/
FastMNMF_DP.py
executable file
·321 lines (260 loc) · 14.8 KB
/
FastMNMF_DP.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
#! /usr/bin/env python3
# coding: utf-8
import sys, os
import numpy as np
import chainer
from chainer import functions as chf
import pickle as pic
from configure_FastModel import *
from FastFCA import FastFCA
try:
from chainer import cuda
FLAG_GPU_Available = True
except:
print("---Warning--- You cannot use GPU acceleration because chainer or cupy is not installed")
class FastMNMF_DP(FastFCA):
""" Blind Speech Enhancement Using Fast Multichannel NMF with a Deep Speech Prior (FastMNMF_DP)
X_FTM: the observed complex spectrogram
Q_FMM: diagonalizer that converts a spatial covariance matrix (SCM) to a diagonal matrix
G_NFM: diagonal elements of the diagonalized SCMs (N means the number of all sources)
W_noise_NnFK: basis vectors for noise sources (Nn means the number of noise sources)
H_noise_NnKT: activations for noise sources
Z_speech_DT: latent variables for speech
power_speech_FT: power spectra of speech that is the output of DNN(Z_speech_DT)
lambda_NFT: power spectral densities of each source
lambda_NFT[0] = U_F * V_T * power_speech_FT
lambda_NFT[1:] = W_noise_NnFK @ H_noise_NnKT
Qx_power_FTM: power spectra of Qx
Y_FTM: \sum_n lambda_NFT G_NFM
"""
def __init__(self, speech_VAE=None, n_noise=1, n_Z_iteration=30, n_latent=16, n_basis_noise=2, xp=np, init_SCM="unit", mode_update_Z="sampling", normalize_encoder_input=True):
""" initialize FastMNMF_DP
Parameters:
-----------
n_noise: int
the number of noise sources
speech_VAE: VAE
trained speech VAE network (necessary if you use VAE as speech model)
n_latent: int
the dimension of latent variable Z
n_basis_noise: int
the number of bases of each noise source
init_SCM: str
how to initialize covariance matrix {unit, obs, ILRMA}
mode_update_Z: str
how to update latent variable Z {sampling, backprop}
"""
super(FastMNMF_DP, self).__init__(n_source=n_noise+1, xp=xp, init_SCM=init_SCM)
self.n_source, self.n_speech, self.n_noise = n_noise+1, 1, n_noise
self.speech_VAE = speech_VAE
self.n_Z_iteration = n_Z_iteration
self.n_basis_noise = n_basis_noise
self.n_latent = n_latent
self.mode_update_Z = mode_update_Z
self.normalize_encoder_input = normalize_encoder_input
self.method_name = "FastMNMF_DP"
def set_parameter(self, n_noise=None, n_iteration=None, n_Z_iteration=None, n_basis_noise=None, init_SCM=None, mode_update_Z=None):
""" set parameters
Parameters:
-----------
n_noise: int
the number of sources
n_iteration: int
the number of iteration
n_Z_iteration: int
the number of iteration of updating Z in each iteration
n_basis_noise: int
the number of basis of noise sources
init_SCM: str
how to initialize covariance matrix {unit, obs, ILRMA}
mode_update_Z: str
how to update latent variable Z {sampling, backprop}
"""
if n_noise != None:
self.n_noise = n_noise
self.n_source = n_noise + 1
if n_iteration != None:
self.n_iteration = n_iteration
if n_Z_iteration != None:
self.n_Z_iteration = n_Z_iteration
if n_basis_noise != None:
self.n_basis_noise = n_basis_noise
if init_SCM != None:
self.init_SCM = init_SCM
if mode_update_Z != None:
self.mode_update_Z = mode_update_Z
def initialize_PSD(self):
"""
initialize parameters related to power spectral density (PSD)
