view SegEval.py @ 19:890cfe424f4a tip

added annotations

#!/usr/bin/env python
# encoding: utf-8
"""
SegEval.py

The main segmentation program.

Created by mi tian on 2015-04-02.
Copyright (c) 2015 __MyCompanyName__. All rights reserved.
"""

# Load starndard python libs
import sys, os, optparse, csv
from itertools import combinations
from os.path import join, isdir, isfile, abspath, dirname, basename, split, splitext
from copy import copy

import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
import numpy as np
import scipy as sp
from scipy.signal import correlate2d, convolve2d, filtfilt, resample
from sklearn.decomposition import PCA
from sklearn.mixture import GMM
from sklearn.cluster import KMeans
from sklearn.preprocessing import normalize
from sklearn.metrics.pairwise import pairwise_distances

# Load dependencies
from utils.SegUtil import getMean, getStd, getDelta, getSSM, reduceSSM, upSample, normaliseFeature, normaliseArray
from utils.PeakPickerUtil import PeakPicker
from utils.gmmdist import *
from utils.GmmMetrics import GmmDistance
from utils.RankClustering import rClustering
from utils.kmeans import Kmeans
# from utils.PathTracker import PathTracker
from utils.OnsetPlotProc import onset_plot, plot_on

# Load bourdary retrieval utilities
import cnmf as cnmf_S
import foote as foote_S
import sf as sf_S
import fmc2d as fmc2d_S
import novelty as novelty_S

# Algorithm params
# cnmf
h = 8               # Size of median filter for features in C-NMF
R = 12              # Size of the median filter for the activation matrix C-NMF
rank = 4            # Rank of decomposition for the boundaries
rank_labels = 16     # Rank of decomposition for the labels
R_labels = 4        # 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():
	op = optparse.OptionParser()
	# IO options
	op.add_option('-g', '--gammatonegram-features', action="store", dest="GF", default='/Volumes/c4dm-03/people/mit/features/gammatonegram/qupujicheng/2048', type="str", help="Loading gammatone features from.." )
	op.add_option('-s', '--spectrogram-features', action="store", dest="SF", default='/Volumes/c4dm-03/people/mit/features/spectrogram/qupujicheng/2048', type="str", help="Loading spectral features from.." )
	op.add_option('-t', '--tempogram-features', action="store", dest="TF", default='/Volumes/c4dm-03/people/mit/features/tempogram/qupujicheng/tempo_features_6s', type="str", help="Loading tempogram features from.." )
	op.add_option('-a', '--annotations', action="store", dest="GT", default='/Volumes/c4dm-03/people/mit/annotation/qupujicheng/lowercase', type="str", help="Loading annotation files from.. ")
	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 ")
	op.add_option('-d', '--dataset', action="store", dest="DATASET", default='qupujicheng', type="str", help="Specify datasets")

	# parameterization options
	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')." )
	op.add_option('-x', '--experiment', action="store", dest="EXPERIMENT", type='choice', choices=['all',  'individual', 'fuse_feature', 'fuse_ssm', 'fuse_novelty', 'fuse_bounds'], default='all', help="Specify experiment to execute." )
	
	# Plot/print/mode options
	op.add_option('-p', '--plot', action="store_true", dest="PLOT", default=False, help="Save plots")
	op.add_option('-e', '--test-mode', action="store_true", dest="TEST", default=False, help="Test mode")
	op.add_option('-v', '--verbose-mode', action="store_true", dest="VERBOSE", default=False, help="Print results in verbose mode.")
	
	return op.parse_args()
options, args = parse_args()

class FeatureObj() :
	__slots__ = ['key', 'audio', 'timestamps', 'gammatone_features', 'tempo_features', 'timbre_features', 'harmonic_features', 'gammatone_ssm', 'tempo_ssm', 'timbre_features', 'harmonic_ssm', 'ssm_timestamps']

class AudioObj():
	__slots__ = ['name', 'feature_list', 'gt', 'label', 'gammatone_features', 'tempo_features', 'timbre_features', 'harmonic_features', 'combined_features',\
	'gammatone_ssm', 'tempo_ssm', 'timbre_ssm', 'harmonic_ssm', 'combined_ssm', 'ssm', 'ssm_timestamps', 'timestamps']

class EvalObj():
	__slots__ = ['TP', 'FP', 'FN', 'P', 'R', 'F', 'AD', 'DA', 'detection']

	
class SSMseg(object):
	'''The main segmentation object'''
	def __init__(self):
		self.SampleRate = 44100.0
		self.NqHz = self.SampleRate/2
		self.previousSample = 0.0
		self.featureWindow = 6.0
		self.featureStep = 3.0
		self.kernel_size = 100 # Adjust this param according to the feature resolution.pq
		self.blockSize = 2048.0
		self.stepSize = 1024.0

		'''NOTE: Match the following params with those used for feature extraction!'''
		
