Question

我正在尝试使用Keras中的Inception V3对1000个类进行多标签图像分类。在我的第一个版本的模型（M1）中，我只是重新训练倒数第二层并且得到了0.0152的验证损失。但是当我训练更多层（倒数第二个和卷积）时，比如模型M2，验证损失将稳定在0.0180。当我训练更多图层时，我预计验证损失会减少。

为了加快训练速度，我增加了从M1到M2的倒数第二层权重。我已经尝试过优化器SGD（lr = 0.0001）和“adam”。我被困在这里。有人可以帮助我或指出我的代码中的缺陷吗？

base_model = applications.InceptionV3(weights = "imagenet", include_top=False)

#Adding FC layers to base Inception model
x = base_model.output
x = AveragePooling2D((8,8), strides=(8,8))(x)
x = Flatten()(x)
x = Dense(1024, activation="relu")(x)
x = Dropout(0.5)(x)
out = Dense(1000, activation="sigmoid")(x)
model = Model(inputs=base_model.input, outputs=out)

#loading weights of M1 i.e. only penultimate layer trained model and adding to our new model M2 before training
model_temp = load_model("model_ppath_M1")
mwt = model_temp.get_weights()


#defining layers to be trained
for layer in model.layers[:279]:
    layer.trainable = False
for layer in model.layers[279:]:
    layer.trainable = True

#adding pre trained model M1 weights to this model to speed up training
model.layers[313].set_weights([mwt[0], mwt[1]])  
model.layers[315].set_weights([mwt[2], mwt[3]]) 

model.compile(optimizer="adam",  metrics=["accuracy"], loss='binary_crossentropy')

model.fit_generator("added appropriate data generator function here")

训练以前的Inception V3卷积层会增加损失

0 个答案: