| 1 |
#!/usr/bin/python |
| 2 |
|
| 3 |
# |
| 4 |
# Copyright (c) 2017, Jean-Marc Valin |
| 5 |
# |
| 6 |
# All rights reserved. |
| 7 |
# |
| 8 |
# Redistribution and use in source and binary forms, with or without |
| 9 |
# modification, are permitted provided that the following conditions |
| 10 |
# are met: |
| 11 |
# |
| 12 |
# - Redistributions of source code must retain the above copyright |
| 13 |
# notice, this list of conditions and the following disclaimer. |
| 14 |
# |
| 15 |
# - Redistributions in binary form must reproduce the above copyright |
| 16 |
# notice, this list of conditions and the following disclaimer in the |
| 17 |
# documentation and/or other materials provided with the distribution. |
| 18 |
# |
| 19 |
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 20 |
# ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 21 |
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 22 |
# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR |
| 23 |
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
| 24 |
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
| 25 |
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
| 26 |
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
| 27 |
# LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
| 28 |
# NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 29 |
# SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 30 |
|
| 31 |
from __future__ import print_function |
| 32 |
|
| 33 |
import keras |
| 34 |
from keras.models import Sequential |
| 35 |
from keras.models import Model |
| 36 |
from keras.layers import Input |
| 37 |
from keras.layers import Dense |
| 38 |
from keras.layers import LSTM |
| 39 |
from keras.layers import GRU |
| 40 |
from keras.layers import SimpleRNN |
| 41 |
from keras.layers import Dropout |
| 42 |
from keras.layers import concatenate |
| 43 |
from keras import losses |
| 44 |
from keras import regularizers |
| 45 |
from keras.constraints import min_max_norm |
| 46 |
import h5py |
| 47 |
|
| 48 |
from keras.constraints import Constraint |
| 49 |
from keras import backend as K |
| 50 |
import numpy as np |
| 51 |
|
| 52 |
#import tensorflow as tf |
| 53 |
#from keras.backend.tensorflow_backend import set_session |
| 54 |
#config = tf.ConfigProto() |
| 55 |
#config.gpu_options.per_process_gpu_memory_fraction = 0.42 |
| 56 |
#set_session(tf.Session(config=config)) |
| 57 |
|
| 58 |
|
| 59 |
def my_crossentropy(y_true, y_pred): |
| 60 |
return K.mean(2*K.abs(y_true-0.5) * K.binary_crossentropy(y_pred, y_true), axis=-1) |
| 61 |
|
| 62 |
def mymask(y_true): |
| 63 |
return K.minimum(y_true+1., 1.) |
| 64 |
|
| 65 |
def msse(y_true, y_pred): |
| 66 |
return K.mean(mymask(y_true) * K.square(K.sqrt(y_pred) - K.sqrt(y_true)), axis=-1) |
| 67 |
|
| 68 |
def mycost(y_true, y_pred): |
| 69 |
return K.mean(mymask(y_true) * (10*K.square(K.square(K.sqrt(y_pred) - K.sqrt(y_true))) + K.square(K.sqrt(y_pred) - K.sqrt(y_true)) + 0.01*K.binary_crossentropy(y_pred, y_true)), axis=-1) |
| 70 |
|
| 71 |
def my_accuracy(y_true, y_pred): |
| 72 |
return K.mean(2*K.abs(y_true-0.5) * K.equal(y_true, K.round(y_pred)), axis=-1) |
| 73 |
|
| 74 |
class WeightClip(Constraint): |
| 75 |
'''Clips the weights incident to each hidden unit to be inside a range |
| 76 |
''' |
| 77 |
def __init__(self, c=2): |
| 78 |
self.c = c |
| 79 |
|
| 80 |
def __call__(self, p): |
| 81 |
return K.clip(p, -self.c, self.c) |
| 82 |
|
| 83 |