		'''NOTE: Unlike spectrogram ones, Gammatone features are extracted without taking an FFT. The windowing is done under the purpose of chunking
		the audio to facilitate the gammatone filtering with the specified blockSize and stepSize. The resulting gammatonegram is aggregated every
		gammatoneLen without overlap.'''
		self.gammatoneLen = 2048
		self.gammatoneBandGroups = [0, 2, 6, 10, 13, 17, 20]
		self.nGammatoneBands = 20
		self.lowFreq = 100
		self.highFreq = self.SampleRate / 4
		
		'''Settings for extracting tempogram features.'''
		self.tempoWindow = 6.0
		self.bpmBands = [30, 45, 60, 80, 100, 120, 180, 240, 400, 600]
		
		'''Peak picking settings for novelty based method'''
		self.threshold = 10
		self.confidence_threshold = 0.4
		self.delta_threshold = 0.0
		self.backtracking_threshold = 1.9
		self.polyfitting_on = True
		self.medfilter_on = True
		self.LPfilter_on = True
		self.whitening_on = True
		self.aCoeffs = [1.0000, -0.5949, 0.2348]
		self.bCoeffs = [0.1600,	 0.3200, 0.1600]
		self.cutoff = 0.34
		self.medianWin = 10
		self.lin = 0.5		
		if plot_on : onset_plot.reset()
			
	def pairwiseF(self, annotation, detection, tolerance=3.0, combine=1.0, idx2time=None):
		'''Pairwise F measure evaluation of detection rates.'''
		
		res = EvalObj()
		res.TP, res.FP, res.FN = 0, 0, 0	
		res.P, res.R, res.F = 0.0, 0.0, 0.0
		res.AD, res.DA = 0.0, 0.0
		
		if len(detection) == 0:
			return res

		if idx2time != None:
			# Map detected idxs to real time
			detection.sort()
			if detection[-1] >= len(idx2time): 
				detection = detection[:-len(np.array(detection)[np.array(detection)-len(idx2time)>=0])]
			detection = [idx2time[int(i)] for i in detection]
		detection = np.append(detection, annotation[-1])
		res.detection = detection
		
		gt = len(annotation)	# Total number of ground truth data points
		dt = len(detection) # Total number of experimental data points
		foundIdx = []	
		D_AD = np.zeros(gt)
		D_DA = np.zeros(dt)
		
		for dtIdx in xrange(dt):
			D_DA[dtIdx] = np.min(abs(detection[dtIdx] - annotation))
		
		for gtIdx in xrange(gt):
			D_AD[gtIdx] = np.min(abs(annotation[gtIdx] - detection))
			for dtIdx in xrange(dt):
				if (annotation[gtIdx] >= detection[dtIdx] - tolerance/2.0) and (annotation[gtIdx] <= detection[dtIdx] + tolerance/2.0):
					foundIdx.append(gtIdx)
					continue
		foundIdx = list(set(foundIdx))		
		res.TP = len(foundIdx)
		# res.FP = dt - res.TP
		res.FP = max(0, dt - res.TP)	
		res.FN = gt - res.TP

		res.AD = np.mean(D_AD)		
		res.DA = np.mean(D_DA)	
		
		if res.TP == 0:
			return res

		res.P = res.TP / float(res.TP+res.FP)
		res.R = res.TP / float(res.TP+res.FN)
		res.F = 2 * res.P * res.R / (res.P + res.R)
		return res
		
	def writeIndividualHeader(self, filename):
		'''Write header of output files for individual features.'''
		
		with open(filename, 'a') as f:
			csvwriter = csv.writer(f, delimiter=',')
			csvwriter.writerow(['audio', 'harmonic_tp_05', 'harmonic_fp_05', 'harmonic_fn_05', 'harmonic_P_05', \
			'harmonic_R_05', 'harmonic_F_05', 'harmonic_AD_05', 'harmonic_DA_05', 'harmonic_tp_3', 'harmonic_fp_3', 'harmonic_fn_3', 'harmonic_P_3', 'harmonic_R_3', 'harmonic_F_3', 'harmonic_AD_3', 'harmonic_DA_3', \
			'timbre_tp_05', 'timbre_fp_05', 'timbre_fn_05', 'timbre_P_05', 'timbre_R_05', 'timbre_F_05', 'timbre_AD_05', 'timbre_DA_05', 'timbre_tp_3', 'timbre_fp_3', 'timbre_fn_3', 'timbre_P_3', 'timbre_R_3', \
			'timbre_F_3', 'timbre_AD_3', 'timbre_DA_3',	 'tempo_tp_05', 'tempo_fp_05', 'tempo_fn_05', 'tempo_P_05', 'tempo_R_05', 'tempo_F_05', 'tempo_AD_05', 'tempo_DA_05', \
			'tempo_tp_3', 'tempo_fp_3', 'tempo_fn_3', 'tempo_P_3', 'tempo_R_3', 'tempo_F_3', 'tempo_AD_3', 'tempo_DA_3'])
	
	def writeIndividualRes(self, filename, ao_name, harmonic_res_05, harmonic_res_3, timbre_res_05, timbre_res_3, tempo_res_05, tempo_res_3):
		'''Write result of single detection for individual features.'''
		
