deep-clustering/visualization.py at master · jcsilva/deep-clustering · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
# -*- coding: utf-8 -*-
"""
Created on Thu Oct  6 15:48:14 2016

@author: valterf
"""
from feats import stft
from config import FRAME_RATE

from itertools import permutations

from sklearn.cluster import KMeans
import soundfile as sf
import numpy as np
import matplotlib.pyplot as plt


def print_examples(wavpaths, nnet, db_threshold=None,
                   source_intensities=None,
                   ignore_background=False,
                   pred_index=0):
    """
    Visualizes the difference between the neural network output and the ground
    truth from a list of files to be mixed up.

    wavpaths is a list of wavfile paths which should all have sampling rate.

    nnet is a loaded pre-trained model using the "load_model" function from
    predict.py

    db_threshold is the backgorund silence threshold, if you wish to treat
    background silence as a "different source"

    source_intensities is a list of floats which weights the source intensities
    of each file from wavpaths by multiplying its amp. It should have the same
    length of wavpaths

    ignore_background is used only if db_threshold is specified. It forces all
    background silence points to the same region, forcing the KMeans algorithm
    to treat this point as a centroid and as a cluster.

    pred_index is a parameter for multiple output networks, and should be
    ignored for now.
    """
    k = len(wavpaths)
    kk = k
    if db_threshold is not None:
        kk += 1
    freq = int(nnet.input.get_shape()[2])
    if(isinstance(nnet.output, list)):
        K = int(nnet.output[pred_index].get_shape()[2]) // freq
    else:
        K = int(nnet.output.get_shape()[2]) // freq
    sigsum = None
    specs = []
    sigs = []

    def get_logspec(sig):
        return np.log10(np.absolute(stft(sig)) + 1e-7)

    for i, wavpath in enumerate(wavpaths):
        sig, rate = sf.read(wavpath)
        if rate != FRAME_RATE:
            raise Exception("Config specifies " + str(FRAME_RATE) +
                            "Hz as sample rate, but file " + str(wavpath) +
                            "is in " + str(rate) + "Hz.")
        sig = sig - np.mean(sig)
        sig = sig/np.max(np.abs(sig))
        if source_intensities is not None:
            sig *= source_intensities[i]
        if sigsum is None:
            sigsum = sig
        else:
            sigsum = sigsum[:len(sig)] + sig[:len(sigsum)]
        sigs.append(sig)
    for sig in sigs:
        specs.append(get_logspec(sig[:len(sigsum)]))
    specs = np.transpose(np.array(specs), (1, 2, 0))
    sigsum = sigsum - np.mean(sigsum)
    sigsum = sigsum/np.max(np.abs(sigsum))
    mag = get_logspec(sigsum)
    X = mag.reshape((1,) + mag.shape)
    if(isinstance(nnet.output, list)):
        V = nnet.predict(X)[pred_index]
    else:
        V = nnet.predict(X)
    x = X.reshape((-1, freq))
    if db_threshold is not None:
        v = V.reshape((-1, freq, K))
        m = np.max(x) - db_threshold / 20.
        if ignore_background:
            v[x < m] = 0
    v = V.reshape((-1, K))
    km = KMeans(kk)
    eg = km.fit_predict(v)
    ref = np.argmax(specs.reshape((x.size, k)), axis=1)
    if db_threshold is not None:
        ref[(x < m).reshape(ref.shape)] = kk - 1

    # Permute classes for oracle alignment
    eg_p = np.zeros((x.size, kk))
    ref_p = np.zeros((x.size, kk))
    for i in range(kk):
        eg_p[eg == i, i] = 1
        ref_p[ref == i, i] = 1
    p = None
    s = np.float('Inf')
    for pp in permutations(list(range(kk))):
        ss = np.sum(np.abs(ref_p - eg_p[:, pp]))
        if ss < s:
            s = ss
            p = pp
    eg_p = eg_p[:, p]
    eg = np.argmax(eg_p, axis=1)

    imshape = x.shape + (3,)
    img = np.zeros((x.size, 3))
    for i in range(min(k, 3)):
        img[eg == i, i] = 1
    if db_threshold is not None:
        img[eg == kk - 1] = [0, 0, 0]
    img = img.reshape(imshape)

    img2 = np.zeros((x.size, 3))
    for i in range(min(k, 3)):
        img2[ref == i, i] = 1
    if db_threshold is not None:
        img2[ref == kk - 1] = [0, 0, 0]
    img2 = img2.reshape(imshape)

    img3 = x
    img3 -= np.min(img3)
    img3 **= 3

    fig, (ax1, ax2, ax3) = plt.subplots(3, 1)
    ax1.imshow(img.swapaxes(0, 1), origin='lower')
    ax2.imshow(img2.swapaxes(0, 1), origin='lower')
    ax3.imshow(img3.swapaxes(0, 1), origin='lower', cmap='afmhot')