def get_config(self): |
| 84 |
return {'name': self.__class__.__name__, |
| 85 |
'c': self.c} |
| 86 |
|
| 87 |
reg = 0.000001 |
| 88 |
constraint = WeightClip(0.499) |
| 89 |
|
| 90 |
print('Build model...') |
| 91 |
main_input = Input(shape=(None, 42), name='main_input') |
| 92 |
tmp = Dense(24, activation='tanh', name='input_dense', kernel_constraint=constraint, bias_constraint=constraint)(main_input) |
| 93 |
vad_gru = GRU(24, activation='tanh', recurrent_activation='sigmoid', return_sequences=True, name='vad_gru', kernel_regularizer=regularizers.l2(reg), recurrent_regularizer=regularizers.l2(reg), kernel_constraint=constraint, recurrent_constraint=constraint, bias_constraint=constraint)(tmp) |
| 94 |
vad_output = Dense(1, activation='sigmoid', name='vad_output', kernel_constraint=constraint, bias_constraint=constraint)(vad_gru) |
| 95 |
noise_input = keras.layers.concatenate([tmp, vad_gru, main_input]) |
| 96 |
noise_gru = GRU(48, activation='relu', recurrent_activation='sigmoid', return_sequences=True, name='noise_gru', kernel_regularizer=regularizers.l2(reg), recurrent_regularizer=regularizers.l2(reg), kernel_constraint=constraint, recurrent_constraint=constraint, bias_constraint=constraint)(noise_input) |
| 97 |
denoise_input = keras.layers.concatenate([vad_gru, noise_gru, main_input]) |
| 98 |
|
| 99 |
denoise_gru = GRU(96, activation='tanh', recurrent_activation='sigmoid', return_sequences=True, name='denoise_gru', kernel_regularizer=regularizers.l2(reg), recurrent_regularizer=regularizers.l2(reg), kernel_constraint=constraint, recurrent_constraint=constraint, bias_constraint=constraint)(denoise_input) |
| 100 |
|
| 101 |
denoise_output = Dense(22, activation='sigmoid', name='denoise_output', kernel_constraint=constraint, bias_constraint=constraint)(denoise_gru) |
| 102 |
|
| 103 |
model = Model(inputs=main_input, outputs=[denoise_output, vad_output]) |
| 104 |
|
| 105 |
model.compile(loss=[mycost, my_crossentropy], |
| 106 |
metrics=[msse], |
| 107 |
optimizer='adam', loss_weights=[10, 0.5]) |
| 108 |
|
| 109 |
|
| 110 |
batch_size = 32 |
| 111 |
|
| 112 |
print('Loading data...') |
| 113 |
with h5py.File('training.h5', 'r') as hf: |
| 114 |
all_data = hf['data'][:] |
| 115 |
print('done.') |
| 116 |
|
| 117 |
window_size = 2000 |
| 118 |
|
| 119 |
nb_sequences = len(all_data)//window_size |
| 120 |
print(nb_sequences, ' sequences') |
| 121 |
x_train = all_data[:nb_sequences*window_size, :42] |
| 122 |
x_train = np.reshape(x_train, (nb_sequences, window_size, 42)) |
| 123 |
|
| 124 |
y_train = np.copy(all_data[:nb_sequences*window_size, 42:64]) |
| 125 |
y_train = np.reshape(y_train, (nb_sequences, window_size, 22)) |
| 126 |
|
| 127 |
noise_train = np.copy(all_data[:nb_sequences*window_size, 64:86]) |
| 128 |
noise_train = np.reshape(noise_train, (nb_sequences, window_size, 22)) |
| 129 |
|
| 130 |
vad_train = np.copy(all_data[:nb_sequences*window_size, 86:87]) |
| 131 |
vad_train = np.reshape(vad_train, (nb_sequences, window_size, 1)) |
| 132 |
|
| 133 |
all_data = 0; |
| 134 |
#x_train = x_train.astype('float32') |
| 135 |
#y_train = y_train.astype('float32') |
| 136 |
|
| 137 |
print(len(x_train), 'train sequences. x shape =', x_train.shape, 'y shape = ', y_train.shape) |
| 138 |
|
| 139 |
print('Train...') |
| 140 |
model.fit(x_train, [y_train, vad_train], |
| 141 |
batch_size=batch_size, |
| 142 |
epochs=120, |
| 143 |
validation_split=0.1) |
| 144 |
model.save("weights.hdf5") |