		with open(filename, 'a') as f:
			csvwriter = csv.writer(f, delimiter=',')
			csvwriter.writerow([ao_name, harmonic_res_05.TP, harmonic_res_05.FP, harmonic_res_05.FN, harmonic_res_05.P, harmonic_res_05.R, harmonic_res_05.F, harmonic_res_05.AD, harmonic_res_05.DA, \
			harmonic_res_3.TP, harmonic_res_3.FP, harmonic_res_3.FN, harmonic_res_3.P, harmonic_res_3.R, harmonic_res_3.F, harmonic_res_3.AD, harmonic_res_3.DA, timbre_res_05.TP, timbre_res_05.FP, \
			timbre_res_05.FN, timbre_res_05.P, timbre_res_05.R, timbre_res_05.F, timbre_res_05.AD, timbre_res_05.DA, timbre_res_3.TP, timbre_res_3.FP, timbre_res_3.FN, timbre_res_3.P, timbre_res_3.R, timbre_res_3.F, \
			timbre_res_3.AD, timbre_res_3.DA, tempo_res_05.TP, tempo_res_05.FP, tempo_res_05.FN, tempo_res_05.P, tempo_res_05.R, tempo_res_05.F, tempo_res_05.AD, tempo_res_05.DA, tempo_res_3.TP, tempo_res_3.FP, \
			tempo_res_3.FN, tempo_res_3.P, tempo_res_3.R, tempo_res_3.F, tempo_res_3.AD, tempo_res_3.DA])
		
	def writeCombinedHeader(self, filename):
		'''Write header of output files for combined features.'''
		
		with open(filename, 'a') as f:
			csvwriter = csv.writer(f, delimiter=',')
			csvwriter.writerow(['audio', 'hm_tb_P_0.5', 'hm_tb_R_0.5', 'hm_tb_F_0.5', 'hm_tb_P_3', 'hm_tb_R_3', 'hm_tb_F_3', 'hm_tp_P_0.5', 'hm_tp_R_0.5',\
			'hm_tp_F_0.5', 'hm_tp_P_3', 'hm_tp_R_3', 'hm_tp_F_3', 'tb_tp_P_0.5', 'tb_tp_R_0.5', 'tb_tp_F_0.5', 'tb_tp_P_3', 'tb_tp_R_3', 'tb_tp_F_3',\
			'hm_tb_tp_P_0.5', 'hm_tb_tp_R_0.5', 'hm_tb_tp_F_0.5', 'hm_tb_tp_P_3', 'hm_tb_tp_R_3', 'hm_tb_tp_F_3'])
		
	def writeCombinedRes(self, filename, ao_name, hm_tb_res_05, hm_tb_res_3, hm_tp_res_05, hm_tp_res_3, tb_tp_res_05, tb_tp_res_3, hm_tb_tp_res_05, hm_tb_tp_res_3):
		'''Write result of single detection for combined features.'''
		
		with open(filename, 'a') as f:
			csvwriter = csv.writer(f, delimiter=',')
			csvwriter.writerow([ao_name, hm_tb_res_05.P, hm_tb_res_05.R, hm_tb_res_05.F, hm_tb_res_3.P, hm_tb_res_3.R, hm_tb_res_3.F,\
			hm_tp_res_05.P, hm_tp_res_05.R, hm_tp_res_05.F, hm_tp_res_3.P, hm_tp_res_3.R, hm_tp_res_3.F, tb_tp_res_05.P, tb_tp_res_05.R, tb_tp_res_05.F, tb_tp_res_3.P, tb_tp_res_3.R, tb_tp_res_3.F, \
			hm_tb_tp_res_05.P, hm_tb_tp_res_05.R, hm_tb_tp_res_05.F, hm_tb_tp_res_3.P, hm_tb_tp_res_3.R, hm_tb_tp_res_3.F])
	
	def removeDuplicates(self, bounds, tol=1.0):		
		'''Remove duplicates by averaging boundaries located in the tolerance window.'''
		new_bounds = []
		bounds = list(set(bounds))
		bounds.sort()
		tol_win = int(tol * self.SampleRate / self.stepSize)
		
		bound_idx = 0
		nBounds = len(bounds)
		
		while bound_idx < nBounds:
			start = bounds[bound_idx]
			cnt = 1
			temp = [start]
			while (bound_idx+cnt < nBounds and (bounds[bound_idx+cnt] - start <= tol_win)):
				temp.append(bounds[bound_idx+cnt])
				cnt += 1
			
			new_bounds.append(int(np.mean(temp)))
			bound_idx += cnt
		
		# print 'new_bounds', nBounds, len(new_bounds)
		return new_bounds
	
	def selectBounds(self, nc, bounds, thresh=0.5):	
		'''Select bounds with nc value above thresh.'''
		# return list(np.array(bounds)[np.where(np.array(nc)>thresh)[0]])	
		bounds_keep = []
		nc = normaliseArray(nc)
		for i, x in enumerate(bounds):
			if nc[x] >= thresh:
				bounds_keep.append(x)
		# print 'bounds_keep', len(bounds), len(bounds_keep)		
		return bounds_keep
			
