/
feature_transforms.py
191 lines (165 loc) · 7.58 KB
/
feature_transforms.py
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
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
from __future__ import division, print_function
import numpy as np
import pandas as pd
import scipy.stats as stats
import sklearn.neighbors as nn
import utils
data_dir = ''
def compute(segment_dict, features):
"""
Args:
segment_dict (dict): dictionary of song segments, containing a list of
segment ids (values) for a set of unique song identifiers (keys).
"""
data_dict = {}
# compute features
for feature in features:
print('computing ' + feature + '...')
feature_name, first_order_aggregates, second_order_aggregates = parse_feature(feature)
corpus_features = []
for song_id in segment_dict.keys():
song_features = []
for segment in segment_dict[song_id]:
raw_features = utils.read_feature([data_dir, feature_name, segment], skip_cols='auto')
segment_features = first_order(raw_features, first_order_aggregates, verbose=False)
song_features.append(segment_features)
if 'song' in second_order_aggregates:
song_features = second_order(song_features, second_order_aggregates, verbose=False)
corpus_features.extend(song_features)
if 'corpus' in second_order_aggregates:
# print(' in: len(corpus_features) = {}, corpus_features[0] = {}'.format(len(corpus_features), corpus_features[0]))
corpus_features = second_order(corpus_features, second_order_aggregates, verbose=False)
# print(' out: len(corpus_features) = {}, corpus_features[0] = {}'.format(len(corpus_features), corpus_features[0]))
data_dict[feature] = np.squeeze(corpus_features)
# add segment ids
song_ids = []
segments = []
for song_id in segment_dict.keys():
for segment in segment_dict[song_id]:
song_ids.append(song_id)
segments.append(segment)
data_dict['song.id'] = np.array(song_ids)
data_dict['segment.id'] = np.array(segments)
# convert to dataframe
return pd.DataFrame(data_dict)
# FEATURE COMPUTATION
def parse_feature(feature):
""" Parse feature string into
(feature name, [1st order aggregates], [2nd order aggregates]).
'Grammar':
- feature name and aggregates are separated by dots, e.g. 'mfcc.entropy'
- feature name is first and contains no dots
- first order and second order aggregates are separated by one of 2 keywords:
'corpus' or 'song'
Ex.:
>>> parse_features('loudness.mean.song.pdf.log')
('loudness', ['mean'], ['song', 'pdf', 'log'])
"""
s = np.array(feature.split('.'))
split_points = (s == 'corpus') | (s == 'song')
split_points = np.nonzero(split_points)[0] if any(split_points) else [len(s)]
return s[0], s[1:split_points[0]].tolist(), s[split_points[-1]:].tolist()
def first_order(feature, aggregates, verbose=False):
if not type(aggregates) is list:
aggregates = [aggregates]
for aggregate in aggregates:
if verbose:
print(' first order computation: ' + aggregate)
if aggregate == 'log':
feature = np.log(feature)
elif aggregate == 'sqrt':
feature = np.sqrt(feature)
elif aggregate == 'minlog':
feature = np.log(1 - feature)
elif aggregate == 'minsqrt':
feature = np.sqrt(1 - feature)
elif aggregate == 'mean':
# feature = np.mean(feature, axis=0)
feature = np.nanmean(feature, axis=0)
elif aggregate == 'var':
feature = np.var(feature, axis=0)
elif aggregate == 'std':
# feature = np.std(feature, axis=0)
feature = np.nanstd(feature, axis=0)
elif aggregate == 'stdmean':
feature = np.hstack([np.mean(feature, axis=0), np.std(feature, axis=0)])
elif aggregate == 'cov':
feature = np.flatten(np.cov(feature, axis=0))
elif aggregate == 'totvar':
feature = np.array([np.mean(np.var(feature, axis=0))])
elif aggregate == 'totstd':
feature = np.array([np.mean(np.std(feature, axis=0))])
elif aggregate == 'entropy':
feature = feature.flatten()
feature = np.array([stats.entropy(feature)])
elif aggregate == 'normentropy':
feature = feature.flatten()
feature = np.array([stats.entropy(feature) / np.log(feature.size)])
elif aggregate == 'information':
feature = - np.log(feature)
return feature
def second_order(features, aggregates, verbose=False):
if not type(aggregates) is list:
aggregates = [aggregates]
features = np.asarray(features)
for aggregate in aggregates:
if verbose and not (aggregate == 'song' or aggregate == 'corpus'):
print(' second order computation ({}): {}'.format(aggregates[0], aggregate))
if aggregate == 'log':
features = np.log(features)
elif aggregate == 'sqrt':
features = np.sqrt(features)
elif aggregate == 'square':
features = np.array(features)**2
elif aggregate == 'minlog':
features = np.log(1 - np.array(features))
elif aggregate == 'minsqrt':
features = np.sqrt(1 - np.array(features))
elif aggregate == 'logit':
features = np.log(np.array(features)) - np.log(1 - np.array(features))
elif aggregate == 'kld':
m = np.sum(features, axis=0)
m /= np.sum(m)
features = [stats.entropy(f.flatten(), m.flatten()) for f in features]
elif aggregate == 'tau':
m = np.sum(features, axis=0)
m /= np.sum(m)
features = [stats.kendalltau(f.flatten(), m.flatten())[0] for f in features]
elif aggregate == 'dot':
m = np.sum(features, axis=0)
features = [np.dot(f.flatten(), m.flatten()) for f in features]
elif aggregate == 'corr':
m = np.sum(features, axis=0)
features = [np.correlate(f.flatten(), m.flatten()) for f in features]
elif aggregate == 'crossentropy' or aggregate == 'information':
m = np.sum(features, axis=0)
m = m.flatten()/np.sum(m)
features = [-np.nansum(np.log(m) * f.flatten()/np.sum(f)) for f in features]
elif aggregate == 'pdf':
n, d = features.shape
finite_rows = np.all(np.isfinite(features), axis=1)
features = features[finite_rows]
s = np.std(features)
bw_factor = n**(-1./(5))*s if d == 1 and s > 0.0 else 1.0
kde = nn.KernelDensity(bandwidth=bw_factor)
kde.fit(features)
scores = kde.score_samples(features)
features = np.zeros((n,))
features[finite_rows] = np.exp(scores)
elif aggregate == 'indeppdf':
# above for independent dimensions: fit each dim and add log scores
kde = nn.KernelDensity(bandwidth=1.0)
scores = np.zeros(len(features))
for feat_dim in features.T:
feat_dim = feat_dim.reshape([-1, 1])
kde.fit(feat_dim)
scores += kde.score_samples(feat_dim)
features = np.exp(scores)
elif aggregate == 'cdf':
f0 = np.min(features)
kde = stats.gaussian_kde(features)
features = [kde.integrate_box(f0, f) for f in features]
elif aggregate == 'rank':
features = (np.argsort(np.argsort(features)) + 0.5) * (1.0 / len(features))
# features = [np.squeeze(f) for f in features]
return features