[keras] 模型保存、加载、model类方法、打印各层权重

红太狼 2022-05-13 11:36 339阅读 0赞

#### 1.模型保存 ####

`model.save_model()`可以保存网络结构权重以及优化器的参数  
`model.save_weights()` 仅仅保存权重

#### 2.模型加载 ####

`from keras.models import load_model`  
`load_model()`只能load 由save\_model保存的，将模型和weight全load进来

filepath: String, path to the saved model.
     custom_objects: Optional dictionary mapping names
         (strings) to custom classes or functions to be
         considered during deserialization.
     compile: Boolean, whether to compile the model
         after loading.

`model.load_weights(self, filepath, by_name=False):`  
在加载权重之前，model必须编译好

metrics = ['accuracy']
            if self.nb_classes >= 10:
                metrics.append('top_k_categorical_accuracy')
    
           # self.input_shape = (seq_length, features_length)
            self.model,self.original_model = self.zf_model()
            optimizer = SGD(lr=1e-3)
            self.model.compile(loss='categorical_crossentropy', optimizer=optimizer,
                               metrics=metrics) #

#### 3.sequential 和functional ####

序列式模型只能有单输入单输出，函数式模型可以有多个输入输出

#### 4.model类 ####

因为是继承， model对象有 container和layer的所有方法，可以用model对象访问下面三个类的所有方法

<table> 
 <thead> 
  <tr> 
   <th>Model(Container)</th> 
   <th>container</th> 
   <th>layer</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>fit</td> 
   <td>summary</td> 
   <td>get_input_at(node_index)</td> 
  </tr> 
  <tr> 
   <td>evaluate</td> 
   <td>get_layer</td> 
   <td>get_config()</td> 
  </tr> 
  <tr> 
   <td>predict</td> 
   <td>get_weights</td> 
   <td>compute_mask(x, mask)</td> 
  </tr> 
  <tr> 
   <td>train on batch</td> 
   <td>set_weights</td> 
   <td>get_input_mask_at(node_index)</td> 
  </tr> 
  <tr> 
   <td>test_on_batch</td> 
   <td>get_config</td> 
   <td>get_output_at(node_index)</td> 
  </tr> 
  <tr> 
   <td>predict_on_batch</td> 
   <td>compute_output_shape</td> 
   <td></td> 
  </tr> 
  <tr> 
   <td>evaluate_generator</td> 
   <td></td> 
   <td></td> 
  </tr> 
  <tr> 
   <td>predict_generator</td> 
   <td></td> 
   <td></td> 
  </tr> 
 </tbody> 
</table>

Container

类属性，不是函数
            name
            inputs
            outputs
            input_layers
            output_layers
            input_spec 
            trainable (boolean)
            input_shape
            output_shape
            inbound_nodes: list of nodes
            outbound_nodes: list of nodes
            trainable_weights (list of variables)
            non_trainable_weights (list of variables)

layer.get\_weights返回的是没有名字的权重array，Model.get\_weights() 是他们的拼接，也没有名字，利用layer.weights 可以访问到后台的变量

#### 5.打印各层权重 ####

for layer in model.layers:
            for weight in layer.weights:
                print weight.name,weight.shape
            #打印各层名字，权重的形状
    block14_sepconv1/pointwise_kernel:0 (1, 1, 1024, 1536)
    block14_sepconv1_bn/gamma:0 (1536,)
    block14_sepconv1_bn/beta:0 (1536,)
    block14_sepconv1_bn/moving_mean:0 (1536,
    conv_att/bias:0 (5,)
    linear_1/kernel:0 (2048, 256)
    linear_1/bias:0 (256,)
    linear_2/kernel:0 (2048, 256)
    linear_2/bias:0 (256,)
    linear_3/kernel:0 (2048, 256)
    linear_3/bias:0 (256,)
    linear_4/kernel:0 (2048, 256)
    linear_4/bias:0 (256,)
    linear_5/kernel:0 (2048, 256)
    linear_5/bias:0 (256,)
    rgb_softmax/kernel:0 (1280, 60)
    rgb_softmax/bias:0 (60,)

from keras.applications.vgg16 import VGG16
    # model.layers  ,layer.weights
    model = VGG16()
    names = [weight.name for layer in model.layers for weight in layer.weights]
    weights = model.get_weights()
    for name, weight in zip(names, weights):
        print(name, weight.shape)