	def process(self):
		'''For the aggregated input features, discard a propertion each time as the pairwise distances within the feature space descending. 
		In the meanwhile evaluate the segmentation result and track the trend of perfomance changing by measuring the feature selection 
		threshold - segmentation f measure curve. 
		'''
		
		peak_picker = PeakPicker()
		peak_picker.params.alpha = 9.0 # Alpha norm
		peak_picker.params.delta = self.delta_threshold # Adaptive thresholding delta
		peak_picker.params.QuadThresh_a = (100 - self.threshold) / 1000.0
		peak_picker.params.QuadThresh_b = 0.0
		peak_picker.params.QuadThresh_c = (100 - self.threshold) / 1500.0
		peak_picker.params.rawSensitivity = 20
		peak_picker.params.aCoeffs = self.aCoeffs 
		peak_picker.params.bCoeffs = self.bCoeffs
		peak_picker.params.preWin = self.medianWin
		peak_picker.params.postWin = self.medianWin + 1
		peak_picker.params.LP_on = self.LPfilter_on
		peak_picker.params.Medfilt_on = self.medfilter_on
		peak_picker.params.Polyfit_on = self.polyfitting_on		
		peak_picker.params.isMedianPositive = False

		# Settings used for feature extraction
		feature_window_frame = int(self.SampleRate / self.gammatoneLen * self.featureWindow)
		feature_step_frame = int(0.5 * self.SampleRate / self.gammatoneLen * self.featureStep)
		aggregation_window, aggregation_step = 100, 50
		featureRate = float(self.SampleRate) / self.stepSize
		
		audio_files = [x for x in os.listdir(options.GT) if not x.startswith(".") ]
		if options.TEST:
			audio_files = audio_files[:1]
		audio_files.sort()
		audio_list = []

		gammatone_feature_list = [i for i in os.listdir(options.GF) if not i.startswith('.')]
		gammatone_feature_list = ['contrast6', 'rolloff4', 'dct']
		tempo_feature_list = [i for i in os.listdir(options.TF) if not i.startswith('.')]
		tempo_feature_list = ['ti', 'tir']
		timbre_feature_list = ['mfcc_harmonic']
		harmonic_feature_list = ['chromagram_harmonic']

		gammatone_feature_list = [join(options.GF, f) for f in gammatone_feature_list]
		timbre_feature_list = [join(options.SF, f) for f in timbre_feature_list]
		tempo_feature_list = [join(options.TF, f) for f in tempo_feature_list]
		harmonic_feature_list = [join(options.SF, f) for f in harmonic_feature_list]
		
		
		# Prepare output files.
		outfile1 = join(options.OUTPUT, 'individual_novelty.csv')
		outfile2 = join(options.OUTPUT, 'individual_cnmf.csv')
		outfile3 = join(options.OUTPUT, 'individual_sf.csv')
		
		outfile4 = join(options.OUTPUT, 'combinedFeatures_novelty.csv')
		outfile5 = join(options.OUTPUT, 'combinedFeatures_cnmf.csv')
		outfile6 = join(options.OUTPUT, 'combinedFeatures_sf.csv')

		# outfile7 = join(options.OUTPUT, 'combinedSSM_novelty.csv')
		# 
		# outfile8 = join(options.OUTPUT, 'combinedBounds_novelty.csv')
		# outfile9 = join(options.OUTPUT, 'combinedBounds_sf.csv')
		
		# self.writeIndividualHeader(outfile1)
		# self.writeIndividualHeader(outfile2)
		# self.writeIndividualHeader(outfile3)
		# 
		self.writeCombinedHeader(outfile4)			
		self.writeCombinedHeader(outfile5)
		self.writeCombinedHeader(outfile6)
		# self.writeCombinedHeader(outfile7)
		# self.writeCombinedHeader(outfile8)
		# self.writeCombinedHeader(outfile9)
		
		# For each audio file, load specific features
		for audio in audio_files:
			ao = AudioObj()
			ao.name = splitext(audio)[0]

			# Load annotations for specified audio collection.
			if options.DATASET == 'qupujicheng':
				annotation_file = join(options.GT, ao.name+'.csv') # qupujicheng
				ao.gt = np.genfromtxt(annotation_file, usecols=0, delimiter=',')	
				ao.label = np.genfromtxt(annotation_file, usecols=1, delimiter=',', dtype=str)
			elif options.DATASET == 'salami':	
				annotation_file = join(options.GT, ao.name+'.txt') # iso, salami
				ao.gt = np.genfromtxt(annotation_file, usecols=0)	
				ao.label = np.genfromtxt(annotation_file, usecols=1, dtype=str)
			else:
				annotation_file = join(options.GT, ao.name+'.lab') # beatles
				ao.gt = np.genfromtxt(annotation_file, usecols=(0,1))
				ao.gt = np.unique(np.ndarray.flatten(ao.gt))
				ao.label = np.genfromtxt(annotation_file, usecols=2, dtype=str)

			gammatone_featureset, timbre_featureset, tempo_featureset, harmonic_featureset = [], [], [], []
			
