-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathevaluate_music_separation.py
286 lines (222 loc) · 12.9 KB
/
evaluate_music_separation.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
import argparse
import json
import torch
import os
import numpy as np
import scipy.io.wavfile as wavfile
import statistics as stat
from utils import si_sdr, beamformer_max_di, \
beamformer_max_re, beamformer_max_sdr, zen_to_ele, azi_to_0_2pi_range
from network import DemucsDirection, center_trim
from pathlib import Path
def flatten(t):
return [item for sublist in t for item in sublist]
def save_audio(save_folder, method, azi_angle, zen_angle, waveform):
Path(save_folder).mkdir(parents = True, exist_ok = True)
output_path = os.path.join(save_folder, method + '_waveform_azi_' + "{:.2f}".format(
azi_angle * 180 / np.pi) + '_zen_' + "{:.2f}".format(zen_angle * 180 / np.pi) + '.wav')
aux_waveform = waveform * np.iinfo(np.int16).max
wavfile.write(output_path, 44100, aux_waveform.astype(np.int16))
def forward_pass(model, mixed_data, conditioning_direction, beamformer_audio, args):
ambi_mixes = mixed_data.float().unsqueeze(0) # Batch size is 1
ambi_mixes = ambi_mixes.to(args.device)
conditioning_direction = conditioning_direction.float().unsqueeze(0)
conditioning_direction = conditioning_direction.to(args.device)
beamformer_audio = beamformer_audio.float().unsqueeze(0)
beamformer_audio = beamformer_audio.to(args.device)
output_signal = model(ambi_mixes, conditioning_direction, beamformer_audio)
output_signal = center_trim(output_signal, ambi_mixes)
output_signal = torch.squeeze(output_signal, dim = 1)
output_np = output_signal.detach().cpu().numpy()
return output_np
def forward_beamformer(bf_type, input_signal, aux):
if bf_type == 'max_di':
beamformer = beamformer_max_di
if bf_type == 'max_re':
beamformer = beamformer_max_re
if bf_type == 'max_sdr':
beamformer = beamformer_max_sdr
return beamformer(input_signal, aux)
def get_items(curr_dir, ambiorder):
with open(Path(curr_dir) / 'metadata.json') as json_file:
metadata = json.load(json_file)
# Iterate over different sources
source_positions = []
source_audios = []
for key in sorted(metadata.keys()):
# get source audio
gt_audio_files = sorted(
list(Path(curr_dir).rglob(key + ".wav")))
assert (len(gt_audio_files) > 0)
_, gt_waveform = wavfile.read(gt_audio_files[0])
gt_waveform = gt_waveform.astype(np.float)
is_all_zero = np.all((gt_waveform == 0))
if not is_all_zero:
rms = np.sqrt(np.mean(gt_waveform ** 2))
gt_waveform = gt_waveform * (0.1 / rms) # desired rms is 0.1
gt_waveform = gt_waveform.T.copy() # MxT numpy array
source_audios.append(gt_waveform)
# get source position
source_azi_angle = metadata[key]['panning_angles'][0]
source_zen_angle = metadata[key]['panning_angles'][1]
source_positions.append([source_azi_angle, source_zen_angle])
# get mixture
mix_path = os.path.join(curr_dir, "mix.wav")
rate, mixture_waveform = wavfile.read(mix_path)
mixture_waveform = mixture_waveform.astype(np.float)
mix_is_all_zero = np.all((mixture_waveform[:, 0] == 0))
if not mix_is_all_zero:
mixture_waveform = mixture_waveform / np.amax(np.abs(mixture_waveform[:, 0])) / np.sqrt(2 * ambiorder + 1)
return mixture_waveform, source_positions, source_audios, sorted(metadata.keys())
def main(args):
print("result path will be")
print(args.result_dir)
print('\n')
device = torch.device('cuda') if args.use_cuda else torch.device('cpu')
args.device = device
n_channels = (args.ambiorder + 1) ** 2
if args.ambimode == 'implicit':
model = DemucsDirection(n_audio_channels = (args.ambiorder + 1) ** 2, ambimode = args.ambimode)
elif args.ambimode == 'mixed':
model = DemucsDirection(n_audio_channels = 5, ambimode = args.ambimode)
model.load_state_dict(torch.load(args.model_checkpoint), strict = True)
