def trans_audio_2_wav(audio_path):
audio_format = audio_path.split('.')[-1]
song = AudioSegment.from_file(audio_path, audio_format)
song = song.set_channels(1)
out_path = audio_path.replace('Data/', 'Data/wav/')
out_path = out_path.replace(audio_format, 'wav')
if_no_create_it(out_path)
song.export(out_path, 'wav')
return 0
def build_cnn_model(onDataset='dataset.h5', featureX='X', targetY='Y', sub_combine_list=[]):
onDataset_name = onDataset.split('/')[-1].split('.')[0]
X_train, X_test, Y_train, Y_test, X_val, Y_val = load_train_test_data(onDataset, featureX, targetY)
print 'len Xtrain=================>>>>>>>>>>>>>>>>>>', len(X_train)
model_save_path = 'models/cnn_' + onDataset_name + '_model.json'
model_weights_save_path = model_save_path.replace('.json', '.h5')
print 'weight:', model_weights_save_path
if LOAD_MODEL_FLAG and os.path.isfile(model_save_path) and os.path.isfile(model_weights_save_path):
print model_save_path
model = load_model(model_save_path)
model = load_model_weights(model_weights_save_path, model)
else:
print 'create model'
model = Sequential()
#parm
# nb_filter
# nb_row
# nb_col
nb_filter_1, nb_row_1, nb_col_1, drop_rate = 63, 1, 10, 0.2
model.add(Conv2D(nb_filter_1, nb_row_1, nb_col_1, input_shape=(1, 2582, 63), activation='relu'))
# shapes: [?,nb_row_1,2582+1-nb_col_1,nb_filter_1]
model.add(Dropout(drop_rate))
model.add(MaxPooling2D(pool_size=(1, 20)))
# shapes: [?,1,(2582+1-nb_col_1+1)/2,nb_filter_1]
model.add(Dropout(drop_rate))
model.add(Conv2D(63, 1, 8, activation='relu'))
model.add(Dropout(drop_rate))
model.add(MaxPooling2D(pool_size=(1, 20)))
model.add(Dropout(drop_rate))
model.add(Flatten())
model.add(Dense(200, activation='relu'))
model.add(Dropout(drop_rate))
model.add(Dense(50, activation='relu'))
model.add(Dropout(drop_rate))
model.add(Dense(20, init='normal', activation='sigmoid'))
model.compile(loss='categorical_crossentropy', optimizer='adam',
metrics=['accuracy', 'precision', 'recall', 'fmeasure']) # accuracy mae fmeasure precision recall
# train model ...
print "train..."
earlyStopping = EarlyStopping(monitor='val_acc',patience=100)
print numpy.shape(X_test)
model.fit(X_train, Y_train, verbose=2, shuffle=True, callbacks=[earlyStopping], nb_epoch=EPOCH,
validation_data=(X_val, Y_val), batch_size=BATCH_SIZE)
if_no_create_it(model_save_path)
save_model(model, model_save_path)
if_no_create_it(model_weights_save_path)
save_model_weights(model, model_weights_save_path)
loss_and_metrics = model.evaluate(X_test, Y_test, verbose=2, batch_size=BATCH_SIZE)
print '==============='
print 'loss_metrics: ', loss_and_metrics
return loss_and_metrics[1]
def build_lstm_model(onDataset='dataset.h5', featureX='X', targetY='Y'):
onDataset_name = onDataset.split('/')[-1].split('.')[0]
X_train, X_test, Y_train, Y_test, X_val, Y_val = load_train_test_data(
onDataset, featureX, targetY)
print 'len Xtrain=================>>>>>>>>>>>>>>>>>>', len(X_train)
INPUT_DIM = len(X_train[0][0])
FEATRE_DIM = INPUT_DIM
model_save_path = 'models/lstm_' + featureX + '_' + onDataset_name + '_model.json'
if_no_create_it(model_save_path)
model_weights_save_path = model_save_path.replace('.json', '.h5')
if_no_create_it(model_weights_save_path)
print 'weight:', model_weights_save_path
if LOAD_MODEL_FLAG and os.path.isfile(model_save_path) and os.path.isfile(
model_weights_save_path):
print model_save_path
model = load_model(model_save_path)
model = load_model_weights(model_weights_save_path, model)
else:
# create model
model = Sequential()
model.add(LSTM(500, input_shape=(TIME_SIZE, FEATRE_DIM)))
model.add(Dropout(.5))
model.add(Dense(3, init='normal', activation='softmax'))
# compile model
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy', 'precision', 'recall', 'fmeasure'
]) # accuracy mae fmeasure precision recall
# train model ...
print "train..."
earlyStopping = EarlyStopping(monitor='val_loss',
patience=5,
verbose=0,
min_delta=0.02)
# model.fit(X_train, Y_train, validation_split=0.2, verbose=2, callbacks=[earlyStopping], nb_epoch=EPOCH,
# batch_size=BATCH_SIZE)
model.fit(X_train,
Y_train,
verbose=2,
shuffle=True,
callbacks=[earlyStopping],
nb_epoch=EPOCH,
validation_data=(X_val, Y_val))
save_model(model, model_save_path)
save_model_weights(model, model_weights_save_path)
loss_and_metrics = model.evaluate(X_test, Y_test, verbose=0)
print '==============='
print 'loss_metrics: ', loss_and_metrics
return loss_and_metrics[1]
def main():
model_path = 'kerasModel/dnn.model'
if_no_create_it(model_path)
retrain = 1
if retrain != 1 and os.path.isfile(model_path):
# load json and create model
json_file = open(model_path, 'r')
loaded_model_json = json_file.read()
json_file.close()
model = model_from_json(loaded_model_json)
# load weights into new model
model.load_weights(model_path.replace('.model', '.weights'))
print("Loaded model from disk")
else:
model = create_baseline()
print model.summary()
callbacks = [EarlyStopping(monitor='val_loss', patience=2, verbose=0)]
model.fit(trainX,
trainY,
batch_size=100,
nb_epoch=10,
validation_data=(validX, validY),
callbacks=callbacks,
verbose=1)
if retrain == 1 or not os.path.isfile(model_path):
# serialize model to JSON
model_json = model.to_json()
with open(model_path, "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights(model_path.replace('.model', '.weights'))
print("Saved model to disk")
pre_testY = model.predict_classes(testX)
print '\n acc: ', get_acc(pre_testY, testY)
return 0