W, H, U, V, Z
"""
power_observation_FT = (self.xp.abs(self.X_FTM) ** 2).mean(axis=2)
shape = 2
self.W_noise_NnFK = self.xp.random.dirichlet(np.ones(self.n_freq)*shape, size=[self.n_noise, self.n_basis_noise]).transpose(0, 2, 1)
self.H_noise_NnKT = self.xp.random.gamma(shape, (power_observation_FT.mean() * self.n_freq * self.n_mic / (self.n_noise * self.n_basis_noise)) / shape, size=[self.n_noise, self.n_basis_noise, self.n_time])
self.H_noise_NnKT[self.H_noise_NnKT < EPS] = EPS
self.U_F = self.xp.ones(self.n_freq) / self.n_freq
self.V_T = self.xp.ones(self.n_time)
if self.normalize_encoder_input:
power_observation_FT = power_observation_FT / power_observation_FT.sum(axis=0).mean()
self.Z_speech_DT = self.speech_VAE.encode_cupy(power_observation_FT.astype(self.xp.float32))
self.z_link_speech = Z_link(self.Z_speech_DT.T)
self.z_optimizer_speech = chainer.optimizers.Adam().setup(self.z_link_speech)
self.power_speech_FT = self.speech_VAE.decode_cupy(self.Z_speech_DT)
self.lambda_NFT = self.xp.zeros([self.n_source, self.n_freq, self.n_time])
self.lambda_NFT[0] = self.U_F[:, None] * self.V_T[None] * self.power_speech_FT
self.lambda_NFT[1:] = self.W_noise_NnFK @ self.H_noise_NnKT
def make_fileName_suffix(self):
self.fileName_suffix = "S={}-it={}-itZ={}-Ln={}-D={}-init={}-latent={}".format(self.n_source, self.n_iteration, self.n_Z_iteration, self.n_basis_noise, self.n_latent, self.init_SCM, self.mode_update_Z)
if hasattr(self, "name_DNN"):
self.fileName_suffix += "-DNN={}".format(self.name_DNN)
if hasattr(self, "file_id"):
self.fileName_suffix += "-ID={}".format(self.file_id)
else:
print("====================\n\nWarning: Please set self.file_id\n\n====================")
print("parameter:", self.fileName_suffix)
def update(self):
self.update_UV()
self.update_Z_speech()
self.update_WH_noise()
self.update_CovarianceDiagElement()
self.udpate_Diagonalizer()
self.normalize()
def normalize(self):
phi_F = self.xp.sum(self.Q_FMM * self.Q_FMM.conj(), axis=(1, 2)).real / self.n_mic
self.Q_FMM = self.Q_FMM / self.xp.sqrt(phi_F)[:, None, None]
self.G_NFM = self.G_NFM / phi_F[None, :, None]
mu_NF = (self.G_NFM).sum(axis=2).real
self.G_NFM = self.G_NFM / mu_NF[:, :, None]
self.U_F = self.U_F * mu_NF[0]
self.W_noise_NnFK = self.W_noise_NnFK * mu_NF[1:][:, :, None]
nu = self.U_F.sum()
self.U_F = self.U_F / nu
self.V_T = nu * self.V_T
self.lambda_NFT[0] = self.U_F[:, None] * self.V_T[None] * self.power_speech_FT
nu_NnK = self.W_noise_NnFK.sum(axis=1)
self.W_noise_NnFK = self.W_noise_NnFK / nu_NnK[:, None]
self.H_noise_NnKT = self.H_noise_NnKT * nu_NnK[:, :, None]
self.lambda_NFT[1:] = self.W_noise_NnFK @ self.H_noise_NnKT + EPS
self.reset_variable()
def update_WH_noise(self):
tmp1_NnFT = (self.G_NFM[1, :, None] * (self.Qx_power_FTM / (self.Y_FTM ** 2))[None]).sum(axis=3)
tmp2_NnFT = (self.G_NFM[1, :, None] / self.Y_FTM[None]).sum(axis=3)
a_W = (self.H_noise_NnKT[:, None] * tmp1_NnFT[:, :, None]).sum(axis=3) # N F K T M
b_W = (self.H_noise_NnKT[:, None] * tmp2_NnFT[:, :, None]).sum(axis=3)
a_H = (self.W_noise_NnFK[..., None] * tmp1_NnFT[:, :, None] ).sum(axis=1) # N F K T M
b_H = (self.W_noise_NnFK[..., None] * tmp2_NnFT[:, :, None]).sum(axis=1) # N F K T M
self.W_noise_NnFK = self.W_noise_NnFK * self.xp.sqrt(a_W / b_W)
