forked from jthickstun/pytorch_musicnet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
musicnet.py
228 lines (191 loc) · 8.94 KB
/
musicnet.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
from __future__ import print_function
from subprocess import call
import torch.utils.data as data
import os,mmap
import os.path
import pickle
import errno
import csv
import numpy as np
import torch
from intervaltree import IntervalTree
from scipy.io import wavfile
sz_float = 4 # size of a float
epsilon = 10e-8 # fudge factor for normalization
class MusicNet(data.Dataset):
"""`MusicNet <http://homes.cs.washington.edu/~thickstn/musicnet.html>`_ Dataset.
Args:
root (string): Root directory of dataset
train (bool, optional): If True, creates dataset from ``train_data``,
otherwise from ``test_data``.
download (bool, optional): If true, downloads the dataset from the internet and
puts it in root directory. If dataset is already downloaded, it is not
downloaded again.
mmap (bool, optional): If true, mmap the dataset for faster access times.
normalize (bool, optional): If true, rescale input vectors to unit norm.
window (int, optional): Size in samples of a data point.
pitch_shift (int,optional): Integral pitch-shifting transformations.
jitter (int, optional): Continuous pitch-jitter transformations.
epoch_size (int, optional): Designated Number of samples for an "epoch"
"""
url = 'https://homes.cs.washington.edu/~thickstn/media/musicnet.tar.gz'
raw_folder = 'raw'
train_data, train_labels, train_tree = 'train_data', 'train_labels', 'train_tree.pckl'
test_data, test_labels, test_tree = 'test_data', 'test_labels', 'test_tree.pckl'
extracted_folders = [train_data,train_labels,test_data,test_labels]
def __init__(self, root, train=True, download=False, mmap=True, normalize=True, window=16384, pitch_shift=0, jitter=0., epoch_size=100000):
self.mmap = mmap
self.normalize = normalize
self.window = window
self.pitch_shift = pitch_shift
self.jitter = jitter
self.size = epoch_size
self.m = 128
self.root = os.path.expanduser(root)
if download:
self.download()
if not self._check_exists():
raise RuntimeError('Dataset not found.' +
' You can use download=True to download it')
if train:
self.data_path = os.path.join(self.root, self.train_data)
labels_path = os.path.join(self.root, self.train_labels, self.train_tree)
else:
self.data_path = os.path.join(self.root, self.test_data)
labels_path = os.path.join(self.root, self.test_labels, self.test_tree)
with open(labels_path) as f:
self.labels = pickle.load(f)
self.rec_ids = self.labels.keys()
self.records = dict()
self.open_files = []
def __enter__(self):
for record in os.listdir(self.data_path):
if not record.endswith('.npy'): continue
if self.mmap:
fd = os.open(os.path.join(self.data_path, record), os.O_RDONLY)
buff = mmap.mmap(fd, 0, mmap.MAP_SHARED, mmap.PROT_READ)
self.records[int(record[:-4])] = (buff, len(buff)/sz_float)
self.open_files.append(fd)
else:
f = open(os.path.join(self.data_path, record))
self.records[int(record[:-4])] = (os.path.join(self.data_path, record),os.fstat(f.fileno()).st_size/sz_float)
f.close()
def __exit__(self, *args):
if self.mmap:
for mm in self.records.values():
mm[0].close()
for fd in self.open_files:
os.close(fd)
self.records = dict()
self.open_files = []
def access(self,rec_id,s,shift=0,jitter=0):
"""
Args:
rec_id (int): MusicNet id of the requested recording
s (int): Position of the requested data point
shift (int, optional): Integral pitch-shift data transformation
jitter (float, optional): Continuous pitch-jitter data transformation
Returns:
tuple: (audio, target) where target is a binary vector indicating notes on at the center of the audio.
