from keras.callbacks import LearningRateScheduler
from keras.layers import Multiply
import keras
from keras.models import Sequential
from keras.models import Model
from keras.layers import Input
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import GRU
from keras.layers import SimpleRNN
from keras.layers import Dropout
from keras.layers import concatenate
from keras import losses
from keras import regularizers
from keras.constraints import min_max_norm
import h5py
from keras.constraints import Constraint
from keras import backend as K
import numpy as np
import os
import sys
import tensorflow as tf
# def generateDataset():
mix, sr = librosa.load("./mix.wav", sr=16000)
clean,sr = librosa.load("./clean.wav", sr=16000)
win_length = 256
hop_length = 128
nfft = 512
mix_spectrum = librosa.stft(mix, win_length=win_length, hop_length=hop_length, n_fft=nfft)
clean_spectrum = librosa.stft(clean, win_length=win_length, hop_length=hop_length, n_fft=nfft)
mix_mag = np.abs(mix_spectrum).T
clean_mag = np.abs(clean_spectrum).T
mix_phase = np.angle(mix_spectrum).T
clean_phase = np.angle(clean_spectrum).T
frame_num = mix_mag.shape[0] - 4
feature = np.zeros([frame_num, 257*5])
k = 0
for i in range(frame_num - 4):
frame = mix_mag[k:k+5]
feature[i] = np.reshape(frame, 257*5)
k += 1
#PSM
snr = np.divide(clean_mag, mix_mag) * np.cos(clean_phase - mix_phase)
label = snr[2:-2]
ref = np.cos(clean_phase - mix_phase)
ref = ref[2:-2]
ss = StandardScaler()
feature = ss.fit_transform(feature)
print(feature.shape)
print(label.shape)
window_size = 2000
nb_sequences = feature.shape[0]//window_size
feature = feature[:nb_sequences*window_size, :1285]
feature = np.reshape(feature, (nb_sequences, window_size, 1285))
label = label[:nb_sequences*window_size, :257]
label = np.reshape(label, (nb_sequences, window_size, 257))
ref = ref[:nb_sequences*window_size, :257]
ref = np.reshape(ref, (nb_sequences, window_size, 257))
ref = ref * label
print(feature.shape)
print(label.shape)
def scheduler(epochs):
# 每隔1个epoch,学习率减小为原来的1/10
lr = K.get_value(model.optimizer.lr)
if epochs < 20:
K.set_value(model.optimizer.lr, 0.00001)
if epochs > 20:
K.set_value(model.optimizer.lr, 0.00001)
lr = K.get_value(model.optimizer.lr)
print("lr changed to {}".format(lr ))
return K.get_value(model.optimizer.lr)
reduce_lr = LearningRateScheduler(scheduler)
def mycost(k_input):
def custom_loss(y_true, y_pred):
k = tf.convert_to_tensor(ref)
bbb = K.mean( K.square(K.sqrt(y_pred) - K.sqrt( k)), axis=-1 )
return bbb
return custom_loss
feature_input = Input(shape=(None, 1285), name = 'feature_input')
dense1 = Dense(2048)(feature_input)
norm1 = BatchNormalization()(dense1)
relu1 = LeakyReLU(alpha=0.1)(norm1)
drop1 = Dropout(0.1)(relu1)
dense2 = Dense(2048)(drop1)
norm2 = BatchNormalization()(dense2)
relu2 = LeakyReLU(alpha=0.1)(norm2)
drop2 = Dropout(0.1)(relu2)
dense3 = Dense(2048)(drop2)
norm3 = BatchNormalization()(dense3)
relu3 = LeakyReLU(alpha=0.1)(norm3)
drop3 = Dropout(0.1)(relu3)
dense4 = Dense(257)(drop3)
norm4 = BatchNormalization()(dense4)
out = Activation('sigmoid')(norm4)
model = Model([feature_input] , out)
# My Cost
model.compile(optimizer='adam',
loss= mycost(model.input[1]),
metrics=['mse'])
midWeightPath = './mid_weights.hdf5'
endEpochHookCB = keras.callbacks.ModelCheckpoint( midWeightPath,
monitor = 'val_loss',
save_weights_only = True,
verbose = 1,
save_best_only = False,
period = 1 )
if os.path.exists(midWeightPath):
model.load_weights( midWeightPath )
print("checkpoint_loaded")
model.fit([feature] ,
label,
batch_size=64, epochs=10,
callbacks =[endEpochHookCB,reduce_lr])
model.save("./model.hdf5")