如何从KERAS应用程序修改内置模型

发布于 2025-02-05 17:57:25 字数 1323 浏览 4 评论 0原文

我正在一些小数据集上训练预先训练的KERAS应用模型，然后使用Grad-CAM可视化图像中每一层中发生的情况。基本上，我想看看哪些层负责识别图像中的兴趣点。之后，我想做的是删除对我的目标没有贡献的层，并查看所得的体系结构是否提供相同的指标，或者它的效率降低。

我找不到如何从内置KERAS应用程序模型中删除某些层。有人知道这是否可能吗？如果是，如何？

更新

最终解决了它。这是代码：


initial_model = tf.keras.applications.VGG16(weights = 'imagenet',include_top = False)
initial_model.trainable = False

#adding all the layers from the architecture into a list
layers = [l for l in initial_model.layers]

#building a new model with the existing layers with their 'imagenet' weights
modified_model = keras.Sequential()
inputs = keras.Input(shape = (None,None,3))
modified_model.add(inputs)

#removing the last block from my VGG16 architecture and
for layer in layers:
  if 'block5' not in layer.name:
    layer._inbound_nodes = []
    layer._outbound_nodes = []
    modified_model.add(layer)


modified_model.add(keras.layers.GlobalAveragePooling2D())
#my data has 38 classes, this last layer depends on the number of classes
modified_model.add(keras.layers.Dense(38, activation="softmax"))

modified_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])


history = modified_model.fit(ds_train,
                    epochs = 40,
                    validation_data = ds_valid,
                    verbose = 1)

原文

I am training the pre-trained keras application models on some small dataset and then visualize what happens in each layer of the model to the images using Grad-CAM. Basically, I want to see which layers are in charge of recognizing the points of interest in the image.
What I want to do after that is to remove the layers that do not contribute to my goal and see if the resulting architecture delivers the same metrics or it will become less efficient.

I wasn't able to find how to remove certain layers from a built-in Keras application model. does any one know if this is possible and if yes, how?

Update

Finally solved it. Here is the code:


initial_model = tf.keras.applications.VGG16(weights = 'imagenet',include_top = False)
initial_model.trainable = False

#adding all the layers from the architecture into a list
layers = [l for l in initial_model.layers]

#building a new model with the existing layers with their 'imagenet' weights
modified_model = keras.Sequential()
inputs = keras.Input(shape = (None,None,3))
modified_model.add(inputs)

#removing the last block from my VGG16 architecture and
for layer in layers:
  if 'block5' not in layer.name:
    layer._inbound_nodes = []
    layer._outbound_nodes = []
    modified_model.add(layer)


modified_model.add(keras.layers.GlobalAveragePooling2D())
#my data has 38 classes, this last layer depends on the number of classes
modified_model.add(keras.layers.Dense(38, activation="softmax"))

modified_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])


history = modified_model.fit(ds_train,
                    epochs = 40,
                    validation_data = ds_valid,
                    verbose = 1)

分享到QQ

分享到微博