代码之家 › 专栏 › 技术社区 › Austin A

TensorFlow和Keras的转移学习问题

vgg-net conv-neural-network keras tensorflow python

Austin A · 技术社区 · 6 年前

我一直在努力重新创造 this blog post . 这篇文章非常全面,而且 code 通过协作共享。

我要做的是从预训练的VGG19网络中提取层,并创建一个以这些层作为输出的新网络。然而,当我组装新的网络时,它非常类似于vg19网络,并且似乎包含我没有提取的层。下面是一个例子。

import tensorflow as tf
from tensorflow.python.keras import models

## Create network based on VGG19 arch with pretrained weights
vgg = tf.keras.applications.vgg19.VGG19(include_top=False, weights='imagenet')
vgg.trainable = False

当我们查看vg19的摘要时,我们看到了我们所期望的体系结构。

vgg.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_2 (InputLayer)         (None, None, None, 3)     0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, None, None, 64)    1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, None, None, 64)    36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, None, None, 64)    0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, None, None, 128)   73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, None, None, 128)   147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, None, None, 128)   0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, None, None, 256)   295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, None, None, 256)   590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, None, None, 256)   590080    
_________________________________________________________________
block3_conv4 (Conv2D)        (None, None, None, 256)   590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, None, None, 256)   0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, None, None, 512)   1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block4_conv4 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, None, None, 512)   0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block5_conv3 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block5_conv4 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, None, None, 512)   0         
=================================================================
Total params: 20,024,384
Trainable params: 0
Non-trainable params: 20,024,384
_________________________________________________________________

然后,我们提取层并创建一个新模型

## Layers to extract
content_layers = ['block5_conv2'] 
style_layers = ['block1_conv1','block2_conv1','block3_conv1','block4_conv1','block5_conv1']
## Get output layers corresponding to style and content layers 
style_outputs = [vgg.get_layer(name).output for name in style_layers]
content_outputs = [vgg.get_layer(name).output for name in content_layers]
model_outputs = style_outputs + content_outputs

new_model = models.Model(vgg.input, model_outputs)

什么时候? new_model 是创造出来的,我相信我们应该有一个更小的模型。然而,对模型的总结表明,新模型非常接近于原始模型(它包含了vg19中22个层中的19个),并且它包含了我们没有提取的层。

new_model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_2 (InputLayer)         (None, None, None, 3)     0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, None, None, 64)    1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, None, None, 64)    36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, None, None, 64)    0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, None, None, 128)   73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, None, None, 128)   147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, None, None, 128)   0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, None, None, 256)   295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, None, None, 256)   590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, None, None, 256)   590080    
_________________________________________________________________
block3_conv4 (Conv2D)        (None, None, None, 256)   590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, None, None, 256)   0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, None, None, 512)   1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block4_conv4 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, None, None, 512)   0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, None, None, 512)   2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, None, None, 512)   2359808   
=================================================================
Total params: 15,304,768
Trainable params: 15,304,768
Non-trainable params: 0
_________________________________________________________________

所以我的问题是…

为什么我没有提取的图层会出现在 纽克模型 . 这些是通过模型的实例化过程推断出来的吗? per the docs ?这似乎太接近VGG19体系结构,无法推断。
从我对喀拉斯的理解来看 Model (functional API) ,传递多个输出层应该创建具有多个输出的模型,但是,新模型似乎是连续的,并且只有一个输出层。是这样吗?

1 回复 | 直到 6 年前

umutto 6 年前

< Buff行情>

为什么我没有提取的图层显示在 new_model>code>中?


   < /块引用>
   
    这是因为当您使用
    
     model s.model(vgg.input,model_output s)创建模型时,
     
      the“intermediate”layers between
      
       vgg.input
      
      and the output layers are included a s well.这是预期的方式,因为vgg是这样构造的。
     
    
   
   
    
     
      例如,如果您要以这种方式创建模型:
      
       models.model(vgg.input,vgg.get_layer('block2_pool').
       
        每个中间层之间的
        
         input_1.
         
          and
          
           block2_pool.
           
            will be included since the input
            
             has to.
             
              flow through them before reasing
              
               block2_pool.
               
