-
Notifications
You must be signed in to change notification settings - Fork 19
/
deploy.py
202 lines (168 loc) · 8.82 KB
/
deploy.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
"""Description
"""
import sys, os, tempfile, argparse
import tensorflow as tf
import numpy as np
from definitions import *
from feeder import SampleReader
from model import SptAudioGen, SptAudioGenParams
from pyutils.iolib.audio import save_wav
import myutils
def parse_arguments():
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument('model_dir', help='Directory containing model snapshot.')
# Inputs
parser.add_argument('input_folder', default='', help='Folder with input sample.')
parser.add_argument('video', default='', help='High resolution video.')
parser.add_argument('--deploy_start', default=0., type=float)
parser.add_argument('--deploy_duration', default=10., type=float)
# Outputs
parser.add_argument('--output_fn', default='output', help='Basename for output files.')
parser.add_argument('--save_ambix', action='store_true', help='Output ambix video file.')
parser.add_argument('--save_video', action='store_true', help='Output video file.')
parser.add_argument('--overlay_map', action='store_true', help='Overlay spherical map.')
parser.add_argument('--VR', action='store_true', help='360 video.')
parser.add_argument('--gpu', type=int, default=0, help="GPU id")
args = parser.parse_args(sys.argv[1:])
if args.deploy_duration <= 0:
args.deploy_duration = None
return args
class W2XYZ(object):
def __init__(self, model_dir):
print('\n' + '='*30 + ' ARGUMENTS ' + '='*30)
params = myutils.load_params(model_dir)
for k, v in params.__dict__.iteritems():
print('TRAIN | {}: {}'.format(k, v))
self.params = params
self.duration = 0.1
self.batch_size = 10
min_t = min([params.context, self.duration, 1. / params.video_rate])
# Model
num_sep = params.num_sep_tracks if params.separation != NO_SEPARATION else 1
net_params = SptAudioGenParams(sep_num_tracks=num_sep, ctx_feats_fc_units=params.context_units,
loc_fc_units=params.loc_units, sep_freq_mask_fc_units=params.freq_mask_units,
sep_fft_window=params.fft_window)
self.model = SptAudioGen(ambi_order=params.ambi_order,
audio_rate=params.audio_rate,
video_rate=params.video_rate,
context=params.context,
sample_duration=self.duration,
encoders=params.encoders,
separation=params.separation,
params=net_params)
self.audio_size = self.model.snd_dur + self.model.snd_contx - 1
self.video_size = int(self.duration * params.video_rate)
shape = (self.batch_size, self.audio_size, 1)
self.tba = {AUDIO: tf.placeholder(dtype=tf.float32, shape=shape)}
if VIDEO in params.encoders:
shape = (self.batch_size, self.video_size, 224, 448, 3)
self.tba[VIDEO] = tf.placeholder(dtype=tf.float32, shape=shape)
if FLOW in params.encoders:
shape = (self.batch_size, self.video_size, 224, 448, 3)
self.tba[FLOW] = tf.placeholder(dtype=tf.float32, shape=shape)
self.ambi_pred_t = self.model.inference_ops(is_training=False, **self.tba)
saver = tf.train.Saver()
config = tf.ConfigProto(
allow_soft_placement=True,
gpu_options=tf.GPUOptions(allow_growth=True)
)
self.sess = tf.Session(config=config)
print('Loading model...')
print(tf.train.latest_checkpoint(model_dir))
saver.restore(self.sess, tf.train.latest_checkpoint(model_dir))
#saver.restore(self.sess, model_dir+'/model.ckpt')
def deploy(self, input_folder, deploy_start, deploy_duration):
reader = SampleReader(input_folder,
ambi_order=self.params.ambi_order,
audio_rate=self.params.audio_rate,
video_rate=self.params.video_rate,
context=self.params.context,
duration=self.duration,
return_video=VIDEO in self.params.encoders,
img_prep=myutils.img_prep_fcn(),
return_flow=FLOW in self.params.encoders,
start_time=deploy_start,
sample_duration=deploy_duration,
skip_silence_thr=None,
shuffle=False,
random_rotations=False,
skip_rate=None)
dt = reader.chunks_t[0] - deploy_start
reader.chunks_t = [t - dt for t in reader.chunks_t]
print('Generating ambisonics...')