			for feature in timbre_feature_list:
				for f in os.listdir(feature):
					if f[:f.find('_vamp')]==ao.name: 
						timbre_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[:,0:14])
						break
			if len(timbre_feature_list) > 1:
				n_frame = np.min([x.shape[0] for x in timbre_featureset])
				timbre_featureset = [x[:n_frame,:] for x in timbre_featureset]
				ao.timbre_features = np.hstack((timbre_featureset))
			else:
				ao.timbre_features = timbre_featureset[0]
			for feature in tempo_feature_list:
				for f in os.listdir(feature):
					if f[:f.find('_vamp')]==ao.name: 
						tempo_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[1:,1:])
						ao.tempo_timestamps = np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[1:,0]
						break
			if len(tempo_feature_list) > 1:
				n_frame = np.min([x.shape[0] for x in tempo_featureset])
				tempo_featureset = [x[:n_frame,:] for x in tempo_featureset]
				ao.tempo_features = np.hstack((tempo_featureset))
			else:
				ao.tempo_features = tempo_featureset[0]
			for feature in harmonic_feature_list:
				for f in os.listdir(feature):
					if f[:f.find('_vamp')]==ao.name: 
						harmonic_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[:,1:])
						break
			if len(harmonic_feature_list) > 1:
				n_frame = np.min([x.shape[0] for x in harmonic_featureset])
				harmonic_featureset = [x[:n_frame,:] for x in harmonic_featureset]
				ao.harmonic_features = np.hstack((harmonic_featureset))
			else:
				ao.harmonic_features = harmonic_featureset[0]

			# Get aggregated features for computing ssm 
			aggregation_window, aggregation_step = 20,10
			featureRate = float(self.SampleRate) / self.stepSize
			pca = PCA(n_components=6)
			
			# Resample and normalise features
			# step = ao.tempo_features.shape[0]
			# ao.timbre_features = resample(ao.timbre_features, step)
			ao.timbre_features = normaliseFeature(ao.timbre_features)
			# ao.harmonic_features = resample(ao.harmonic_features, step)
			ao.harmonic_features = normaliseFeature(ao.harmonic_features)
			nFrames = np.min([ao.timbre_features.shapes[0], ao.harmonic_features.shapes[0]])
			ao.timbre_features = ao.timbre_features[:nFrames, :]
			ao.harmonic_features = ao.harmonic_features[:nFrames, :]
			ao.tempo_features = normaliseFeature(ao.tempo_features)
			ao.tempo_features = upSample(ao.tempo_features, nFrames)
			ao.timestamps = np.array(map(lambda x: x / featureRate, np.arange(0, nFrames)))
			
			step = nFrames / 10
			ao.timbre_features = resample(ao.timbre_features, step)
			ao.harmonic_features = resample(ao.harmonic_features, step)
			ao.tempo_features = resample(ao.tempo_features, step)
			
			pca.fit(ao.tempo_features)
			ao.tempo_features = pca.transform(ao.tempo_features)
			ao.tempo_ssm = getSSM(ao.tempo_features)
			
			pca.fit(ao.timbre_features)
			ao.timbre_features = pca.transform(ao.timbre_features)
			ao.timbre_ssm = getSSM(ao.timbre_features)

			pca.fit(ao.harmonic_features)
			ao.harmonic_features = pca.transform(ao.harmonic_features)
			ao.harmonic_ssm = getSSM(ao.harmonic_features)
			

			
			############################################################################################################################################	
			# Experiment 1: segmentation using individual features.
			
			timbre_novelty, smoothed_timbre_novelty, timbre_novelty_idxs = novelty_S.process(ao.timbre_ssm, peak_picker, self.kernel_size)
			tempo_novelty, smoothed_tempo_novelty, tempo_novelty_idxs = novelty_S.process(ao.tempo_ssm, peak_picker, self.kernel_size)
			harmonic_novelty, smoothed_harmonic_novelty, harmonic_novelty_idxs = novelty_S.process(ao.harmonic_ssm, peak_picker, self.kernel_size)
						
			timbre_cnmf_idxs = cnmf_S.segmentation(ao.timbre_features, rank=rank, R=R, h=h, niter=300)[-1]
			tempo_cnmf_idxs = cnmf_S.segmentation(ao.tempo_features, rank=rank, R=R, h=h, niter=300)[-1]
			harmonic_cnmf_idxs = cnmf_S.segmentation(ao.harmonic_features, rank=rank, R=R, h=h, niter=300)[-1]
			
			timbre_sf_nc, timbre_sf_idxs = sf_S.segmentation(ao.timbre_features)
			tempo_sf_nc, tempo_sf_idxs = sf_S.segmentation(ao.tempo_features)
			harmonic_sf_nc, harmonic_sf_idxs = sf_S.segmentation(ao.harmonic_features)
				
