-
Notifications
You must be signed in to change notification settings - Fork 19
/
feeder.py
435 lines (372 loc) · 17.2 KB
/
feeder.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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
import os
import random
import threading
import sys
import numpy as np
import skimage.io as sio
import tensorflow as tf
from definitions import *
from pyutils.iolib.audio import load_wav
class FilenameProvider(object):
def __init__(self, directory,
subset_fn=None,
num_epochs=1,
shuffle=False):
self.directory = directory
self.sample_ids = os.listdir(directory)
assert len(self.sample_ids) > 0, 'Dataset directory is empty.'
if subset_fn is not None:
assert os.path.exists(subset_fn)
subset = open(subset_fn).read().splitlines()
self.sample_ids = [y for y in self.sample_ids if y in subset]
self.num_epochs, self.epoch = num_epochs, 0
self.num_samples = len(self.sample_ids)
self.shuffle = shuffle
self.head = -1
def get_next_sample(self):
self.head = (self.head + 1) % self.num_samples
if self.head == 0:
self.epoch += 1
if self.epoch > self.num_epochs:
return None
if self.shuffle:
random.shuffle(self.sample_ids)
return self.sample_ids[self.head]
def loop_samples(self):
while True:
yid = self.get_next_sample()
if yid is None:
break
yield yid
class AudioReader(object):
def __init__(self, audio_folder, rate=None, ambi_order=1):
from scikits.audiolab import Sndfile
self.audio_folder = audio_folder
fns = os.listdir(audio_folder)
self.num_files = len(fns)
fp = Sndfile(os.path.join(self.audio_folder, fns[0]), 'r')
self.rate = float(fp.samplerate) if rate is None else rate
self.num_channels = min((fp.channels, (ambi_order+1)**2))
self.duration = self.num_files
self.num_frames = int(self.duration * rate)
def get(self, start_time, size, rotation=None):
# Check if padding is necessary
start_frame = int(start_time * self.rate)
pad_before, pad_after = 0, 0
if start_frame < 0:
pad_before = abs(start_frame)
size -= pad_before
start_time, start_frame = 0., 0
if start_frame + size > self.num_frames:
pad_after = start_frame + size - self.num_frames
size -= pad_after
# Load audio
index = range(int(start_time), min(int(np.ceil(start_time + size / float(self.rate))), self.num_files))
fns = ['{}/{:06d}.wav'.format(self.audio_folder, i) for i in index]
chunk = [load_wav(fn, self.rate)[0] for fn in fns]
chunk = np.concatenate(chunk, axis=0) if len(chunk) > 1 else chunk[0]
ss = int((start_time - int(start_time)) * self.rate)
chunk = chunk[ss:ss + size, :self.num_channels]
# Pad
if pad_before > 0:
pad = np.zeros((pad_before, self.num_channels))
chunk = np.concatenate((pad, chunk), axis=0)
if pad_after > 0:
pad = np.zeros((pad_after, self.num_channels))
chunk = np.concatenate((chunk, pad), axis=0)
# Apply rotation
if rotation is not None:
assert -np.pi <= rotation < np.pi
c = np.cos(rotation)
s = np.sin(rotation)
rot_mtx = np.array([[1, 0, 0, 0], # W' = W
[0, c, 0, s], # Y' = X sin + Y cos
[0, 0, 1, 0], # Z' = Z
[0, -s, 0, c]]) # X' = X cos - Y sin
chunk = np.dot(chunk, rot_mtx.T)
return chunk
class VideoReader(object):
def __init__(self, video_folder, rate=None, img_prep=None):
raw_rate = 10.