ss = self.model.snd_contx / 2
mono, ambi_pred, sep_channels, sep_mask, weights, biases = [], [], [], [], [], []
while True:
batch = []
for _ in range(self.batch_size):
chunk = reader.get()
if chunk is None:
break
batch.append(chunk)
if not batch:
break
vids = [b['id'] for b in batch]
n_samples = len(vids)
ambix = np.stack([b['ambix'] for b in batch], axis=0)
if n_samples != self.batch_size:
ambix = np.concatenate([ambix, np.zeros((self.batch_size - n_samples, ambix.shape[1], ambix.shape[2]))], axis=0)
feed_dict = {self.tba[AUDIO]: ambix[:, :, :1]}
if VIDEO in self.params.encoders:
video = np.stack([b['video'] for b in batch], axis=0)
if n_samples != self.batch_size:
video = np.concatenate([video, np.zeros((self.batch_size - n_samples, video.shape[1], video.shape[2], video.shape[3], video.shape[4]))], axis=0)
feed_dict[self.tba[VIDEO]] = video
if FLOW in self.params.encoders:
flow = np.stack([b['flow'] for b in batch], axis=0)
if n_samples != self.batch_size:
flow = np.concatenate([flow, np.zeros((self.batch_size - n_samples, flow.shape[1], flow.shape[2], flow.shape[3], flow.shape[4]))], axis=0)
feed_dict[self.tba[FLOW]] = flow
ambi_pred_chk = self.sess.run(self.ambi_pred_t, feed_dict=feed_dict)
n_frames = n_samples * ambi_pred_chk.shape[1]
n_out = ambi_pred_chk.shape[2]
ambi_pred_chk = np.copy(ambi_pred_chk[:n_samples]).reshape((n_frames, n_out))
ambi_pred.append(ambi_pred_chk)
mono.append(np.copy(ambix[:n_samples, ss:ss + self.model.snd_dur, :1]).reshape((-1, 1)))
mono = np.concatenate(mono, 0)
ambi_pred = np.concatenate((mono, np.concatenate(ambi_pred, 0)), 1)
return ambi_pred
def main(args):
os.environ["CUDA_VISIBLE_DEVICES"] = "%d" % args.gpu
tmp_ambix_fn = tempfile.mktemp(prefix='/tmp/', suffix='.wav')
tmp_video_fn = tempfile.mktemp(prefix='/tmp/', suffix='.mp4')
model = W2XYZ(args.model_dir)
ambi_pred = model.deploy(args.input_folder, args.deploy_start, args.deploy_duration)
# dur_t = model.model.duration
# snd1 = model.deploy(args.input_folder, args.deploy_start - dur_t/2, args.deploy_duration + dur_t)
# hann1 = np.hanning(model.model.snd_dur)
# hann1 = np.tile(hann1, snd1.shape[0]/hann1.size)[:, np.newaxis]
# ss = model.model.snd_dur/2
# t = int(args.deploy_duration * model.params.audio_rate)
# snd1 = snd1[ss:ss+t]
# hann1 = hann1[ss:ss+t]
# snd2 = model.deploy(args.input_folder, args.deploy_start, args.deploy_duration + dur_t)
# hann2 = np.hanning(model.model.snd_dur)
# hann2 = np.tile(hann2, snd2.shape[0]/hann2.size)[:, np.newaxis]
# ss = 0
# t = int(args.deploy_duration * model.params.audio_rate)
# snd2 = snd2[ss:ss+t]
# hann2 = hann2[ss:ss+t]
# ambi_pred = (snd1 * hann1 + snd2 * hann2) / (hann1 + hann2)
# Save ambisonics
save_wav(tmp_ambix_fn, ambi_pred, model.params.audio_rate)
if args.save_ambix:
print('Saving ambisonics wav...')
cmd = 'ffmpeg -y -i {} -strict -2 {}'.format(tmp_ambix_fn, args.output_fn)
os.system(cmd)
if args.save_video:
print('Saving video...')
cmd = 'ffmpeg -y -ss {} -i {} -t {} {}'.format(args.deploy_start, args.video, args.deploy_duration, tmp_video_fn)
os.system(cmd)
myutils.gen_360video(tmp_ambix_fn, tmp_video_fn, args.output_fn, overlay_map=args.overlay_map, inject_meta=args.VR, binauralize=not args.VR)
os.remove(tmp_video_fn)
os.remove(tmp_ambix_fn)
if __name__ == '__main__':
print(os.getcwd())
main(parse_arguments())