model.train = False
model.to(device)
all_dirs = sorted(list(Path(args.test_dir).glob('[0-9]*')))
si_sdr_nn = {'vocals': [], 'drums': [], 'bass': []}
si_sdr_beamformer_max_di = {'vocals': [], 'drums': [], 'bass': []}
si_sdr_beamformer_max_re = {'vocals': [], 'drums': [], 'bass': []}
si_sdr_beamformer_max_sdr = {'vocals': [], 'drums': [], 'bass': []}
si_sdr_omnimix = {'vocals': [], 'drums': [],
'bass': []} # For baseline, we consider the omni mix as the separated source
si_sdr_stats = {'median': {'nn': {'vocals': None, 'drums': None, 'bass': None, 'all': None},
'beamformer_max_di': {'vocals': None, 'drums': None, 'bass': None, 'all': None},
'beamformer_max_re': {'vocals': None, 'drums': None, 'bass': None, 'all': None},
'beamformer_max_sdr': {'vocals': None, 'drums': None, 'bass': None, 'all': None},
'omnimix': {'vocals': None, 'drums': None, 'bass': None, 'all': None}}}
checkpoint_folder = os.path.split(os.path.normpath(args.model_checkpoint))[0]
if args.save_audios:
print("Audios will be saved here:")
print(os.path.join(checkpoint_folder, 'test_audio_samples'))
print('\n')
cmt = 0
for idx in range(0, len(all_dirs)):
print(idx)
curr_dir = all_dirs[idx]
# Loads the data
mixed_data, source_positions, source_audios, sources_name = get_items(curr_dir, args.ambiorder)
mixed_data = mixed_data[:, 0:n_channels]
idx_folder = os.path.basename(os.path.normpath(curr_dir))
omni_mix = mixed_data[:, 0]
if args.save_audios:
save_folder = os.path.join(checkpoint_folder, 'test_musdb_audio_samples', idx_folder)
Path(save_folder).mkdir(parents = True, exist_ok = True)
output_path = os.path.join(save_folder, 'omnimix.wav')
aux_omni_mix = omni_mix * np.iinfo(np.int16).max
wavfile.write(output_path, 44100, aux_omni_mix.astype(np.int16))
for [azi_angle, zen_angle], gt_waveform, key in zip(source_positions, source_audios, sources_name):
if args.save_audios:
idx_folder = os.path.basename(os.path.normpath(curr_dir))
save_folder = os.path.join(checkpoint_folder, 'test_musdb_audio_samples', idx_folder)
save_audio(save_folder, 'gt', azi_angle, zen_angle, gt_waveform)
azi_angle_beamformer = azi_to_0_2pi_range(azi_angle)
ele_angle_beamformer = zen_to_ele(zen_angle)
# beamformer_max_di output at this location
beamformer_max_di_audio = forward_beamformer('max_di', mixed_data,
np.array((azi_angle_beamformer, ele_angle_beamformer)))[:, 0]
if args.save_audios:
save_audio(save_folder, 'max_di', azi_angle, zen_angle, beamformer_max_di_audio)
# beamformer_max_re output at this location
beamformer_max_re_audio = forward_beamformer('max_re', mixed_data,
np.array((azi_angle_beamformer, ele_angle_beamformer)))[:, 0]
beamformer_audio = beamformer_max_re_audio.copy()
beamformer_audio = np.expand_dims(beamformer_audio, axis=0)
if args.save_audios:
save_audio(save_folder, 'max_re', azi_angle, zen_angle, beamformer_max_re_audio)
# beamformer_max_sdr output at this location
beamformer_max_sdr_audio, singular_matrix = forward_beamformer('max_sdr', mixed_data, gt_waveform)
if singular_matrix:
cmt += 1
print("Total singular matrices = " + str(cmt))
print('\n')
if args.save_audios and not singular_matrix:
save_audio(save_folder, 'max_sdr', azi_angle, zen_angle, beamformer_max_sdr_audio)
# neural network audio output at this location
azi_normalized = (azi_angle + np.pi) / np.pi - 1
zen_normalized = 2 * zen_angle / np.pi - 1
conditioning_direction = np.asarray([azi_normalized, zen_normalized])
conditioning_direction = torch.tensor(conditioning_direction).float()
nn_mixed_data = mixed_data
if args.ambimode == 'mixed':
rms = np.sqrt(np.mean(beamformer_audio ** 2))
if rms != 0:
beamformer_audio = beamformer_audio * (0.1 / rms)
nn_mixed_data = mixed_data[:, 0:4]
nn_mixed_data = torch.tensor(nn_mixed_data.T).float()
nn_beamformer_audio = torch.tensor(beamformer_audio).float()