			# Evaluate and write results.
			harmonic_novelty_05 = self.pairwiseF(ao.gt, harmonic_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			harmonic_novelty_3 = self.pairwiseF(ao.gt, harmonic_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			tempo_novelty_05 = self.pairwiseF(ao.gt, tempo_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			tempo_novelty_3 = self.pairwiseF(ao.gt, tempo_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			timbre_novelty_05 = self.pairwiseF(ao.gt, timbre_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			timbre_novelty_3 = self.pairwiseF(ao.gt, timbre_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			
			harmonic_cnmf_05 = self.pairwiseF(ao.gt, harmonic_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			harmonic_cnmf_3 = self.pairwiseF(ao.gt, harmonic_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			tempo_cnmf_05 = self.pairwiseF(ao.gt, tempo_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			tempo_cnmf_3 = self.pairwiseF(ao.gt, tempo_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			timbre_cnmf_05 = self.pairwiseF(ao.gt, timbre_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			timbre_cnmf_3 = self.pairwiseF(ao.gt, timbre_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			
			harmonic_sf_05 = self.pairwiseF(ao.gt, harmonic_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			harmonic_sf_3 = self.pairwiseF(ao.gt, harmonic_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			tempo_sf_05 = self.pairwiseF(ao.gt, tempo_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			tempo_sf_3 = self.pairwiseF(ao.gt, tempo_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			timbre_sf_05 = self.pairwiseF(ao.gt, timbre_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			timbre_sf_3 = self.pairwiseF(ao.gt, timbre_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			
			self.writeIndividualRes(outfile1, ao.name, harmonic_novelty_05, harmonic_novelty_3, tempo_novelty_05, tempo_novelty_3, timbre_novelty_05, timbre_novelty_3)
			self.writeIndividualRes(outfile2, ao.name, harmonic_cnmf_05, harmonic_cnmf_3, tempo_cnmf_05, tempo_cnmf_3, timbre_cnmf_05, timbre_cnmf_3)
			self.writeIndividualRes(outfile3, ao.name, harmonic_sf_05, harmonic_sf_3, tempo_sf_05, tempo_sf_3, timbre_sf_05, timbre_sf_3)
			
			
			############################################################################################################################################
			# Experiment 2: segmentation using combined features.
			
			# Dumping features.			
			hm_tb = np.hstack([ao.harmonic_features, ao.timbre_features])
			hm_tp = np.hstack([ao.harmonic_features, ao.tempo_features])
			tb_tp = np.hstack([ao.timbre_features, ao.tempo_features])
			hm_tb_tp = np.hstack([ao.harmonic_features, ao.timbre_features, ao.tempo_features])
			
			hm_tb_feature_ssm = getSSM(hm_tb)
			hm_tp_feature_ssm = getSSM(hm_tp)
			tb_tp_feature_ssm = getSSM(tb_tp)
			hm_tb_tp_feature_ssm = getSSM(hm_tb_tp)
			
			# Evaluting and writing results.
			hm_tb_novelty_idxs = novelty_S.process(hm_tb_feature_ssm)[-1]
			hm_tp_novelty_idxs = novelty_S.process(hm_tp_feature_ssm)[-1]
			tb_tp_novelty_idxs = novelty_S.process(tb_tp_feature_ssm)[-1]
			hm_tb_tp_novelty_idxs = novelty_S.process(hm_tb_tp_feature_ssm)[-1]
			
			hm_tb_sf_idxs = sf_S.segmentation(hm_tb)[-1]
			hm_tp_sf_idxs = sf_S.segmentation(hm_tp)[-1]
			tb_tp_sf_idxs = sf_S.segmentation(tb_tp)[-1]
			hm_tb_tp_sf_idxs = sf_S.segmentation(hm_tb_tp)[-1]
			
			hm_tb_cnmf_idxs = cnmf_S.segmentation(hm_tb, rank=4, R=R, h=h, niter=300)[-1]
			hm_tp_cnmf_idxs = cnmf_S.segmentation(hm_tp, rank=4, R=R, h=h, niter=300)[-1]
			tb_tp_cnmf_idxs = cnmf_S.segmentation(tb_tp, rank=4, R=R, h=h, niter=300)[-1]
			hm_tb_tp_cnmf_idxs = cnmf_S.segmentation(hm_tb_tp, rank=6, R=R, h=h, niter=300)[-1]
			
			hm_tb_novelty_05 = self.pairwiseF(ao.gt, hm_tb_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tp_novelty_05 = self.pairwiseF(ao.gt, hm_tp_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			tb_tp_novelty_05 = self.pairwiseF(ao.gt, tb_tp_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tb_tp_novelty_05 = self.pairwiseF(ao.gt, hm_tb_tp_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			
			hm_tb_novelty_3 = self.pairwiseF(ao.gt, hm_tb_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tp_novelty_3 = self.pairwiseF(ao.gt, hm_tp_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			tb_tp_novelty_3 = self.pairwiseF(ao.gt, tb_tp_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tb_tp_novelty_3 = self.pairwiseF(ao.gt, hm_tb_tp_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
					