self.H_noise_NnKT = self.H_noise_NnKT * self.xp.sqrt(a_H / b_H)
self.lambda_NFT[1:] = self.W_noise_NnFK @ self.H_noise_NnKT + EPS
self.Y_FTM = (self.lambda_NFT[..., None] * self.G_NFM[:, :, None]).sum(axis=0)
def update_UV(self):
a_1 = ((self.V_T[None] * self.power_speech_FT)[:, :, None] * self.Qx_power_FTM * self.G_NFM[0, :, None] / (self.Y_FTM ** 2)).sum(axis=2).sum(axis=1).real
b_1 = ((self.V_T[None] * self.power_speech_FT)[:, :, None] * self.G_NFM[0, :, None] / self.Y_FTM).sum(axis=2).sum(axis=1).real
self.U_F = self.U_F * self.xp.sqrt(a_1 / b_1)
self.lambda_NFT[0] = self.U_F[:, None] * self.V_T[None] * self.power_speech_FT
self.Y_FTM = (self.lambda_NFT[..., None] * self.G_NFM[:, :, None]).sum(axis=0)
a_1 = ((self.U_F[:, None] * self.power_speech_FT)[:, :, None] * self.Qx_power_FTM * self.G_NFM[0, :, None] / (self.Y_FTM ** 2)).sum(axis=2).sum(axis=0).real
b_1 = ((self.U_F[:, None] * self.power_speech_FT)[:, :, None] * self.G_NFM[0, :, None] / self.Y_FTM).sum(axis=2).sum(axis=0).real
self.V_T = self.V_T * self.xp.sqrt(a_1 / b_1)
self.lambda_NFT[0] = self.U_F[:, None] * self.V_T[None] * self.power_speech_FT
self.Y_FTM = (self.lambda_NFT[..., None] * self.G_NFM[:, :, None]).sum(axis=0)
def loss_func_Z(self, z, vae, n): # for update Z by backprop
power_tmp_FT = chf.exp(vae.decode(z).T) + EPS
Y_tmp_FTM = power_tmp_FT[:, :, None] * self.UVG_FTM+ self.WHG_noise_FTM
return chf.sum(chf.log(Y_tmp_FTM) + self.Qx_power_FTM / Y_tmp_FTM ) / (self.n_freq * self.n_mic)
def update_Z_speech(self, var_propose_distribution=1e-4):
"""
Parameters:
var_propose_distribution: float
the variance of the propose distribution
Results:
self.Z_speech_DT: self.xp.array [ n_latent x T ]
the latent variable of each speech
"""
self.WHG_noise_FTM = (self.lambda_NFT[1:][..., None] * self.G_NFM[1:, :, None]).sum(axis=0)
self.UVG_FTM = (self.U_F[:, None] * self.V_T[None])[:, :, None] * self.G_NFM[0, :, None]
if "backprop" in self.mode_update_Z: # acceptance rate is calculated from likelihood
for it in range(self.n_Z_iteration):
with chainer.using_config('train', False):
self.z_optimizer_speech.update(self.loss_func_Z, self.z_link_speech.z, self.speech_VAE, 0)
self.Z_speech_DT = self.z_link_speech.z.data.T
self.power_speech_FT = self.speech_VAE.decode_cupy(self.Z_speech_DT)
if "sampling" in self.mode_update_Z:
log_var = self.xp.log(self.xp.ones_like(self.Z_speech_DT).astype(self.xp.float32) * var_propose_distribution)
Z_speech_old_DT = self.Z_speech_DT
power_old_FTM = self.speech_VAE.decode_cupy(Z_speech_old_DT)[:, :, None]
for it in range(self.n_Z_iteration):
Z_speech_new_DT = chf.gaussian(Z_speech_old_DT, log_var).data
lambda_old_FTM = power_old_FTM * self.UVG_FTM + self.WHG_noise_FTM
power_new_FTM = self.speech_VAE.decode_cupy(Z_speech_new_DT)[:, :, None]
lambda_new_FTM = power_new_FTM * self.UVG_FTM + self.WHG_noise_FTM
acceptance_rate = self.xp.exp((self.Qx_power_FTM * (1 / lambda_old_FTM - 1 / lambda_new_FTM)).sum(axis=2).sum(axis=0) + self.xp.log( ( lambda_old_FTM / lambda_new_FTM ).prod(axis=2).prod(axis=0) ) )
accept_flag = self.xp.random.random([self.n_time]) < acceptance_rate
Z_speech_old_DT[:, accept_flag] = Z_speech_new_DT[:, accept_flag]
power_old_FTM[:, accept_flag] = power_new_FTM[:, accept_flag]
self.Z_speech_DT = Z_speech_old_DT
