-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathexamine_msd_latent_tsne.py
154 lines (113 loc) · 4.57 KB
/
examine_msd_latent_tsne.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
# Examine the Million Song Dataset. Running this script will visualize each Echo
# Nest Taste Profile songs' latent features with t-SNE.
import os, subprocess
import matplotlib.pyplot as plt
from matplotlib.text import Annotation
import numpy as np
import scipy
from sklearn.preprocessing import StandardScaler
from sklearn.manifold import TSNE
from sklearn.cluster import MiniBatchKMeans, AgglomerativeClustering
from krotos.msd.latent.features import LatentFeatures
from krotos.paths import ROOT_PATH
from krotos.debug import report
report("Get LatentFeatures instance...")
lf = LatentFeatures()
report("Scaling latent features")
latents = lf.Y
# latents = StandardScaler().fit_transform(latents)
samples = 5000
report("Selecting {0} random samples...".format(samples))
sample_idxs = np.random.choice(latents.shape[0], size=samples, replace=False)
latents = latents[sample_idxs, :]
report("Performing t-SNE embedding...")
distance = scipy.spatial.distance.squareform(scipy.spatial.distance.pdist(latents, metric='cosine'))
model = TSNE(n_components=2, method='barnes_hut', metric='precomputed')
embedding = model.fit_transform(distance)
report("Performing clustering...")
n_clusters = 10
db = AgglomerativeClustering(
n_clusters=n_clusters,
linkage='average',
affinity='cosine'
).fit(latents)
cluster_labels = db.labels_
cmap = plt.get_cmap('jet', n_clusters)
report("Getting norms...")
norms = np.linalg.norm(latents, axis=1)
# min_max_norms = (norms - np.min(norms)) / (np.max(norms) - np.min(norms))
report("Retrieving song labels...")
song_labels = {}
echonest = lf._echonest
unique_tracks_path = os.path.join(ROOT_PATH, 'msd/resources/unique_tracks.txt')
if not os.path.exists(unique_tracks_path): raise Exception("unique_tracks.txt not found.")
with open(unique_tracks_path, 'r') as unique_tracks:
i = 0
for line in unique_tracks:
_, song_id, artist, track = line.strip().split("<SEP>")
song_labels[song_id] = (artist + ' - ' + track)
i += 1
if (i % 5000 == 0):
report("{0:7d} song labels...".format(i), sameline=True)
report('')
sid_mismatches_path = os.path.join(ROOT_PATH, 'msd/resources/sid_mismatches.txt')
if not os.path.exists(sid_mismatches_path): raise Exception("sid_mismatches.txt not found.")
with open(sid_mismatches_path, 'r') as sid_mismatches:
i = 0
for line in sid_mismatches:
song_labels[line[8:26]] = "<bad data: mismatched song>"
i += 1
if (i % 100 == 0):
report("{0:5d} erroneous song labels noted...".format(i), sameline=True)
report('')
report("Plotting...")
fig = plt.figure()
ax = fig.add_subplot(111)
ax.scatter(embedding[:, 0], embedding[:, 1], c=cluster_labels, cmap=cmap, s=100, alpha=0.5, linewidths=0.0, picker=True)
last_annotation = None
last_latent = None
last_ind = None
def report_top(idx, n):
report("\tGlobal closest songs:")
closest = lf.closest(lf.Y[idx], n=n, ordered=True)
i = 0
for echonest_id, score, norm in closest:
i += 1
report("\t\t{0}\t{1:7.5} ({2:7.5}): {3}".format(i, score, norm, song_labels[echonest_id]))
def onpick(event):
global last_annotation
global last_latent
global last_ind
ind = event.ind[0]
idx = sample_idxs[ind]
x, y = embedding[ind]
track_id_echonest = echonest.get_track_id(idx)
label = unicode(song_labels[track_id_echonest], errors='ignore')
if last_ind == ind:
report_top(idx, 100)
return
last_ind = ind
if last_latent is not None:
s = np.dot(last_latent, latents[ind]) / (np.linalg.norm(last_latent) * np.linalg.norm(latents[ind]))
d = np.linalg.norm(last_latent - latents[ind])
report('Cosine similarity: {0: 7.5f}'.format(s))
report('L2 distance: {0: .5f}'.format(d))
last_latent = latents[ind]
report('Selected point at (x={1: 7.3f}, y={2: 7.3f}, norm={3: 7.3f}): \t{0}.'.format(label, x, y, norms[ind]))
if last_annotation is not None: last_annotation.remove()
last_annotation = Annotation(
label,
xy=(x, y),
xytext=(-20, 20),
textcoords='offset points',
ha='right',
va='bottom',
bbox=dict(boxstyle = 'round,pad=0.5', fc='white', alpha=0.6),
arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0', alpha=0.4)
)
ax.add_artist(last_annotation)
fig.canvas.draw()
fig.canvas.mpl_connect('pick_event', onpick)
report("Displaying plot")
plt.show()
report('')