-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_lstm_sequence.py
134 lines (121 loc) · 6.04 KB
/
test_lstm_sequence.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
import tensorflow as tf
from tensorflow.keras.models import Sequential, Model, load_model
from tensorflow.keras.layers import Input, Dense, GRU, LSTM, Multiply
from tensorflow.keras import backend as K
import numpy as np
import glob
import os
import matplotlib.pyplot as plt
block_length = 0.050#->500ms
voice_max_length = int(0.5/block_length)#->10s
frame_length = 512
model_name = "noise_model_lstm_sequence"
batch_size = 32
epochs = 10
print("voice_max_length:", voice_max_length)
def audioToTensor(filepath:str):
#print("audioToTensor:filepath:", filepath)
audio_binary = tf.io.read_file(filepath)
audio, audioSR = tf.audio.decode_wav(audio_binary)
audioSR = tf.get_static_value(audioSR)
audio = tf.squeeze(audio, axis=-1)
frame_step = int(audioSR * 0.008)#16000*0.008=128
spectrogram = tf.signal.stft(audio, frame_length=frame_length, frame_step=frame_step)
spect_image = tf.math.imag(spectrogram)
spect_real = tf.math.real(spectrogram)
spect_sign = tf.sign(spect_real)
spect_real = tf.abs(spect_real)
return spect_real, spect_image, spect_sign, audioSR
def spectToOscillo(spect_real, spect_sign, spect_image, audioSR):
frame_step = int(audioSR * 0.008)
#spect_real = pow(10, spect_real/20)#power value
spect_real*=spect_sign
spect_all = tf.complex(spect_real, spect_image)
inverse_stft = tf.signal.inverse_stft(spect_all, frame_length=frame_length, frame_step=frame_step, window_fn=tf.signal.inverse_stft_window_fn(frame_step))
return inverse_stft
clear_files = glob.glob("data/clear/*.wav")
x_train = []
x_train_count = 0
for i, path_clear in enumerate(clear_files):
spectNoisy, _, _, audioNoisySR = audioToTensor(path_clear)
x_train.append((path_clear, len(spectNoisy)-voice_max_length))
x_train_count+=len(spectNoisy)//voice_max_length
print("x_train_count:", x_train_count)
class MySequence(tf.keras.utils.Sequence):
def __init__(self, x_train, x_train_count, batch_size):
self.x_train= x_train
self.x_train_count = x_train_count
self.batch_size = batch_size
def __len__(self):
return self.x_train_count//self.batch_size
def __getitem__(self, idx):
batch_x_train = np.zeros((batch_size, voice_max_length, int(frame_length/2+1)))
batch_y_train = np.zeros((batch_size, voice_max_length, int(frame_length/2+1)))
current_size = 0
while current_size < batch_size:
path_clear, _ = self.x_train[(idx * self.batch_size + current_size)%len(clear_files)]
path_noisy = path_clear.replace("clear", "noisy")
spectNoisy, _, _, audioNoisySR = audioToTensor(path_noisy)
spectClear, _, _, audioClearSR = audioToTensor(path_clear)
for k in range(0, min(len(spectNoisy), len(spectClear))-voice_max_length):
batch_x_train[current_size] = spectNoisy[k*voice_max_length:(k+1)*voice_max_length]
batch_y_train[current_size] = spectClear[k*voice_max_length:(k+1)*voice_max_length]
current_size+=1
if current_size>=batch_size:
break
return batch_x_train, batch_y_train
print('Build model...')
if os.path.exists(model_name):
print("Load: " + model_name)
model = load_model(model_name)
else:
main_input = Input(shape=(None, int(frame_length/2+1)), name='main_input')
x = main_input
x = LSTM(256, activation='tanh', recurrent_activation='sigmoid', return_sequences=True)(x)
x = Dense(int(frame_length/2+1), activation='sigmoid')(x)
x = Multiply()([x, main_input])
model = Model(inputs=main_input, outputs=x)
tf.keras.utils.plot_model(model, to_file='model_lstm_sequence.png', show_shapes=True)
model.compile(loss='mse', metrics='mse', optimizer=tf.keras.optimizers.Adam(learning_rate=0.001))#Adam, SGD, Adagrad
print('Train...')
history = model.fit(MySequence(x_train, x_train_count, batch_size), epochs=epochs, steps_per_epoch=(x_train_count//batch_size))
model.save(model_name)
metrics = history.history
plt.plot(history.epoch, metrics['mse'])
plt.legend(['mse'])
plt.savefig("learning-lstm_sequence.png")
plt.show()
plt.close()
print('Evaluate...')
total_loss = 0
for i, path_clear in enumerate(clear_files):
path_noisy = path_clear.replace("clear", "noisy")
spectNoisy, _, _, _ = audioToTensor(path_noisy)
spectClear, _, _, _ = audioToTensor(path_clear)
input = np.zeros((len(spectNoisy)//voice_max_length, voice_max_length, int(frame_length/2+1)))
for i in range(0, len(spectNoisy)//voice_max_length):
input[i] = spectNoisy[i*voice_max_length:(i+1)*voice_max_length]
result = model.predict(input)
result = np.reshape(result, (result.shape[0]*result.shape[1], result.shape[2]))
loss = np.mean(tf.keras.losses.mean_squared_error(spectClear[0:len(result)], result).numpy())
total_loss+=loss
print(path_noisy, "->", loss)
print("total_loss:", total_loss/len(clear_files))
for test_path in [('data/noisy/book_00000_chp_0009_reader_06709_0_---1_cCGK4M.wav'), ('data/noisy/book_00000_chp_0009_reader_06709_1_---1_cCGK4M.wav')]:
print("test_string: ", test_path)
spect_real, spect_image, spect_sign, audioSR = audioToTensor(test_path)
print("spect_real:", spect_real)
input = np.zeros((len(spect_real)//voice_max_length, voice_max_length, int(frame_length/2+1)))
print("len(spect_real):", len(spect_real))
print("test_audio_abs.shape:", input.shape)
for i in range(0, len(spect_real)//voice_max_length):
input[i] = spect_real[i*voice_max_length:(i+1)*voice_max_length]
print("input.shape:", input.shape)
result = model.predict(input)
result = np.reshape(result, (result.shape[0]*result.shape[1], result.shape[2]))
print("result.shape:", result.shape)
oscillo = spectToOscillo(spect_real=result, spect_sign=spect_sign[0:len(result)], spect_image=spect_image[0:len(result)], audioSR=16000)
oscillo = tf.expand_dims(oscillo, axis=-1)
audio_string = tf.audio.encode_wav(oscillo, sample_rate=audioSR)
clear_filepath = test_path.replace("data/noisy/", "test_")
tf.io.write_file(clear_filepath, contents=audio_string)