self.z_link_speech.z = chainer.Parameter(self.Z_speech_DT.T)
self.power_speech_FT = self.speech_VAE.decode_cupy(self.Z_speech_DT)
self.lambda_NFT[0] = self.U_F[:, None] * self.V_T[None] * self.power_speech_FT
self.Y_FTM = (self.lambda_NFT[..., None] * self.G_NFM[:, :, None]).sum(axis=0)
def save_parameter(self, fileName):
param_list = [self.lambda_NFT, self.G_NFM, self.Q_FMM, self.U_F, self.V_T, self.Z_speech_DT, self.W_noise_NnFK, self.H_noise_NnKT]
if self.xp != np:
param_list = [self.convert_to_NumpyArray(param) for param in param_list]
pic.dump(param_list, open(fileName, "wb"))
def load_parameter(self, fileName):
param_list = pic.load(open(fileName, "rb"))
if self.xp != np:
param_list = [cuda.to_gpu(param) for param in param_list]
self.lambda_NFT, self.G_NFM, self.Q_FMM, self.U_F, self.V_T, self.Z_speech_DT, self.W_noise_NnFK, self.H_noise_NnKT = param_list
class Z_link(chainer.link.Link):
def __init__(self, z):
super(Z_link, self).__init__()
with self.init_scope():
self.z = chainer.Parameter(z)
if __name__ == "__main__":
import soundfile as sf
import librosa
import sys, os
from chainer import serializers
import argparse
parser = argparse.ArgumentParser()
parser.add_argument( 'input_fileName', type= str, help='filename of the multichannel observed signals')
parser.add_argument( '--file_id', type= str, default="None", help='file id')
parser.add_argument( '--gpu', type= int, default= 0, help='GPU ID')
parser.add_argument( '--n_latent', type= int, default= 16, help='dimention of encoded vector')
parser.add_argument( '--n_noise', type= int, default= 1, help='number of noise')
parser.add_argument('--n_basis_noise', type= int, default= 64, help='number of basis of noise (MODE_noise=NMF)')
parser.add_argument( '--init_SCM', type= str, default= "obs", help='unit, obs, ILRMA')
parser.add_argument( '--n_iteration', type= int, default= 100, help='number of iteration')
parser.add_argument('--n_Z_iteration', type= int, default= 30, help='number of update Z iteration')
parser.add_argument('--mode_update_Z', type= str, default="sampling", help='sampling, sampling2, backprop, backprop2, hybrid, hybrid2')
args = parser.parse_args()
if args.gpu < 0:
import numpy as xp
else:
import cupy as xp
print("Use GPU " + str(args.gpu))
chainer.cuda.get_device_from_id(args.gpu).use()
sys.path.append("../DeepSpeechPrior")
import network_VAE
model_fileName = "../DeepSpeechPrior/model-VAE-best-scale=gamma-D={}.npz".format(args.n_latent)
speech_VAE = network_VAE.VAE(n_latent=args.n_latent)
serializers.load_npz(model_fileName, speech_VAE)
name_DNN = "VAE"
if xp != np:
speech_VAE.to_gpu()
wav, fs = sf.read(args.input_fileName)
wav = wav.T
M = len(wav)
for m in range(M):
tmp = librosa.core.stft(wav[m], n_fft=1024, hop_length=256)
if m == 0:
spec = np.zeros([tmp.shape[0], tmp.shape[1], M], dtype=np.complex)
spec[:, :, m] = tmp
separater = FastMNMF_DP(n_noise=args.n_noise, speech_VAE=speech_VAE, n_Z_iteration=args.n_Z_iteration, n_basis_noise=args.n_basis_noise, xp=xp, init_SCM=args.init_SCM, mode_update_Z=args.mode_update_Z)
separater.load_spectrogram(spec)
separater.file_id = args.file_id
separater.name_DNN = name_DNN
separater.solve(n_iteration=args.n_iteration, save_likelihood=False, save_parameter=False, save_path="./", interval_save_parameter=25)