def __init__(self, modelh5=None):
if modelh5 == None:
mpath = os.path.join(os.getcwd(), 'CNNevaluation', 'regular',
'regi=1', 'SPEC_Regular_8_epoch13.h5')
else:
mpath = modelh5
self.model = load_model(mpath,
custom_objects={
'FocalLoss': focal_loss,
'focal_loss_fixed': focal_loss()
})
optimizer = optimizers.Adadelta()
self.model.compile(optimizer=optimizer,
loss=[focal_loss(alpha=0.25, gamma=2)])
def loadbyFolder(self, foldername):
filename = [
os.path.join(foldername, file) for file in os.listdir(foldername)
if 'weights' not in file
]
if len(filename) == 1:
print('Start building model..')
self.model = load_model(filename[0],
custom_objects={
'FocalLoss': focal_loss,
'focal_loss_fixed': focal_loss()
})
optimizer = optimizers.Adadelta()
self.model.compile(optimizer=optimizer,
loss=[focal_loss(alpha=0.25, gamma=2)])
else:
print('Zero or multiple files are detected!')
assert (0)
def CNNForest(self, feature_type, module_type, layer_counts):
feature_length = 200 if feature_type == 'SPEC' else 26
input_shape = (AUDIO_LENGTH, feature_length, 1)
dictmap = {
'Regular': ReguBlock,
'Residual': ResiBlock,
'Inception': XcepBlock
}
Module = dictmap[module_type]
X_input = Input(name='the_input', shape=input_shape)
level_h1 = Module(32)(X_input)
level_m1 = MaxPooling2D(pool_size=2, strides=None,
padding="valid")(level_h1) # 池化层
level_h2 = Module(64)(level_m1)
level_m2 = MaxPooling2D(pool_size=2, strides=None,
padding="valid")(level_h2) # 池化层
level_h3 = Module(128)(level_m2)
level_m3 = MaxPooling2D(pool_size=2, strides=None,
padding="valid")(level_h3) # 池化层
level_h4 = Module(256)(level_m3)
level_m4 = MaxPooling2D(pool_size=2, strides=None,
padding="valid")(level_h4) # 池化层
level_s5 = Module(512)(level_m4)
if feature_type == 'SPEC':
level_s5 = MaxPooling2D(pool_size=2, strides=None,
padding="valid")(level_s5) # 池化层
layers = [level_m1, level_m2, level_m3, level_m4, level_s5]
output = layers[layer_counts // 2 - 1]
flayer = GlobalAveragePooling2D()(output)
fc2 = Dense(CLASS_NUM, use_bias=True,
kernel_initializer='he_normal')(flayer) # 全连接层
y_pred = Activation('softmax', name='Activation0')(fc2)
model = Model(inputs=X_input, outputs=y_pred)
optimizer = optimizers.Adadelta()
model.compile(optimizer=optimizer,
loss=[focal_loss(alpha=0.25, gamma=2)])
print('{} cnn model with the {} feature and {} layers are estabished.'.
format(module_type, feature_type, str(layer_counts)))
modelname = feature_type + '_' + module_type + '_' + str(layer_counts)
return model, modelname
three.append(spath)
spath = os.path.join(os.getcwd(),'log&&materials','20843728498872_2018_01_08_10_42_58.wav')
three.append(spath)
spath = os.path.join(os.getcwd(),'log&&materials','199481741554988_2018_10_20_00_04_25.wav')
three.append(spath)
# n1path = os.path.join(os.getcwd(),'log&&materials','noisy_sounds','199481741554988_2018_10_20_00_00_00.wav')
for spath in three:
signal,fs = read_wav_data(spath)
spec = GetFrequencyFeatures(signal,fs)
img = spec.reshape(1,spec.shape[0], spec.shape[1], 1)
#2. load model
# mpath = os.path.join(os.getcwd(),'CNNevaluation','residual','resi=1','SPEC_Residual_8_epoch14.h5')
# mpath = os.path.join(os.getcwd(),'CNNdesign','spec+inception','10','SPEC_Inception_10_epoch4.h5')
mpath = os.path.join(os.getcwd(),'CNNevaluation','regular','regi=0','spec_Regular_8_epoch12.h5')
model = load_model(mpath,custom_objects={'FocalLoss': focal_loss,'focal_loss_fixed': focal_loss()})
optimizer = optimizers.Adadelta()
model.compile(optimizer=optimizer, loss=[focal_loss(alpha=0.25, gamma=2)])
#3. try to predictverbose=0
print(np.argmax(model.predict(img)))
print(model.predict(img,verbose=1))
# print(np.argmax(model.predict(img)[0]))
# print(model.layers[27])
# print(model.layers[28])
# print(model.layers[29])
bowel_output = model.output[:,1]
# print(bowel_output)
last_conv_layer = model.layers[28]
grads = K.gradients(bowel_output, last_conv_layer.output)[0]
pooled_grads = K.mean(grads, axis=(0, 1, 2))
iterate = K.function([model.input], [pooled_grads, last_conv_layer.output[0]])
def ModelTrainingSetting(self, model):
optimizer = optimizers.Adadelta()
model.compile(optimizer=optimizer,
loss=[focal_loss(alpha=0.25, gamma=2)])
