mi@0: """
mi@0: 2D-FMC for segmentation, published here:
mi@0: 
mi@0: Nieto, O., & Bello, J. P. (2014). Music Segment Similarity Using 2D-Fourier
mi@0: Magnitude Coefficients. In Proc. of the 39th IEEE International Conference on
mi@0: Acoustics Speech and Signal Processing (pp. 664-668). Florence, Italy.
mi@0: """
mi@0: 
mi@0: __author__ = "Oriol Nieto"
mi@0: __copyright__ = "Copyright 2014, Music and Audio Research Lab (MARL)"
mi@0: __license__ = "GPL"
mi@0: __version__ = "1.0"
mi@0: __email__ = "oriol@nyu.edu"
mi@0: 
mi@0: import numpy as np
mi@0: import scipy.cluster.vq as vq
mi@0: import logging
mi@0: 
mi@0: # Local stuff
mi@0: from utils import utils_2dfmc as utils2d
mi@0: from utils.xmeans import XMeans
mi@0: 
mi@0: 
mi@0: MIN_LEN = 4     # Minimum lenght for the segments
mi@0: 
mi@0: 
mi@0: def get_pcp_segments(PCP, bound_idxs):
mi@0:     """Returns a set of segments defined by the bound_idxs."""
mi@0:     pcp_segments = []
mi@0:     for i in xrange(len(bound_idxs)-1):
mi@0:         pcp_segments.append(PCP[bound_idxs[i]:bound_idxs[i+1], :])
mi@0:     return pcp_segments
mi@0: 
mi@0: 
mi@0: def pcp_segments_to_2dfmc_fixed(pcp_segments, N=75):
mi@0:     """From a list of PCP segments, return a list of 2D-Fourier Magnitude
mi@0:         Coefs using a fixed segment size (N) and aggregating."""
mi@0: 
mi@0:     fmcs = []
mi@0:     for pcp_segment in pcp_segments:
mi@0:         X = []
mi@0: 
mi@0:         # Append so that we never lose a segment
mi@0:         skip = False
mi@0:         while pcp_segment.shape[0] < MIN_LEN:
mi@0:             try:
mi@0:                 pcp_segment = np.vstack((pcp_segment,
mi@0:                                          pcp_segment[-1][np.newaxis, :]))
mi@0:             except:
mi@0:                 logging.warning("Error: Can't stack PCP arrays, "
mi@0:                                 "skipping segment")
mi@0:                 skip = True
mi@0:                 break
mi@0: 
mi@0:         if skip:
mi@0:             continue
mi@0: 
mi@0:         curr_len = pcp_segment.shape[0]
mi@0: 
mi@0:         if curr_len > N:
mi@0:             # Sub segment if greater than minimum
mi@0:             for i in xrange(curr_len - N + 1):
mi@0:                 sub_segment = pcp_segment[i:i + N]
mi@0:                 X.append(utils2d.compute_ffmc2d(sub_segment))
mi@0: 
mi@0:             # Aggregate
mi@0:             X = np.max(np.asarray(X), axis=0)
mi@0: 
mi@0:             fmcs.append(X)
mi@0: 
mi@0:         elif curr_len <= N:
mi@0:             # Zero-pad
mi@0:             X = np.zeros((N, pcp_segment.shape[1]))
mi@0:             X[:curr_len, :] = pcp_segment
mi@0: 
mi@0:             # 2D-FMC
mi@0:             fmcs.append(utils2d.compute_ffmc2d(X))
mi@0: 
mi@0:     return np.asarray(fmcs)
mi@0: 
mi@0: 
mi@0: def compute_labels_kmeans(fmcs, k=6):
mi@0:     # Removing the higher frequencies seem to yield better results
mi@0:     fmcs = fmcs[:, fmcs.shape[1]/2:]
mi@0: 
mi@0:     fmcs = np.log1p(fmcs)
mi@0:     wfmcs = vq.whiten(fmcs)
mi@0: 
mi@0:     dic, dist = vq.kmeans(wfmcs, k, iter=100)
mi@0:     labels, dist = vq.vq(wfmcs, dic)
mi@0: 
mi@0:     return labels
mi@0: 
mi@0: 
mi@0: def compute_similarity(PCP, bound_idxs, xmeans=False, k=5, N=32):
mi@0:     """Main function to compute the segment similarity of file file_struct."""
mi@0: 
mi@0:     # Get PCP segments
mi@0:     pcp_segments = get_pcp_segments(PCP, bound_idxs)
mi@0: 
mi@0:     # Get the 2d-FMCs segments
mi@0:     fmcs = pcp_segments_to_2dfmc_fixed(pcp_segments, N=N)
mi@0:     if fmcs == [] or fmcs is None:
mi@0:         return np.arange(len(bound_idxs) - 1)
mi@0: 
mi@0:     # Compute the labels using kmeans
mi@0:     if xmeans:
mi@0:         xm = XMeans(fmcs, plot=False)
mi@0:         k = xm.estimate_K_knee(th=0.01, maxK=8)
mi@0:     est_labels = compute_labels_kmeans(fmcs, k=k)
mi@0: 
mi@0:     # Plot results
mi@0:     #plot_pcp_wgt(PCP, bound_idxs)
mi@0: 
mi@0:     return est_labels