self.video_folder = video_folder
self.rate = rate if rate is not None else raw_rate
self.img_prep = img_prep if img_prep is not None else lambda x: x
frame_fns = [fn for fn in os.listdir(video_folder) if fn.endswith('.jpg')]
self.num_frames = len(frame_fns)
self.duration = self.num_frames / raw_rate
img = sio.imread(os.path.join(video_folder, frame_fns[0]))
self.frame_shape = self.img_prep(img).shape
def get_by_index(self, start_time, size, rotation=None):
ss = max(int(start_time * self.rate), 0)
chunk = []
for fno in range(ss, ss+size):
fn = os.path.join(self.video_folder, '{:06d}.jpg'.format(fno))
frame = self.img_prep(sio.imread(fn))
chunk.append(frame)
chunk = np.stack(chunk, 0) if len(chunk) > 1 else chunk[0][np.newaxis]
if rotation is not None:
roll = -int(rotation / (2. * np.pi) * self.frame_shape[1])
chunk = np.roll(chunk, roll, axis=2)
return chunk
class FlowReader(object):
def __init__(self, flow_dir, flow_lims_fn, rate=None, flow_prep=None):
self.reader = VideoReader(flow_dir, rate=rate)
self.lims = np.load(flow_lims_fn)
self.rate = self.reader.rate
self.duration = self.reader.duration
self.flow_prep = flow_prep if flow_prep is not None else lambda x: x
dummy_img = self.flow_prep(np.zeros(self.reader.frame_shape[:2], dtype=np.float32))
self.frame_shape = dummy_img.shape + (1,)
self.dtype = dummy_img.dtype
def get_by_index(self, start_time, size, rotation=None):
chunk = self.reader.get_by_index(start_time, size, rotation)
chunk = chunk.astype(np.float32)
ss = max(int(start_time * self.rate), 0)
t = chunk.shape[0]
m_min = self.lims[ss:ss+t, 0].reshape((-1, 1, 1))
m_max = self.lims[ss:ss+t, 1].reshape((-1, 1, 1))
chunk[:, :, :, 2] *= (m_max - m_min) / 255.
chunk[:, :, :, 2] += m_min
chunk[:, :, :, 0] *= (2 * np.pi) / 255.
chunk[:, :, :, 1] = chunk[:, :, :, 2] * np.sin(chunk[:, :, :, 0])
chunk[:, :, :, 0] = chunk[:, :, :, 2] * np.cos(chunk[:, :, :, 0])
return chunk
class SampleReader(object):
""" Sample reader that preprocesses one sample (ambisonics, video)."""
def __init__(self, folder,
ambi_order=1,
audio_rate=48000,
video_rate=10,
context=1.0,
duration=0.1,
return_video=True,
img_prep=None,
return_flow=False,
flow_prep=None,
skip_silence_thr=None,
shuffle=True,
start_time=0.5,
sample_duration=None,
skip_rate=None,
random_rotations=True,
num_threads=1,
thread_id=0):
a2v = float(audio_rate) / video_rate
snd_dur = duration * audio_rate
vid_dur = duration * video_rate
snd_ctx = context * audio_rate
self.video_id = os.path.split(folder)[-1]
# Check input settings
assert a2v==int(a2v)
assert float(snd_dur)==int(snd_dur)
assert float(vid_dur)==int(vid_dur)
assert float(snd_ctx)==int(snd_ctx)
# Readers
self.audio_reader = AudioReader(os.path.join(folder, 'ambix'), audio_rate, ambi_order)
self.video_reader = VideoReader(os.path.join(folder, 'video'), video_rate, img_prep)
if return_flow:
flow_dir = os.path.join(folder, 'flow')
flow_lims = os.path.join(folder, 'flow', 'flow_limits.npy')
self.flow_reader = FlowReader(flow_dir, flow_lims, video_rate, flow_prep)
# Store arguments
self.folder = folder
self.duration = duration
self.context = context
self.audio_rate = audio_rate
self.video_rate = video_rate
self.audio_size = int(snd_dur) + int(snd_ctx) - 1
self.video_size = int(vid_dur)
self.video_shape = self.video_reader.frame_shape
self.return_video = return_video
self.return_flow = return_flow
self.random_rotations = random_rotations
# If is not training, iterate through video, else extract random time frames
audio_pow_fn = os.path.join(folder, 'audio_pow.lst')