			hm_tb_sf_05 = self.pairwiseF(ao.gt, hm_tb_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tp_sf_05 = self.pairwiseF(ao.gt, hm_tp_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			tb_tp_sf_05 = self.pairwiseF(ao.gt, tb_tp_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tb_tp_sf_05 = self.pairwiseF(ao.gt, hm_tb_tp_sf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			
			hm_tb_sf_3 = self.pairwiseF(ao.gt, hm_tb_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tp_sf_3 = self.pairwiseF(ao.gt, hm_tp_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			tb_tp_sf_3 = self.pairwiseF(ao.gt, tb_tp_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tb_tp_sf_3 = self.pairwiseF(ao.gt, hm_tb_tp_sf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			
			hm_tb_cnmf_05 = self.pairwiseF(ao.gt, hm_tb_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tp_cnmf_05 = self.pairwiseF(ao.gt, hm_tp_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			tb_tp_cnmf_05 = self.pairwiseF(ao.gt, tb_tp_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tb_tp_cnmf_05 = self.pairwiseF(ao.gt, hm_tb_tp_cnmf_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			
			hm_tb_cnmf_3 = self.pairwiseF(ao.gt, hm_tb_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tp_cnmf_3 = self.pairwiseF(ao.gt, hm_tp_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			tb_tp_cnmf_3 = self.pairwiseF(ao.gt, tb_tp_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			hm_tb_tp_cnmf_3 = self.pairwiseF(ao.gt, hm_tb_tp_cnmf_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
					
			
			self.writeCombinedRes(outfile4, ao.name, hm_tb_novelty_05, hm_tb_novelty_3, hm_tp_novelty_05, hm_tp_novelty_3, tb_tp_novelty_05, tb_tp_novelty_3, hm_tb_tp_novelty_05, hm_tb_tp_novelty_3)
			self.writeCombinedRes(outfile5, ao.name, hm_tb_cnmf_05, hm_tb_cnmf_3, hm_tp_cnmf_05, hm_tp_cnmf_3, tb_tp_cnmf_05, tb_tp_cnmf_3, hm_tb_tp_cnmf_05, hm_tb_tp_cnmf_3)
			self.writeCombinedRes(outfile6, ao.name, hm_tb_sf_05, hm_tb_sf_3, hm_tp_sf_05, hm_tp_sf_3, tb_tp_sf_05, tb_tp_sf_3, hm_tb_tp_sf_05, hm_tb_tp_sf_3)
			
			
			# ############################################################################################################################################
			# # Experiment 3: late fusion -- segmentation using combined ssms.
			# 
			# hm_tb_ssm = self.lin * ao.harmonic_ssm + (1-self.lin) * ao.timbre_ssm
			# hm_tp_ssm = self.lin * ao.harmonic_ssm + (1-self.lin) * ao.tempo_ssm
			# tb_tp_ssm = self.lin * ao.timbre_ssm + (1-self.lin) * ao.tempo_ssm
			# hm_tb_tp_ssm = (ao.harmonic_ssm + ao.timbre_ssm + ao.tempo_ssm) / 3.0
			# 
			# hm_tb_ssm_novelty_idxs = novelty_S.process(hm_tb_ssm)[-1]
			# hm_tp_ssm_novelty_idxs = novelty_S.process(hm_tp_ssm)[-1]
			# tb_tp_ssm_novelty_idxs = novelty_S.process(tb_tp_ssm)[-1]
			# hm_tb_tp_ssm_novelty_idxs = novelty_S.process(hm_tb_tp_ssm)[-1]
			# 
			# hm_tb_ssm_novelty_05 = self.pairwiseF(ao.gt, hm_tb_ssm_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tp_ssm_novelty_05 = self.pairwiseF(ao.gt, hm_tp_ssm_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# tb_tp_ssm_novelty_05 = self.pairwiseF(ao.gt, tb_tp_ssm_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tb_tp_ssm_novelty_05 = self.pairwiseF(ao.gt, hm_tb_tp_ssm_novelty_idxs, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# 
			# hm_tb_ssm_novelty_3 = self.pairwiseF(ao.gt, hm_tb_ssm_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tp_ssm_novelty_3 = self.pairwiseF(ao.gt, hm_tp_ssm_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# tb_tp_ssm_novelty_3 = self.pairwiseF(ao.gt, tb_tp_ssm_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tb_tp_ssm_novelty_3 = self.pairwiseF(ao.gt, hm_tb_tp_ssm_novelty_idxs, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# 
			# self.writeCombinedRes(outfile7, ao.name, hm_tb_ssm_novelty_05, hm_tb_ssm_novelty_3, hm_tp_ssm_novelty_05, hm_tp_ssm_novelty_3, tb_tp_ssm_novelty_05, tb_tp_ssm_novelty_3, hm_tb_tp_ssm_novelty_05, hm_tb_tp_ssm_novelty_3)
			