                。下面是一个部分的VGG图,可以帮助解决这一问题。
               
              
             
            
           
          
         
        
       
      
     
     
      
       
        
         
          
           
            
             
              
               
               
              
             
            
           
          
         
        
       
      
     
     
      
       
        
         
          
           
            
             
              
               
                
                 
                
               
              
             
            
           
          
         
        
       
      
     
     
      
       
        
         
          
           
            
             
              
               
                现在,-如果我没有误解-如果你想创建一个不包含中间层的模型(这可能会很糟糕),你必须自己创建一个。函数式API在这方面非常有用。在
                
                 documentation上有示例,但您要做的事情的要点如下:
                
               
               
               
               from keras.layers import conv2d,input

x_输入=输入(形状=(28,28,1,))
block1_conv1=conv2d(64,(3,3),padding='same')(x_输入)
block2_conv2=conv2d(128,(3,3),padding='same')(x_输入)
…

new_model=models.model(x_input,[block1_conv1,block2_conv2,…])
< /代码> 

< Buff行情>
<开始=“2”>
然而,新模型似乎是连续的,只有一个输出层。是这样吗?

< /块引用>

不,您的模型有多个您想要的输出。model.summary()should have display which layers are connected to what(which will help understanding the structure),but I think there is a small bug with some versions that prevents.在任何情况下,您都可以通过检查new_model.output


[<tf.tensor'block1_conv1/relu:0'shape=???,64)dtype=float32>,
<tf.tensor'block2_conv1/relu:0'形状=???,128)dtype=float32>,
<tf.tensor'block3_conv1/relu:0'形状=???,256)dtype=float32>,
<tf.tensor'block4_conv1/relu:0'形状=???,512)dtype=float32>,
<tf.tensor'block5_conv1/relu:0'形状=???,512)dtype=float32>,
<tf.tensor'block5_conv2/relu:0'形状=???,512)dtype=float32>]
< /代码> 

在new_model.summary()中按顺序打印它。is just a design choice,as it would get hairy with complex models.

使用创建模型models.Model(vgg.input, model_outputs)中间层vgg.input输出层也包括在内。这是预期的方式,因为vgg是这样构造的。


例如,如果要以这种方式创建模型:models.Model(vgg.input, vgg.get_layer('block2_pool')每个中间层之间input_1和block2_pool因为输入不得不在到达之前流过它们块2A池. 下面是一个部分的VGG图,可以帮助实现这一点。



现在,-如果我没有误解-如果你想创建一个不包含中间层的模型(这可能会很糟糕),你必须自己创建一个。函数式API在这方面非常有用。有一些关于documentation但你想做的事情的要点如下:

from keras.layers import Conv2D, Input

x_input = Input(shape=(28, 28, 1,))
block1_conv1 = Conv2D(64, (3, 3), padding='same')(x_input)
block2_conv2 = Conv2D(128, (3, 3), padding='same')(x_input)
...

new_model = models.Model(x_input, [block1_conv1, block2_conv2, ...])




…然而,新模型似乎是连续的,只有一个输出层。是这样吗?



不,您的模型有多个您想要的输出。model.summary()应该有显示哪些层与什么连接(这有助于理解结构),但我相信有一个小错误,有些版本阻止了这一点。在任何情况下,您都可以通过检查看到您的模型有多个输出。new_model.output,应该给你:

[<tf.Tensor 'block1_conv1/Relu:0' shape=(?, ?, ?, 64) dtype=float32>,
 <tf.Tensor 'block2_conv1/Relu:0' shape=(?, ?, ?, 128) dtype=float32>,
 <tf.Tensor 'block3_conv1/Relu:0' shape=(?, ?, ?, 256) dtype=float32>,
 <tf.Tensor 'block4_conv1/Relu:0' shape=(?, ?, ?, 512) dtype=float32>,
 <tf.Tensor 'block5_conv1/Relu:0' shape=(?, ?, ?, 512) dtype=float32>,
 <tf.Tensor 'block5_conv2/Relu:0' shape=(?, ?, ?, 512) dtype=float32>]


按顺序打印new_model.summary()只是一个设计选择,因为复杂的模型会让它变得毛茸茸的。