make_sequence.py

# coding: utf-8

import numpy as np
from numpy.lib.stride_tricks import as_strided

def make_sequence(y, frame_length, hop_length=1):
    '''Slice a time series into overlapping frames.
    
    code derived from:
    https://librosa.github.io/librosa/_modules/librosa/util/utils.html

    This implementation uses low-level stride manipulation to avoid
    redundant copies of the time series data.

    code test:

    # code derived from https://librosa.github.io/librosa/_modules/librosa/util/utils.html

    from numpy.lib.stride_tricks import as_strided

    frame_length = 360
    hop_length = 1
    n_frames = 1 + int((len(y) - frame_length) / hop_length)
    n_vars = y.shape[1]

    yy = as_strided(y, shape=(n_frames,frame_length,n_vars),
                   strides=(y.itemsize*n_vars,y.itemsize*n_vars,y.itemsize*hop_length))

    print(yy[0,1,0])
    print(yy[1,0,0])
    assert np.all(yy[1:,:-1,:] == yy[:-1,1:,:])

    '''

    assert isinstance(y, np.ndarray),type(y)

    assert len(y) >= frame_length,len(y)

    assert hop_length >= 1,hop_length

    assert y.flags['C_CONTIGUOUS'],y.flags

    # Compute the number of frames that will fit. The end may get truncated.
    n_frames = 1 + int((len(y) - frame_length) / hop_length)
    n_vars = y.shape[1]

    # 1st and 2nd stride is n_vars sample
    # 3rd stride is `hop_length` samples
    y_frames = as_strided(y, shape=(n_frames,frame_length,n_vars),
               strides=(y.itemsize*n_vars,y.itemsize*n_vars,y.itemsize*hop_length))

    return y_frames