nn_predicted_audio = forward_pass(model, nn_mixed_data, conditioning_direction, nn_beamformer_audio, args)
nn_predicted_audio = nn_predicted_audio[0, 0, :]
if args.save_audios:
save_audio(save_folder, 'nn_' + args.ambimode, azi_angle, zen_angle, nn_predicted_audio)
is_all_zero = np.all((gt_waveform == 0))
if not is_all_zero:
si_sdr_nn[key].append(si_sdr(nn_predicted_audio, gt_waveform))
si_sdr_beamformer_max_di[key].append(si_sdr(beamformer_max_di_audio, gt_waveform))
si_sdr_beamformer_max_re[key].append(si_sdr(beamformer_max_re_audio, gt_waveform))
if not singular_matrix:
si_sdr_beamformer_max_sdr[key].append(si_sdr(beamformer_max_sdr_audio, gt_waveform))
si_sdr_omnimix[key].append(si_sdr(omni_mix, gt_waveform))
Path(args.result_dir).mkdir(parents = True, exist_ok = True)
json_path = os.path.join(args.result_dir, 'si_sdr_nn.json')
with open(json_path, 'w') as fp:
json.dump(si_sdr_nn, fp)
json_path = os.path.join(args.result_dir, 'si_sdr_beamformer_max_di.json')
with open(json_path, 'w') as fp:
json.dump(si_sdr_beamformer_max_di, fp)
json_path = os.path.join(args.result_dir, 'si_sdr_beamformer_max_re.json')
with open(json_path, 'w') as fp:
json.dump(si_sdr_beamformer_max_re, fp)
json_path = os.path.join(args.result_dir, 'si_sdr_beamformer_max_sdr.json')
with open(json_path, 'w') as fp:
json.dump(si_sdr_beamformer_max_sdr, fp)
json_path = os.path.join(args.result_dir, 'si_sdr_omnimix.json')
with open(json_path, 'w') as fp:
json.dump(si_sdr_omnimix, fp)
for key in ['vocals', 'drums', 'bass']:
si_sdr_stats['median']['nn'][key] = stat.median(si_sdr_nn[key])
si_sdr_stats['median']['beamformer_max_di'][key] = stat.median(si_sdr_beamformer_max_di[key])
si_sdr_stats['median']['beamformer_max_re'][key] = stat.median(si_sdr_beamformer_max_re[key])
si_sdr_stats['median']['beamformer_max_sdr'][key] = stat.median(si_sdr_beamformer_max_sdr[key])
si_sdr_stats['median']['omnimix'][key] = stat.median(si_sdr_omnimix[key])
si_sdr_stats['median']['nn']['all'] = stat.median(
flatten([si_sdr_nn['vocals'], si_sdr_nn['drums'], si_sdr_nn['bass']]))
si_sdr_stats['median']['beamformer_max_di']['all'] = stat.median(flatten(
[si_sdr_beamformer_max_di['vocals'], si_sdr_beamformer_max_di['drums'], si_sdr_beamformer_max_di['bass']]))
si_sdr_stats['median']['beamformer_max_re']['all'] = stat.median(flatten(
[si_sdr_beamformer_max_re['vocals'], si_sdr_beamformer_max_re['drums'], si_sdr_beamformer_max_re['bass']]))
si_sdr_stats['median']['beamformer_max_sdr']['all'] = stat.median(flatten(
[si_sdr_beamformer_max_sdr['vocals'], si_sdr_beamformer_max_sdr['drums'], si_sdr_beamformer_max_sdr['bass']]))
si_sdr_stats['median']['omnimix']['all'] = stat.median(
flatten([si_sdr_omnimix['vocals'], si_sdr_omnimix['drums'], si_sdr_omnimix['bass']]))
json_path = os.path.join(args.result_dir, 'si_sdr_stats.json')
with open(json_path, 'w') as fp:
json.dump(si_sdr_stats, fp)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('test_dir', type = str,
help = "Path to the testing directory")
parser.add_argument('model_checkpoint', type = str,
help = "Path to the model file")
parser.add_argument('--use_cuda', dest = 'use_cuda', action = 'store_true',
help = "Whether to use cuda")
parser.add_argument('--ambiorder', type = int, default = 4,
help = "Ambisonics order")
parser.add_argument('--ambimode', type = str, default = 'implicit',
help = "Ambisonics mode. 'implicit': raw Ambisonics mixture as input. "
"'mixed': raw first order Ambisonics mixture and bf concatenated.")
parser.add_argument('--save_audios', dest = 'save_audios', action = 'store_true',
help = "Whether to save predicted audios")
parser.add_argument('--result_dir', dest = 'result_dir', type = str,
help = "Path for the si_sdr results")
print(parser.parse_args())
main(parser.parse_args())