if options.LABEL == 'cnmf': + gammatone_bound_labels = cnmf_S.compute_labels(gammatone_bound_idxs, est_bound_idxs, nFrames) + timbre_bound_labels = cnmf_S.compute_labels(timbre_bound_idxs, est_bound_idxs, nFrames) + tempo_bound_labels = cnmf_S.compute_labels(tempo_bound_idxs, est_bound_idxs, nFrames) + harmonic_bound_labels = cnmf_S.compute_labels(harmonic_bound_idxs, est_bound_idxs, nFrames) + + gammatone_detection = [0.0] + [ao.ssm_timestamps[int(np.rint(i))] for i in gammatone_novelty_peaks] + timbre_detection = [0.0] + [ao.ssm_timestamps[int(np.rint(i))] for i in timbre_novelty_peaks] + harmonic_detection = [0.0] + [ao.ssm_timestamps[int(np.rint(i))] for i in harmonic_novelty_peaks] + tempo_detection = [0.0] + [ao.ssm_timestamps[int(np.rint(i))] for i in tempo_novelty_peaks] + + # Experiment 2: Trying combined features using the best boundary retrieval method ao_featureset = [ao.gammatone_features, ao.harmonic_features, ao.timbre_features, ao.tempo_features] feature_sel = [int(x) for x in options.FEATURES if x.isdigit()] ao_featureset = [ao_featureset[i] for i in feature_sel] - gammatone_novelty, smoothed_gammatone_novelty, gammatone_novelty_peaks = getNoveltyPeaks(ao.gammatone_ssm, self.kernel_size, peak_picker) - timbre_novelty, smoothed_timbre_novelty, timbre_novelty_peaks = getNoveltyPeaks(ao.timbre_ssm, self.kernel_size, peak_picker) - tempo_novelty, smoothed_harmonic_novelty, harmonic_novelty_peaks = getNoveltyPeaks(ao.tempo_ssm, self.kernel_size, peak_picker) - harmonic_novelty, smoothed_tempo_novelty, tempo_novelty_peaks = getNoveltyPeaks(ao.harmonic_ssm, self.kernel_size, peak_picker) - - # Peak picking from the novelty curve - smoothed_gammatone_novelty, gammatone_novelty_peaks = peak_picker.process(gammatone_novelty) - gammatone_detection = [ao.ssm_timestamps[int(np.rint(i))] for i in gammatone_novelty_peaks] - smoothed_timbre_novelty, timbre_novelty_peaks = peak_picker.process(timbre_novelty) - timbre_detection = [ao.ssm_timestamps[int(np.rint(i))] for i in timbre_novelty_peaks] - smoothed_harmonic_novelty, harmonic_novelty_peaks = peak_picker.process(harmonic_novelty) - harmonic_detection = [ao.ssm_timestamps[int(np.rint(i))] for i in harmonic_novelty_peaks] - smoothed_tempo_novelty, tempo_novelty_peaks = peak_picker.process(tempo_novelty) - tempo_detection = [ao.ssm_timestamps[int(np.rint(i))] for i in tempo_novelty_peaks] - - if (len(gammatone_novelty_peaks) == 0 or len(harmonic_novelty_peaks)== 0 or len(timbre_novelty_peaks) == 0 or len(tempo_novelty_peaks) == 0): - print ao.name, len(gammatone_novelty_peaks), len(harmonic_novelty_peaks), len(timbre_novelty_peaks), len(tempo_novelty_peaks) - - smoothed_gammatone_novelty -= np.min(smoothed_gammatone_novelty) - smoothed_harmonic_novelty -= np.min(smoothed_harmonic_novelty) - smoothed_timbre_novelty -= np.min(smoothed_timbre_novelty) - smoothed_tempo_novelty -= np.min(smoothed_tempo_novelty) - combined_sdf = (np.array(smoothed_gammatone_novelty) + np.array(smoothed_harmonic_novelty) + np.array(smoothed_timbre_novelty) + np.array(smoothed_tempo_novelty)) - def main():
--- a/cnmf.py Thu Apr 02 18:09:27 2015 +0100 +++ b/cnmf.py Thu Apr 02 22:16:38 2015 +0100 @@ -115,7 +115,7 @@ """ # Filter - X = utils.median_filter(X, M=h) + X = SegUtil.median_filter(X, M=h) X = X.T # Find non filtered boundaries
