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
|
# coding: utf-8
from tensorflow.python.client import device_lib
import tensorflow as tf
import librosa
import os
# from IPython.display import Audio, display
import numpy as np
# import matplotlib.pyplot as plt
import sys
# get_ipython().magic(u'matplotlib inline')
# ### Load style and content
# In[5]:
if len(sys.argv) < 4:
print "python nsatf.py content.wav style.wav output.wav alpha"
sys.exit()
CONTENT_FILENAME = sys.argv[1]
STYLE_FILENAME = sys.argv[2]
OUTPUT_FILENAME = sys.argv[3]
if len(sys.argv) == 5:
ALPHA = float(sys.argv[4] or "1e-3")
else:
ALPHA = 1e-3
device_ids = [device.name for device in device_lib.list_local_devices()]
if '/device:GPU:0' in device_ids:
DEVICE = '/device:GPU:0'
else:
DEVICE = '/device:CPU:0'
print DEVICE
# In[6]:
# display(Audio(CONTENT_FILENAME))
# display(Audio(STYLE_FILENAME))
# In[7]:
# Reads wav file and produces spectrum
# Fourier phases are ignored
N_FFT = 2048
def read_audio_spectum(filename):
print 'load ' + filename
x, fs = librosa.load(filename, 44100)
S = librosa.stft(x, N_FFT)
p = np.angle(S)
S = np.log1p(np.abs(S[:,:1020]))
return S, fs
# In[8]:
a_content, fs = read_audio_spectum(CONTENT_FILENAME)
a_style, fs = read_audio_spectum(STYLE_FILENAME)
hs = a_content.shape[1]
ms = a_style.shape[1]
if hs > ms:
a_style = np.lib.pad(a_style, ((0,0), (0, hs - ms)), 'constant', constant_values=(0, 0))
else:
a_content = np.lib.pad(a_content, ((0,0), (0, ms - hs)), 'constant', constant_values=(0, 0))
print a_content.shape
print a_style.shape
hs = a_content.shape[0]
ms = a_style.shape[0]
if hs > ms:
a_style = np.lib.pad(a_style, ((0, hs - ms), (0,0)), 'constant', constant_values=(0, 0))
else:
a_content = np.lib.pad(a_content, ((0, ms - hs), (0,0)), 'constant', constant_values=(0, 0))
print a_content.shape
print a_style.shape
N_SAMPLES = a_style.shape[1]
N_CHANNELS = a_style.shape[0]
# ### Visualize spectrograms for content and style tracks
# In[9]:
"""
plt.figure(figsize=(10, 5))
plt.subplot(1, 2, 1)
plt.title('Content')
plt.imshow(a_content[:400,:])
plt.subplot(1, 2, 2)
plt.title('Style')
plt.imshow(a_style[:400,:])
plt.show()
"""
# ### Compute content and style feats
# In[10]:
N_FILTERS = 4096
a_content_tf = np.ascontiguousarray(a_content.T[None,None,:,:])
a_style_tf = np.ascontiguousarray(a_style.T[None,None,:,:])
# filter shape is "[filter_height, filter_width, in_channels, out_channels]"
std = np.sqrt(2) * np.sqrt(2.0 / ((N_CHANNELS + N_FILTERS) * 10))
kernel = np.random.randn(1, 10, N_CHANNELS, N_FILTERS)*std
g = tf.Graph()
with g.as_default(), g.device(DEVICE), tf.Session() as sess:
# data shape is "[batch, in_height, in_width, in_channels]",
x = tf.placeholder('float32', [1,1,N_SAMPLES,N_CHANNELS], name="x")
kernel_tf = tf.constant(kernel, name="kernel", dtype='float32')
conv = tf.nn.conv2d(
x,
kernel_tf,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
net = tf.nn.relu(conv)
content_features = net.eval(feed_dict={x: a_content_tf})
style_features = net.eval(feed_dict={x: a_style_tf})
features = np.reshape(style_features, (-1, N_FILTERS))
style_gram = np.matmul(features.T, features) / N_SAMPLES
# ### Optimize
# In[14]:
from sys import stderr
iterations = 100
result = None
with tf.Graph().as_default():
# Build graph with variable input
# x = tf.Variable(np.zeros([1,1,N_SAMPLES,N_CHANNELS], dtype=np.float32), name="x")
x = tf.Variable(np.random.randn(1,1,N_SAMPLES,N_CHANNELS).astype(np.float32)*1e-3, name="x")
kernel_tf = tf.constant(kernel, name="kernel", dtype='float32')
conv = tf.nn.conv2d(
x,
kernel_tf,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
net = tf.nn.relu(conv)
content_loss = ALPHA * 2 * tf.nn.l2_loss( net - content_features )
style_loss = 0
_, height, width, number = map(lambda i: i.value, net.get_shape())
size = height * width * number
feats = tf.reshape(net, (-1, number))
gram = tf.matmul(tf.transpose(feats), feats) / N_SAMPLES
style_loss = 2 * tf.nn.l2_loss(gram - style_gram)
# Overall loss
loss = content_loss + style_loss
opt = tf.contrib.opt.ScipyOptimizerInterface(
loss, method='L-BFGS-B', options={'maxiter': iterations})
# Optimization
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print('Started optimization.')
opt.minimize(sess)
print 'Final loss:', loss.eval()
result = x.eval()
# ### Invert spectrogram and save the result
# In[15]:
a = np.zeros_like(a_content)
a[:N_CHANNELS,:] = np.exp(result[0,0].T) - 1
# This code is supposed to do phase reconstruction
p = 2 * np.pi * np.random.random_sample(a.shape) - np.pi
for i in range(500):
S = a * np.exp(1j*p)
x = librosa.istft(S)
p = np.angle(librosa.stft(x, N_FFT))
librosa.output.write_wav(OUTPUT_FILENAME, x, fs)
# In[16]:
#print OUTPUT_FILENAME
#display(Audio(OUTPUT_FILENAME))
# ### Visualize spectrograms
# In[17]:
"""
plt.figure(figsize=(15,5))
plt.subplot(1,3,1)
plt.title('Content')
plt.imshow(a_content[:400,:])
plt.subplot(1,3,2)
plt.title('Style')
plt.imshow(a_style[:400,:])
plt.subplot(1,3,3)
plt.title('Result')
plt.imshow(a[:400,:])
plt.show()
"""
# In[ ]:
|