"""
scale = 2.**((shift+jitter)/12.)
if self.mmap:
x = np.frombuffer(self.records[rec_id][0][s*sz_float:int(s+scale*self.window)*sz_float], dtype=np.float32).copy()
else:
fid,_ = self.records[rec_id]
with open(fid, 'rb') as f:
f.seek(s*sz_float, os.SEEK_SET)
x = np.fromfile(f, dtype=np.float32, count=int(scale*self.window))
if self.normalize: x /= np.linalg.norm(x) + epsilon
xp = np.arange(self.window,dtype=np.float32)
x = np.interp(scale*xp,np.arange(len(x),dtype=np.float32),x).astype(np.float32)
y = np.zeros(self.m,dtype=np.float32)
for label in self.labels[rec_id][s+scale*self.window/2]:
y[label.data[1]+shift] = 1
return x,y
def __getitem__(self, index):
"""
Args:
index (int): (ignored by this dataset; a random data point is returned)
Returns:
tuple: (audio, target) where target is a binary vector indicating notes on at the center of the audio.
"""
shift = 0
if self.pitch_shift> 0:
shift = np.random.randint(-self.pitch_shift,self.pitch_shift)
jitter = 0.
if self.jitter > 0:
jitter = np.random.uniform(-self.jitter,self.jitter)
rec_id = self.rec_ids[np.random.randint(0,len(self.rec_ids))]
s = np.random.randint(0,self.records[rec_id][1]-(2.**((shift+jitter)/12.))*self.window)
return self.access(rec_id,s,shift,jitter)
def __len__(self):
return self.size
def _check_exists(self):
return os.path.exists(os.path.join(self.root, self.train_data)) and \
os.path.exists(os.path.join(self.root, self.test_data)) and \
os.path.exists(os.path.join(self.root, self.train_labels, self.train_tree)) and \
os.path.exists(os.path.join(self.root, self.test_labels, self.test_tree))
def download(self):
"""Download the MusicNet data if it doesn't exist in ``raw_folder`` already."""
from six.moves import urllib
import gzip
if self._check_exists():
return
# download files
try:
os.makedirs(os.path.join(self.root, self.raw_folder))
except OSError as e:
if e.errno == errno.EEXIST:
pass
else:
raise
filename = self.url.rpartition('/')[2]
file_path = os.path.join(self.root, self.raw_folder, filename)
if not os.path.exists(file_path):
print('Downloading ' + self.url)
data = urllib.request.urlopen(self.url)
with open(file_path, 'wb') as f:
f.write(data.read())
if not all(map(lambda f: os.path.exists(os.path.join(self.root, f)), self.extracted_folders)):
print('Extracting ' + filename)
if call(["tar", "-xvvf", file_path]) != 0:
raise OSError("Failed tarball extraction")
# process and save as torch files
print('Processing...')
self.process_data(self.test_data)
trees = self.process_labels(self.test_labels)
with open(os.path.join(self.root, self.test_labels, self.test_tree), 'wb') as f:
pickle.dump(trees, f)
self.process_data(self.train_data)
trees = self.process_labels(self.train_labels)
with open(os.path.join(self.root, self.train_labels, self.train_tree), 'wb') as f:
pickle.dump(trees, f)
print('Download Complete')
# write out wavfiles as arrays for direct mmap access
def process_data(self, path):
for item in os.listdir(os.path.join(self.root,path)):
if not item.endswith('.wav'): continue
uid = int(item[:-4])
_, data = wavfile.read(os.path.join(self.root,path,item))
np.save(os.path.join(self.root,path,item[:-4]),data)
# wite out labels in intervaltrees for fast access
def process_labels(self, path):
trees = dict()
for item in os.listdir(os.path.join(self.root,path)):
if not item.endswith('.csv'): continue
uid = int(item[:-4])
tree = IntervalTree()
with open(os.path.join(self.root,path,item), 'rb') as f:
reader = csv.DictReader(f, delimiter=',')
for label in reader:
start_time = int(label['start_time'])
end_time = int(label['end_time'])
instrument = int(label['instrument'])
note = int(label['note'])
start_beat = float(label['start_beat'])
end_beat = float(label['end_beat'])
note_value = label['note_value']
tree[start_time:end_time] = (instrument,note,start_beat,end_beat,note_value)
trees[uid] = tree
return trees