--- a/foote.py Thu Apr 02 18:09:27 2015 +0100 +++ b/foote.py Thu Apr 02 22:16:38 2015 +0100 @@ -36,16 +36,16 @@ Array containing the indices of the boundaries. """ # Filter - F = utils.median_filter(F, M=M) + F = SegUtil.median_filter(F, M=M) # Self Similarity Matrix - S = utils.compute_ssm(F) + S = SegUtil.compute_ssm(F) # Compute gaussian kernel - G = utils.compute_gaussian_krnl(Mg) + G = SegUtil.compute_gaussian_krnl(Mg) # Compute the novelty curve - nc = utils.compute_nc(S, G) + nc = SegUtil.compute_nc(S, G) # Find peaks in the novelty curve - return utils.pick_peaks(nc, L=L, plot=plot) + return SegUtil.pick_peaks(nc, L=L, plot=plot)
--- a/novelty.py Thu Apr 02 18:09:27 2015 +0100 +++ b/novelty.py Thu Apr 02 22:16:38 2015 +0100 @@ -10,6 +10,7 @@ import sys, os import numpy as np from scipy.signal import correlate2d, convolve2d +import matplotlib.pyplot as plt # from utils.PeakPickerUtil import PeakPicker @@ -60,9 +61,14 @@ return kernel -def getNoveltyPeaks(ssm, kernel_size, peak_picker, normalise=False): +def process(ssm, kernel_size, peak_picker, normalise=False, plot=False): '''Detect segment boundaries in the ssm.''' novelty = getNoveltyCurve(ssm, kernel_size, normalise=False) smoothed_novelty, novelty_peaks = peak_picker.process(novelty) - return novelty, smoothed_novelty, novelty_peaks \ No newline at end of file + if plot: + plot_detection(smoothed_novelty, novelty_peaks) + return novelty, smoothed_novelty, novelty_peaks + +def plot_detection(smoothed_novelty, novelty_peaks): + pass \ No newline at end of file
--- a/sf.py Thu Apr 02 18:09:27 2015 +0100 +++ b/sf.py Thu Apr 02 22:16:38 2015 +0100 @@ -28,7 +28,6 @@ # Local stuff from utils import SegUtil - def median_filter(X, M=8): """Median filter along the first axis of the feature matrix X.""" for i in xrange(X.shape[1]): @@ -125,7 +124,7 @@ E = embedded_space(F, m) # Recurrence matrix - R = utils.recurrence_matrix(E.T, k=k * int(F.shape[0]), + R = SegUtil.recurrence_matrix(E.T, k=k * int(F.shape[0]), width=0, # zeros from the diagonal metric="seuclidean", sym=True).astype(np.float32)
--- a/utils/SegProperties.py Thu Apr 02 18:09:27 2015 +0100 +++ b/utils/SegProperties.py Thu Apr 02 22:16:38 2015 +0100 @@ -7,8 +7,11 @@ Copyright (c) 2015 __MyCompanyName__. All rights reserved. """ -import sys -import os +import sys, os +import numpy as np +from sklearn.metrics.pairwise import pairwise_distances +from gmmdist import * +from GmmMetrics import GmmDistance class FeatureGMM(object): '''Represent segment candidates using single GMMs and compute pairwise distances.''' @@ -39,7 +42,7 @@ gf=GaussianFeature(feature[int(num * stepsize) : int((num * stepsize) + win_len), :],2) # f.write("\n%s" %str(gf)) gaussian_list.append(gf) - tsi = int(floor( num * stepsize + 1)) + tsi = int(np.floor( num * stepsize + 1)) gaussian_timestamps.append(self.timestamp[tsi]) # f.close()
--- a/utils/SegUtil.py Thu Apr 02 18:09:27 2015 +0100 +++ b/utils/SegUtil.py Thu Apr 02 22:16:38 2015 +0100 @@ -14,13 +14,14 @@ import copy import numpy as np -import os +import os, sys import scipy from scipy.spatial import distance from scipy.ndimage import filters, zoom from scipy import signal import pylab as plt from scipy.spatial.distance import squareform, pdist +from scipy.ndimage.filters import * def lognormalize_chroma(C):