我在Keras建立了一个squeezenet模型。使用标准化的自定义数据集。验证准确度似乎在各个时期都没有提高，训练准确度也没有提高。这是我的代码
我尝试在第二层之后添加maxpooling。输入图像为50x50。
我也尝试使用SGD。将学习率更改为1e-5和1e-4。如github中建议的那样，添加了更多的卷积层。似乎没有什么改变验证准确性。
def build_squeezenet(input_shape):
input_layer = Input(shape = input_shape)
out = Conv2D(96, kernel_size =(3, 3), activation='relu')(input_layer)
out = MaxPooling2D(pool_size=2, strides=None, padding='valid')(out)
# no mxpooling done since small images
out = fire_module(out,squeeze =16 , expansion =64)
out = fire_module(out,squeeze = 16 , expansion=64)
out = fire_module(out,squeeze= 32,expansion = 128)
out = fire_module(out,squeeze= 32,expansion = 128)
out = fire_module(out , squeeze = 48 ,expansion=192)
out = fire_module(out,squeeze = 48 , expansion = 192)
out = fire_module(out,squeeze = 64 , expansion = 256)
out = fire_module(out,squeeze = 64 , expansion = 256)
out = Dropout(0.2)(out)
out = Conv2D(128,kernel_size=(1,1),padding='valid',activation='relu')(out)
out = Conv2D(2,kernel_size=(1,1),padding='valid',activation='relu')(out)
out = GlobalAveragePooling2D()(out)
out = Dense(2, activation="softmax")(out)
model = Model(input_layer, out, name='squeezenet')
return model
def fire_module(input_layer, squeeze =16 , expansion =32):
# a definition for fire module et.al https://github.com/rcmalli/keras-squeezenet/
fire_sq = Conv2D(squeeze,1,1,activation='relu')(input_layer)
fire_exp1 = Conv2D(expansion ,kernel_size=(1,1),activation='relu',padding='valid')(fire_sq)
fire_exp2 = Conv2D(expansion ,kernel_size=(3,3),activation='relu',padding='same')(fire_sq)
out = concatenate([fire_exp1,fire_exp2], axis=3)
return out
Model training commands
model_dw = build_squeezenet(input_shape= (50,50,1))
model_dw.compile(optimizer=Adam(lr=1e-3), loss='categorical_crossentropy',metrics=['accuracy'])
model_dw.fit(x=train_images,
y=train_labels,
batch_size=25,
epochs=10,
verbose=1,
validation_split=0.2,
validation_data=(val_images,val_labels),
shuffle=True)