chunks_t = [float(l.strip().split()[0]) for l in open(audio_pow_fn)]
chunks_pow = [float(l.strip().split()[1]) for l in open(audio_pow_fn)]
if skip_rate is not None:
num_chunks = len(chunks_t)
chunks_t = [chunks_t[i] for i in range(0, num_chunks, skip_rate)]
chunks_pow = [chunks_pow[i] for i in range(0, num_chunks, skip_rate)]
if skip_silence_thr is not None:
chunks_t = [chunks_t[i] for i in range(len(chunks_t)) if chunks_pow[i]>skip_silence_thr]
if start_time > 0.5:
chunks_t = [chunks_t[i] for i in range(len(chunks_t)) if chunks_t[i]>=start_time]
if sample_duration is not None:
chunks_t = [chunks_t[i] for i in range(len(chunks_t)) if chunks_t[i]<start_time+sample_duration]
if num_threads > 1:
lims = np.linspace(0, len(chunks_t), num_threads+1).astype(int)
chunks_t = chunks_t[lims[thread_id]:lims[thread_id+1]]
if shuffle:
random.shuffle(chunks_t)
self.chunks_t = chunks_t
self.head = -1
def get(self):
self.head += 1
if self.head >= len(self.chunks_t):
return None
self.cur_t = self.chunks_t[self.head]
cur_t = self.cur_t
rotation = random.random() * 2 * np.pi - np.pi if self.random_rotations else None
chunks = {'id': self.video_id + ' ' + str(cur_t)}
# Audio
audio_ss = cur_t - self.context / 2
chunks['ambix'] = self.audio_reader.get(audio_ss, self.audio_size, rotation)
assert chunks['ambix'] is not None, 'Could not get ambix data for file {} (sec: {})'.format(self.folder, audio_ss)
# Video
if self.return_video:
chunks['video'] = self.video_reader.get_by_index(cur_t, self.video_size, rotation)
assert chunks['video'] is not None, 'Could not get video data for file {} (frame: {})'.format(self.folder, cur_t)
# Flow
if self.return_flow:
chunks['flow'] = self.flow_reader.get_by_index(cur_t, self.video_size, rotation)
assert chunks['flow'] is not None, 'Could not get flow data for file {} (frame: {})'.format(self.folder, cur_t)
return chunks
def loop_chunks(self, n=np.inf):
k = 0
while True:
k += 1
if k > n:
break
chunks = self.get()
if chunks is None:
break
else:
yield chunks
class Feeder(object):
""" Background feeder that preprocesses audio and video files
and enqueues them into a TensorFlow queue."""
def __init__(self, sample_dir,
subset_fn=None,
ambi_order=1,
audio_rate=48000,
video_rate=10,
context=1.0,
duration=0.1,
return_video=True,
frame_size=None,
img_prep=None,
return_flow=False,
flow_prep=None,
queue_size=32,
n_threads=1,
for_eval=False):
self.sample_dir, self.subset_fn = sample_dir, subset_fn
self.ambi_order = ambi_order
self.audio_rate, self.video_rate = audio_rate, video_rate
self.context, self.duration = context, duration
self.return_video = return_video
self.img_prep = img_prep
self.return_flow = return_flow
self.flow_prep = flow_prep
self.n_threads, self.threads = n_threads, []
self.for_eval = for_eval
self.skip_silence_thr = None if for_eval else (0.01 if 'REC-Street' in self.subset_fn else 0.2)
audio_layouts = 'meta/audio_layouts.txt'
masks = {'WXYZ': np.array([1., 1., 1., 1.]), 'WXY': np.array([1., 1., 0., 1.])}
self.channel_mask = {l.split()[0]: masks[l.split()[1]] for l in open(audio_layouts).read().splitlines()}
# Placeholders
snd_ctx = int(context * audio_rate)
snd_dur = int(duration * audio_rate)
snd_shape = (snd_dur + snd_ctx - 1, int(ambi_order+1) ** 2)
vid_dur = int(duration * video_rate)
vid_shape = (vid_dur, frame_size[0], frame_size[1], 3)
names = ['id', 'ambix', 'audio_mask']
shapes = [(), snd_shape, ((self.ambi_order+1)**2,)]
dtypes = [tf.string, tf.float32, tf.float32]
if return_video:
names += ['video']
