python - 如何使用 tf.keras.Model.summary 查看父模型中子模型的层？

Question

我有一个 tf.keras.Model 的子类模型，代码如下

import tensorflow as tf


class Mymodel(tf.keras.Model):

    def __init__(self, classes, backbone_model, *args, **kwargs):
        super(Mymodel, self).__init__(self, args, kwargs)
        self.backbone = backbone_model
        self.classify_layer = tf.keras.layers.Dense(classes,activation='sigmoid')

    def call(self, inputs):
        x = self.backbone(inputs)
        x = self.classify_layer(x)
        return x

inputs = tf.keras.Input(shape=(224, 224, 3))
model = Mymodel(inputs=inputs, classes=61, 
                backbone_model=tf.keras.applications.MobileNet())
model.build(input_shape=(20, 224, 224, 3))
model.summary()

结果是:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
mobilenet_1.00_224 (Model)   (None, 1000)              4253864   
_________________________________________________________________
dense (Dense)                multiple                  61061     
=================================================================
Total params: 4,314,925
Trainable params: 4,293,037
Non-trainable params: 21,888
_________________________________________________________________

但我想查看 mobilenet 的所有层，然后我尝试提取 mobilenet 的所有层并放入模型中:

import tensorflow as tf


class Mymodel(tf.keras.Model):

    def __init__(self, classes, backbone_model, *args, **kwargs):
        super(Mymodel, self).__init__(self, args, kwargs)
        self.backbone = backbone_model
        self.classify_layer = tf.keras.layers.Dense(classes,activation='sigmoid')

    def my_process_layers(self,inputs):
        layers = self.backbone.layers
        tmp_x = inputs
        for i in range(1,len(layers)):
            tmp_x = layers[i](tmp_x)
        return tmp_x

    def call(self, inputs):
        x = self.my_process_layers(inputs)
        x = self.classify_layer(x)
        return x

inputs = tf.keras.Input(shape=(224, 224, 3))
model = Mymodel(inputs=inputs, classes=61, 
                backbone_model=tf.keras.applications.MobileNet())
model.build(input_shape=(20, 224, 224, 3))
model.summary()

那么结果没有改变。

    _________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
mobilenet_1.00_224 (Model)   (None, 1000)              4253864   
_________________________________________________________________
dense (Dense)                multiple                  61061     
=================================================================
Total params: 4,314,925
Trainable params: 4,293,037
Non-trainable params: 21,888
_________________________________________________________________

然后我尝试提取一层插入到模型中:

import tensorflow as tf


class Mymodel(tf.keras.Model):

    def __init__(self, classes, backbone_model, *args, **kwargs):
        super(Mymodel, self).__init__(self, args, kwargs)
        self.backbone = backbone_model
        self.classify_layer = tf.keras.layers.Dense(classes,activation='sigmoid')

    def call(self, inputs):
        x = self.backbone.layers[1](inputs)
        x = self.classify_layer(x)
        return x

inputs = tf.keras.Input(shape=(224, 224, 3))
model = Mymodel(inputs=inputs, classes=61, 
                backbone_model=tf.keras.applications.MobileNet())
model.build(input_shape=(20, 224, 224, 3))
model.summary()

它也没有改变。我很困惑。

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
mobilenet_1.00_224 (Model)   (None, 1000)              4253864   
_________________________________________________________________
dense (Dense)                multiple                  244       
=================================================================
Total params: 4,254,108
Trainable params: 4,232,220
Non-trainable params: 21,888
_________________________________________________________________

但是我发现dense层的参数改变了，我不知道发生了什么。

Answer 1

@Ioannis 的答案非常好，但不幸的是它放弃了 keras ' Model Subclassing ' 问题中存在的结构。如果像我一样，您想要保留此模型子类化并仍然在摘要中显示所有层，则可以使用 for 循环分支到更复杂模型的所有各个层:

class MyMobileNet(tf.keras.Sequential):
    def __init__(self, input_shape=(224, 224, 3), classes=61):
        super(MyMobileNet, self).__init__()
        self.backbone_model = [layer for layer in
               tf.keras.applications.MobileNet(input_shape, include_top=False, pooling='avg').layers]
        self.classificator = tf.keras.layers.Dense(classes,activation='sigmoid', name='classificator')

    def call(self, inputs):
        x = inputs
        for layer in self.backbone_model:
            x = layer(x)
        x = self.classificator(x)
        return x
model = MyMobileNet()

之后我们可以直接构建模型并调用summary:

model.build(input_shape=(None, 224, 224, 3))
model.summary()

>
Model: "my_mobile_net"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv1_pad (ZeroPadding2D)    (None, 225, 225, 3)       0         
_________________________________________________________________
conv1 (Conv2D)               (None, 112, 112, 32)      864       
_________________________________________________________________
conv1_bn (BatchNormalization (None, 112, 112, 32)      128       
_________________________________________________________________
....
....
conv_pw_13 (Conv2D)          (None, 7, 7, 1024)        1048576   
_________________________________________________________________
conv_pw_13_bn (BatchNormaliz (None, 7, 7, 1024)        4096      
_________________________________________________________________
conv_pw_13_relu (ReLU)       (None, 7, 7, 1024)        0         
_________________________________________________________________
global_average_pooling2d_13  (None, 1024)              0         
_________________________________________________________________
classificator (Dense)        multiple                  62525     
=================================================================
Total params: 3,291,389
Trainable params: 3,269,501
Non-trainable params: 21,888
_________________________________________________________________