			# ############################################################################################################################################
			# # Experiment 4: late fusion -- segmentation using combined boundaries.
			# hm_novelty_bounds = self.selectBounds(smoothed_harmonic_novelty, harmonic_novelty_idxs, self.confidence_threshold)
			# tb_novelty_bounds = self.selectBounds(smoothed_timbre_novelty, timbre_novelty_idxs, self.confidence_threshold)
			# tp_novelty_bounds = self.selectBounds(smoothed_tempo_novelty, timbre_novelty_idxs, self.confidence_threshold)
			# 
			# hm_tb_novelty_bounds = hm_novelty_bounds + tb_novelty_bounds
			# hm_tp_novelty_bounds = hm_novelty_bounds + tp_novelty_bounds
			# tb_tp_novelty_bounds = tb_novelty_bounds + tp_novelty_bounds
			# hm_tb_tp_novelty_bounds = hm_novelty_bounds + tb_novelty_bounds + tp_novelty_bounds
			# 
			# hm_tb_novelty_bounds = self.removeDuplicates(hm_tb_novelty_bounds, tol=1.0)
			# hm_tp_novelty_bounds = self.removeDuplicates(hm_tp_novelty_bounds, tol=1.0)
			# tb_tp_novelty_bounds = self.removeDuplicates(tb_tp_novelty_bounds, tol=1.0)
			# hm_tb_tp_novelty_bounds = self.removeDuplicates(hm_tb_tp_novelty_bounds, tol=1.0)
			# 
			# hm_sf_bounds = self.selectBounds(harmonic_sf_nc, harmonic_sf_idxs, self.confidence_threshold)
			# tb_sf_bounds = self.selectBounds(timbre_sf_nc, timbre_sf_idxs, self.confidence_threshold)
			# tp_sf_bounds = self.selectBounds(tempo_sf_nc, tempo_sf_idxs, self.confidence_threshold)
			# 
			# hm_tb_sf_bounds = hm_sf_bounds + tb_sf_bounds
			# hm_tp_sf_bounds = hm_sf_bounds + tp_sf_bounds
			# tb_tp_sf_bounds = tb_sf_bounds + tp_sf_bounds
			# hm_tb_tp_sf_bounds = hm_sf_bounds + tb_sf_bounds + tp_sf_bounds
			# 
			# hm_tb_sf_bounds = self.removeDuplicates(hm_tb_sf_bounds, tol=1.0)
			# hm_tp_sf_bounds = self.removeDuplicates(hm_tp_sf_bounds, tol=1.0)
			# tb_tp_sf_bounds = self.removeDuplicates(tb_tp_sf_bounds, tol=1.0)
			# hm_tb_tp_sf_bounds = self.removeDuplicates(hm_tb_tp_sf_bounds, tol=1.0)
			# 
			# 
			# hm_tb_novelty_bounds_05 = self.pairwiseF(ao.gt, hm_tb_novelty_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tp_novelty_bounds_05 = self.pairwiseF(ao.gt, hm_tp_novelty_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# tb_tp_novelty_bounds_05 = self.pairwiseF(ao.gt, tb_tp_novelty_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tb_tp_novelty_bounds_05 = self.pairwiseF(ao.gt, hm_tb_tp_novelty_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# 
			# hm_tb_sf_bounds_05 = self.pairwiseF(ao.gt, hm_tb_sf_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tp_sf_bounds_05 = self.pairwiseF(ao.gt, hm_tp_sf_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# tb_tp_sf_bounds_05 = self.pairwiseF(ao.gt, tb_tp_sf_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tb_tp_sf_bounds_05 = self.pairwiseF(ao.gt, hm_tb_tp_sf_bounds, tolerance=0.5, combine=1.0, idx2time=ao.ssm_timestamps)
			# 
			# hm_tb_novelty_bounds_3 = self.pairwiseF(ao.gt, hm_tb_novelty_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tp_novelty_bounds_3 = self.pairwiseF(ao.gt, hm_tp_novelty_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# tb_tp_novelty_bounds_3 = self.pairwiseF(ao.gt, tb_tp_novelty_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tb_tp_novelty_bounds_3 = self.pairwiseF(ao.gt, hm_tb_tp_novelty_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# 
			# hm_tb_sf_bounds_3 = self.pairwiseF(ao.gt, hm_tb_sf_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tp_sf_bounds_3 = self.pairwiseF(ao.gt, hm_tp_sf_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# tb_tp_sf_bounds_3 = self.pairwiseF(ao.gt, tb_tp_sf_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# hm_tb_tp_sf_bounds_3 = self.pairwiseF(ao.gt, hm_tb_tp_sf_bounds, tolerance=3, combine=1.0, idx2time=ao.ssm_timestamps)
			# 
			# self.writeCombinedRes(outfile8, ao.name, hm_tb_novelty_bounds_05, hm_tb_novelty_bounds_3, hm_tp_novelty_bounds_05, hm_tp_novelty_bounds_3, tb_tp_novelty_bounds_05, tb_tp_novelty_bounds_3, hm_tb_tp_novelty_bounds_05, hm_tb_tp_novelty_bounds_3)
			# self.writeCombinedRes(outfile9, ao.name, hm_tb_sf_bounds_05, hm_tb_sf_bounds_3, hm_tp_sf_bounds_05, hm_tp_sf_bounds_3, tb_tp_sf_bounds_05, tb_tp_sf_bounds_3, hm_tb_tp_sf_bounds_05, hm_tb_tp_sf_bounds_3)
			


def main():
	
	segmenter = SSMseg()
	segmenter.process()


if __name__ == '__main__':
	main()