shapes += [vid_shape]
dtypes += [tf.float32]
if return_flow:
names += ['flow']
shapes += [vid_shape]
dtypes += [tf.float32]
self.tba = {m: tf.placeholder(dtype=t, shape=s) for m, s, t in zip(names, shapes, dtypes)}
# Setup tf queue
self.queue = tf.PaddingFIFOQueue(queue_size, names=names, dtypes=dtypes, shapes=shapes)
self.enqueue = self.queue.enqueue(self.tba)
self.queue_state = self.queue.size()
# Print feeder state
fn_provider = FilenameProvider(self.sample_dir, subset_fn=self.subset_fn, num_epochs=1)
n_chunks = 0
for yid in fn_provider.loop_samples():
folder = os.path.join(self.sample_dir, yid)
reader = SampleReader(folder, skip_silence_thr=self.skip_silence_thr,
skip_rate=10 if self.for_eval else None)
n_chunks += len(reader.chunks_t)
print('\n'+'='*20, 'Feeder', '='*20)
print('{:20s} | {}'.format('Input directory', fn_provider.directory))
print('{:20s} | {}'.format('# videos', fn_provider.num_samples))
print('{:20s} | {}'.format('# chunks', n_chunks))
print('{:20s} | {}'.format('# threads', self.n_threads))
print('{:20s} | {}'.format('Mode', 'eval' if self.for_eval else 'train'))
print('{:20s} | {}'.format('Video fps', video_rate))
print('{:20s} | {} frames, {} secs'.format('Video context', 0, 0))
print('{:20s} | {} frames, {} secs'.format('Video duration', vid_dur, duration))
print('{:20s} | {}'.format('Audio rate', audio_rate))
print('{:20s} | {} frames, {} secs'.format('Audio context', snd_ctx, context))
print('{:20s} | {} frames, {} secs'.format('Audio duration', snd_dur, duration))
print('\nFeeder output tensors')
for m, s, t in zip(names, shapes, dtypes):
print(' * {:10s} | {:20s} | {:10s}'.format(m, str(s), str(t)))
sys.stdout.flush()
def dequeue(self, num_elements):
return self.queue.dequeue_many(num_elements)
def thread_main(self, sess, thread_id, num_threads):
thread = threading.currentThread()
fn_provider = FilenameProvider(self.sample_dir, subset_fn=self.subset_fn,
num_epochs=1 if self.for_eval else np.inf,
shuffle=not self.for_eval)
NUM_SAMPLING = np.inf if self.for_eval else 5
SKIP_RATE = 10 if self.for_eval else None
thread_id = thread_id if self.for_eval else 0
num_threads = num_threads if self.for_eval else 1
for yid in fn_provider.loop_samples():
# Start readers
folder = os.path.join(self.sample_dir, yid)
reader = SampleReader(folder,
ambi_order=self.ambi_order,
audio_rate=self.audio_rate,
video_rate=self.video_rate,
context=self.context,
duration=self.duration,
return_video=self.return_video,
img_prep=self.img_prep,
return_flow=self.return_flow,
flow_prep=self.flow_prep,
skip_silence_thr=self.skip_silence_thr,
shuffle=not self.for_eval,
random_rotations=not self.for_eval,
skip_rate=SKIP_RATE,
thread_id=thread_id,
num_threads=num_threads)
# Feed data into tf queue
for chunk in reader.loop_chunks(NUM_SAMPLING):
feed_dict = {self.tba[n]: chunk[n] for n in chunk}
feed_dict[self.tba['audio_mask']] = self.channel_mask[yid]
if not thread.should_stop:
sess.run(self.enqueue, feed_dict=feed_dict)
else:
return
def done(self, sess):
for t in self.threads:
if t.isAlive():
return False
qsize = sess.run(self.queue_state)
if qsize >= 32:
return False
return True
def start_threads(self, sess):
# Launch feeding threads
for i in range(self.n_threads):
thread = threading.Thread(target=self.thread_main, args=(sess, i, self.n_threads))
thread.should_stop = False
thread.daemon = True # Thread will close when parent quits.
thread.start()
self.threads.append(thread)
return self.threads
def join(self):
for t in self.threads:
t.should_stop = True
t.join()